AnotherFoxGuy/webrtc-audio-processing
1
0
Fork 0
Code Pull Requests Activity

Update to current webrtc library

Browse Source 
This is from the upstream library commit id
3326535126e435f1ba647885ce43a8f0f3d317eb, corresponding to Chromium
88.0.4290.1.
This commit is contained in:
Arun Raghavan
2020-10-12 18:08:02 -04:00
parent b1b02581d3
commit bcec8b0b21
859 changed files with 76187 additions and 49580 deletions
Download Patch File Download Diff File Expand all files Collapse all files

801
webrtc/modules/audio_processing/BUILD.gn
View File

@ -6,281 +6,606 @@
# in the file PATENTS. All contributing project authors may
# be found in the AUTHORS file in the root of the source tree.
import("//build/config/arm.gni")
import("//third_party/protobuf/proto_library.gni")
import("../../build/webrtc.gni")
declare_args() {
# Outputs some low-level debug files.
aec_debug_dump = false
# Disables the usual mode where we trust the reported system delay
# values the AEC receives. The corresponding define is set appropriately
# in the code, but it can be force-enabled here for testing.
aec_untrusted_delay_for_testing = false
import("../../webrtc.gni")
if (rtc_enable_protobuf) {
import("//third_party/protobuf/proto_library.gni")
}
source_set("audio_processing") {
config("apm_debug_dump") {
if (apm_debug_dump) {
defines = [ "WEBRTC_APM_DEBUG_DUMP=1" ]
} else {
defines = [ "WEBRTC_APM_DEBUG_DUMP=0" ]
}
}
rtc_library("config") {
visibility = [ ":*" ]
sources = [
"include/config.cc",
"include/config.h",
]
deps = [ "../../rtc_base/system:rtc_export" ]
}
rtc_library("api") {
visibility = [ "*" ]
sources = [
"include/audio_processing.cc",
"include/audio_processing.h",
]
deps = [
":audio_frame_view",
":audio_processing_statistics",
":config",
"../../api:array_view",
"../../api:scoped_refptr",
"../../api/audio:aec3_config",
"../../api/audio:audio_frame_api",
"../../api/audio:echo_control",
"../../rtc_base:deprecation",
"../../rtc_base:rtc_base_approved",
"../../rtc_base/system:arch",
"../../rtc_base/system:file_wrapper",
"../../rtc_base/system:rtc_export",
"agc:gain_control_interface",
]
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
}
rtc_library("audio_frame_proxies") {
visibility = [ "*" ]
sources = [
"include/audio_frame_proxies.cc",
"include/audio_frame_proxies.h",
]
deps = [
":api",
":audio_frame_view",
"../../api/audio:audio_frame_api",
]
}
rtc_library("audio_buffer") {
visibility = [ "*" ]
configs += [ ":apm_debug_dump" ]
sources = [
"aec/aec_core.c",
"aec/aec_core.h",
"aec/aec_core_internal.h",
"aec/aec_rdft.c",
"aec/aec_rdft.h",
"aec/aec_resampler.c",
"aec/aec_resampler.h",
"aec/echo_cancellation.c",
"aec/echo_cancellation_internal.h",
"aec/include/echo_cancellation.h",
"aecm/aecm_core.c",
"aecm/aecm_core.h",
"aecm/echo_control_mobile.c",
"aecm/include/echo_control_mobile.h",
"agc/agc.cc",
"agc/agc.h",
"agc/agc_manager_direct.cc",
"agc/agc_manager_direct.h",
"agc/gain_map_internal.h",
"agc/histogram.cc",
"agc/histogram.h",
"agc/legacy/analog_agc.c",
"agc/legacy/analog_agc.h",
"agc/legacy/digital_agc.c",
"agc/legacy/digital_agc.h",
"agc/legacy/gain_control.h",
"agc/utility.cc",
"agc/utility.h",
"audio_buffer.cc",
"audio_buffer.h",
"audio_processing_impl.cc",
"audio_processing_impl.h",
"beamformer/array_util.cc",
"beamformer/array_util.h",
"beamformer/beamformer.h",
"beamformer/complex_matrix.h",
"beamformer/covariance_matrix_generator.cc",
"beamformer/covariance_matrix_generator.h",
"beamformer/matrix.h",
"beamformer/nonlinear_beamformer.cc",
"beamformer/nonlinear_beamformer.h",
"common.h",
"echo_cancellation_impl.cc",
"echo_cancellation_impl.h",
"echo_control_mobile_impl.cc",
"echo_control_mobile_impl.h",
"gain_control_impl.cc",
"gain_control_impl.h",
"high_pass_filter_impl.cc",
"high_pass_filter_impl.h",
"include/audio_processing.h",
"intelligibility/intelligibility_enhancer.cc",
"intelligibility/intelligibility_enhancer.h",
"intelligibility/intelligibility_utils.cc",
"intelligibility/intelligibility_utils.h",
"level_estimator_impl.cc",
"level_estimator_impl.h",
"logging/aec_logging.h",
"logging/aec_logging_file_handling.cc",
"logging/aec_logging_file_handling.h",
"noise_suppression_impl.cc",
"noise_suppression_impl.h",
"processing_component.cc",
"processing_component.h",
"rms_level.cc",
"rms_level.h",
"splitting_filter.cc",
"splitting_filter.h",
"three_band_filter_bank.cc",
"three_band_filter_bank.h",
"transient/common.h",
"transient/daubechies_8_wavelet_coeffs.h",
"transient/dyadic_decimator.h",
"transient/moving_moments.cc",
"transient/moving_moments.h",
"transient/transient_detector.cc",
"transient/transient_detector.h",
"transient/transient_suppressor.cc",
"transient/transient_suppressor.h",
"transient/wpd_node.cc",
"transient/wpd_node.h",
"transient/wpd_tree.cc",
"transient/wpd_tree.h",
"typing_detection.cc",
"typing_detection.h",
"utility/delay_estimator.c",
"utility/delay_estimator.h",
"utility/delay_estimator_internal.h",
"utility/delay_estimator_wrapper.c",
"utility/delay_estimator_wrapper.h",
"vad/common.h",
"vad/gmm.cc",
"vad/gmm.h",
"vad/noise_gmm_tables.h",
"vad/pitch_based_vad.cc",
"vad/pitch_based_vad.h",
"vad/pitch_internal.cc",
"vad/pitch_internal.h",
"vad/pole_zero_filter.cc",
"vad/pole_zero_filter.h",
"vad/standalone_vad.cc",
"vad/standalone_vad.h",
"vad/vad_audio_proc.cc",
"vad/vad_audio_proc.h",
"vad/vad_audio_proc_internal.h",
"vad/vad_circular_buffer.cc",
"vad/vad_circular_buffer.h",
"vad/voice_activity_detector.cc",
"vad/voice_activity_detector.h",
"vad/voice_gmm_tables.h",
"voice_detection_impl.cc",
"voice_detection_impl.h",
]
configs += [ "../..:common_config" ]
public_configs = [ "../..:common_inherited_config" ]
defines = []
deps = [
":api",
"../../api:array_view",
"../../common_audio",
"../../common_audio:common_audio_c",
"../../rtc_base:checks",
]
}
rtc_library("high_pass_filter") {
visibility = [ "*" ]
sources = [
"high_pass_filter.cc",
"high_pass_filter.h",
]
defines = []
deps = [
":audio_buffer",
"../../api:array_view",
"../../rtc_base:checks",
"utility:cascaded_biquad_filter",
]
}
rtc_source_set("aec_dump_interface") {
visibility = [ "*" ]
sources = [
"include/aec_dump.cc",
"include/aec_dump.h",
]
deps = [
":api",
":audio_frame_view",
"../../rtc_base:deprecation",
]
}
rtc_library("audio_processing") {
visibility = [ "*" ]
configs += [ ":apm_debug_dump" ]
sources = [
"audio_processing_builder_impl.cc",
"audio_processing_impl.cc",
"audio_processing_impl.h",
"common.h",
"echo_control_mobile_impl.cc",
"echo_control_mobile_impl.h",
"echo_detector/circular_buffer.cc",
"echo_detector/circular_buffer.h",
"echo_detector/mean_variance_estimator.cc",
"echo_detector/mean_variance_estimator.h",
"echo_detector/moving_max.cc",
"echo_detector/moving_max.h",
"echo_detector/normalized_covariance_estimator.cc",
"echo_detector/normalized_covariance_estimator.h",
"gain_control_impl.cc",
"gain_control_impl.h",
"gain_controller2.cc",
"gain_controller2.h",
"level_estimator.cc",
"level_estimator.h",
"render_queue_item_verifier.h",
"residual_echo_detector.cc",
"residual_echo_detector.h",
"typing_detection.cc",
"typing_detection.h",
]
defines = []
deps = [
"../..:webrtc_common",
"../audio_coding:isac",
":aec_dump_interface",
":api",
":apm_logging",
":audio_buffer",
":audio_frame_proxies",
":audio_frame_view",
":audio_processing_statistics",
":config",
":high_pass_filter",
":optionally_built_submodule_creators",
":rms_level",
":voice_detection",
"../../api:array_view",
"../../api:function_view",
"../../api/audio:aec3_config",
"../../api/audio:audio_frame_api",
"../../api/audio:echo_control",
"../../audio/utility:audio_frame_operations",
"../../common_audio:common_audio_c",
"../../common_audio/third_party/ooura:fft_size_256",
"../../rtc_base:checks",
"../../rtc_base:deprecation",
"../../rtc_base:gtest_prod",
"../../rtc_base:ignore_wundef",
"../../rtc_base:refcount",
"../../rtc_base:safe_minmax",
"../../rtc_base:sanitizer",
"../../rtc_base/synchronization:mutex",
"../../rtc_base/system:rtc_export",
"../../system_wrappers",
"../../system_wrappers:field_trial",
"../../system_wrappers:metrics",
"aec3",
"aec_dump:aec_dump",
"aecm:aecm_core",
"agc",
"agc:gain_control_interface",
"agc:legacy_agc",
"agc2:adaptive_digital",
"agc2:fixed_digital",
"agc2:gain_applier",
"ns",
"transient:transient_suppressor_api",
"vad",
]
if (aec_debug_dump) {
defines += [ "WEBRTC_AEC_DEBUG_DUMP" ]
}
if (aec_untrusted_delay_for_testing) {
defines += [ "WEBRTC_UNTRUSTED_DELAY" ]
}
if (rtc_enable_protobuf) {
defines += [ "WEBRTC_AUDIOPROC_DEBUG_DUMP" ]
deps += [ ":audioproc_debug_proto" ]
}
if (rtc_prefer_fixed_point) {
defines += [ "WEBRTC_NS_FIXED" ]
sources += [
"ns/include/noise_suppression_x.h",
"ns/noise_suppression_x.c",
"ns/nsx_core.c",
"ns/nsx_core.h",
"ns/nsx_defines.h",
]
if (current_cpu == "mipsel") {
sources += [ "ns/nsx_core_mips.c" ]
} else {
sources += [ "ns/nsx_core_c.c" ]
}
} else {
defines += [ "WEBRTC_NS_FLOAT" ]
sources += [
"ns/defines.h",
"ns/include/noise_suppression.h",
"ns/noise_suppression.c",
"ns/ns_core.c",
"ns/ns_core.h",
"ns/windows_private.h",
]
}
if (current_cpu == "x86" || current_cpu == "x64") {
deps += [ ":audio_processing_sse2" ]
}
if (rtc_build_with_neon) {
deps += [ ":audio_processing_neon" ]
}
if (current_cpu == "mipsel") {
sources += [ "aecm/aecm_core_mips.c" ]
if (mips_float_abi == "hard") {
sources += [
"aec/aec_core_mips.c",
"aec/aec_rdft_mips.c",
]
}
} else {
sources += [ "aecm/aecm_core_c.c" ]
}
if (is_win) {
cflags = [
# TODO(jschuh): Bug 1348: fix this warning.
"/wd4267", # size_t to int truncations
]
}
if (is_clang) {
# Suppress warnings from Chrome's Clang plugins.
# See http://code.google.com/p/webrtc/issues/detail?id=163 for details.
configs -= [ "//build/config/clang:find_bad_constructs" ]
}
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
deps += [
"../../base:rtc_base_approved",
"../../common_audio",
"../../common_audio:fir_filter",
"../../common_audio:fir_filter_factory",
"../../rtc_base:rtc_base_approved",
"../../system_wrappers",
]
if (rtc_enable_protobuf) {
deps += [ "aec_dump:aec_dump_impl" ]
} else {
deps += [ "aec_dump:null_aec_dump_factory" ]
}
}
rtc_library("voice_detection") {
sources = [
"voice_detection.cc",
"voice_detection.h",
]
deps = [
":api",
":audio_buffer",
"../../api/audio:audio_frame_api",
"../../common_audio:common_audio_c",
"../../rtc_base:checks",
]
}
rtc_library("optionally_built_submodule_creators") {
sources = [
"optionally_built_submodule_creators.cc",
"optionally_built_submodule_creators.h",
]
deps = [
"transient:transient_suppressor_api",
"transient:transient_suppressor_impl",
]
}
rtc_source_set("rms_level") {
visibility = [ "*" ]
sources = [
"rms_level.cc",
"rms_level.h",
]
deps = [
"../../api:array_view",
"../../rtc_base:checks",
]
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
}
rtc_library("audio_processing_statistics") {
visibility = [ "*" ]
sources = [
"include/audio_processing_statistics.cc",
"include/audio_processing_statistics.h",
]
deps = [ "../../rtc_base/system:rtc_export" ]
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
}
rtc_source_set("audio_frame_view") {
sources = [ "include/audio_frame_view.h" ]
deps = [ "../../api:array_view" ]
}
if (rtc_enable_protobuf) {
proto_library("audioproc_debug_proto") {
sources = [
"debug.proto",
]
sources = [ "debug.proto" ]
proto_out_dir = "webrtc/audio_processing"
proto_out_dir = "modules/audio_processing"
}
}
if (current_cpu == "x86" || current_cpu == "x64") {
source_set("audio_processing_sse2") {
sources = [
"aec/aec_core_sse2.c",
"aec/aec_rdft_sse2.c",
]
if (is_posix) {
cflags = [ "-msse2" ]
}
configs += [ "../..:common_config" ]
public_configs = [ "../..:common_inherited_config" ]
}
rtc_library("apm_logging") {
configs += [ ":apm_debug_dump" ]
sources = [
"logging/apm_data_dumper.cc",
"logging/apm_data_dumper.h",
]
deps = [
"../../api:array_view",
"../../common_audio",
"../../rtc_base:checks",
"../../rtc_base:rtc_base_approved",
]
defines = []
}
if (rtc_build_with_neon) {
source_set("audio_processing_neon") {
sources = [
"aec/aec_core_neon.c",
"aec/aec_rdft_neon.c",
"aecm/aecm_core_neon.c",
"ns/nsx_core_neon.c",
if (rtc_include_tests) {
rtc_source_set("mocks") {
testonly = true
sources = [ "include/mock_audio_processing.h" ]
deps = [
":aec_dump_interface",
":api",
":audio_buffer",
":audio_processing",
":audio_processing_statistics",
"../../test:test_support",
]
}
group("audio_processing_tests") {
testonly = true
deps = [
":audioproc_test_utils",
"transient:click_annotate",
"transient:transient_suppression_test",
]
if (current_cpu != "arm64") {
# Enable compilation for the NEON instruction set. This is needed
# since //build/config/arm.gni only enables NEON for iOS, not Android.
# This provides the same functionality as webrtc/build/arm_neon.gypi.
configs -= [ "//build/config/compiler:compiler_arm_fpu" ]
cflags = [ "-mfpu=neon" ]
if (rtc_enable_protobuf) {
deps += [
":audioproc_unittest_proto",
"aec_dump:aec_dump_unittests",
"test/conversational_speech",
"test/py_quality_assessment",
]
}
}
rtc_library("audio_processing_unittests") {
testonly = true
configs += [ ":apm_debug_dump" ]
sources = [
"audio_buffer_unittest.cc",
"audio_frame_view_unittest.cc",
"config_unittest.cc",
"echo_control_mobile_unittest.cc",
"gain_controller2_unittest.cc",
"splitting_filter_unittest.cc",
"test/fake_recording_device_unittest.cc",
]
deps = [
":analog_mic_simulation",
":api",
":apm_logging",
":audio_buffer",
":audio_frame_view",
":audio_processing",
":audioproc_test_utils",
":config",
":high_pass_filter",
":mocks",
":voice_detection",
"../../api:array_view",
"../../api:scoped_refptr",
"../../api/audio:aec3_config",
"../../api/audio:aec3_factory",
"../../common_audio",
"../../common_audio:common_audio_c",
"../../rtc_base",
"../../rtc_base:checks",
"../../rtc_base:gtest_prod",
"../../rtc_base:ignore_wundef",
"../../rtc_base:protobuf_utils",
"../../rtc_base:rtc_base_approved",
"../../rtc_base:rtc_base_tests_utils",
"../../rtc_base:safe_minmax",
"../../rtc_base:task_queue_for_test",
"../../rtc_base/synchronization:mutex",
"../../rtc_base/system:arch",
"../../rtc_base/system:file_wrapper",
"../../system_wrappers",
"../../test:fileutils",
"../../test:rtc_expect_death",
"../../test:test_support",
"../audio_coding:neteq_input_audio_tools",
"aec_dump:mock_aec_dump_unittests",
"agc:agc_unittests",
"agc2:adaptive_digital_unittests",
"agc2:biquad_filter_unittests",
"agc2:fixed_digital_unittests",
"agc2:noise_estimator_unittests",
"agc2:rnn_vad_with_level_unittests",
"agc2:test_utils",
"agc2/rnn_vad:unittests",
"test/conversational_speech:unittest",
"transient:transient_suppression_unittests",
"utility:legacy_delay_estimator_unittest",
"utility:pffft_wrapper_unittest",
"vad:vad_unittests",
"//testing/gtest",
]
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
defines = []
if (rtc_prefer_fixed_point) {
defines += [ "WEBRTC_AUDIOPROC_FIXED_PROFILE" ]
} else {
defines += [ "WEBRTC_AUDIOPROC_FLOAT_PROFILE" ]
}
# Disable LTO on NEON targets due to compiler bug.
# TODO(fdegans): Enable this. See crbug.com/408997.
if (rtc_use_lto) {
cflags -= [
"-flto",
"-ffat-lto-objects",
if (rtc_enable_protobuf) {
defines += [ "WEBRTC_AUDIOPROC_DEBUG_DUMP" ]
deps += [
":audioproc_debug_proto",
":audioproc_protobuf_utils",
":audioproc_test_utils",
":audioproc_unittest_proto",
":optionally_built_submodule_creators",
":rms_level",
":runtime_settings_protobuf_utils",
"../../api/audio:audio_frame_api",
"../../api/audio:echo_control",
"../../rtc_base:rtc_base_tests_utils",
"../../rtc_base:rtc_task_queue",
"aec_dump",
"aec_dump:aec_dump_unittests",
]
absl_deps += [ "//third_party/abseil-cpp/absl/flags:flag" ]
sources += [
"audio_processing_impl_locking_unittest.cc",
"audio_processing_impl_unittest.cc",
"audio_processing_unittest.cc",
"echo_control_mobile_bit_exact_unittest.cc",
"echo_detector/circular_buffer_unittest.cc",
"echo_detector/mean_variance_estimator_unittest.cc",
"echo_detector/moving_max_unittest.cc",
"echo_detector/normalized_covariance_estimator_unittest.cc",
"gain_control_unittest.cc",
"high_pass_filter_unittest.cc",
"level_estimator_unittest.cc",
"residual_echo_detector_unittest.cc",
"rms_level_unittest.cc",
"test/debug_dump_replayer.cc",
"test/debug_dump_replayer.h",
"test/debug_dump_test.cc",
"test/echo_canceller_test_tools.cc",
"test/echo_canceller_test_tools.h",
"test/echo_canceller_test_tools_unittest.cc",
"test/echo_control_mock.h",
"test/test_utils.h",
"voice_detection_unittest.cc",
]
}
}
rtc_library("audio_processing_perf_tests") {
testonly = true
configs += [ ":apm_debug_dump" ]
sources = [ "audio_processing_performance_unittest.cc" ]
deps = [
":audio_processing",
":audioproc_test_utils",
"../../api:array_view",
"../../rtc_base:protobuf_utils",
"../../rtc_base:rtc_base_approved",
"../../system_wrappers",
"../../test:perf_test",
"../../test:test_support",
]
}
rtc_library("analog_mic_simulation") {
sources = [
"test/fake_recording_device.cc",
"test/fake_recording_device.h",
]
deps = [
"../../api:array_view",
"../../api/audio:audio_frame_api",
"../../common_audio",
"../../rtc_base:checks",
"../../rtc_base:rtc_base_approved",
"../../rtc_base:safe_minmax",
"agc:gain_map",
]
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
}
if (rtc_enable_protobuf) {
rtc_library("audioproc_f_impl") {
testonly = true
configs += [ ":apm_debug_dump" ]
sources = [
"test/aec_dump_based_simulator.cc",
"test/aec_dump_based_simulator.h",
"test/api_call_statistics.cc",
"test/api_call_statistics.h",
"test/audio_processing_simulator.cc",
"test/audio_processing_simulator.h",
"test/audioproc_float_impl.cc",
"test/audioproc_float_impl.h",
"test/wav_based_simulator.cc",
"test/wav_based_simulator.h",
]
deps = [
":analog_mic_simulation",
":api",
":apm_logging",
":audio_processing",
":audioproc_debug_proto",
":audioproc_protobuf_utils",
":audioproc_test_utils",
":runtime_settings_protobuf_utils",
"../../api/audio:aec3_config_json",
"../../api/audio:aec3_factory",
"../../common_audio",
"../../rtc_base:checks",
"../../rtc_base:ignore_wundef",
"../../rtc_base:protobuf_utils",
"../../rtc_base:rtc_base_approved",
"../../rtc_base:rtc_json",
"../../rtc_base:task_queue_for_test",
"../../rtc_base/system:file_wrapper",
"../../system_wrappers",
"../../system_wrappers:field_trial",
"../../test:test_support",
"aec_dump",
"aec_dump:aec_dump_impl",
"//testing/gtest",
]
absl_deps = [
"//third_party/abseil-cpp/absl/flags:flag",
"//third_party/abseil-cpp/absl/flags:parse",
"//third_party/abseil-cpp/absl/strings",
"//third_party/abseil-cpp/absl/types:optional",
]
} # audioproc_f_impl
}
if (rtc_enable_protobuf) {
proto_library("audioproc_unittest_proto") {
sources = [ "test/unittest.proto" ]
proto_out_dir = "modules/audio_processing/test"
}
rtc_library("audioproc_protobuf_utils") {
sources = [
"test/protobuf_utils.cc",
"test/protobuf_utils.h",
]
deps = [
":audioproc_debug_proto",
"../../rtc_base:checks",
"../../rtc_base:ignore_wundef",
"../../rtc_base:protobuf_utils",
"../../rtc_base:rtc_base_approved",
"../../rtc_base/system:arch",
]
}
configs += [ "../..:common_config" ]
public_configs = [ "../..:common_inherited_config" ]
rtc_library("runtime_settings_protobuf_utils") {
testonly = true
sources = [
"test/runtime_setting_util.cc",
"test/runtime_setting_util.h",
]
deps = [
"../../common_audio",
]
deps = [
":api",
":audioproc_debug_proto",
":audioproc_protobuf_utils",
"../../rtc_base:checks",
]
}
}
}
rtc_library("audioproc_test_utils") {
visibility = [ "*" ]
testonly = true
sources = [
"test/audio_buffer_tools.cc",
"test/audio_buffer_tools.h",
"test/audio_processing_builder_for_testing.cc",
"test/audio_processing_builder_for_testing.h",
"test/bitexactness_tools.cc",
"test/bitexactness_tools.h",
"test/performance_timer.cc",
"test/performance_timer.h",
"test/simulator_buffers.cc",
"test/simulator_buffers.h",
"test/test_utils.cc",
"test/test_utils.h",
]
configs += [ ":apm_debug_dump" ]
deps = [
":api",
":audio_buffer",
":audio_processing",
"../../api:array_view",
"../../api/audio:audio_frame_api",
"../../common_audio",
"../../rtc_base:checks",
"../../rtc_base:rtc_base_approved",
"../../rtc_base/system:arch",
"../../system_wrappers",
"../../test:fileutils",
"../../test:test_support",
"../audio_coding:neteq_input_audio_tools",
"//testing/gtest",
]
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
}

32
webrtc/modules/audio_processing/aec/aec_common.h
View File

@ -1,32 +0,0 @@
/*
* Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_COMMON_H_
#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_COMMON_H_
#include "webrtc/typedefs.h"
#ifdef _MSC_VER /* visual c++ */
#define ALIGN16_BEG __declspec(align(16))
#define ALIGN16_END
#else /* gcc or icc */
#define ALIGN16_BEG
#define ALIGN16_END __attribute__((aligned(16)))
#endif
extern ALIGN16_BEG const float ALIGN16_END WebRtcAec_sqrtHanning[65];
extern ALIGN16_BEG const float ALIGN16_END WebRtcAec_weightCurve[65];
extern ALIGN16_BEG const float ALIGN16_END WebRtcAec_overDriveCurve[65];
extern const float WebRtcAec_kExtendedSmoothingCoefficients[2][2];
extern const float WebRtcAec_kNormalSmoothingCoefficients[2][2];
extern const float WebRtcAec_kMinFarendPSD;
#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_COMMON_H_

1929
webrtc/modules/audio_processing/aec/aec_core.c
View File

File diff suppressed because it is too large Load Diff

129
webrtc/modules/audio_processing/aec/aec_core.h
View File

@ -1,129 +0,0 @@
/*
* Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
/*
* Specifies the interface for the AEC core.
*/
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_CORE_H_
#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_CORE_H_
#include <stddef.h>
#include "webrtc/typedefs.h"
#define FRAME_LEN 80
#define PART_LEN 64 // Length of partition
#define PART_LEN1 (PART_LEN + 1) // Unique fft coefficients
#define PART_LEN2 (PART_LEN * 2) // Length of partition * 2
#define NUM_HIGH_BANDS_MAX 2 // Max number of high bands
typedef float complex_t[2];
// For performance reasons, some arrays of complex numbers are replaced by twice
// as long arrays of float, all the real parts followed by all the imaginary
// ones (complex_t[SIZE] -> float[2][SIZE]). This allows SIMD optimizations and
// is better than two arrays (one for the real parts and one for the imaginary
// parts) as this other way would require two pointers instead of one and cause
// extra register spilling. This also allows the offsets to be calculated at
// compile time.
// Metrics
enum {
kOffsetLevel = -100
};
typedef struct Stats {
float instant;
float average;
float min;
float max;
float sum;
float hisum;
float himean;
int counter;
int hicounter;
} Stats;
typedef struct AecCore AecCore;
AecCore* WebRtcAec_CreateAec(); // Returns NULL on error.
void WebRtcAec_FreeAec(AecCore* aec);
int WebRtcAec_InitAec(AecCore* aec, int sampFreq);
void WebRtcAec_InitAec_SSE2(void);
#if defined(MIPS_FPU_LE)
void WebRtcAec_InitAec_mips(void);
#endif
#if defined(WEBRTC_DETECT_NEON) || defined(WEBRTC_HAS_NEON)
void WebRtcAec_InitAec_neon(void);
#endif
void WebRtcAec_BufferFarendPartition(AecCore* aec, const float* farend);
void WebRtcAec_ProcessFrames(AecCore* aec,
const float* const* nearend,
size_t num_bands,
size_t num_samples,
int knownDelay,
float* const* out);
// A helper function to call WebRtc_MoveReadPtr() for all far-end buffers.
// Returns the number of elements moved, and adjusts |system_delay| by the
// corresponding amount in ms.
int WebRtcAec_MoveFarReadPtr(AecCore* aec, int elements);
// Calculates the median, standard deviation and amount of poor values among the
// delay estimates aggregated up to the first call to the function. After that
// first call the metrics are aggregated and updated every second. With poor
// values we mean values that most likely will cause the AEC to perform poorly.
// TODO(bjornv): Consider changing tests and tools to handle constant
// constant aggregation window throughout the session instead.
int WebRtcAec_GetDelayMetricsCore(AecCore* self, int* median, int* std,
float* fraction_poor_delays);
// Returns the echo state (1: echo, 0: no echo).
int WebRtcAec_echo_state(AecCore* self);
// Gets statistics of the echo metrics ERL, ERLE, A_NLP.
void WebRtcAec_GetEchoStats(AecCore* self,
Stats* erl,
Stats* erle,
Stats* a_nlp);
#ifdef WEBRTC_AEC_DEBUG_DUMP
void* WebRtcAec_far_time_buf(AecCore* self);
#endif
// Sets local configuration modes.
void WebRtcAec_SetConfigCore(AecCore* self,
int nlp_mode,
int metrics_mode,
int delay_logging);
// Non-zero enables, zero disables.
void WebRtcAec_enable_delay_agnostic(AecCore* self, int enable);
// Returns non-zero if delay agnostic (i.e., signal based delay estimation) is
// enabled and zero if disabled.
int WebRtcAec_delay_agnostic_enabled(AecCore* self);
// Enables or disables extended filter mode. Non-zero enables, zero disables.
void WebRtcAec_enable_extended_filter(AecCore* self, int enable);
// Returns non-zero if extended filter mode is enabled and zero if disabled.
int WebRtcAec_extended_filter_enabled(AecCore* self);
// Returns the current |system_delay|, i.e., the buffered difference between
// far-end and near-end.
int WebRtcAec_system_delay(AecCore* self);
// Sets the |system_delay| to |value|. Note that if the value is changed
// improperly, there can be a performance regression. So it should be used with
// care.
void WebRtcAec_SetSystemDelay(AecCore* self, int delay);
#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_CORE_H_

202
webrtc/modules/audio_processing/aec/aec_core_internal.h
View File

@ -1,202 +0,0 @@
/*
* Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_CORE_INTERNAL_H_
#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_CORE_INTERNAL_H_
#include "webrtc/common_audio/ring_buffer.h"
#include "webrtc/common_audio/wav_file.h"
#include "webrtc/modules/audio_processing/aec/aec_common.h"
#include "webrtc/modules/audio_processing/aec/aec_core.h"
#include "webrtc/typedefs.h"
// Number of partitions for the extended filter mode. The first one is an enum
// to be used in array declarations, as it represents the maximum filter length.
enum {
kExtendedNumPartitions = 32
};
static const int kNormalNumPartitions = 12;
// Delay estimator constants, used for logging and delay compensation if
// if reported delays are disabled.
enum {
kLookaheadBlocks = 15
};
enum {
// 500 ms for 16 kHz which is equivalent with the limit of reported delays.
kHistorySizeBlocks = 125
};
// Extended filter adaptation parameters.
// TODO(ajm): No narrowband tuning yet.
static const float kExtendedMu = 0.4f;
static const float kExtendedErrorThreshold = 1.0e-6f;
typedef struct PowerLevel {
float sfrsum;
int sfrcounter;
float framelevel;
float frsum;
int frcounter;
float minlevel;
float averagelevel;
} PowerLevel;
struct AecCore {
int farBufWritePos, farBufReadPos;
int knownDelay;
int inSamples, outSamples;
int delayEstCtr;
RingBuffer* nearFrBuf;
RingBuffer* outFrBuf;
RingBuffer* nearFrBufH[NUM_HIGH_BANDS_MAX];
RingBuffer* outFrBufH[NUM_HIGH_BANDS_MAX];
float dBuf[PART_LEN2]; // nearend
float eBuf[PART_LEN2]; // error
float dBufH[NUM_HIGH_BANDS_MAX][PART_LEN2]; // nearend
float xPow[PART_LEN1];
float dPow[PART_LEN1];
float dMinPow[PART_LEN1];
float dInitMinPow[PART_LEN1];
float* noisePow;
float xfBuf[2][kExtendedNumPartitions * PART_LEN1]; // farend fft buffer
float wfBuf[2][kExtendedNumPartitions * PART_LEN1]; // filter fft
complex_t sde[PART_LEN1]; // cross-psd of nearend and error
complex_t sxd[PART_LEN1]; // cross-psd of farend and nearend
// Farend windowed fft buffer.
complex_t xfwBuf[kExtendedNumPartitions * PART_LEN1];
float sx[PART_LEN1], sd[PART_LEN1], se[PART_LEN1]; // far, near, error psd
float hNs[PART_LEN1];
float hNlFbMin, hNlFbLocalMin;
float hNlXdAvgMin;
int hNlNewMin, hNlMinCtr;
float overDrive, overDriveSm;
int nlp_mode;
float outBuf[PART_LEN];
int delayIdx;
short stNearState, echoState;
short divergeState;
int xfBufBlockPos;
RingBuffer* far_buf;
RingBuffer* far_buf_windowed;
int system_delay; // Current system delay buffered in AEC.
int mult; // sampling frequency multiple
int sampFreq;
size_t num_bands;
uint32_t seed;
float normal_mu; // stepsize
float normal_error_threshold; // error threshold
int noiseEstCtr;
PowerLevel farlevel;
PowerLevel nearlevel;
PowerLevel linoutlevel;
PowerLevel nlpoutlevel;
int metricsMode;
int stateCounter;
Stats erl;
Stats erle;
Stats aNlp;
Stats rerl;
// Quantities to control H band scaling for SWB input
int freq_avg_ic; // initial bin for averaging nlp gain
int flag_Hband_cn; // for comfort noise
float cn_scale_Hband; // scale for comfort noise in H band
int delay_metrics_delivered;
int delay_histogram[kHistorySizeBlocks];
int num_delay_values;
int delay_median;
int delay_std;
float fraction_poor_delays;
int delay_logging_enabled;
void* delay_estimator_farend;
void* delay_estimator;
// Variables associated with delay correction through signal based delay
// estimation feedback.
int signal_delay_correction;
int previous_delay;
int delay_correction_count;
int shift_offset;
float delay_quality_threshold;
int frame_count;
// 0 = delay agnostic mode (signal based delay correction) disabled.
// Otherwise enabled.
int delay_agnostic_enabled;
// 1 = extended filter mode enabled, 0 = disabled.
int extended_filter_enabled;
// Runtime selection of number of filter partitions.
int num_partitions;
#ifdef WEBRTC_AEC_DEBUG_DUMP
// Sequence number of this AEC instance, so that different instances can
// choose different dump file names.
int instance_index;
// Number of times we've restarted dumping; used to pick new dump file names
// each time.
int debug_dump_count;
RingBuffer* far_time_buf;
rtc_WavWriter* farFile;
rtc_WavWriter* nearFile;
rtc_WavWriter* outFile;
rtc_WavWriter* outLinearFile;
FILE* e_fft_file;
#endif
};
typedef void (*WebRtcAecFilterFar)(AecCore* aec, float yf[2][PART_LEN1]);
extern WebRtcAecFilterFar WebRtcAec_FilterFar;
typedef void (*WebRtcAecScaleErrorSignal)(AecCore* aec, float ef[2][PART_LEN1]);
extern WebRtcAecScaleErrorSignal WebRtcAec_ScaleErrorSignal;
typedef void (*WebRtcAecFilterAdaptation)(AecCore* aec,
float* fft,
float ef[2][PART_LEN1]);
extern WebRtcAecFilterAdaptation WebRtcAec_FilterAdaptation;
typedef void (*WebRtcAecOverdriveAndSuppress)(AecCore* aec,
float hNl[PART_LEN1],
const float hNlFb,
float efw[2][PART_LEN1]);
extern WebRtcAecOverdriveAndSuppress WebRtcAec_OverdriveAndSuppress;
typedef void (*WebRtcAecComfortNoise)(AecCore* aec,
float efw[2][PART_LEN1],
complex_t* comfortNoiseHband,
const float* noisePow,
const float* lambda);
extern WebRtcAecComfortNoise WebRtcAec_ComfortNoise;
typedef void (*WebRtcAecSubBandCoherence)(AecCore* aec,
float efw[2][PART_LEN1],
float xfw[2][PART_LEN1],
float* fft,
float* cohde,
float* cohxd);
extern WebRtcAecSubBandCoherence WebRtcAec_SubbandCoherence;
#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_CORE_INTERNAL_H_

774
webrtc/modules/audio_processing/aec/aec_core_mips.c
View File

@ -1,774 +0,0 @@
/*
* Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
/*
* The core AEC algorithm, which is presented with time-aligned signals.
*/
#include "webrtc/modules/audio_processing/aec/aec_core.h"
#include <math.h>
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include "webrtc/modules/audio_processing/aec/aec_core_internal.h"
#include "webrtc/modules/audio_processing/aec/aec_rdft.h"
static const int flagHbandCn = 1; // flag for adding comfort noise in H band
extern const float WebRtcAec_weightCurve[65];
extern const float WebRtcAec_overDriveCurve[65];
void WebRtcAec_ComfortNoise_mips(AecCore* aec,
float efw[2][PART_LEN1],
complex_t* comfortNoiseHband,
const float* noisePow,
const float* lambda) {
int i, num;
float rand[PART_LEN];
float noise, noiseAvg, tmp, tmpAvg;
int16_t randW16[PART_LEN];
complex_t u[PART_LEN1];
const float pi2 = 6.28318530717959f;
const float pi2t = pi2 / 32768;
// Generate a uniform random array on [0 1]
WebRtcSpl_RandUArray(randW16, PART_LEN, &aec->seed);
int16_t* randWptr = randW16;
float randTemp, randTemp2, randTemp3, randTemp4;
int32_t tmp1s, tmp2s, tmp3s, tmp4s;
for (i = 0; i < PART_LEN; i+=4) {
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"lh %[tmp1s], 0(%[randWptr]) \n\t"
"lh %[tmp2s], 2(%[randWptr]) \n\t"
"lh %[tmp3s], 4(%[randWptr]) \n\t"
"lh %[tmp4s], 6(%[randWptr]) \n\t"
"mtc1 %[tmp1s], %[randTemp] \n\t"
"mtc1 %[tmp2s], %[randTemp2] \n\t"
"mtc1 %[tmp3s], %[randTemp3] \n\t"
"mtc1 %[tmp4s], %[randTemp4] \n\t"
"cvt.s.w %[randTemp], %[randTemp] \n\t"
"cvt.s.w %[randTemp2], %[randTemp2] \n\t"
"cvt.s.w %[randTemp3], %[randTemp3] \n\t"
"cvt.s.w %[randTemp4], %[randTemp4] \n\t"
"addiu %[randWptr], %[randWptr], 8 \n\t"
"mul.s %[randTemp], %[randTemp], %[pi2t] \n\t"
"mul.s %[randTemp2], %[randTemp2], %[pi2t] \n\t"
"mul.s %[randTemp3], %[randTemp3], %[pi2t] \n\t"
"mul.s %[randTemp4], %[randTemp4], %[pi2t] \n\t"
".set pop \n\t"
: [randWptr] "+r" (randWptr), [randTemp] "=&f" (randTemp),
[randTemp2] "=&f" (randTemp2), [randTemp3] "=&f" (randTemp3),
[randTemp4] "=&f" (randTemp4), [tmp1s] "=&r" (tmp1s),
[tmp2s] "=&r" (tmp2s), [tmp3s] "=&r" (tmp3s),
[tmp4s] "=&r" (tmp4s)
: [pi2t] "f" (pi2t)
: "memory"
);
u[i+1][0] = cosf(randTemp);
u[i+1][1] = sinf(randTemp);
u[i+2][0] = cosf(randTemp2);
u[i+2][1] = sinf(randTemp2);
u[i+3][0] = cosf(randTemp3);
u[i+3][1] = sinf(randTemp3);
u[i+4][0] = cosf(randTemp4);
u[i+4][1] = sinf(randTemp4);
}
// Reject LF noise
float* u_ptr = &u[1][0];
float noise2, noise3, noise4;
float tmp1f, tmp2f, tmp3f, tmp4f, tmp5f, tmp6f, tmp7f, tmp8f;
u[0][0] = 0;
u[0][1] = 0;
for (i = 1; i < PART_LEN1; i+=4) {
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"lwc1 %[noise], 4(%[noisePow]) \n\t"
"lwc1 %[noise2], 8(%[noisePow]) \n\t"
"lwc1 %[noise3], 12(%[noisePow]) \n\t"
"lwc1 %[noise4], 16(%[noisePow]) \n\t"
"sqrt.s %[noise], %[noise] \n\t"
"sqrt.s %[noise2], %[noise2] \n\t"
"sqrt.s %[noise3], %[noise3] \n\t"
"sqrt.s %[noise4], %[noise4] \n\t"
"lwc1 %[tmp1f], 0(%[u_ptr]) \n\t"
"lwc1 %[tmp2f], 4(%[u_ptr]) \n\t"
"lwc1 %[tmp3f], 8(%[u_ptr]) \n\t"
"lwc1 %[tmp4f], 12(%[u_ptr]) \n\t"
"lwc1 %[tmp5f], 16(%[u_ptr]) \n\t"
"lwc1 %[tmp6f], 20(%[u_ptr]) \n\t"
"lwc1 %[tmp7f], 24(%[u_ptr]) \n\t"
"lwc1 %[tmp8f], 28(%[u_ptr]) \n\t"
"addiu %[noisePow], %[noisePow], 16 \n\t"
"mul.s %[tmp1f], %[tmp1f], %[noise] \n\t"
"mul.s %[tmp2f], %[tmp2f], %[noise] \n\t"
"mul.s %[tmp3f], %[tmp3f], %[noise2] \n\t"
"mul.s %[tmp4f], %[tmp4f], %[noise2] \n\t"
"mul.s %[tmp5f], %[tmp5f], %[noise3] \n\t"
"mul.s %[tmp6f], %[tmp6f], %[noise3] \n\t"
"swc1 %[tmp1f], 0(%[u_ptr]) \n\t"
"swc1 %[tmp3f], 8(%[u_ptr]) \n\t"
"mul.s %[tmp8f], %[tmp8f], %[noise4] \n\t"
"mul.s %[tmp7f], %[tmp7f], %[noise4] \n\t"
"neg.s %[tmp2f] \n\t"
"neg.s %[tmp4f] \n\t"
"neg.s %[tmp6f] \n\t"
"neg.s %[tmp8f] \n\t"
"swc1 %[tmp5f], 16(%[u_ptr]) \n\t"
"swc1 %[tmp7f], 24(%[u_ptr]) \n\t"
"swc1 %[tmp2f], 4(%[u_ptr]) \n\t"
"swc1 %[tmp4f], 12(%[u_ptr]) \n\t"
"swc1 %[tmp6f], 20(%[u_ptr]) \n\t"
"swc1 %[tmp8f], 28(%[u_ptr]) \n\t"
"addiu %[u_ptr], %[u_ptr], 32 \n\t"
".set pop \n\t"
: [u_ptr] "+r" (u_ptr), [noisePow] "+r" (noisePow),
[noise] "=&f" (noise), [noise2] "=&f" (noise2),
[noise3] "=&f" (noise3), [noise4] "=&f" (noise4),
[tmp1f] "=&f" (tmp1f), [tmp2f] "=&f" (tmp2f),
[tmp3f] "=&f" (tmp3f), [tmp4f] "=&f" (tmp4f),
[tmp5f] "=&f" (tmp5f), [tmp6f] "=&f" (tmp6f),
[tmp7f] "=&f" (tmp7f), [tmp8f] "=&f" (tmp8f)
:
: "memory"
);
}
u[PART_LEN][1] = 0;
noisePow -= PART_LEN;
u_ptr = &u[0][0];
float* u_ptr_end = &u[PART_LEN][0];
float* efw_ptr_0 = &efw[0][0];
float* efw_ptr_1 = &efw[1][0];
float tmp9f, tmp10f;
const float tmp1c = 1.0;
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"1: \n\t"
"lwc1 %[tmp1f], 0(%[lambda]) \n\t"
"lwc1 %[tmp6f], 4(%[lambda]) \n\t"
"addiu %[lambda], %[lambda], 8 \n\t"
"c.lt.s %[tmp1f], %[tmp1c] \n\t"
"bc1f 4f \n\t"
" nop \n\t"
"c.lt.s %[tmp6f], %[tmp1c] \n\t"
"bc1f 3f \n\t"
" nop \n\t"
"2: \n\t"
"mul.s %[tmp1f], %[tmp1f], %[tmp1f] \n\t"
"mul.s %[tmp6f], %[tmp6f], %[tmp6f] \n\t"
"sub.s %[tmp1f], %[tmp1c], %[tmp1f] \n\t"
"sub.s %[tmp6f], %[tmp1c], %[tmp6f] \n\t"
"sqrt.s %[tmp1f], %[tmp1f] \n\t"
"sqrt.s %[tmp6f], %[tmp6f] \n\t"
"lwc1 %[tmp2f], 0(%[efw_ptr_0]) \n\t"
"lwc1 %[tmp3f], 0(%[u_ptr]) \n\t"
"lwc1 %[tmp7f], 4(%[efw_ptr_0]) \n\t"
"lwc1 %[tmp8f], 8(%[u_ptr]) \n\t"
"lwc1 %[tmp4f], 0(%[efw_ptr_1]) \n\t"
"lwc1 %[tmp5f], 4(%[u_ptr]) \n\t"
"lwc1 %[tmp9f], 4(%[efw_ptr_1]) \n\t"
"lwc1 %[tmp10f], 12(%[u_ptr]) \n\t"
#if !defined(MIPS32_R2_LE)
"mul.s %[tmp3f], %[tmp1f], %[tmp3f] \n\t"
"add.s %[tmp2f], %[tmp2f], %[tmp3f] \n\t"
"mul.s %[tmp3f], %[tmp1f], %[tmp5f] \n\t"
"add.s %[tmp4f], %[tmp4f], %[tmp3f] \n\t"
"mul.s %[tmp3f], %[tmp6f], %[tmp8f] \n\t"
"add.s %[tmp7f], %[tmp7f], %[tmp3f] \n\t"
"mul.s %[tmp3f], %[tmp6f], %[tmp10f] \n\t"
"add.s %[tmp9f], %[tmp9f], %[tmp3f] \n\t"
#else // #if !defined(MIPS32_R2_LE)
"madd.s %[tmp2f], %[tmp2f], %[tmp1f], %[tmp3f] \n\t"
"madd.s %[tmp4f], %[tmp4f], %[tmp1f], %[tmp5f] \n\t"
"madd.s %[tmp7f], %[tmp7f], %[tmp6f], %[tmp8f] \n\t"
"madd.s %[tmp9f], %[tmp9f], %[tmp6f], %[tmp10f] \n\t"
#endif // #if !defined(MIPS32_R2_LE)
"swc1 %[tmp2f], 0(%[efw_ptr_0]) \n\t"
"swc1 %[tmp4f], 0(%[efw_ptr_1]) \n\t"
"swc1 %[tmp7f], 4(%[efw_ptr_0]) \n\t"
"b 5f \n\t"
" swc1 %[tmp9f], 4(%[efw_ptr_1]) \n\t"
"3: \n\t"
"mul.s %[tmp1f], %[tmp1f], %[tmp1f] \n\t"
"sub.s %[tmp1f], %[tmp1c], %[tmp1f] \n\t"
"sqrt.s %[tmp1f], %[tmp1f] \n\t"
"lwc1 %[tmp2f], 0(%[efw_ptr_0]) \n\t"
"lwc1 %[tmp3f], 0(%[u_ptr]) \n\t"
"lwc1 %[tmp4f], 0(%[efw_ptr_1]) \n\t"
"lwc1 %[tmp5f], 4(%[u_ptr]) \n\t"
#if !defined(MIPS32_R2_LE)
"mul.s %[tmp3f], %[tmp1f], %[tmp3f] \n\t"
"add.s %[tmp2f], %[tmp2f], %[tmp3f] \n\t"
"mul.s %[tmp3f], %[tmp1f], %[tmp5f] \n\t"
"add.s %[tmp4f], %[tmp4f], %[tmp3f] \n\t"
#else // #if !defined(MIPS32_R2_LE)
"madd.s %[tmp2f], %[tmp2f], %[tmp1f], %[tmp3f] \n\t"
"madd.s %[tmp4f], %[tmp4f], %[tmp1f], %[tmp5f] \n\t"
#endif // #if !defined(MIPS32_R2_LE)
"swc1 %[tmp2f], 0(%[efw_ptr_0]) \n\t"
"b 5f \n\t"
" swc1 %[tmp4f], 0(%[efw_ptr_1]) \n\t"
"4: \n\t"
"c.lt.s %[tmp6f], %[tmp1c] \n\t"
"bc1f 5f \n\t"
" nop \n\t"
"mul.s %[tmp6f], %[tmp6f], %[tmp6f] \n\t"
"sub.s %[tmp6f], %[tmp1c], %[tmp6f] \n\t"
"sqrt.s %[tmp6f], %[tmp6f] \n\t"
"lwc1 %[tmp7f], 4(%[efw_ptr_0]) \n\t"
"lwc1 %[tmp8f], 8(%[u_ptr]) \n\t"
"lwc1 %[tmp9f], 4(%[efw_ptr_1]) \n\t"
"lwc1 %[tmp10f], 12(%[u_ptr]) \n\t"
#if !defined(MIPS32_R2_LE)
"mul.s %[tmp3f], %[tmp6f], %[tmp8f] \n\t"
"add.s %[tmp7f], %[tmp7f], %[tmp3f] \n\t"
"mul.s %[tmp3f], %[tmp6f], %[tmp10f] \n\t"
"add.s %[tmp9f], %[tmp9f], %[tmp3f] \n\t"
#else // #if !defined(MIPS32_R2_LE)
"madd.s %[tmp7f], %[tmp7f], %[tmp6f], %[tmp8f] \n\t"
"madd.s %[tmp9f], %[tmp9f], %[tmp6f], %[tmp10f] \n\t"
#endif // #if !defined(MIPS32_R2_LE)
"swc1 %[tmp7f], 4(%[efw_ptr_0]) \n\t"
"swc1 %[tmp9f], 4(%[efw_ptr_1]) \n\t"
"5: \n\t"
"addiu %[u_ptr], %[u_ptr], 16 \n\t"
"addiu %[efw_ptr_0], %[efw_ptr_0], 8 \n\t"
"bne %[u_ptr], %[u_ptr_end], 1b \n\t"
" addiu %[efw_ptr_1], %[efw_ptr_1], 8 \n\t"
".set pop \n\t"
: [lambda] "+r" (lambda), [u_ptr] "+r" (u_ptr),
[efw_ptr_0] "+r" (efw_ptr_0), [efw_ptr_1] "+r" (efw_ptr_1),
[tmp1f] "=&f" (tmp1f), [tmp2f] "=&f" (tmp2f), [tmp3f] "=&f" (tmp3f),
[tmp4f] "=&f" (tmp4f), [tmp5f] "=&f" (tmp5f),
[tmp6f] "=&f" (tmp6f), [tmp7f] "=&f" (tmp7f), [tmp8f] "=&f" (tmp8f),
[tmp9f] "=&f" (tmp9f), [tmp10f] "=&f" (tmp10f)
: [tmp1c] "f" (tmp1c), [u_ptr_end] "r" (u_ptr_end)
: "memory"
);
lambda -= PART_LEN;
tmp = sqrtf(WEBRTC_SPL_MAX(1 - lambda[PART_LEN] * lambda[PART_LEN], 0));
//tmp = 1 - lambda[i];
efw[0][PART_LEN] += tmp * u[PART_LEN][0];
efw[1][PART_LEN] += tmp * u[PART_LEN][1];
// For H band comfort noise
// TODO: don't compute noise and "tmp" twice. Use the previous results.
noiseAvg = 0.0;
tmpAvg = 0.0;
num = 0;
if ((aec->sampFreq == 32000 || aec->sampFreq == 48000) && flagHbandCn == 1) {
for (i = 0; i < PART_LEN; i++) {
rand[i] = ((float)randW16[i]) / 32768;
}
// average noise scale
// average over second half of freq spectrum (i.e., 4->8khz)
// TODO: we shouldn't need num. We know how many elements we're summing.
for (i = PART_LEN1 >> 1; i < PART_LEN1; i++) {
num++;
noiseAvg += sqrtf(noisePow[i]);
}
noiseAvg /= (float)num;
// average nlp scale
// average over second half of freq spectrum (i.e., 4->8khz)
// TODO: we shouldn't need num. We know how many elements we're summing.
num = 0;
for (i = PART_LEN1 >> 1; i < PART_LEN1; i++) {
num++;
tmpAvg += sqrtf(WEBRTC_SPL_MAX(1 - lambda[i] * lambda[i], 0));
}
tmpAvg /= (float)num;
// Use average noise for H band
// TODO: we should probably have a new random vector here.
// Reject LF noise
u[0][0] = 0;
u[0][1] = 0;
for (i = 1; i < PART_LEN1; i++) {
tmp = pi2 * rand[i - 1];
// Use average noise for H band
u[i][0] = noiseAvg * (float)cos(tmp);
u[i][1] = -noiseAvg * (float)sin(tmp);
}
u[PART_LEN][1] = 0;
for (i = 0; i < PART_LEN1; i++) {
// Use average NLP weight for H band
comfortNoiseHband[i][0] = tmpAvg * u[i][0];
comfortNoiseHband[i][1] = tmpAvg * u[i][1];
}
}
}
void WebRtcAec_FilterFar_mips(AecCore* aec, float yf[2][PART_LEN1]) {
int i;
for (i = 0; i < aec->num_partitions; i++) {
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1;
// Check for wrap
if (i + aec->xfBufBlockPos >= aec->num_partitions) {
xPos -= aec->num_partitions * (PART_LEN1);
}
float* yf0 = yf[0];
float* yf1 = yf[1];
float* aRe = aec->xfBuf[0] + xPos;
float* aIm = aec->xfBuf[1] + xPos;
float* bRe = aec->wfBuf[0] + pos;
float* bIm = aec->wfBuf[1] + pos;
float f0, f1, f2, f3, f4, f5, f6, f7, f8, f9, f10, f11, f12, f13;
int len = PART_LEN1 >> 1;
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"1: \n\t"
"lwc1 %[f0], 0(%[aRe]) \n\t"
"lwc1 %[f1], 0(%[bRe]) \n\t"
"lwc1 %[f2], 0(%[bIm]) \n\t"
"lwc1 %[f3], 0(%[aIm]) \n\t"
"lwc1 %[f4], 4(%[aRe]) \n\t"
"lwc1 %[f5], 4(%[bRe]) \n\t"
"lwc1 %[f6], 4(%[bIm]) \n\t"
"mul.s %[f8], %[f0], %[f1] \n\t"
"mul.s %[f0], %[f0], %[f2] \n\t"
"mul.s %[f9], %[f4], %[f5] \n\t"
"mul.s %[f4], %[f4], %[f6] \n\t"
"lwc1 %[f7], 4(%[aIm]) \n\t"
#if !defined(MIPS32_R2_LE)
"mul.s %[f12], %[f2], %[f3] \n\t"
"mul.s %[f1], %[f3], %[f1] \n\t"
"mul.s %[f11], %[f6], %[f7] \n\t"
"addiu %[aRe], %[aRe], 8 \n\t"
"addiu %[aIm], %[aIm], 8 \n\t"
"addiu %[len], %[len], -1 \n\t"
"sub.s %[f8], %[f8], %[f12] \n\t"
"mul.s %[f12], %[f7], %[f5] \n\t"
"lwc1 %[f2], 0(%[yf0]) \n\t"
"add.s %[f1], %[f0], %[f1] \n\t"
"lwc1 %[f3], 0(%[yf1]) \n\t"
"sub.s %[f9], %[f9], %[f11] \n\t"
"lwc1 %[f6], 4(%[yf0]) \n\t"
"add.s %[f4], %[f4], %[f12] \n\t"
#else // #if !defined(MIPS32_R2_LE)
"addiu %[aRe], %[aRe], 8 \n\t"
"addiu %[aIm], %[aIm], 8 \n\t"
"addiu %[len], %[len], -1 \n\t"
"nmsub.s %[f8], %[f8], %[f2], %[f3] \n\t"
"lwc1 %[f2], 0(%[yf0]) \n\t"
"madd.s %[f1], %[f0], %[f3], %[f1] \n\t"
"lwc1 %[f3], 0(%[yf1]) \n\t"
"nmsub.s %[f9], %[f9], %[f6], %[f7] \n\t"
"lwc1 %[f6], 4(%[yf0]) \n\t"
"madd.s %[f4], %[f4], %[f7], %[f5] \n\t"
#endif // #if !defined(MIPS32_R2_LE)
"lwc1 %[f5], 4(%[yf1]) \n\t"
"add.s %[f2], %[f2], %[f8] \n\t"
"addiu %[bRe], %[bRe], 8 \n\t"
"addiu %[bIm], %[bIm], 8 \n\t"
"add.s %[f3], %[f3], %[f1] \n\t"
"add.s %[f6], %[f6], %[f9] \n\t"
"add.s %[f5], %[f5], %[f4] \n\t"
"swc1 %[f2], 0(%[yf0]) \n\t"
"swc1 %[f3], 0(%[yf1]) \n\t"
"swc1 %[f6], 4(%[yf0]) \n\t"
"swc1 %[f5], 4(%[yf1]) \n\t"
"addiu %[yf0], %[yf0], 8 \n\t"
"bgtz %[len], 1b \n\t"
" addiu %[yf1], %[yf1], 8 \n\t"
"lwc1 %[f0], 0(%[aRe]) \n\t"
"lwc1 %[f1], 0(%[bRe]) \n\t"
"lwc1 %[f2], 0(%[bIm]) \n\t"
"lwc1 %[f3], 0(%[aIm]) \n\t"
"mul.s %[f8], %[f0], %[f1] \n\t"
"mul.s %[f0], %[f0], %[f2] \n\t"
#if !defined(MIPS32_R2_LE)
"mul.s %[f12], %[f2], %[f3] \n\t"
"mul.s %[f1], %[f3], %[f1] \n\t"
"sub.s %[f8], %[f8], %[f12] \n\t"
"lwc1 %[f2], 0(%[yf0]) \n\t"
"add.s %[f1], %[f0], %[f1] \n\t"
"lwc1 %[f3], 0(%[yf1]) \n\t"
#else // #if !defined(MIPS32_R2_LE)
"nmsub.s %[f8], %[f8], %[f2], %[f3] \n\t"
"lwc1 %[f2], 0(%[yf0]) \n\t"
"madd.s %[f1], %[f0], %[f3], %[f1] \n\t"
"lwc1 %[f3], 0(%[yf1]) \n\t"
#endif // #if !defined(MIPS32_R2_LE)
"add.s %[f2], %[f2], %[f8] \n\t"
"add.s %[f3], %[f3], %[f1] \n\t"
"swc1 %[f2], 0(%[yf0]) \n\t"
"swc1 %[f3], 0(%[yf1]) \n\t"
".set pop \n\t"
: [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2),
[f3] "=&f" (f3), [f4] "=&f" (f4), [f5] "=&f" (f5),
[f6] "=&f" (f6), [f7] "=&f" (f7), [f8] "=&f" (f8),
[f9] "=&f" (f9), [f10] "=&f" (f10), [f11] "=&f" (f11),
[f12] "=&f" (f12), [f13] "=&f" (f13), [aRe] "+r" (aRe),
[aIm] "+r" (aIm), [bRe] "+r" (bRe), [bIm] "+r" (bIm),
[yf0] "+r" (yf0), [yf1] "+r" (yf1), [len] "+r" (len)
:
: "memory"
);
}
}
void WebRtcAec_FilterAdaptation_mips(AecCore* aec,
float* fft,
float ef[2][PART_LEN1]) {
int i;
for (i = 0; i < aec->num_partitions; i++) {
int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1);
int pos;
// Check for wrap
if (i + aec->xfBufBlockPos >= aec->num_partitions) {
xPos -= aec->num_partitions * PART_LEN1;
}
pos = i * PART_LEN1;
float* aRe = aec->xfBuf[0] + xPos;
float* aIm = aec->xfBuf[1] + xPos;
float* bRe = ef[0];
float* bIm = ef[1];
float* fft_tmp;
float f0, f1, f2, f3, f4, f5, f6 ,f7, f8, f9, f10, f11, f12;
int len = PART_LEN >> 1;
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"addiu %[fft_tmp], %[fft], 0 \n\t"
"1: \n\t"
"lwc1 %[f0], 0(%[aRe]) \n\t"
"lwc1 %[f1], 0(%[bRe]) \n\t"
"lwc1 %[f2], 0(%[bIm]) \n\t"
"lwc1 %[f4], 4(%[aRe]) \n\t"
"lwc1 %[f5], 4(%[bRe]) \n\t"
"lwc1 %[f6], 4(%[bIm]) \n\t"
"addiu %[aRe], %[aRe], 8 \n\t"
"addiu %[bRe], %[bRe], 8 \n\t"
"mul.s %[f8], %[f0], %[f1] \n\t"
"mul.s %[f0], %[f0], %[f2] \n\t"
"lwc1 %[f3], 0(%[aIm]) \n\t"
"mul.s %[f9], %[f4], %[f5] \n\t"
"lwc1 %[f7], 4(%[aIm]) \n\t"
"mul.s %[f4], %[f4], %[f6] \n\t"
#if !defined(MIPS32_R2_LE)
"mul.s %[f10], %[f3], %[f2] \n\t"
"mul.s %[f1], %[f3], %[f1] \n\t"
"mul.s %[f11], %[f7], %[f6] \n\t"
"mul.s %[f5], %[f7], %[f5] \n\t"
"addiu %[aIm], %[aIm], 8 \n\t"
"addiu %[bIm], %[bIm], 8 \n\t"
"addiu %[len], %[len], -1 \n\t"
"add.s %[f8], %[f8], %[f10] \n\t"
"sub.s %[f1], %[f0], %[f1] \n\t"
"add.s %[f9], %[f9], %[f11] \n\t"
"sub.s %[f5], %[f4], %[f5] \n\t"
#else // #if !defined(MIPS32_R2_LE)
"addiu %[aIm], %[aIm], 8 \n\t"
"addiu %[bIm], %[bIm], 8 \n\t"
"addiu %[len], %[len], -1 \n\t"
"madd.s %[f8], %[f8], %[f3], %[f2] \n\t"
"nmsub.s %[f1], %[f0], %[f3], %[f1] \n\t"
"madd.s %[f9], %[f9], %[f7], %[f6] \n\t"
"nmsub.s %[f5], %[f4], %[f7], %[f5] \n\t"
#endif // #if !defined(MIPS32_R2_LE)
"swc1 %[f8], 0(%[fft_tmp]) \n\t"
"swc1 %[f1], 4(%[fft_tmp]) \n\t"
"swc1 %[f9], 8(%[fft_tmp]) \n\t"
"swc1 %[f5], 12(%[fft_tmp]) \n\t"
"bgtz %[len], 1b \n\t"
" addiu %[fft_tmp], %[fft_tmp], 16 \n\t"
"lwc1 %[f0], 0(%[aRe]) \n\t"
"lwc1 %[f1], 0(%[bRe]) \n\t"
"lwc1 %[f2], 0(%[bIm]) \n\t"
"lwc1 %[f3], 0(%[aIm]) \n\t"
"mul.s %[f8], %[f0], %[f1] \n\t"
#if !defined(MIPS32_R2_LE)
"mul.s %[f10], %[f3], %[f2] \n\t"
"add.s %[f8], %[f8], %[f10] \n\t"
#else // #if !defined(MIPS32_R2_LE)
"madd.s %[f8], %[f8], %[f3], %[f2] \n\t"
#endif // #if !defined(MIPS32_R2_LE)
"swc1 %[f8], 4(%[fft]) \n\t"
".set pop \n\t"
: [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2),
[f3] "=&f" (f3), [f4] "=&f" (f4), [f5] "=&f" (f5),
[f6] "=&f" (f6), [f7] "=&f" (f7), [f8] "=&f" (f8),
[f9] "=&f" (f9), [f10] "=&f" (f10), [f11] "=&f" (f11),
[f12] "=&f" (f12), [aRe] "+r" (aRe), [aIm] "+r" (aIm),
[bRe] "+r" (bRe), [bIm] "+r" (bIm), [fft_tmp] "=&r" (fft_tmp),
[len] "+r" (len)
: [fft] "r" (fft)
: "memory"
);
aec_rdft_inverse_128(fft);
memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN);
// fft scaling
{
float scale = 2.0f / PART_LEN2;
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"addiu %[fft_tmp], %[fft], 0 \n\t"
"addiu %[len], $zero, 8 \n\t"
"1: \n\t"
"addiu %[len], %[len], -1 \n\t"
"lwc1 %[f0], 0(%[fft_tmp]) \n\t"
"lwc1 %[f1], 4(%[fft_tmp]) \n\t"
"lwc1 %[f2], 8(%[fft_tmp]) \n\t"
"lwc1 %[f3], 12(%[fft_tmp]) \n\t"
"mul.s %[f0], %[f0], %[scale] \n\t"
"mul.s %[f1], %[f1], %[scale] \n\t"
"mul.s %[f2], %[f2], %[scale] \n\t"
"mul.s %[f3], %[f3], %[scale] \n\t"
"lwc1 %[f4], 16(%[fft_tmp]) \n\t"
"lwc1 %[f5], 20(%[fft_tmp]) \n\t"
"lwc1 %[f6], 24(%[fft_tmp]) \n\t"
"lwc1 %[f7], 28(%[fft_tmp]) \n\t"
"mul.s %[f4], %[f4], %[scale] \n\t"
"mul.s %[f5], %[f5], %[scale] \n\t"
"mul.s %[f6], %[f6], %[scale] \n\t"
"mul.s %[f7], %[f7], %[scale] \n\t"
"swc1 %[f0], 0(%[fft_tmp]) \n\t"
"swc1 %[f1], 4(%[fft_tmp]) \n\t"
"swc1 %[f2], 8(%[fft_tmp]) \n\t"
"swc1 %[f3], 12(%[fft_tmp]) \n\t"
"swc1 %[f4], 16(%[fft_tmp]) \n\t"
"swc1 %[f5], 20(%[fft_tmp]) \n\t"
"swc1 %[f6], 24(%[fft_tmp]) \n\t"
"swc1 %[f7], 28(%[fft_tmp]) \n\t"
"bgtz %[len], 1b \n\t"
" addiu %[fft_tmp], %[fft_tmp], 32 \n\t"
".set pop \n\t"
: [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2),
[f3] "=&f" (f3), [f4] "=&f" (f4), [f5] "=&f" (f5),
[f6] "=&f" (f6), [f7] "=&f" (f7), [len] "=&r" (len),
[fft_tmp] "=&r" (fft_tmp)
: [scale] "f" (scale), [fft] "r" (fft)
: "memory"
);
}
aec_rdft_forward_128(fft);
aRe = aec->wfBuf[0] + pos;
aIm = aec->wfBuf[1] + pos;
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"addiu %[fft_tmp], %[fft], 0 \n\t"
"addiu %[len], $zero, 31 \n\t"
"lwc1 %[f0], 0(%[aRe]) \n\t"
"lwc1 %[f1], 0(%[fft_tmp]) \n\t"
"lwc1 %[f2], 256(%[aRe]) \n\t"
"lwc1 %[f3], 4(%[fft_tmp]) \n\t"
"lwc1 %[f4], 4(%[aRe]) \n\t"
"lwc1 %[f5], 8(%[fft_tmp]) \n\t"
"lwc1 %[f6], 4(%[aIm]) \n\t"
"lwc1 %[f7], 12(%[fft_tmp]) \n\t"
"add.s %[f0], %[f0], %[f1] \n\t"
"add.s %[f2], %[f2], %[f3] \n\t"
"add.s %[f4], %[f4], %[f5] \n\t"
"add.s %[f6], %[f6], %[f7] \n\t"
"addiu %[fft_tmp], %[fft_tmp], 16 \n\t"
"swc1 %[f0], 0(%[aRe]) \n\t"
"swc1 %[f2], 256(%[aRe]) \n\t"
"swc1 %[f4], 4(%[aRe]) \n\t"
"addiu %[aRe], %[aRe], 8 \n\t"
"swc1 %[f6], 4(%[aIm]) \n\t"
"addiu %[aIm], %[aIm], 8 \n\t"
"1: \n\t"
"lwc1 %[f0], 0(%[aRe]) \n\t"
"lwc1 %[f1], 0(%[fft_tmp]) \n\t"
"lwc1 %[f2], 0(%[aIm]) \n\t"
"lwc1 %[f3], 4(%[fft_tmp]) \n\t"
"lwc1 %[f4], 4(%[aRe]) \n\t"
"lwc1 %[f5], 8(%[fft_tmp]) \n\t"
"lwc1 %[f6], 4(%[aIm]) \n\t"
"lwc1 %[f7], 12(%[fft_tmp]) \n\t"
"add.s %[f0], %[f0], %[f1] \n\t"
"add.s %[f2], %[f2], %[f3] \n\t"
"add.s %[f4], %[f4], %[f5] \n\t"
"add.s %[f6], %[f6], %[f7] \n\t"
"addiu %[len], %[len], -1 \n\t"
"addiu %[fft_tmp], %[fft_tmp], 16 \n\t"
"swc1 %[f0], 0(%[aRe]) \n\t"
"swc1 %[f2], 0(%[aIm]) \n\t"
"swc1 %[f4], 4(%[aRe]) \n\t"
"addiu %[aRe], %[aRe], 8 \n\t"
"swc1 %[f6], 4(%[aIm]) \n\t"
"bgtz %[len], 1b \n\t"
" addiu %[aIm], %[aIm], 8 \n\t"
".set pop \n\t"
: [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2),
[f3] "=&f" (f3), [f4] "=&f" (f4), [f5] "=&f" (f5),
[f6] "=&f" (f6), [f7] "=&f" (f7), [len] "=&r" (len),
[fft_tmp] "=&r" (fft_tmp), [aRe] "+r" (aRe), [aIm] "+r" (aIm)
: [fft] "r" (fft)
: "memory"
);
}
}
void WebRtcAec_OverdriveAndSuppress_mips(AecCore* aec,
float hNl[PART_LEN1],
const float hNlFb,
float efw[2][PART_LEN1]) {
int i;
const float one = 1.0;
float* p_hNl;
float* p_efw0;
float* p_efw1;
float* p_WebRtcAec_wC;
float temp1, temp2, temp3, temp4;
p_hNl = &hNl[0];
p_efw0 = &efw[0][0];
p_efw1 = &efw[1][0];
p_WebRtcAec_wC = (float*)&WebRtcAec_weightCurve[0];
for (i = 0; i < PART_LEN1; i++) {
// Weight subbands
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"lwc1 %[temp1], 0(%[p_hNl]) \n\t"
"lwc1 %[temp2], 0(%[p_wC]) \n\t"
"c.lt.s %[hNlFb], %[temp1] \n\t"
"bc1f 1f \n\t"
" mul.s %[temp3], %[temp2], %[hNlFb] \n\t"
"sub.s %[temp4], %[one], %[temp2] \n\t"
#if !defined(MIPS32_R2_LE)
"mul.s %[temp1], %[temp1], %[temp4] \n\t"
"add.s %[temp1], %[temp3], %[temp1] \n\t"
#else // #if !defined(MIPS32_R2_LE)
"madd.s %[temp1], %[temp3], %[temp1], %[temp4] \n\t"
#endif // #if !defined(MIPS32_R2_LE)
"swc1 %[temp1], 0(%[p_hNl]) \n\t"
"1: \n\t"
"addiu %[p_wC], %[p_wC], 4 \n\t"
".set pop \n\t"
: [temp1] "=&f" (temp1), [temp2] "=&f" (temp2), [temp3] "=&f" (temp3),
[temp4] "=&f" (temp4), [p_wC] "+r" (p_WebRtcAec_wC)
: [hNlFb] "f" (hNlFb), [one] "f" (one), [p_hNl] "r" (p_hNl)
: "memory"
);
hNl[i] = powf(hNl[i], aec->overDriveSm * WebRtcAec_overDriveCurve[i]);
__asm __volatile (
"lwc1 %[temp1], 0(%[p_hNl]) \n\t"
"lwc1 %[temp3], 0(%[p_efw1]) \n\t"
"lwc1 %[temp2], 0(%[p_efw0]) \n\t"
"addiu %[p_hNl], %[p_hNl], 4 \n\t"
"mul.s %[temp3], %[temp3], %[temp1] \n\t"
"mul.s %[temp2], %[temp2], %[temp1] \n\t"
"addiu %[p_efw0], %[p_efw0], 4 \n\t"
"addiu %[p_efw1], %[p_efw1], 4 \n\t"
"neg.s %[temp4], %[temp3] \n\t"
"swc1 %[temp2], -4(%[p_efw0]) \n\t"
"swc1 %[temp4], -4(%[p_efw1]) \n\t"
: [temp1] "=&f" (temp1), [temp2] "=&f" (temp2), [temp3] "=&f" (temp3),
[temp4] "=&f" (temp4), [p_efw0] "+r" (p_efw0), [p_efw1] "+r" (p_efw1),
[p_hNl] "+r" (p_hNl)
:
: "memory"
);
}
}
void WebRtcAec_ScaleErrorSignal_mips(AecCore* aec, float ef[2][PART_LEN1]) {
const float mu = aec->extended_filter_enabled ? kExtendedMu : aec->normal_mu;
const float error_threshold = aec->extended_filter_enabled
? kExtendedErrorThreshold
: aec->normal_error_threshold;
int len = (PART_LEN1);
float* ef0 = ef[0];
float* ef1 = ef[1];
float* xPow = aec->xPow;
float fac1 = 1e-10f;
float err_th2 = error_threshold * error_threshold;
float f0, f1, f2;
#if !defined(MIPS32_R2_LE)
float f3;
#endif
__asm __volatile (
".set push \n\t"
".set noreorder \n\t"
"1: \n\t"
"lwc1 %[f0], 0(%[xPow]) \n\t"
"lwc1 %[f1], 0(%[ef0]) \n\t"
"lwc1 %[f2], 0(%[ef1]) \n\t"
"add.s %[f0], %[f0], %[fac1] \n\t"
"div.s %[f1], %[f1], %[f0] \n\t"
"div.s %[f2], %[f2], %[f0] \n\t"
"mul.s %[f0], %[f1], %[f1] \n\t"
#if defined(MIPS32_R2_LE)
"madd.s %[f0], %[f0], %[f2], %[f2] \n\t"
#else
"mul.s %[f3], %[f2], %[f2] \n\t"
"add.s %[f0], %[f0], %[f3] \n\t"
#endif
"c.le.s %[f0], %[err_th2] \n\t"
"nop \n\t"
"bc1t 2f \n\t"
" nop \n\t"
"sqrt.s %[f0], %[f0] \n\t"
"add.s %[f0], %[f0], %[fac1] \n\t"
"div.s %[f0], %[err_th], %[f0] \n\t"
"mul.s %[f1], %[f1], %[f0] \n\t"
"mul.s %[f2], %[f2], %[f0] \n\t"
"2: \n\t"
"mul.s %[f1], %[f1], %[mu] \n\t"
"mul.s %[f2], %[f2], %[mu] \n\t"
"swc1 %[f1], 0(%[ef0]) \n\t"
"swc1 %[f2], 0(%[ef1]) \n\t"
"addiu %[len], %[len], -1 \n\t"
"addiu %[xPow], %[xPow], 4 \n\t"
"addiu %[ef0], %[ef0], 4 \n\t"
"bgtz %[len], 1b \n\t"
" addiu %[ef1], %[ef1], 4 \n\t"
".set pop \n\t"
: [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2),
#if !defined(MIPS32_R2_LE)
[f3] "=&f" (f3),
#endif
[xPow] "+r" (xPow), [ef0] "+r" (ef0), [ef1] "+r" (ef1),
[len] "+r" (len)
: [fac1] "f" (fac1), [err_th2] "f" (err_th2), [mu] "f" (mu),
[err_th] "f" (error_threshold)
: "memory"
);
}
void WebRtcAec_InitAec_mips(void) {
WebRtcAec_FilterFar = WebRtcAec_FilterFar_mips;
WebRtcAec_FilterAdaptation = WebRtcAec_FilterAdaptation_mips;
WebRtcAec_ScaleErrorSignal = WebRtcAec_ScaleErrorSignal_mips;
WebRtcAec_ComfortNoise = WebRtcAec_ComfortNoise_mips;
WebRtcAec_OverdriveAndSuppress = WebRtcAec_OverdriveAndSuppress_mips;
}

736
webrtc/modules/audio_processing/aec/aec_core_neon.c
View File

@ -1,736 +0,0 @@
/*
* Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
/*
* The core AEC algorithm, neon version of speed-critical functions.
*
* Based on aec_core_sse2.c.
*/
#include <arm_neon.h>
#include <math.h>
#include <string.h> // memset
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include "webrtc/modules/audio_processing/aec/aec_common.h"
#include "webrtc/modules/audio_processing/aec/aec_core_internal.h"
#include "webrtc/modules/audio_processing/aec/aec_rdft.h"
enum { kShiftExponentIntoTopMantissa = 8 };
enum { kFloatExponentShift = 23 };
__inline static float MulRe(float aRe, float aIm, float bRe, float bIm) {
return aRe * bRe - aIm * bIm;
}
__inline static float MulIm(float aRe, float aIm, float bRe, float bIm) {
return aRe * bIm + aIm * bRe;
}
static void FilterFarNEON(AecCore* aec, float yf[2][PART_LEN1]) {
int i;
const int num_partitions = aec->num_partitions;
for (i = 0; i < num_partitions; i++) {
int j;
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1;
// Check for wrap
if (i + aec->xfBufBlockPos >= num_partitions) {
xPos -= num_partitions * PART_LEN1;
}
// vectorized code (four at once)
for (j = 0; j + 3 < PART_LEN1; j += 4) {
const float32x4_t xfBuf_re = vld1q_f32(&aec->xfBuf[0][xPos + j]);
const float32x4_t xfBuf_im = vld1q_f32(&aec->xfBuf[1][xPos + j]);
const float32x4_t wfBuf_re = vld1q_f32(&aec->wfBuf[0][pos + j]);
const float32x4_t wfBuf_im = vld1q_f32(&aec->wfBuf[1][pos + j]);
const float32x4_t yf_re = vld1q_f32(&yf[0][j]);
const float32x4_t yf_im = vld1q_f32(&yf[1][j]);
const float32x4_t a = vmulq_f32(xfBuf_re, wfBuf_re);
const float32x4_t e = vmlsq_f32(a, xfBuf_im, wfBuf_im);
const float32x4_t c = vmulq_f32(xfBuf_re, wfBuf_im);
const float32x4_t f = vmlaq_f32(c, xfBuf_im, wfBuf_re);
const float32x4_t g = vaddq_f32(yf_re, e);
const float32x4_t h = vaddq_f32(yf_im, f);
vst1q_f32(&yf[0][j], g);
vst1q_f32(&yf[1][j], h);
}
// scalar code for the remaining items.
for (; j < PART_LEN1; j++) {
yf[0][j] += MulRe(aec->xfBuf[0][xPos + j],
aec->xfBuf[1][xPos + j],
aec->wfBuf[0][pos + j],
aec->wfBuf[1][pos + j]);
yf[1][j] += MulIm(aec->xfBuf[0][xPos + j],
aec->xfBuf[1][xPos + j],
aec->wfBuf[0][pos + j],
aec->wfBuf[1][pos + j]);
}
}
}
// ARM64's arm_neon.h has already defined vdivq_f32 vsqrtq_f32.
#if !defined (WEBRTC_ARCH_ARM64)
static float32x4_t vdivq_f32(float32x4_t a, float32x4_t b) {
int i;
float32x4_t x = vrecpeq_f32(b);
// from arm documentation
// The Newton-Raphson iteration:
// x[n+1] = x[n] * (2 - d * x[n])
// converges to (1/d) if x0 is the result of VRECPE applied to d.
//
// Note: The precision did not improve after 2 iterations.
for (i = 0; i < 2; i++) {
x = vmulq_f32(vrecpsq_f32(b, x), x);
}
// a/b = a*(1/b)
return vmulq_f32(a, x);
}
static float32x4_t vsqrtq_f32(float32x4_t s) {
int i;
float32x4_t x = vrsqrteq_f32(s);
// Code to handle sqrt(0).
// If the input to sqrtf() is zero, a zero will be returned.
// If the input to vrsqrteq_f32() is zero, positive infinity is returned.
const uint32x4_t vec_p_inf = vdupq_n_u32(0x7F800000);
// check for divide by zero
const uint32x4_t div_by_zero = vceqq_u32(vec_p_inf, vreinterpretq_u32_f32(x));
// zero out the positive infinity results
x = vreinterpretq_f32_u32(vandq_u32(vmvnq_u32(div_by_zero),
vreinterpretq_u32_f32(x)));
// from arm documentation
// The Newton-Raphson iteration:
// x[n+1] = x[n] * (3 - d * (x[n] * x[n])) / 2)
// converges to (1/√d) if x0 is the result of VRSQRTE applied to d.
//
// Note: The precision did not improve after 2 iterations.
for (i = 0; i < 2; i++) {
x = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, x), s), x);
}
// sqrt(s) = s * 1/sqrt(s)
return vmulq_f32(s, x);;
}
#endif // WEBRTC_ARCH_ARM64
static void ScaleErrorSignalNEON(AecCore* aec, float ef[2][PART_LEN1]) {
const float mu = aec->extended_filter_enabled ? kExtendedMu : aec->normal_mu;
const float error_threshold = aec->extended_filter_enabled ?
kExtendedErrorThreshold : aec->normal_error_threshold;
const float32x4_t k1e_10f = vdupq_n_f32(1e-10f);
const float32x4_t kMu = vmovq_n_f32(mu);
const float32x4_t kThresh = vmovq_n_f32(error_threshold);
int i;
// vectorized code (four at once)
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t xPow = vld1q_f32(&aec->xPow[i]);
const float32x4_t ef_re_base = vld1q_f32(&ef[0][i]);
const float32x4_t ef_im_base = vld1q_f32(&ef[1][i]);
const float32x4_t xPowPlus = vaddq_f32(xPow, k1e_10f);
float32x4_t ef_re = vdivq_f32(ef_re_base, xPowPlus);
float32x4_t ef_im = vdivq_f32(ef_im_base, xPowPlus);
const float32x4_t ef_re2 = vmulq_f32(ef_re, ef_re);
const float32x4_t ef_sum2 = vmlaq_f32(ef_re2, ef_im, ef_im);
const float32x4_t absEf = vsqrtq_f32(ef_sum2);
const uint32x4_t bigger = vcgtq_f32(absEf, kThresh);
const float32x4_t absEfPlus = vaddq_f32(absEf, k1e_10f);
const float32x4_t absEfInv = vdivq_f32(kThresh, absEfPlus);
uint32x4_t ef_re_if = vreinterpretq_u32_f32(vmulq_f32(ef_re, absEfInv));
uint32x4_t ef_im_if = vreinterpretq_u32_f32(vmulq_f32(ef_im, absEfInv));
uint32x4_t ef_re_u32 = vandq_u32(vmvnq_u32(bigger),
vreinterpretq_u32_f32(ef_re));
uint32x4_t ef_im_u32 = vandq_u32(vmvnq_u32(bigger),
vreinterpretq_u32_f32(ef_im));
ef_re_if = vandq_u32(bigger, ef_re_if);
ef_im_if = vandq_u32(bigger, ef_im_if);
ef_re_u32 = vorrq_u32(ef_re_u32, ef_re_if);
ef_im_u32 = vorrq_u32(ef_im_u32, ef_im_if);
ef_re = vmulq_f32(vreinterpretq_f32_u32(ef_re_u32), kMu);
ef_im = vmulq_f32(vreinterpretq_f32_u32(ef_im_u32), kMu);
vst1q_f32(&ef[0][i], ef_re);
vst1q_f32(&ef[1][i], ef_im);
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
float abs_ef;
ef[0][i] /= (aec->xPow[i] + 1e-10f);
ef[1][i] /= (aec->xPow[i] + 1e-10f);
abs_ef = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]);
if (abs_ef > error_threshold) {
abs_ef = error_threshold / (abs_ef + 1e-10f);
ef[0][i] *= abs_ef;
ef[1][i] *= abs_ef;
}
// Stepsize factor
ef[0][i] *= mu;
ef[1][i] *= mu;
}
}
static void FilterAdaptationNEON(AecCore* aec,
float* fft,
float ef[2][PART_LEN1]) {
int i;
const int num_partitions = aec->num_partitions;
for (i = 0; i < num_partitions; i++) {
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1;
int j;
// Check for wrap
if (i + aec->xfBufBlockPos >= num_partitions) {
xPos -= num_partitions * PART_LEN1;
}
// Process the whole array...
for (j = 0; j < PART_LEN; j += 4) {
// Load xfBuf and ef.
const float32x4_t xfBuf_re = vld1q_f32(&aec->xfBuf[0][xPos + j]);
const float32x4_t xfBuf_im = vld1q_f32(&aec->xfBuf[1][xPos + j]);
const float32x4_t ef_re = vld1q_f32(&ef[0][j]);
const float32x4_t ef_im = vld1q_f32(&ef[1][j]);
// Calculate the product of conjugate(xfBuf) by ef.
// re(conjugate(a) * b) = aRe * bRe + aIm * bIm
// im(conjugate(a) * b)= aRe * bIm - aIm * bRe
const float32x4_t a = vmulq_f32(xfBuf_re, ef_re);
const float32x4_t e = vmlaq_f32(a, xfBuf_im, ef_im);
const float32x4_t c = vmulq_f32(xfBuf_re, ef_im);
const float32x4_t f = vmlsq_f32(c, xfBuf_im, ef_re);
// Interleave real and imaginary parts.
const float32x4x2_t g_n_h = vzipq_f32(e, f);
// Store
vst1q_f32(&fft[2 * j + 0], g_n_h.val[0]);
vst1q_f32(&fft[2 * j + 4], g_n_h.val[1]);
}
// ... and fixup the first imaginary entry.
fft[1] = MulRe(aec->xfBuf[0][xPos + PART_LEN],
-aec->xfBuf[1][xPos + PART_LEN],
ef[0][PART_LEN],
ef[1][PART_LEN]);
aec_rdft_inverse_128(fft);
memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN);
// fft scaling
{
const float scale = 2.0f / PART_LEN2;
const float32x4_t scale_ps = vmovq_n_f32(scale);
for (j = 0; j < PART_LEN; j += 4) {
const float32x4_t fft_ps = vld1q_f32(&fft[j]);
const float32x4_t fft_scale = vmulq_f32(fft_ps, scale_ps);
vst1q_f32(&fft[j], fft_scale);
}
}
aec_rdft_forward_128(fft);
{
const float wt1 = aec->wfBuf[1][pos];
aec->wfBuf[0][pos + PART_LEN] += fft[1];
for (j = 0; j < PART_LEN; j += 4) {
float32x4_t wtBuf_re = vld1q_f32(&aec->wfBuf[0][pos + j]);
float32x4_t wtBuf_im = vld1q_f32(&aec->wfBuf[1][pos + j]);
const float32x4_t fft0 = vld1q_f32(&fft[2 * j + 0]);
const float32x4_t fft4 = vld1q_f32(&fft[2 * j + 4]);
const float32x4x2_t fft_re_im = vuzpq_f32(fft0, fft4);
wtBuf_re = vaddq_f32(wtBuf_re, fft_re_im.val[0]);
wtBuf_im = vaddq_f32(wtBuf_im, fft_re_im.val[1]);
vst1q_f32(&aec->wfBuf[0][pos + j], wtBuf_re);
vst1q_f32(&aec->wfBuf[1][pos + j], wtBuf_im);
}
aec->wfBuf[1][pos] = wt1;
}
}
}
static float32x4_t vpowq_f32(float32x4_t a, float32x4_t b) {
// a^b = exp2(b * log2(a))
// exp2(x) and log2(x) are calculated using polynomial approximations.
float32x4_t log2_a, b_log2_a, a_exp_b;
// Calculate log2(x), x = a.
{
// To calculate log2(x), we decompose x like this:
// x = y * 2^n
// n is an integer
// y is in the [1.0, 2.0) range
//
// log2(x) = log2(y) + n
// n can be evaluated by playing with float representation.
// log2(y) in a small range can be approximated, this code uses an order
// five polynomial approximation. The coefficients have been
// estimated with the Remez algorithm and the resulting
// polynomial has a maximum relative error of 0.00086%.
// Compute n.
// This is done by masking the exponent, shifting it into the top bit of
// the mantissa, putting eight into the biased exponent (to shift/
// compensate the fact that the exponent has been shifted in the top/
// fractional part and finally getting rid of the implicit leading one
// from the mantissa by substracting it out.
const uint32x4_t vec_float_exponent_mask = vdupq_n_u32(0x7F800000);
const uint32x4_t vec_eight_biased_exponent = vdupq_n_u32(0x43800000);
const uint32x4_t vec_implicit_leading_one = vdupq_n_u32(0x43BF8000);
const uint32x4_t two_n = vandq_u32(vreinterpretq_u32_f32(a),
vec_float_exponent_mask);
const uint32x4_t n_1 = vshrq_n_u32(two_n, kShiftExponentIntoTopMantissa);
const uint32x4_t n_0 = vorrq_u32(n_1, vec_eight_biased_exponent);
const float32x4_t n =
vsubq_f32(vreinterpretq_f32_u32(n_0),
vreinterpretq_f32_u32(vec_implicit_leading_one));
// Compute y.
const uint32x4_t vec_mantissa_mask = vdupq_n_u32(0x007FFFFF);
const uint32x4_t vec_zero_biased_exponent_is_one = vdupq_n_u32(0x3F800000);
const uint32x4_t mantissa = vandq_u32(vreinterpretq_u32_f32(a),
vec_mantissa_mask);
const float32x4_t y =
vreinterpretq_f32_u32(vorrq_u32(mantissa,
vec_zero_biased_exponent_is_one));
// Approximate log2(y) ~= (y - 1) * pol5(y).
// pol5(y) = C5 * y^5 + C4 * y^4 + C3 * y^3 + C2 * y^2 + C1 * y + C0
const float32x4_t C5 = vdupq_n_f32(-3.4436006e-2f);
const float32x4_t C4 = vdupq_n_f32(3.1821337e-1f);
const float32x4_t C3 = vdupq_n_f32(-1.2315303f);
const float32x4_t C2 = vdupq_n_f32(2.5988452f);
const float32x4_t C1 = vdupq_n_f32(-3.3241990f);
const float32x4_t C0 = vdupq_n_f32(3.1157899f);
float32x4_t pol5_y = C5;
pol5_y = vmlaq_f32(C4, y, pol5_y);
pol5_y = vmlaq_f32(C3, y, pol5_y);
pol5_y = vmlaq_f32(C2, y, pol5_y);
pol5_y = vmlaq_f32(C1, y, pol5_y);
pol5_y = vmlaq_f32(C0, y, pol5_y);
const float32x4_t y_minus_one =
vsubq_f32(y, vreinterpretq_f32_u32(vec_zero_biased_exponent_is_one));
const float32x4_t log2_y = vmulq_f32(y_minus_one, pol5_y);
// Combine parts.
log2_a = vaddq_f32(n, log2_y);
}
// b * log2(a)
b_log2_a = vmulq_f32(b, log2_a);
// Calculate exp2(x), x = b * log2(a).
{
// To calculate 2^x, we decompose x like this:
// x = n + y
// n is an integer, the value of x - 0.5 rounded down, therefore
// y is in the [0.5, 1.5) range
//
// 2^x = 2^n * 2^y
// 2^n can be evaluated by playing with float representation.
// 2^y in a small range can be approximated, this code uses an order two
// polynomial approximation. The coefficients have been estimated
// with the Remez algorithm and the resulting polynomial has a
// maximum relative error of 0.17%.
// To avoid over/underflow, we reduce the range of input to ]-127, 129].
const float32x4_t max_input = vdupq_n_f32(129.f);
const float32x4_t min_input = vdupq_n_f32(-126.99999f);
const float32x4_t x_min = vminq_f32(b_log2_a, max_input);
const float32x4_t x_max = vmaxq_f32(x_min, min_input);
// Compute n.
const float32x4_t half = vdupq_n_f32(0.5f);
const float32x4_t x_minus_half = vsubq_f32(x_max, half);
const int32x4_t x_minus_half_floor = vcvtq_s32_f32(x_minus_half);
// Compute 2^n.
const int32x4_t float_exponent_bias = vdupq_n_s32(127);
const int32x4_t two_n_exponent =
vaddq_s32(x_minus_half_floor, float_exponent_bias);
const float32x4_t two_n =
vreinterpretq_f32_s32(vshlq_n_s32(two_n_exponent, kFloatExponentShift));
// Compute y.
const float32x4_t y = vsubq_f32(x_max, vcvtq_f32_s32(x_minus_half_floor));
// Approximate 2^y ~= C2 * y^2 + C1 * y + C0.
const float32x4_t C2 = vdupq_n_f32(3.3718944e-1f);
const float32x4_t C1 = vdupq_n_f32(6.5763628e-1f);
const float32x4_t C0 = vdupq_n_f32(1.0017247f);
float32x4_t exp2_y = C2;
exp2_y = vmlaq_f32(C1, y, exp2_y);
exp2_y = vmlaq_f32(C0, y, exp2_y);
// Combine parts.
a_exp_b = vmulq_f32(exp2_y, two_n);
}
return a_exp_b;
}
static void OverdriveAndSuppressNEON(AecCore* aec,
float hNl[PART_LEN1],
const float hNlFb,
float efw[2][PART_LEN1]) {
int i;
const float32x4_t vec_hNlFb = vmovq_n_f32(hNlFb);
const float32x4_t vec_one = vdupq_n_f32(1.0f);
const float32x4_t vec_minus_one = vdupq_n_f32(-1.0f);
const float32x4_t vec_overDriveSm = vmovq_n_f32(aec->overDriveSm);
// vectorized code (four at once)
for (i = 0; i + 3 < PART_LEN1; i += 4) {
// Weight subbands
float32x4_t vec_hNl = vld1q_f32(&hNl[i]);
const float32x4_t vec_weightCurve = vld1q_f32(&WebRtcAec_weightCurve[i]);
const uint32x4_t bigger = vcgtq_f32(vec_hNl, vec_hNlFb);
const float32x4_t vec_weightCurve_hNlFb = vmulq_f32(vec_weightCurve,
vec_hNlFb);
const float32x4_t vec_one_weightCurve = vsubq_f32(vec_one, vec_weightCurve);
const float32x4_t vec_one_weightCurve_hNl = vmulq_f32(vec_one_weightCurve,
vec_hNl);
const uint32x4_t vec_if0 = vandq_u32(vmvnq_u32(bigger),
vreinterpretq_u32_f32(vec_hNl));
const float32x4_t vec_one_weightCurve_add =
vaddq_f32(vec_weightCurve_hNlFb, vec_one_weightCurve_hNl);
const uint32x4_t vec_if1 =
vandq_u32(bigger, vreinterpretq_u32_f32(vec_one_weightCurve_add));
vec_hNl = vreinterpretq_f32_u32(vorrq_u32(vec_if0, vec_if1));
{
const float32x4_t vec_overDriveCurve =
vld1q_f32(&WebRtcAec_overDriveCurve[i]);
const float32x4_t vec_overDriveSm_overDriveCurve =
vmulq_f32(vec_overDriveSm, vec_overDriveCurve);
vec_hNl = vpowq_f32(vec_hNl, vec_overDriveSm_overDriveCurve);
vst1q_f32(&hNl[i], vec_hNl);
}
// Suppress error signal
{
float32x4_t vec_efw_re = vld1q_f32(&efw[0][i]);
float32x4_t vec_efw_im = vld1q_f32(&efw[1][i]);
vec_efw_re = vmulq_f32(vec_efw_re, vec_hNl);
vec_efw_im = vmulq_f32(vec_efw_im, vec_hNl);
// Ooura fft returns incorrect sign on imaginary component. It matters
// here because we are making an additive change with comfort noise.
vec_efw_im = vmulq_f32(vec_efw_im, vec_minus_one);
vst1q_f32(&efw[0][i], vec_efw_re);
vst1q_f32(&efw[1][i], vec_efw_im);
}
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
// Weight subbands
if (hNl[i] > hNlFb) {
hNl[i] = WebRtcAec_weightCurve[i] * hNlFb +
(1 - WebRtcAec_weightCurve[i]) * hNl[i];
}
hNl[i] = powf(hNl[i], aec->overDriveSm * WebRtcAec_overDriveCurve[i]);
// Suppress error signal
efw[0][i] *= hNl[i];
efw[1][i] *= hNl[i];
// Ooura fft returns incorrect sign on imaginary component. It matters
// here because we are making an additive change with comfort noise.
efw[1][i] *= -1;
}
}
static int PartitionDelay(const AecCore* aec) {
// Measures the energy in each filter partition and returns the partition with
// highest energy.
// TODO(bjornv): Spread computational cost by computing one partition per
// block?
float wfEnMax = 0;
int i;
int delay = 0;
for (i = 0; i < aec->num_partitions; i++) {
int j;
int pos = i * PART_LEN1;
float wfEn = 0;
float32x4_t vec_wfEn = vdupq_n_f32(0.0f);
// vectorized code (four at once)
for (j = 0; j + 3 < PART_LEN1; j += 4) {
const float32x4_t vec_wfBuf0 = vld1q_f32(&aec->wfBuf[0][pos + j]);
const float32x4_t vec_wfBuf1 = vld1q_f32(&aec->wfBuf[1][pos + j]);
vec_wfEn = vmlaq_f32(vec_wfEn, vec_wfBuf0, vec_wfBuf0);
vec_wfEn = vmlaq_f32(vec_wfEn, vec_wfBuf1, vec_wfBuf1);
}
{
float32x2_t vec_total;
// A B C D
vec_total = vpadd_f32(vget_low_f32(vec_wfEn), vget_high_f32(vec_wfEn));
// A+B C+D
vec_total = vpadd_f32(vec_total, vec_total);
// A+B+C+D A+B+C+D
wfEn = vget_lane_f32(vec_total, 0);
}
// scalar code for the remaining items.
for (; j < PART_LEN1; j++) {
wfEn += aec->wfBuf[0][pos + j] * aec->wfBuf[0][pos + j] +
aec->wfBuf[1][pos + j] * aec->wfBuf[1][pos + j];
}
if (wfEn > wfEnMax) {
wfEnMax = wfEn;
delay = i;
}
}
return delay;
}
// Updates the following smoothed Power Spectral Densities (PSD):
// - sd : near-end
// - se : residual echo
// - sx : far-end
// - sde : cross-PSD of near-end and residual echo
// - sxd : cross-PSD of near-end and far-end
//
// In addition to updating the PSDs, also the filter diverge state is determined
// upon actions are taken.
static void SmoothedPSD(AecCore* aec,
float efw[2][PART_LEN1],
float dfw[2][PART_LEN1],
float xfw[2][PART_LEN1]) {
// Power estimate smoothing coefficients.
const float* ptrGCoh = aec->extended_filter_enabled
? WebRtcAec_kExtendedSmoothingCoefficients[aec->mult - 1]
: WebRtcAec_kNormalSmoothingCoefficients[aec->mult - 1];
int i;
float sdSum = 0, seSum = 0;
const float32x4_t vec_15 = vdupq_n_f32(WebRtcAec_kMinFarendPSD);
float32x4_t vec_sdSum = vdupq_n_f32(0.0f);
float32x4_t vec_seSum = vdupq_n_f32(0.0f);
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t vec_dfw0 = vld1q_f32(&dfw[0][i]);
const float32x4_t vec_dfw1 = vld1q_f32(&dfw[1][i]);
const float32x4_t vec_efw0 = vld1q_f32(&efw[0][i]);
const float32x4_t vec_efw1 = vld1q_f32(&efw[1][i]);
const float32x4_t vec_xfw0 = vld1q_f32(&xfw[0][i]);
const float32x4_t vec_xfw1 = vld1q_f32(&xfw[1][i]);
float32x4_t vec_sd = vmulq_n_f32(vld1q_f32(&aec->sd[i]), ptrGCoh[0]);
float32x4_t vec_se = vmulq_n_f32(vld1q_f32(&aec->se[i]), ptrGCoh[0]);
float32x4_t vec_sx = vmulq_n_f32(vld1q_f32(&aec->sx[i]), ptrGCoh[0]);
float32x4_t vec_dfw_sumsq = vmulq_f32(vec_dfw0, vec_dfw0);
float32x4_t vec_efw_sumsq = vmulq_f32(vec_efw0, vec_efw0);
float32x4_t vec_xfw_sumsq = vmulq_f32(vec_xfw0, vec_xfw0);
vec_dfw_sumsq = vmlaq_f32(vec_dfw_sumsq, vec_dfw1, vec_dfw1);
vec_efw_sumsq = vmlaq_f32(vec_efw_sumsq, vec_efw1, vec_efw1);
vec_xfw_sumsq = vmlaq_f32(vec_xfw_sumsq, vec_xfw1, vec_xfw1);
vec_xfw_sumsq = vmaxq_f32(vec_xfw_sumsq, vec_15);
vec_sd = vmlaq_n_f32(vec_sd, vec_dfw_sumsq, ptrGCoh[1]);
vec_se = vmlaq_n_f32(vec_se, vec_efw_sumsq, ptrGCoh[1]);
vec_sx = vmlaq_n_f32(vec_sx, vec_xfw_sumsq, ptrGCoh[1]);
vst1q_f32(&aec->sd[i], vec_sd);
vst1q_f32(&aec->se[i], vec_se);
vst1q_f32(&aec->sx[i], vec_sx);
{
float32x4x2_t vec_sde = vld2q_f32(&aec->sde[i][0]);
float32x4_t vec_dfwefw0011 = vmulq_f32(vec_dfw0, vec_efw0);
float32x4_t vec_dfwefw0110 = vmulq_f32(vec_dfw0, vec_efw1);
vec_sde.val[0] = vmulq_n_f32(vec_sde.val[0], ptrGCoh[0]);
vec_sde.val[1] = vmulq_n_f32(vec_sde.val[1], ptrGCoh[0]);
vec_dfwefw0011 = vmlaq_f32(vec_dfwefw0011, vec_dfw1, vec_efw1);
vec_dfwefw0110 = vmlsq_f32(vec_dfwefw0110, vec_dfw1, vec_efw0);
vec_sde.val[0] = vmlaq_n_f32(vec_sde.val[0], vec_dfwefw0011, ptrGCoh[1]);
vec_sde.val[1] = vmlaq_n_f32(vec_sde.val[1], vec_dfwefw0110, ptrGCoh[1]);
vst2q_f32(&aec->sde[i][0], vec_sde);
}
{
float32x4x2_t vec_sxd = vld2q_f32(&aec->sxd[i][0]);
float32x4_t vec_dfwxfw0011 = vmulq_f32(vec_dfw0, vec_xfw0);
float32x4_t vec_dfwxfw0110 = vmulq_f32(vec_dfw0, vec_xfw1);
vec_sxd.val[0] = vmulq_n_f32(vec_sxd.val[0], ptrGCoh[0]);
vec_sxd.val[1] = vmulq_n_f32(vec_sxd.val[1], ptrGCoh[0]);
vec_dfwxfw0011 = vmlaq_f32(vec_dfwxfw0011, vec_dfw1, vec_xfw1);
vec_dfwxfw0110 = vmlsq_f32(vec_dfwxfw0110, vec_dfw1, vec_xfw0);
vec_sxd.val[0] = vmlaq_n_f32(vec_sxd.val[0], vec_dfwxfw0011, ptrGCoh[1]);
vec_sxd.val[1] = vmlaq_n_f32(vec_sxd.val[1], vec_dfwxfw0110, ptrGCoh[1]);
vst2q_f32(&aec->sxd[i][0], vec_sxd);
}
vec_sdSum = vaddq_f32(vec_sdSum, vec_sd);
vec_seSum = vaddq_f32(vec_seSum, vec_se);
}
{
float32x2_t vec_sdSum_total;
float32x2_t vec_seSum_total;
// A B C D
vec_sdSum_total = vpadd_f32(vget_low_f32(vec_sdSum),
vget_high_f32(vec_sdSum));
vec_seSum_total = vpadd_f32(vget_low_f32(vec_seSum),
vget_high_f32(vec_seSum));
// A+B C+D
vec_sdSum_total = vpadd_f32(vec_sdSum_total, vec_sdSum_total);
vec_seSum_total = vpadd_f32(vec_seSum_total, vec_seSum_total);
// A+B+C+D A+B+C+D
sdSum = vget_lane_f32(vec_sdSum_total, 0);
seSum = vget_lane_f32(vec_seSum_total, 0);
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
aec->sd[i] = ptrGCoh[0] * aec->sd[i] +
ptrGCoh[1] * (dfw[0][i] * dfw[0][i] + dfw[1][i] * dfw[1][i]);
aec->se[i] = ptrGCoh[0] * aec->se[i] +
ptrGCoh[1] * (efw[0][i] * efw[0][i] + efw[1][i] * efw[1][i]);
// We threshold here to protect against the ill-effects of a zero farend.
// The threshold is not arbitrarily chosen, but balances protection and
// adverse interaction with the algorithm's tuning.
// TODO(bjornv): investigate further why this is so sensitive.
aec->sx[i] =
ptrGCoh[0] * aec->sx[i] +
ptrGCoh[1] * WEBRTC_SPL_MAX(
xfw[0][i] * xfw[0][i] + xfw[1][i] * xfw[1][i],
WebRtcAec_kMinFarendPSD);
aec->sde[i][0] =
ptrGCoh[0] * aec->sde[i][0] +
ptrGCoh[1] * (dfw[0][i] * efw[0][i] + dfw[1][i] * efw[1][i]);
aec->sde[i][1] =
ptrGCoh[0] * aec->sde[i][1] +
ptrGCoh[1] * (dfw[0][i] * efw[1][i] - dfw[1][i] * efw[0][i]);
aec->sxd[i][0] =
ptrGCoh[0] * aec->sxd[i][0] +
ptrGCoh[1] * (dfw[0][i] * xfw[0][i] + dfw[1][i] * xfw[1][i]);
aec->sxd[i][1] =
ptrGCoh[0] * aec->sxd[i][1] +
ptrGCoh[1] * (dfw[0][i] * xfw[1][i] - dfw[1][i] * xfw[0][i]);
sdSum += aec->sd[i];
seSum += aec->se[i];
}
// Divergent filter safeguard.
aec->divergeState = (aec->divergeState ? 1.05f : 1.0f) * seSum > sdSum;
if (aec->divergeState)
memcpy(efw, dfw, sizeof(efw[0][0]) * 2 * PART_LEN1);
// Reset if error is significantly larger than nearend (13 dB).
if (!aec->extended_filter_enabled && seSum > (19.95f * sdSum))
memset(aec->wfBuf, 0, sizeof(aec->wfBuf));
}
// Window time domain data to be used by the fft.
__inline static void WindowData(float* x_windowed, const float* x) {
int i;
for (i = 0; i < PART_LEN; i += 4) {
const float32x4_t vec_Buf1 = vld1q_f32(&x[i]);
const float32x4_t vec_Buf2 = vld1q_f32(&x[PART_LEN + i]);
const float32x4_t vec_sqrtHanning = vld1q_f32(&WebRtcAec_sqrtHanning[i]);
// A B C D
float32x4_t vec_sqrtHanning_rev =
vld1q_f32(&WebRtcAec_sqrtHanning[PART_LEN - i - 3]);
// B A D C
vec_sqrtHanning_rev = vrev64q_f32(vec_sqrtHanning_rev);
// D C B A
vec_sqrtHanning_rev = vcombine_f32(vget_high_f32(vec_sqrtHanning_rev),
vget_low_f32(vec_sqrtHanning_rev));
vst1q_f32(&x_windowed[i], vmulq_f32(vec_Buf1, vec_sqrtHanning));
vst1q_f32(&x_windowed[PART_LEN + i],
vmulq_f32(vec_Buf2, vec_sqrtHanning_rev));
}
}
// Puts fft output data into a complex valued array.
__inline static void StoreAsComplex(const float* data,
float data_complex[2][PART_LEN1]) {
int i;
for (i = 0; i < PART_LEN; i += 4) {
const float32x4x2_t vec_data = vld2q_f32(&data[2 * i]);
vst1q_f32(&data_complex[0][i], vec_data.val[0]);
vst1q_f32(&data_complex[1][i], vec_data.val[1]);
}
// fix beginning/end values
data_complex[1][0] = 0;
data_complex[1][PART_LEN] = 0;
data_complex[0][0] = data[0];
data_complex[0][PART_LEN] = data[1];
}
static void SubbandCoherenceNEON(AecCore* aec,
float efw[2][PART_LEN1],
float xfw[2][PART_LEN1],
float* fft,
float* cohde,
float* cohxd) {
float dfw[2][PART_LEN1];
int i;
if (aec->delayEstCtr == 0)
aec->delayIdx = PartitionDelay(aec);
// Use delayed far.
memcpy(xfw,
aec->xfwBuf + aec->delayIdx * PART_LEN1,
sizeof(xfw[0][0]) * 2 * PART_LEN1);
// Windowed near fft
WindowData(fft, aec->dBuf);
aec_rdft_forward_128(fft);
StoreAsComplex(fft, dfw);
// Windowed error fft
WindowData(fft, aec->eBuf);
aec_rdft_forward_128(fft);
StoreAsComplex(fft, efw);
SmoothedPSD(aec, efw, dfw, xfw);
{
const float32x4_t vec_1eminus10 = vdupq_n_f32(1e-10f);
// Subband coherence
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t vec_sd = vld1q_f32(&aec->sd[i]);
const float32x4_t vec_se = vld1q_f32(&aec->se[i]);
const float32x4_t vec_sx = vld1q_f32(&aec->sx[i]);
const float32x4_t vec_sdse = vmlaq_f32(vec_1eminus10, vec_sd, vec_se);
const float32x4_t vec_sdsx = vmlaq_f32(vec_1eminus10, vec_sd, vec_sx);
float32x4x2_t vec_sde = vld2q_f32(&aec->sde[i][0]);
float32x4x2_t vec_sxd = vld2q_f32(&aec->sxd[i][0]);
float32x4_t vec_cohde = vmulq_f32(vec_sde.val[0], vec_sde.val[0]);
float32x4_t vec_cohxd = vmulq_f32(vec_sxd.val[0], vec_sxd.val[0]);
vec_cohde = vmlaq_f32(vec_cohde, vec_sde.val[1], vec_sde.val[1]);
vec_cohde = vdivq_f32(vec_cohde, vec_sdse);
vec_cohxd = vmlaq_f32(vec_cohxd, vec_sxd.val[1], vec_sxd.val[1]);
vec_cohxd = vdivq_f32(vec_cohxd, vec_sdsx);
vst1q_f32(&cohde[i], vec_cohde);
vst1q_f32(&cohxd[i], vec_cohxd);
}
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
cohde[i] =
(aec->sde[i][0] * aec->sde[i][0] + aec->sde[i][1] * aec->sde[i][1]) /
(aec->sd[i] * aec->se[i] + 1e-10f);
cohxd[i] =
(aec->sxd[i][0] * aec->sxd[i][0] + aec->sxd[i][1] * aec->sxd[i][1]) /
(aec->sx[i] * aec->sd[i] + 1e-10f);
}
}
void WebRtcAec_InitAec_neon(void) {
WebRtcAec_FilterFar = FilterFarNEON;
WebRtcAec_ScaleErrorSignal = ScaleErrorSignalNEON;
WebRtcAec_FilterAdaptation = FilterAdaptationNEON;
WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressNEON;
WebRtcAec_SubbandCoherence = SubbandCoherenceNEON;
}

731
webrtc/modules/audio_processing/aec/aec_core_sse2.c
View File

@ -1,731 +0,0 @@
/*
* Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
/*
* The core AEC algorithm, SSE2 version of speed-critical functions.
*/
#include <emmintrin.h>
#include <math.h>
#include <string.h> // memset
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include "webrtc/modules/audio_processing/aec/aec_common.h"
#include "webrtc/modules/audio_processing/aec/aec_core_internal.h"
#include "webrtc/modules/audio_processing/aec/aec_rdft.h"
__inline static float MulRe(float aRe, float aIm, float bRe, float bIm) {
return aRe * bRe - aIm * bIm;
}
__inline static float MulIm(float aRe, float aIm, float bRe, float bIm) {
return aRe * bIm + aIm * bRe;
}
static void FilterFarSSE2(AecCore* aec, float yf[2][PART_LEN1]) {
int i;
const int num_partitions = aec->num_partitions;
for (i = 0; i < num_partitions; i++) {
int j;
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1;
// Check for wrap
if (i + aec->xfBufBlockPos >= num_partitions) {
xPos -= num_partitions * (PART_LEN1);
}
// vectorized code (four at once)
for (j = 0; j + 3 < PART_LEN1; j += 4) {
const __m128 xfBuf_re = _mm_loadu_ps(&aec->xfBuf[0][xPos + j]);
const __m128 xfBuf_im = _mm_loadu_ps(&aec->xfBuf[1][xPos + j]);
const __m128 wfBuf_re = _mm_loadu_ps(&aec->wfBuf[0][pos + j]);
const __m128 wfBuf_im = _mm_loadu_ps(&aec->wfBuf[1][pos + j]);
const __m128 yf_re = _mm_loadu_ps(&yf[0][j]);
const __m128 yf_im = _mm_loadu_ps(&yf[1][j]);
const __m128 a = _mm_mul_ps(xfBuf_re, wfBuf_re);
const __m128 b = _mm_mul_ps(xfBuf_im, wfBuf_im);
const __m128 c = _mm_mul_ps(xfBuf_re, wfBuf_im);
const __m128 d = _mm_mul_ps(xfBuf_im, wfBuf_re);
const __m128 e = _mm_sub_ps(a, b);
const __m128 f = _mm_add_ps(c, d);
const __m128 g = _mm_add_ps(yf_re, e);
const __m128 h = _mm_add_ps(yf_im, f);
_mm_storeu_ps(&yf[0][j], g);
_mm_storeu_ps(&yf[1][j], h);
}
// scalar code for the remaining items.
for (; j < PART_LEN1; j++) {
yf[0][j] += MulRe(aec->xfBuf[0][xPos + j],
aec->xfBuf[1][xPos + j],
aec->wfBuf[0][pos + j],
aec->wfBuf[1][pos + j]);
yf[1][j] += MulIm(aec->xfBuf[0][xPos + j],
aec->xfBuf[1][xPos + j],
aec->wfBuf[0][pos + j],
aec->wfBuf[1][pos + j]);
}
}
}
static void ScaleErrorSignalSSE2(AecCore* aec, float ef[2][PART_LEN1]) {
const __m128 k1e_10f = _mm_set1_ps(1e-10f);
const __m128 kMu = aec->extended_filter_enabled ? _mm_set1_ps(kExtendedMu)
: _mm_set1_ps(aec->normal_mu);
const __m128 kThresh = aec->extended_filter_enabled
? _mm_set1_ps(kExtendedErrorThreshold)
: _mm_set1_ps(aec->normal_error_threshold);
int i;
// vectorized code (four at once)
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const __m128 xPow = _mm_loadu_ps(&aec->xPow[i]);
const __m128 ef_re_base = _mm_loadu_ps(&ef[0][i]);
const __m128 ef_im_base = _mm_loadu_ps(&ef[1][i]);
const __m128 xPowPlus = _mm_add_ps(xPow, k1e_10f);
__m128 ef_re = _mm_div_ps(ef_re_base, xPowPlus);
__m128 ef_im = _mm_div_ps(ef_im_base, xPowPlus);
const __m128 ef_re2 = _mm_mul_ps(ef_re, ef_re);
const __m128 ef_im2 = _mm_mul_ps(ef_im, ef_im);
const __m128 ef_sum2 = _mm_add_ps(ef_re2, ef_im2);
const __m128 absEf = _mm_sqrt_ps(ef_sum2);
const __m128 bigger = _mm_cmpgt_ps(absEf, kThresh);
__m128 absEfPlus = _mm_add_ps(absEf, k1e_10f);
const __m128 absEfInv = _mm_div_ps(kThresh, absEfPlus);
__m128 ef_re_if = _mm_mul_ps(ef_re, absEfInv);
__m128 ef_im_if = _mm_mul_ps(ef_im, absEfInv);
ef_re_if = _mm_and_ps(bigger, ef_re_if);
ef_im_if = _mm_and_ps(bigger, ef_im_if);
ef_re = _mm_andnot_ps(bigger, ef_re);
ef_im = _mm_andnot_ps(bigger, ef_im);
ef_re = _mm_or_ps(ef_re, ef_re_if);
ef_im = _mm_or_ps(ef_im, ef_im_if);
ef_re = _mm_mul_ps(ef_re, kMu);
ef_im = _mm_mul_ps(ef_im, kMu);
_mm_storeu_ps(&ef[0][i], ef_re);
_mm_storeu_ps(&ef[1][i], ef_im);
}
// scalar code for the remaining items.
{
const float mu =
aec->extended_filter_enabled ? kExtendedMu : aec->normal_mu;
const float error_threshold = aec->extended_filter_enabled
? kExtendedErrorThreshold
: aec->normal_error_threshold;
for (; i < (PART_LEN1); i++) {
float abs_ef;
ef[0][i] /= (aec->xPow[i] + 1e-10f);
ef[1][i] /= (aec->xPow[i] + 1e-10f);
abs_ef = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]);
if (abs_ef > error_threshold) {
abs_ef = error_threshold / (abs_ef + 1e-10f);
ef[0][i] *= abs_ef;
ef[1][i] *= abs_ef;
}
// Stepsize factor
ef[0][i] *= mu;
ef[1][i] *= mu;
}
}
}
static void FilterAdaptationSSE2(AecCore* aec,
float* fft,
float ef[2][PART_LEN1]) {
int i, j;
const int num_partitions = aec->num_partitions;
for (i = 0; i < num_partitions; i++) {
int xPos = (i + aec->xfBufBlockPos) * (PART_LEN1);
int pos = i * PART_LEN1;
// Check for wrap
if (i + aec->xfBufBlockPos >= num_partitions) {
xPos -= num_partitions * PART_LEN1;
}
// Process the whole array...
for (j = 0; j < PART_LEN; j += 4) {
// Load xfBuf and ef.
const __m128 xfBuf_re = _mm_loadu_ps(&aec->xfBuf[0][xPos + j]);
const __m128 xfBuf_im = _mm_loadu_ps(&aec->xfBuf[1][xPos + j]);
const __m128 ef_re = _mm_loadu_ps(&ef[0][j]);
const __m128 ef_im = _mm_loadu_ps(&ef[1][j]);
// Calculate the product of conjugate(xfBuf) by ef.
// re(conjugate(a) * b) = aRe * bRe + aIm * bIm
// im(conjugate(a) * b)= aRe * bIm - aIm * bRe
const __m128 a = _mm_mul_ps(xfBuf_re, ef_re);
const __m128 b = _mm_mul_ps(xfBuf_im, ef_im);
const __m128 c = _mm_mul_ps(xfBuf_re, ef_im);
const __m128 d = _mm_mul_ps(xfBuf_im, ef_re);
const __m128 e = _mm_add_ps(a, b);
const __m128 f = _mm_sub_ps(c, d);
// Interleave real and imaginary parts.
const __m128 g = _mm_unpacklo_ps(e, f);
const __m128 h = _mm_unpackhi_ps(e, f);
// Store
_mm_storeu_ps(&fft[2 * j + 0], g);
_mm_storeu_ps(&fft[2 * j + 4], h);
}
// ... and fixup the first imaginary entry.
fft[1] = MulRe(aec->xfBuf[0][xPos + PART_LEN],
-aec->xfBuf[1][xPos + PART_LEN],
ef[0][PART_LEN],
ef[1][PART_LEN]);
aec_rdft_inverse_128(fft);
memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN);
// fft scaling
{
float scale = 2.0f / PART_LEN2;
const __m128 scale_ps = _mm_load_ps1(&scale);
for (j = 0; j < PART_LEN; j += 4) {
const __m128 fft_ps = _mm_loadu_ps(&fft[j]);
const __m128 fft_scale = _mm_mul_ps(fft_ps, scale_ps);
_mm_storeu_ps(&fft[j], fft_scale);
}
}
aec_rdft_forward_128(fft);
{
float wt1 = aec->wfBuf[1][pos];
aec->wfBuf[0][pos + PART_LEN] += fft[1];
for (j = 0; j < PART_LEN; j += 4) {
__m128 wtBuf_re = _mm_loadu_ps(&aec->wfBuf[0][pos + j]);
__m128 wtBuf_im = _mm_loadu_ps(&aec->wfBuf[1][pos + j]);
const __m128 fft0 = _mm_loadu_ps(&fft[2 * j + 0]);
const __m128 fft4 = _mm_loadu_ps(&fft[2 * j + 4]);
const __m128 fft_re =
_mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(2, 0, 2, 0));
const __m128 fft_im =
_mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(3, 1, 3, 1));
wtBuf_re = _mm_add_ps(wtBuf_re, fft_re);
wtBuf_im = _mm_add_ps(wtBuf_im, fft_im);
_mm_storeu_ps(&aec->wfBuf[0][pos + j], wtBuf_re);
_mm_storeu_ps(&aec->wfBuf[1][pos + j], wtBuf_im);
}
aec->wfBuf[1][pos] = wt1;
}
}
}
static __m128 mm_pow_ps(__m128 a, __m128 b) {
// a^b = exp2(b * log2(a))
// exp2(x) and log2(x) are calculated using polynomial approximations.
__m128 log2_a, b_log2_a, a_exp_b;
// Calculate log2(x), x = a.
{
// To calculate log2(x), we decompose x like this:
// x = y * 2^n
// n is an integer
// y is in the [1.0, 2.0) range
//
// log2(x) = log2(y) + n
// n can be evaluated by playing with float representation.
// log2(y) in a small range can be approximated, this code uses an order
// five polynomial approximation. The coefficients have been
// estimated with the Remez algorithm and the resulting
// polynomial has a maximum relative error of 0.00086%.
// Compute n.
// This is done by masking the exponent, shifting it into the top bit of
// the mantissa, putting eight into the biased exponent (to shift/
// compensate the fact that the exponent has been shifted in the top/
// fractional part and finally getting rid of the implicit leading one
// from the mantissa by substracting it out.
static const ALIGN16_BEG int float_exponent_mask[4] ALIGN16_END = {
0x7F800000, 0x7F800000, 0x7F800000, 0x7F800000};
static const ALIGN16_BEG int eight_biased_exponent[4] ALIGN16_END = {
0x43800000, 0x43800000, 0x43800000, 0x43800000};
static const ALIGN16_BEG int implicit_leading_one[4] ALIGN16_END = {
0x43BF8000, 0x43BF8000, 0x43BF8000, 0x43BF8000};
static const int shift_exponent_into_top_mantissa = 8;
const __m128 two_n = _mm_and_ps(a, *((__m128*)float_exponent_mask));
const __m128 n_1 = _mm_castsi128_ps(_mm_srli_epi32(
_mm_castps_si128(two_n), shift_exponent_into_top_mantissa));
const __m128 n_0 = _mm_or_ps(n_1, *((__m128*)eight_biased_exponent));
const __m128 n = _mm_sub_ps(n_0, *((__m128*)implicit_leading_one));
// Compute y.
static const ALIGN16_BEG int mantissa_mask[4] ALIGN16_END = {
0x007FFFFF, 0x007FFFFF, 0x007FFFFF, 0x007FFFFF};
static const ALIGN16_BEG int zero_biased_exponent_is_one[4] ALIGN16_END = {
0x3F800000, 0x3F800000, 0x3F800000, 0x3F800000};
const __m128 mantissa = _mm_and_ps(a, *((__m128*)mantissa_mask));
const __m128 y =
_mm_or_ps(mantissa, *((__m128*)zero_biased_exponent_is_one));
// Approximate log2(y) ~= (y - 1) * pol5(y).
// pol5(y) = C5 * y^5 + C4 * y^4 + C3 * y^3 + C2 * y^2 + C1 * y + C0
static const ALIGN16_BEG float ALIGN16_END C5[4] = {
-3.4436006e-2f, -3.4436006e-2f, -3.4436006e-2f, -3.4436006e-2f};
static const ALIGN16_BEG float ALIGN16_END
C4[4] = {3.1821337e-1f, 3.1821337e-1f, 3.1821337e-1f, 3.1821337e-1f};
static const ALIGN16_BEG float ALIGN16_END
C3[4] = {-1.2315303f, -1.2315303f, -1.2315303f, -1.2315303f};
static const ALIGN16_BEG float ALIGN16_END
C2[4] = {2.5988452f, 2.5988452f, 2.5988452f, 2.5988452f};
static const ALIGN16_BEG float ALIGN16_END
C1[4] = {-3.3241990f, -3.3241990f, -3.3241990f, -3.3241990f};
static const ALIGN16_BEG float ALIGN16_END
C0[4] = {3.1157899f, 3.1157899f, 3.1157899f, 3.1157899f};
const __m128 pol5_y_0 = _mm_mul_ps(y, *((__m128*)C5));
const __m128 pol5_y_1 = _mm_add_ps(pol5_y_0, *((__m128*)C4));
const __m128 pol5_y_2 = _mm_mul_ps(pol5_y_1, y);
const __m128 pol5_y_3 = _mm_add_ps(pol5_y_2, *((__m128*)C3));
const __m128 pol5_y_4 = _mm_mul_ps(pol5_y_3, y);
const __m128 pol5_y_5 = _mm_add_ps(pol5_y_4, *((__m128*)C2));
const __m128 pol5_y_6 = _mm_mul_ps(pol5_y_5, y);
const __m128 pol5_y_7 = _mm_add_ps(pol5_y_6, *((__m128*)C1));
const __m128 pol5_y_8 = _mm_mul_ps(pol5_y_7, y);
const __m128 pol5_y = _mm_add_ps(pol5_y_8, *((__m128*)C0));
const __m128 y_minus_one =
_mm_sub_ps(y, *((__m128*)zero_biased_exponent_is_one));
const __m128 log2_y = _mm_mul_ps(y_minus_one, pol5_y);
// Combine parts.
log2_a = _mm_add_ps(n, log2_y);
}
// b * log2(a)
b_log2_a = _mm_mul_ps(b, log2_a);
// Calculate exp2(x), x = b * log2(a).
{
// To calculate 2^x, we decompose x like this:
// x = n + y
// n is an integer, the value of x - 0.5 rounded down, therefore
// y is in the [0.5, 1.5) range
//
// 2^x = 2^n * 2^y
// 2^n can be evaluated by playing with float representation.
// 2^y in a small range can be approximated, this code uses an order two
// polynomial approximation. The coefficients have been estimated
// with the Remez algorithm and the resulting polynomial has a
// maximum relative error of 0.17%.
// To avoid over/underflow, we reduce the range of input to ]-127, 129].
static const ALIGN16_BEG float max_input[4] ALIGN16_END = {129.f, 129.f,
129.f, 129.f};
static const ALIGN16_BEG float min_input[4] ALIGN16_END = {
-126.99999f, -126.99999f, -126.99999f, -126.99999f};
const __m128 x_min = _mm_min_ps(b_log2_a, *((__m128*)max_input));
const __m128 x_max = _mm_max_ps(x_min, *((__m128*)min_input));
// Compute n.
static const ALIGN16_BEG float half[4] ALIGN16_END = {0.5f, 0.5f,
0.5f, 0.5f};
const __m128 x_minus_half = _mm_sub_ps(x_max, *((__m128*)half));
const __m128i x_minus_half_floor = _mm_cvtps_epi32(x_minus_half);
// Compute 2^n.
static const ALIGN16_BEG int float_exponent_bias[4] ALIGN16_END = {
127, 127, 127, 127};
static const int float_exponent_shift = 23;
const __m128i two_n_exponent =
_mm_add_epi32(x_minus_half_floor, *((__m128i*)float_exponent_bias));
const __m128 two_n =
_mm_castsi128_ps(_mm_slli_epi32(two_n_exponent, float_exponent_shift));
// Compute y.
const __m128 y = _mm_sub_ps(x_max, _mm_cvtepi32_ps(x_minus_half_floor));
// Approximate 2^y ~= C2 * y^2 + C1 * y + C0.
static const ALIGN16_BEG float C2[4] ALIGN16_END = {
3.3718944e-1f, 3.3718944e-1f, 3.3718944e-1f, 3.3718944e-1f};
static const ALIGN16_BEG float C1[4] ALIGN16_END = {
6.5763628e-1f, 6.5763628e-1f, 6.5763628e-1f, 6.5763628e-1f};
static const ALIGN16_BEG float C0[4] ALIGN16_END = {1.0017247f, 1.0017247f,
1.0017247f, 1.0017247f};
const __m128 exp2_y_0 = _mm_mul_ps(y, *((__m128*)C2));
const __m128 exp2_y_1 = _mm_add_ps(exp2_y_0, *((__m128*)C1));
const __m128 exp2_y_2 = _mm_mul_ps(exp2_y_1, y);
const __m128 exp2_y = _mm_add_ps(exp2_y_2, *((__m128*)C0));
// Combine parts.
a_exp_b = _mm_mul_ps(exp2_y, two_n);
}
return a_exp_b;
}
static void OverdriveAndSuppressSSE2(AecCore* aec,
float hNl[PART_LEN1],
const float hNlFb,
float efw[2][PART_LEN1]) {
int i;
const __m128 vec_hNlFb = _mm_set1_ps(hNlFb);
const __m128 vec_one = _mm_set1_ps(1.0f);
const __m128 vec_minus_one = _mm_set1_ps(-1.0f);
const __m128 vec_overDriveSm = _mm_set1_ps(aec->overDriveSm);
// vectorized code (four at once)
for (i = 0; i + 3 < PART_LEN1; i += 4) {
// Weight subbands
__m128 vec_hNl = _mm_loadu_ps(&hNl[i]);
const __m128 vec_weightCurve = _mm_loadu_ps(&WebRtcAec_weightCurve[i]);
const __m128 bigger = _mm_cmpgt_ps(vec_hNl, vec_hNlFb);
const __m128 vec_weightCurve_hNlFb = _mm_mul_ps(vec_weightCurve, vec_hNlFb);
const __m128 vec_one_weightCurve = _mm_sub_ps(vec_one, vec_weightCurve);
const __m128 vec_one_weightCurve_hNl =
_mm_mul_ps(vec_one_weightCurve, vec_hNl);
const __m128 vec_if0 = _mm_andnot_ps(bigger, vec_hNl);
const __m128 vec_if1 = _mm_and_ps(
bigger, _mm_add_ps(vec_weightCurve_hNlFb, vec_one_weightCurve_hNl));
vec_hNl = _mm_or_ps(vec_if0, vec_if1);
{
const __m128 vec_overDriveCurve =
_mm_loadu_ps(&WebRtcAec_overDriveCurve[i]);
const __m128 vec_overDriveSm_overDriveCurve =
_mm_mul_ps(vec_overDriveSm, vec_overDriveCurve);
vec_hNl = mm_pow_ps(vec_hNl, vec_overDriveSm_overDriveCurve);
_mm_storeu_ps(&hNl[i], vec_hNl);
}
// Suppress error signal
{
__m128 vec_efw_re = _mm_loadu_ps(&efw[0][i]);
__m128 vec_efw_im = _mm_loadu_ps(&efw[1][i]);
vec_efw_re = _mm_mul_ps(vec_efw_re, vec_hNl);
vec_efw_im = _mm_mul_ps(vec_efw_im, vec_hNl);
// Ooura fft returns incorrect sign on imaginary component. It matters
// here because we are making an additive change with comfort noise.
vec_efw_im = _mm_mul_ps(vec_efw_im, vec_minus_one);
_mm_storeu_ps(&efw[0][i], vec_efw_re);
_mm_storeu_ps(&efw[1][i], vec_efw_im);
}
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
// Weight subbands
if (hNl[i] > hNlFb) {
hNl[i] = WebRtcAec_weightCurve[i] * hNlFb +
(1 - WebRtcAec_weightCurve[i]) * hNl[i];
}
hNl[i] = powf(hNl[i], aec->overDriveSm * WebRtcAec_overDriveCurve[i]);
// Suppress error signal
efw[0][i] *= hNl[i];
efw[1][i] *= hNl[i];
// Ooura fft returns incorrect sign on imaginary component. It matters
// here because we are making an additive change with comfort noise.
efw[1][i] *= -1;
}
}
__inline static void _mm_add_ps_4x1(__m128 sum, float *dst) {
// A+B C+D
sum = _mm_add_ps(sum, _mm_shuffle_ps(sum, sum, _MM_SHUFFLE(0, 0, 3, 2)));
// A+B+C+D A+B+C+D
sum = _mm_add_ps(sum, _mm_shuffle_ps(sum, sum, _MM_SHUFFLE(1, 1, 1, 1)));
_mm_store_ss(dst, sum);
}
static int PartitionDelay(const AecCore* aec) {
// Measures the energy in each filter partition and returns the partition with
// highest energy.
// TODO(bjornv): Spread computational cost by computing one partition per
// block?
float wfEnMax = 0;
int i;
int delay = 0;
for (i = 0; i < aec->num_partitions; i++) {
int j;
int pos = i * PART_LEN1;
float wfEn = 0;
__m128 vec_wfEn = _mm_set1_ps(0.0f);
// vectorized code (four at once)
for (j = 0; j + 3 < PART_LEN1; j += 4) {
const __m128 vec_wfBuf0 = _mm_loadu_ps(&aec->wfBuf[0][pos + j]);
const __m128 vec_wfBuf1 = _mm_loadu_ps(&aec->wfBuf[1][pos + j]);
vec_wfEn = _mm_add_ps(vec_wfEn, _mm_mul_ps(vec_wfBuf0, vec_wfBuf0));
vec_wfEn = _mm_add_ps(vec_wfEn, _mm_mul_ps(vec_wfBuf1, vec_wfBuf1));
}
_mm_add_ps_4x1(vec_wfEn, &wfEn);
// scalar code for the remaining items.
for (; j < PART_LEN1; j++) {
wfEn += aec->wfBuf[0][pos + j] * aec->wfBuf[0][pos + j] +
aec->wfBuf[1][pos + j] * aec->wfBuf[1][pos + j];
}
if (wfEn > wfEnMax) {
wfEnMax = wfEn;
delay = i;
}
}
return delay;
}
// Updates the following smoothed Power Spectral Densities (PSD):
// - sd : near-end
// - se : residual echo
// - sx : far-end
// - sde : cross-PSD of near-end and residual echo
// - sxd : cross-PSD of near-end and far-end
//
// In addition to updating the PSDs, also the filter diverge state is determined
// upon actions are taken.
static void SmoothedPSD(AecCore* aec,
float efw[2][PART_LEN1],
float dfw[2][PART_LEN1],
float xfw[2][PART_LEN1]) {
// Power estimate smoothing coefficients.
const float* ptrGCoh = aec->extended_filter_enabled
? WebRtcAec_kExtendedSmoothingCoefficients[aec->mult - 1]
: WebRtcAec_kNormalSmoothingCoefficients[aec->mult - 1];
int i;
float sdSum = 0, seSum = 0;
const __m128 vec_15 = _mm_set1_ps(WebRtcAec_kMinFarendPSD);
const __m128 vec_GCoh0 = _mm_set1_ps(ptrGCoh[0]);
const __m128 vec_GCoh1 = _mm_set1_ps(ptrGCoh[1]);
__m128 vec_sdSum = _mm_set1_ps(0.0f);
__m128 vec_seSum = _mm_set1_ps(0.0f);
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const __m128 vec_dfw0 = _mm_loadu_ps(&dfw[0][i]);
const __m128 vec_dfw1 = _mm_loadu_ps(&dfw[1][i]);
const __m128 vec_efw0 = _mm_loadu_ps(&efw[0][i]);
const __m128 vec_efw1 = _mm_loadu_ps(&efw[1][i]);
const __m128 vec_xfw0 = _mm_loadu_ps(&xfw[0][i]);
const __m128 vec_xfw1 = _mm_loadu_ps(&xfw[1][i]);
__m128 vec_sd = _mm_mul_ps(_mm_loadu_ps(&aec->sd[i]), vec_GCoh0);
__m128 vec_se = _mm_mul_ps(_mm_loadu_ps(&aec->se[i]), vec_GCoh0);
__m128 vec_sx = _mm_mul_ps(_mm_loadu_ps(&aec->sx[i]), vec_GCoh0);
__m128 vec_dfw_sumsq = _mm_mul_ps(vec_dfw0, vec_dfw0);
__m128 vec_efw_sumsq = _mm_mul_ps(vec_efw0, vec_efw0);
__m128 vec_xfw_sumsq = _mm_mul_ps(vec_xfw0, vec_xfw0);
vec_dfw_sumsq = _mm_add_ps(vec_dfw_sumsq, _mm_mul_ps(vec_dfw1, vec_dfw1));
vec_efw_sumsq = _mm_add_ps(vec_efw_sumsq, _mm_mul_ps(vec_efw1, vec_efw1));
vec_xfw_sumsq = _mm_add_ps(vec_xfw_sumsq, _mm_mul_ps(vec_xfw1, vec_xfw1));
vec_xfw_sumsq = _mm_max_ps(vec_xfw_sumsq, vec_15);
vec_sd = _mm_add_ps(vec_sd, _mm_mul_ps(vec_dfw_sumsq, vec_GCoh1));
vec_se = _mm_add_ps(vec_se, _mm_mul_ps(vec_efw_sumsq, vec_GCoh1));
vec_sx = _mm_add_ps(vec_sx, _mm_mul_ps(vec_xfw_sumsq, vec_GCoh1));
_mm_storeu_ps(&aec->sd[i], vec_sd);
_mm_storeu_ps(&aec->se[i], vec_se);
_mm_storeu_ps(&aec->sx[i], vec_sx);
{
const __m128 vec_3210 = _mm_loadu_ps(&aec->sde[i][0]);
const __m128 vec_7654 = _mm_loadu_ps(&aec->sde[i + 2][0]);
__m128 vec_a = _mm_shuffle_ps(vec_3210, vec_7654,
_MM_SHUFFLE(2, 0, 2, 0));
__m128 vec_b = _mm_shuffle_ps(vec_3210, vec_7654,
_MM_SHUFFLE(3, 1, 3, 1));
__m128 vec_dfwefw0011 = _mm_mul_ps(vec_dfw0, vec_efw0);
__m128 vec_dfwefw0110 = _mm_mul_ps(vec_dfw0, vec_efw1);
vec_a = _mm_mul_ps(vec_a, vec_GCoh0);
vec_b = _mm_mul_ps(vec_b, vec_GCoh0);
vec_dfwefw0011 = _mm_add_ps(vec_dfwefw0011,
_mm_mul_ps(vec_dfw1, vec_efw1));
vec_dfwefw0110 = _mm_sub_ps(vec_dfwefw0110,
_mm_mul_ps(vec_dfw1, vec_efw0));
vec_a = _mm_add_ps(vec_a, _mm_mul_ps(vec_dfwefw0011, vec_GCoh1));
vec_b = _mm_add_ps(vec_b, _mm_mul_ps(vec_dfwefw0110, vec_GCoh1));
_mm_storeu_ps(&aec->sde[i][0], _mm_unpacklo_ps(vec_a, vec_b));
_mm_storeu_ps(&aec->sde[i + 2][0], _mm_unpackhi_ps(vec_a, vec_b));
}
{
const __m128 vec_3210 = _mm_loadu_ps(&aec->sxd[i][0]);
const __m128 vec_7654 = _mm_loadu_ps(&aec->sxd[i + 2][0]);
__m128 vec_a = _mm_shuffle_ps(vec_3210, vec_7654,
_MM_SHUFFLE(2, 0, 2, 0));
__m128 vec_b = _mm_shuffle_ps(vec_3210, vec_7654,
_MM_SHUFFLE(3, 1, 3, 1));
__m128 vec_dfwxfw0011 = _mm_mul_ps(vec_dfw0, vec_xfw0);
__m128 vec_dfwxfw0110 = _mm_mul_ps(vec_dfw0, vec_xfw1);
vec_a = _mm_mul_ps(vec_a, vec_GCoh0);
vec_b = _mm_mul_ps(vec_b, vec_GCoh0);
vec_dfwxfw0011 = _mm_add_ps(vec_dfwxfw0011,
_mm_mul_ps(vec_dfw1, vec_xfw1));
vec_dfwxfw0110 = _mm_sub_ps(vec_dfwxfw0110,
_mm_mul_ps(vec_dfw1, vec_xfw0));
vec_a = _mm_add_ps(vec_a, _mm_mul_ps(vec_dfwxfw0011, vec_GCoh1));
vec_b = _mm_add_ps(vec_b, _mm_mul_ps(vec_dfwxfw0110, vec_GCoh1));
_mm_storeu_ps(&aec->sxd[i][0], _mm_unpacklo_ps(vec_a, vec_b));
_mm_storeu_ps(&aec->sxd[i + 2][0], _mm_unpackhi_ps(vec_a, vec_b));
}
vec_sdSum = _mm_add_ps(vec_sdSum, vec_sd);
vec_seSum = _mm_add_ps(vec_seSum, vec_se);
}
_mm_add_ps_4x1(vec_sdSum, &sdSum);
_mm_add_ps_4x1(vec_seSum, &seSum);
for (; i < PART_LEN1; i++) {
aec->sd[i] = ptrGCoh[0] * aec->sd[i] +
ptrGCoh[1] * (dfw[0][i] * dfw[0][i] + dfw[1][i] * dfw[1][i]);
aec->se[i] = ptrGCoh[0] * aec->se[i] +
ptrGCoh[1] * (efw[0][i] * efw[0][i] + efw[1][i] * efw[1][i]);
// We threshold here to protect against the ill-effects of a zero farend.
// The threshold is not arbitrarily chosen, but balances protection and
// adverse interaction with the algorithm's tuning.
// TODO(bjornv): investigate further why this is so sensitive.
aec->sx[i] =
ptrGCoh[0] * aec->sx[i] +
ptrGCoh[1] * WEBRTC_SPL_MAX(
xfw[0][i] * xfw[0][i] + xfw[1][i] * xfw[1][i],
WebRtcAec_kMinFarendPSD);
aec->sde[i][0] =
ptrGCoh[0] * aec->sde[i][0] +
ptrGCoh[1] * (dfw[0][i] * efw[0][i] + dfw[1][i] * efw[1][i]);
aec->sde[i][1] =
ptrGCoh[0] * aec->sde[i][1] +
ptrGCoh[1] * (dfw[0][i] * efw[1][i] - dfw[1][i] * efw[0][i]);
aec->sxd[i][0] =
ptrGCoh[0] * aec->sxd[i][0] +
ptrGCoh[1] * (dfw[0][i] * xfw[0][i] + dfw[1][i] * xfw[1][i]);
aec->sxd[i][1] =
ptrGCoh[0] * aec->sxd[i][1] +
ptrGCoh[1] * (dfw[0][i] * xfw[1][i] - dfw[1][i] * xfw[0][i]);
sdSum += aec->sd[i];
seSum += aec->se[i];
}
// Divergent filter safeguard.
aec->divergeState = (aec->divergeState ? 1.05f : 1.0f) * seSum > sdSum;
if (aec->divergeState)
memcpy(efw, dfw, sizeof(efw[0][0]) * 2 * PART_LEN1);
// Reset if error is significantly larger than nearend (13 dB).
if (!aec->extended_filter_enabled && seSum > (19.95f * sdSum))
memset(aec->wfBuf, 0, sizeof(aec->wfBuf));
}
// Window time domain data to be used by the fft.
__inline static void WindowData(float* x_windowed, const float* x) {
int i;
for (i = 0; i < PART_LEN; i += 4) {
const __m128 vec_Buf1 = _mm_loadu_ps(&x[i]);
const __m128 vec_Buf2 = _mm_loadu_ps(&x[PART_LEN + i]);
const __m128 vec_sqrtHanning = _mm_load_ps(&WebRtcAec_sqrtHanning[i]);
// A B C D
__m128 vec_sqrtHanning_rev =
_mm_loadu_ps(&WebRtcAec_sqrtHanning[PART_LEN - i - 3]);
// D C B A
vec_sqrtHanning_rev =
_mm_shuffle_ps(vec_sqrtHanning_rev, vec_sqrtHanning_rev,
_MM_SHUFFLE(0, 1, 2, 3));
_mm_storeu_ps(&x_windowed[i], _mm_mul_ps(vec_Buf1, vec_sqrtHanning));
_mm_storeu_ps(&x_windowed[PART_LEN + i],
_mm_mul_ps(vec_Buf2, vec_sqrtHanning_rev));
}
}
// Puts fft output data into a complex valued array.
__inline static void StoreAsComplex(const float* data,
float data_complex[2][PART_LEN1]) {
int i;
for (i = 0; i < PART_LEN; i += 4) {
const __m128 vec_fft0 = _mm_loadu_ps(&data[2 * i]);
const __m128 vec_fft4 = _mm_loadu_ps(&data[2 * i + 4]);
const __m128 vec_a = _mm_shuffle_ps(vec_fft0, vec_fft4,
_MM_SHUFFLE(2, 0, 2, 0));
const __m128 vec_b = _mm_shuffle_ps(vec_fft0, vec_fft4,
_MM_SHUFFLE(3, 1, 3, 1));
_mm_storeu_ps(&data_complex[0][i], vec_a);
_mm_storeu_ps(&data_complex[1][i], vec_b);
}
// fix beginning/end values
data_complex[1][0] = 0;
data_complex[1][PART_LEN] = 0;
data_complex[0][0] = data[0];
data_complex[0][PART_LEN] = data[1];
}
static void SubbandCoherenceSSE2(AecCore* aec,
float efw[2][PART_LEN1],
float xfw[2][PART_LEN1],
float* fft,
float* cohde,
float* cohxd) {
float dfw[2][PART_LEN1];
int i;
if (aec->delayEstCtr == 0)
aec->delayIdx = PartitionDelay(aec);
// Use delayed far.
memcpy(xfw,
aec->xfwBuf + aec->delayIdx * PART_LEN1,
sizeof(xfw[0][0]) * 2 * PART_LEN1);
// Windowed near fft
WindowData(fft, aec->dBuf);
aec_rdft_forward_128(fft);
StoreAsComplex(fft, dfw);
// Windowed error fft
WindowData(fft, aec->eBuf);
aec_rdft_forward_128(fft);
StoreAsComplex(fft, efw);
SmoothedPSD(aec, efw, dfw, xfw);
{
const __m128 vec_1eminus10 = _mm_set1_ps(1e-10f);
// Subband coherence
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const __m128 vec_sd = _mm_loadu_ps(&aec->sd[i]);
const __m128 vec_se = _mm_loadu_ps(&aec->se[i]);
const __m128 vec_sx = _mm_loadu_ps(&aec->sx[i]);
const __m128 vec_sdse = _mm_add_ps(vec_1eminus10,
_mm_mul_ps(vec_sd, vec_se));
const __m128 vec_sdsx = _mm_add_ps(vec_1eminus10,
_mm_mul_ps(vec_sd, vec_sx));
const __m128 vec_sde_3210 = _mm_loadu_ps(&aec->sde[i][0]);
const __m128 vec_sde_7654 = _mm_loadu_ps(&aec->sde[i + 2][0]);
const __m128 vec_sxd_3210 = _mm_loadu_ps(&aec->sxd[i][0]);
const __m128 vec_sxd_7654 = _mm_loadu_ps(&aec->sxd[i + 2][0]);
const __m128 vec_sde_0 = _mm_shuffle_ps(vec_sde_3210, vec_sde_7654,
_MM_SHUFFLE(2, 0, 2, 0));
const __m128 vec_sde_1 = _mm_shuffle_ps(vec_sde_3210, vec_sde_7654,
_MM_SHUFFLE(3, 1, 3, 1));
const __m128 vec_sxd_0 = _mm_shuffle_ps(vec_sxd_3210, vec_sxd_7654,
_MM_SHUFFLE(2, 0, 2, 0));
const __m128 vec_sxd_1 = _mm_shuffle_ps(vec_sxd_3210, vec_sxd_7654,
_MM_SHUFFLE(3, 1, 3, 1));
__m128 vec_cohde = _mm_mul_ps(vec_sde_0, vec_sde_0);
__m128 vec_cohxd = _mm_mul_ps(vec_sxd_0, vec_sxd_0);
vec_cohde = _mm_add_ps(vec_cohde, _mm_mul_ps(vec_sde_1, vec_sde_1));
vec_cohde = _mm_div_ps(vec_cohde, vec_sdse);
vec_cohxd = _mm_add_ps(vec_cohxd, _mm_mul_ps(vec_sxd_1, vec_sxd_1));
vec_cohxd = _mm_div_ps(vec_cohxd, vec_sdsx);
_mm_storeu_ps(&cohde[i], vec_cohde);
_mm_storeu_ps(&cohxd[i], vec_cohxd);
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
cohde[i] =
(aec->sde[i][0] * aec->sde[i][0] + aec->sde[i][1] * aec->sde[i][1]) /
(aec->sd[i] * aec->se[i] + 1e-10f);
cohxd[i] =
(aec->sxd[i][0] * aec->sxd[i][0] + aec->sxd[i][1] * aec->sxd[i][1]) /
(aec->sx[i] * aec->sd[i] + 1e-10f);
}
}
}
void WebRtcAec_InitAec_SSE2(void) {
WebRtcAec_FilterFar = FilterFarSSE2;
WebRtcAec_ScaleErrorSignal = ScaleErrorSignalSSE2;
WebRtcAec_FilterAdaptation = FilterAdaptationSSE2;
WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressSSE2;
WebRtcAec_SubbandCoherence = SubbandCoherenceSSE2;
}

589
webrtc/modules/audio_processing/aec/aec_rdft.c
View File

@ -1,589 +0,0 @@
/*
* http://www.kurims.kyoto-u.ac.jp/~ooura/fft.html
* Copyright Takuya OOURA, 1996-2001
*
* You may use, copy, modify and distribute this code for any purpose (include
* commercial use) and without fee. Please refer to this package when you modify
* this code.
*
* Changes by the WebRTC authors:
* - Trivial type modifications.
* - Minimal code subset to do rdft of length 128.
* - Optimizations because of known length.
*
* All changes are covered by the WebRTC license and IP grant:
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "webrtc/modules/audio_processing/aec/aec_rdft.h"
#include <math.h>
#include "webrtc/system_wrappers/include/cpu_features_wrapper.h"
#include "webrtc/typedefs.h"
// These tables used to be computed at run-time. For example, refer to:
// https://code.google.com/p/webrtc/source/browse/trunk/webrtc/modules/audio_processing/aec/aec_rdft.c?r=6564
// to see the initialization code.
const float rdft_w[64] = {
1.0000000000f, 0.0000000000f, 0.7071067691f, 0.7071067691f,
0.9238795638f, 0.3826834559f, 0.3826834559f, 0.9238795638f,
0.9807852507f, 0.1950903237f, 0.5555702448f, 0.8314695954f,
0.8314695954f, 0.5555702448f, 0.1950903237f, 0.9807852507f,
0.9951847196f, 0.0980171412f, 0.6343933344f, 0.7730104327f,
0.8819212914f, 0.4713967443f, 0.2902846634f, 0.9569403529f,
0.9569403529f, 0.2902846634f, 0.4713967443f, 0.8819212914f,
0.7730104327f, 0.6343933344f, 0.0980171412f, 0.9951847196f,
0.7071067691f, 0.4993977249f, 0.4975923598f, 0.4945882559f,
0.4903926253f, 0.4850156307f, 0.4784701765f, 0.4707720280f,
0.4619397819f, 0.4519946277f, 0.4409606457f, 0.4288643003f,
0.4157347977f, 0.4016037583f, 0.3865052164f, 0.3704755902f,
0.3535533845f, 0.3357794881f, 0.3171966672f, 0.2978496552f,
0.2777851224f, 0.2570513785f, 0.2356983721f, 0.2137775421f,
0.1913417280f, 0.1684449315f, 0.1451423317f, 0.1214900985f,
0.0975451618f, 0.0733652338f, 0.0490085706f, 0.0245338380f,
};
const float rdft_wk3ri_first[16] = {
1.000000000f, 0.000000000f, 0.382683456f, 0.923879564f,
0.831469536f, 0.555570245f, -0.195090353f, 0.980785251f,
0.956940353f, 0.290284693f, 0.098017156f, 0.995184720f,
0.634393334f, 0.773010492f, -0.471396863f, 0.881921172f,
};
const float rdft_wk3ri_second[16] = {
-0.707106769f, 0.707106769f, -0.923879564f, -0.382683456f,
-0.980785251f, 0.195090353f, -0.555570245f, -0.831469536f,
-0.881921172f, 0.471396863f, -0.773010492f, -0.634393334f,
-0.995184720f, -0.098017156f, -0.290284693f, -0.956940353f,
};
ALIGN16_BEG const float ALIGN16_END rdft_wk1r[32] = {
1.000000000f, 1.000000000f, 0.707106769f, 0.707106769f,
0.923879564f, 0.923879564f, 0.382683456f, 0.382683456f,
0.980785251f, 0.980785251f, 0.555570245f, 0.555570245f,
0.831469595f, 0.831469595f, 0.195090324f, 0.195090324f,
0.995184720f, 0.995184720f, 0.634393334f, 0.634393334f,
0.881921291f, 0.881921291f, 0.290284663f, 0.290284663f,
0.956940353f, 0.956940353f, 0.471396744f, 0.471396744f,
0.773010433f, 0.773010433f, 0.098017141f, 0.098017141f,
};
ALIGN16_BEG const float ALIGN16_END rdft_wk2r[32] = {
1.000000000f, 1.000000000f, -0.000000000f, -0.000000000f,
0.707106769f, 0.707106769f, -0.707106769f, -0.707106769f,
0.923879564f, 0.923879564f, -0.382683456f, -0.382683456f,
0.382683456f, 0.382683456f, -0.923879564f, -0.923879564f,
0.980785251f, 0.980785251f, -0.195090324f, -0.195090324f,
0.555570245f, 0.555570245f, -0.831469595f, -0.831469595f,
0.831469595f, 0.831469595f, -0.555570245f, -0.555570245f,
0.195090324f, 0.195090324f, -0.980785251f, -0.980785251f,
};
ALIGN16_BEG const float ALIGN16_END rdft_wk3r[32] = {
1.000000000f, 1.000000000f, -0.707106769f, -0.707106769f,
0.382683456f, 0.382683456f, -0.923879564f, -0.923879564f,
0.831469536f, 0.831469536f, -0.980785251f, -0.980785251f,
-0.195090353f, -0.195090353f, -0.555570245f, -0.555570245f,
0.956940353f, 0.956940353f, -0.881921172f, -0.881921172f,
0.098017156f, 0.098017156f, -0.773010492f, -0.773010492f,
0.634393334f, 0.634393334f, -0.995184720f, -0.995184720f,
-0.471396863f, -0.471396863f, -0.290284693f, -0.290284693f,
};
ALIGN16_BEG const float ALIGN16_END rdft_wk1i[32] = {
-0.000000000f, 0.000000000f, -0.707106769f, 0.707106769f,
-0.382683456f, 0.382683456f, -0.923879564f, 0.923879564f,
-0.195090324f, 0.195090324f, -0.831469595f, 0.831469595f,
-0.555570245f, 0.555570245f, -0.980785251f, 0.980785251f,
-0.098017141f, 0.098017141f, -0.773010433f, 0.773010433f,
-0.471396744f, 0.471396744f, -0.956940353f, 0.956940353f,
-0.290284663f, 0.290284663f, -0.881921291f, 0.881921291f,
-0.634393334f, 0.634393334f, -0.995184720f, 0.995184720f,
};
ALIGN16_BEG const float ALIGN16_END rdft_wk2i[32] = {
-0.000000000f, 0.000000000f, -1.000000000f, 1.000000000f,
-0.707106769f, 0.707106769f, -0.707106769f, 0.707106769f,
-0.382683456f, 0.382683456f, -0.923879564f, 0.923879564f,
-0.923879564f, 0.923879564f, -0.382683456f, 0.382683456f,
-0.195090324f, 0.195090324f, -0.980785251f, 0.980785251f,
-0.831469595f, 0.831469595f, -0.555570245f, 0.555570245f,
-0.555570245f, 0.555570245f, -0.831469595f, 0.831469595f,
-0.980785251f, 0.980785251f, -0.195090324f, 0.195090324f,
};
ALIGN16_BEG const float ALIGN16_END rdft_wk3i[32] = {
-0.000000000f, 0.000000000f, -0.707106769f, 0.707106769f,
-0.923879564f, 0.923879564f, 0.382683456f, -0.382683456f,
-0.555570245f, 0.555570245f, -0.195090353f, 0.195090353f,
-0.980785251f, 0.980785251f, 0.831469536f, -0.831469536f,
-0.290284693f, 0.290284693f, -0.471396863f, 0.471396863f,
-0.995184720f, 0.995184720f, 0.634393334f, -0.634393334f,
-0.773010492f, 0.773010492f, 0.098017156f, -0.098017156f,
-0.881921172f, 0.881921172f, 0.956940353f, -0.956940353f,
};
ALIGN16_BEG const float ALIGN16_END cftmdl_wk1r[4] = {
0.707106769f, 0.707106769f, 0.707106769f, -0.707106769f,
};
static void bitrv2_128_C(float* a) {
/*
Following things have been attempted but are no faster:
(a) Storing the swap indexes in a LUT (index calculations are done
for 'free' while waiting on memory/L1).
(b) Consolidate the load/store of two consecutive floats by a 64 bit
integer (execution is memory/L1 bound).
(c) Do a mix of floats and 64 bit integer to maximize register
utilization (execution is memory/L1 bound).
(d) Replacing ip[i] by ((k<<31)>>25) + ((k >> 1)<<5).
(e) Hard-coding of the offsets to completely eliminates index
calculations.
*/
unsigned int j, j1, k, k1;
float xr, xi, yr, yi;
static const int ip[4] = {0, 64, 32, 96};
for (k = 0; k < 4; k++) {
for (j = 0; j < k; j++) {
j1 = 2 * j + ip[k];
k1 = 2 * k + ip[j];
xr = a[j1 + 0];
xi = a[j1 + 1];
yr = a[k1 + 0];
yi = a[k1 + 1];
a[j1 + 0] = yr;
a[j1 + 1] = yi;
a[k1 + 0] = xr;
a[k1 + 1] = xi;
j1 += 8;
k1 += 16;
xr = a[j1 + 0];
xi = a[j1 + 1];
yr = a[k1 + 0];
yi = a[k1 + 1];
a[j1 + 0] = yr;
a[j1 + 1] = yi;
a[k1 + 0] = xr;
a[k1 + 1] = xi;
j1 += 8;
k1 -= 8;
xr = a[j1 + 0];
xi = a[j1 + 1];
yr = a[k1 + 0];
yi = a[k1 + 1];
a[j1 + 0] = yr;
a[j1 + 1] = yi;
a[k1 + 0] = xr;
a[k1 + 1] = xi;
j1 += 8;
k1 += 16;
xr = a[j1 + 0];
xi = a[j1 + 1];
yr = a[k1 + 0];
yi = a[k1 + 1];
a[j1 + 0] = yr;
a[j1 + 1] = yi;
a[k1 + 0] = xr;
a[k1 + 1] = xi;
}
j1 = 2 * k + 8 + ip[k];
k1 = j1 + 8;
xr = a[j1 + 0];
xi = a[j1 + 1];
yr = a[k1 + 0];
yi = a[k1 + 1];
a[j1 + 0] = yr;
a[j1 + 1] = yi;
a[k1 + 0] = xr;
a[k1 + 1] = xi;
}
}
static void cft1st_128_C(float* a) {
const int n = 128;
int j, k1, k2;
float wk1r, wk1i, wk2r, wk2i, wk3r, wk3i;
float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
// The processing of the first set of elements was simplified in C to avoid
// some operations (multiplication by zero or one, addition of two elements
// multiplied by the same weight, ...).
x0r = a[0] + a[2];
x0i = a[1] + a[3];
x1r = a[0] - a[2];
x1i = a[1] - a[3];
x2r = a[4] + a[6];
x2i = a[5] + a[7];
x3r = a[4] - a[6];
x3i = a[5] - a[7];
a[0] = x0r + x2r;
a[1] = x0i + x2i;
a[4] = x0r - x2r;
a[5] = x0i - x2i;
a[2] = x1r - x3i;
a[3] = x1i + x3r;
a[6] = x1r + x3i;
a[7] = x1i - x3r;
wk1r = rdft_w[2];
x0r = a[8] + a[10];
x0i = a[9] + a[11];
x1r = a[8] - a[10];
x1i = a[9] - a[11];
x2r = a[12] + a[14];
x2i = a[13] + a[15];
x3r = a[12] - a[14];
x3i = a[13] - a[15];
a[8] = x0r + x2r;
a[9] = x0i + x2i;
a[12] = x2i - x0i;
a[13] = x0r - x2r;
x0r = x1r - x3i;
x0i = x1i + x3r;
a[10] = wk1r * (x0r - x0i);
a[11] = wk1r * (x0r + x0i);
x0r = x3i + x1r;
x0i = x3r - x1i;
a[14] = wk1r * (x0i - x0r);
a[15] = wk1r * (x0i + x0r);
k1 = 0;
for (j = 16; j < n; j += 16) {
k1 += 2;
k2 = 2 * k1;
wk2r = rdft_w[k1 + 0];
wk2i = rdft_w[k1 + 1];
wk1r = rdft_w[k2 + 0];
wk1i = rdft_w[k2 + 1];
wk3r = rdft_wk3ri_first[k1 + 0];
wk3i = rdft_wk3ri_first[k1 + 1];
x0r = a[j + 0] + a[j + 2];
x0i = a[j + 1] + a[j + 3];
x1r = a[j + 0] - a[j + 2];
x1i = a[j + 1] - a[j + 3];
x2r = a[j + 4] + a[j + 6];
x2i = a[j + 5] + a[j + 7];
x3r = a[j + 4] - a[j + 6];
x3i = a[j + 5] - a[j + 7];
a[j + 0] = x0r + x2r;
a[j + 1] = x0i + x2i;
x0r -= x2r;
x0i -= x2i;
a[j + 4] = wk2r * x0r - wk2i * x0i;
a[j + 5] = wk2r * x0i + wk2i * x0r;
x0r = x1r - x3i;
x0i = x1i + x3r;
a[j + 2] = wk1r * x0r - wk1i * x0i;
a[j + 3] = wk1r * x0i + wk1i * x0r;
x0r = x1r + x3i;
x0i = x1i - x3r;
a[j + 6] = wk3r * x0r - wk3i * x0i;
a[j + 7] = wk3r * x0i + wk3i * x0r;
wk1r = rdft_w[k2 + 2];
wk1i = rdft_w[k2 + 3];
wk3r = rdft_wk3ri_second[k1 + 0];
wk3i = rdft_wk3ri_second[k1 + 1];
x0r = a[j + 8] + a[j + 10];
x0i = a[j + 9] + a[j + 11];
x1r = a[j + 8] - a[j + 10];
x1i = a[j + 9] - a[j + 11];
x2r = a[j + 12] + a[j + 14];
x2i = a[j + 13] + a[j + 15];
x3r = a[j + 12] - a[j + 14];
x3i = a[j + 13] - a[j + 15];
a[j + 8] = x0r + x2r;
a[j + 9] = x0i + x2i;
x0r -= x2r;
x0i -= x2i;
a[j + 12] = -wk2i * x0r - wk2r * x0i;
a[j + 13] = -wk2i * x0i + wk2r * x0r;
x0r = x1r - x3i;
x0i = x1i + x3r;
a[j + 10] = wk1r * x0r - wk1i * x0i;
a[j + 11] = wk1r * x0i + wk1i * x0r;
x0r = x1r + x3i;
x0i = x1i - x3r;
a[j + 14] = wk3r * x0r - wk3i * x0i;
a[j + 15] = wk3r * x0i + wk3i * x0r;
}
}
static void cftmdl_128_C(float* a) {
const int l = 8;
const int n = 128;
const int m = 32;
int j0, j1, j2, j3, k, k1, k2, m2;
float wk1r, wk1i, wk2r, wk2i, wk3r, wk3i;
float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
for (j0 = 0; j0 < l; j0 += 2) {
j1 = j0 + 8;
j2 = j0 + 16;
j3 = j0 + 24;
x0r = a[j0 + 0] + a[j1 + 0];
x0i = a[j0 + 1] + a[j1 + 1];
x1r = a[j0 + 0] - a[j1 + 0];
x1i = a[j0 + 1] - a[j1 + 1];
x2r = a[j2 + 0] + a[j3 + 0];
x2i = a[j2 + 1] + a[j3 + 1];
x3r = a[j2 + 0] - a[j3 + 0];
x3i = a[j2 + 1] - a[j3 + 1];
a[j0 + 0] = x0r + x2r;
a[j0 + 1] = x0i + x2i;
a[j2 + 0] = x0r - x2r;
a[j2 + 1] = x0i - x2i;
a[j1 + 0] = x1r - x3i;
a[j1 + 1] = x1i + x3r;
a[j3 + 0] = x1r + x3i;
a[j3 + 1] = x1i - x3r;
}
wk1r = rdft_w[2];
for (j0 = m; j0 < l + m; j0 += 2) {
j1 = j0 + 8;
j2 = j0 + 16;
j3 = j0 + 24;
x0r = a[j0 + 0] + a[j1 + 0];
x0i = a[j0 + 1] + a[j1 + 1];
x1r = a[j0 + 0] - a[j1 + 0];
x1i = a[j0 + 1] - a[j1 + 1];
x2r = a[j2 + 0] + a[j3 + 0];
x2i = a[j2 + 1] + a[j3 + 1];
x3r = a[j2 + 0] - a[j3 + 0];
x3i = a[j2 + 1] - a[j3 + 1];
a[j0 + 0] = x0r + x2r;
a[j0 + 1] = x0i + x2i;
a[j2 + 0] = x2i - x0i;
a[j2 + 1] = x0r - x2r;
x0r = x1r - x3i;
x0i = x1i + x3r;
a[j1 + 0] = wk1r * (x0r - x0i);
a[j1 + 1] = wk1r * (x0r + x0i);
x0r = x3i + x1r;
x0i = x3r - x1i;
a[j3 + 0] = wk1r * (x0i - x0r);
a[j3 + 1] = wk1r * (x0i + x0r);
}
k1 = 0;
m2 = 2 * m;
for (k = m2; k < n; k += m2) {
k1 += 2;
k2 = 2 * k1;
wk2r = rdft_w[k1 + 0];
wk2i = rdft_w[k1 + 1];
wk1r = rdft_w[k2 + 0];
wk1i = rdft_w[k2 + 1];
wk3r = rdft_wk3ri_first[k1 + 0];
wk3i = rdft_wk3ri_first[k1 + 1];
for (j0 = k; j0 < l + k; j0 += 2) {
j1 = j0 + 8;
j2 = j0 + 16;
j3 = j0 + 24;
x0r = a[j0 + 0] + a[j1 + 0];
x0i = a[j0 + 1] + a[j1 + 1];
x1r = a[j0 + 0] - a[j1 + 0];
x1i = a[j0 + 1] - a[j1 + 1];
x2r = a[j2 + 0] + a[j3 + 0];
x2i = a[j2 + 1] + a[j3 + 1];
x3r = a[j2 + 0] - a[j3 + 0];
x3i = a[j2 + 1] - a[j3 + 1];
a[j0 + 0] = x0r + x2r;
a[j0 + 1] = x0i + x2i;
x0r -= x2r;
x0i -= x2i;
a[j2 + 0] = wk2r * x0r - wk2i * x0i;
a[j2 + 1] = wk2r * x0i + wk2i * x0r;
x0r = x1r - x3i;
x0i = x1i + x3r;
a[j1 + 0] = wk1r * x0r - wk1i * x0i;
a[j1 + 1] = wk1r * x0i + wk1i * x0r;
x0r = x1r + x3i;
x0i = x1i - x3r;
a[j3 + 0] = wk3r * x0r - wk3i * x0i;
a[j3 + 1] = wk3r * x0i + wk3i * x0r;
}
wk1r = rdft_w[k2 + 2];
wk1i = rdft_w[k2 + 3];
wk3r = rdft_wk3ri_second[k1 + 0];
wk3i = rdft_wk3ri_second[k1 + 1];
for (j0 = k + m; j0 < l + (k + m); j0 += 2) {
j1 = j0 + 8;
j2 = j0 + 16;
j3 = j0 + 24;
x0r = a[j0 + 0] + a[j1 + 0];
x0i = a[j0 + 1] + a[j1 + 1];
x1r = a[j0 + 0] - a[j1 + 0];
x1i = a[j0 + 1] - a[j1 + 1];
x2r = a[j2 + 0] + a[j3 + 0];
x2i = a[j2 + 1] + a[j3 + 1];
x3r = a[j2 + 0] - a[j3 + 0];
x3i = a[j2 + 1] - a[j3 + 1];
a[j0 + 0] = x0r + x2r;
a[j0 + 1] = x0i + x2i;
x0r -= x2r;
x0i -= x2i;
a[j2 + 0] = -wk2i * x0r - wk2r * x0i;
a[j2 + 1] = -wk2i * x0i + wk2r * x0r;
x0r = x1r - x3i;
x0i = x1i + x3r;
a[j1 + 0] = wk1r * x0r - wk1i * x0i;
a[j1 + 1] = wk1r * x0i + wk1i * x0r;
x0r = x1r + x3i;
x0i = x1i - x3r;
a[j3 + 0] = wk3r * x0r - wk3i * x0i;
a[j3 + 1] = wk3r * x0i + wk3i * x0r;
}
}
}
static void cftfsub_128_C(float* a) {
int j, j1, j2, j3, l;
float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
cft1st_128(a);
cftmdl_128(a);
l = 32;
for (j = 0; j < l; j += 2) {
j1 = j + l;
j2 = j1 + l;
j3 = j2 + l;
x0r = a[j] + a[j1];
x0i = a[j + 1] + a[j1 + 1];
x1r = a[j] - a[j1];
x1i = a[j + 1] - a[j1 + 1];
x2r = a[j2] + a[j3];
x2i = a[j2 + 1] + a[j3 + 1];
x3r = a[j2] - a[j3];
x3i = a[j2 + 1] - a[j3 + 1];
a[j] = x0r + x2r;
a[j + 1] = x0i + x2i;
a[j2] = x0r - x2r;
a[j2 + 1] = x0i - x2i;
a[j1] = x1r - x3i;
a[j1 + 1] = x1i + x3r;
a[j3] = x1r + x3i;
a[j3 + 1] = x1i - x3r;
}
}
static void cftbsub_128_C(float* a) {
int j, j1, j2, j3, l;
float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
cft1st_128(a);
cftmdl_128(a);
l = 32;
for (j = 0; j < l; j += 2) {
j1 = j + l;
j2 = j1 + l;
j3 = j2 + l;
x0r = a[j] + a[j1];
x0i = -a[j + 1] - a[j1 + 1];
x1r = a[j] - a[j1];
x1i = -a[j + 1] + a[j1 + 1];
x2r = a[j2] + a[j3];
x2i = a[j2 + 1] + a[j3 + 1];
x3r = a[j2] - a[j3];
x3i = a[j2 + 1] - a[j3 + 1];
a[j] = x0r + x2r;
a[j + 1] = x0i - x2i;
a[j2] = x0r - x2r;
a[j2 + 1] = x0i + x2i;
a[j1] = x1r - x3i;
a[j1 + 1] = x1i - x3r;
a[j3] = x1r + x3i;
a[j3 + 1] = x1i + x3r;
}
}
static void rftfsub_128_C(float* a) {
const float* c = rdft_w + 32;
int j1, j2, k1, k2;
float wkr, wki, xr, xi, yr, yi;
for (j1 = 1, j2 = 2; j2 < 64; j1 += 1, j2 += 2) {
k2 = 128 - j2;
k1 = 32 - j1;
wkr = 0.5f - c[k1];
wki = c[j1];
xr = a[j2 + 0] - a[k2 + 0];
xi = a[j2 + 1] + a[k2 + 1];
yr = wkr * xr - wki * xi;
yi = wkr * xi + wki * xr;
a[j2 + 0] -= yr;
a[j2 + 1] -= yi;
a[k2 + 0] += yr;
a[k2 + 1] -= yi;
}
}
static void rftbsub_128_C(float* a) {
const float* c = rdft_w + 32;
int j1, j2, k1, k2;
float wkr, wki, xr, xi, yr, yi;
a[1] = -a[1];
for (j1 = 1, j2 = 2; j2 < 64; j1 += 1, j2 += 2) {
k2 = 128 - j2;
k1 = 32 - j1;
wkr = 0.5f - c[k1];
wki = c[j1];
xr = a[j2 + 0] - a[k2 + 0];
xi = a[j2 + 1] + a[k2 + 1];
yr = wkr * xr + wki * xi;
yi = wkr * xi - wki * xr;
a[j2 + 0] = a[j2 + 0] - yr;
a[j2 + 1] = yi - a[j2 + 1];
a[k2 + 0] = yr + a[k2 + 0];
a[k2 + 1] = yi - a[k2 + 1];
}
a[65] = -a[65];
}
void aec_rdft_forward_128(float* a) {
float xi;
bitrv2_128(a);
cftfsub_128(a);
rftfsub_128(a);
xi = a[0] - a[1];
a[0] += a[1];
a[1] = xi;
}
void aec_rdft_inverse_128(float* a) {
a[1] = 0.5f * (a[0] - a[1]);
a[0] -= a[1];
rftbsub_128(a);
bitrv2_128(a);
cftbsub_128(a);
}
// code path selection
RftSub128 cft1st_128;
RftSub128 cftmdl_128;
RftSub128 rftfsub_128;
RftSub128 rftbsub_128;
RftSub128 cftfsub_128;
RftSub128 cftbsub_128;
RftSub128 bitrv2_128;
void aec_rdft_init(void) {
cft1st_128 = cft1st_128_C;
cftmdl_128 = cftmdl_128_C;
rftfsub_128 = rftfsub_128_C;
rftbsub_128 = rftbsub_128_C;
cftfsub_128 = cftfsub_128_C;
cftbsub_128 = cftbsub_128_C;
bitrv2_128 = bitrv2_128_C;
#if defined(WEBRTC_ARCH_X86_FAMILY)
if (WebRtc_GetCPUInfo(kSSE2)) {
aec_rdft_init_sse2();
}
#endif
#if defined(MIPS_FPU_LE)
aec_rdft_init_mips();
#endif
#if defined(WEBRTC_HAS_NEON)
aec_rdft_init_neon();
#elif defined(WEBRTC_DETECT_NEON)
if ((WebRtc_GetCPUFeaturesARM() & kCPUFeatureNEON) != 0) {
aec_rdft_init_neon();
}
#endif
}

61
webrtc/modules/audio_processing/aec/aec_rdft.h
View File

@ -1,61 +0,0 @@
/*
* Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC_MAIN_SOURCE_AEC_RDFT_H_
#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_MAIN_SOURCE_AEC_RDFT_H_
#include "webrtc/modules/audio_processing/aec/aec_common.h"
// These intrinsics were unavailable before VS 2008.
// TODO(andrew): move to a common file.
#if defined(_MSC_VER) && _MSC_VER < 1500
#include <emmintrin.h>
static __inline __m128 _mm_castsi128_ps(__m128i a) { return *(__m128*)&a; }
static __inline __m128i _mm_castps_si128(__m128 a) { return *(__m128i*)&a; }
#endif
// Constants shared by all paths (C, SSE2, NEON).
extern const float rdft_w[64];
// Constants used by the C path.
extern const float rdft_wk3ri_first[16];
extern const float rdft_wk3ri_second[16];
// Constants used by SSE2 and NEON but initialized in the C path.
extern ALIGN16_BEG const float ALIGN16_END rdft_wk1r[32];
extern ALIGN16_BEG const float ALIGN16_END rdft_wk2r[32];
extern ALIGN16_BEG const float ALIGN16_END rdft_wk3r[32];
extern ALIGN16_BEG const float ALIGN16_END rdft_wk1i[32];
extern ALIGN16_BEG const float ALIGN16_END rdft_wk2i[32];
extern ALIGN16_BEG const float ALIGN16_END rdft_wk3i[32];
extern ALIGN16_BEG const float ALIGN16_END cftmdl_wk1r[4];
// code path selection function pointers
typedef void (*RftSub128)(float* a);
extern RftSub128 rftfsub_128;
extern RftSub128 rftbsub_128;
extern RftSub128 cft1st_128;
extern RftSub128 cftmdl_128;
extern RftSub128 cftfsub_128;
extern RftSub128 cftbsub_128;
extern RftSub128 bitrv2_128;
// entry points
void aec_rdft_init(void);
void aec_rdft_init_sse2(void);
void aec_rdft_forward_128(float* a);
void aec_rdft_inverse_128(float* a);
#if defined(MIPS_FPU_LE)
void aec_rdft_init_mips(void);
#endif
#if defined(WEBRTC_DETECT_NEON) || defined(WEBRTC_HAS_NEON)
void aec_rdft_init_neon(void);
#endif
#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_MAIN_SOURCE_AEC_RDFT_H_

1187
webrtc/modules/audio_processing/aec/aec_rdft_mips.c
View File

File diff suppressed because it is too large Load Diff

355
webrtc/modules/audio_processing/aec/aec_rdft_neon.c
View File

@ -1,355 +0,0 @@
/*
* Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
/*
* The rdft AEC algorithm, neon version of speed-critical functions.
*
* Based on the sse2 version.
*/
#include "webrtc/modules/audio_processing/aec/aec_rdft.h"
#include <arm_neon.h>
static const ALIGN16_BEG float ALIGN16_END
k_swap_sign[4] = {-1.f, 1.f, -1.f, 1.f};
static void cft1st_128_neon(float* a) {
const float32x4_t vec_swap_sign = vld1q_f32((float32_t*)k_swap_sign);
int j, k2;
for (k2 = 0, j = 0; j < 128; j += 16, k2 += 4) {
float32x4_t a00v = vld1q_f32(&a[j + 0]);
float32x4_t a04v = vld1q_f32(&a[j + 4]);
float32x4_t a08v = vld1q_f32(&a[j + 8]);
float32x4_t a12v = vld1q_f32(&a[j + 12]);
float32x4_t a01v = vcombine_f32(vget_low_f32(a00v), vget_low_f32(a08v));
float32x4_t a23v = vcombine_f32(vget_high_f32(a00v), vget_high_f32(a08v));
float32x4_t a45v = vcombine_f32(vget_low_f32(a04v), vget_low_f32(a12v));
float32x4_t a67v = vcombine_f32(vget_high_f32(a04v), vget_high_f32(a12v));
const float32x4_t wk1rv = vld1q_f32(&rdft_wk1r[k2]);
const float32x4_t wk1iv = vld1q_f32(&rdft_wk1i[k2]);
const float32x4_t wk2rv = vld1q_f32(&rdft_wk2r[k2]);
const float32x4_t wk2iv = vld1q_f32(&rdft_wk2i[k2]);
const float32x4_t wk3rv = vld1q_f32(&rdft_wk3r[k2]);
const float32x4_t wk3iv = vld1q_f32(&rdft_wk3i[k2]);
float32x4_t x0v = vaddq_f32(a01v, a23v);
const float32x4_t x1v = vsubq_f32(a01v, a23v);
const float32x4_t x2v = vaddq_f32(a45v, a67v);
const float32x4_t x3v = vsubq_f32(a45v, a67v);
const float32x4_t x3w = vrev64q_f32(x3v);
float32x4_t x0w;
a01v = vaddq_f32(x0v, x2v);
x0v = vsubq_f32(x0v, x2v);
x0w = vrev64q_f32(x0v);
a45v = vmulq_f32(wk2rv, x0v);
a45v = vmlaq_f32(a45v, wk2iv, x0w);
x0v = vmlaq_f32(x1v, x3w, vec_swap_sign);
x0w = vrev64q_f32(x0v);
a23v = vmulq_f32(wk1rv, x0v);
a23v = vmlaq_f32(a23v, wk1iv, x0w);
x0v = vmlsq_f32(x1v, x3w, vec_swap_sign);
x0w = vrev64q_f32(x0v);
a67v = vmulq_f32(wk3rv, x0v);
a67v = vmlaq_f32(a67v, wk3iv, x0w);
a00v = vcombine_f32(vget_low_f32(a01v), vget_low_f32(a23v));
a04v = vcombine_f32(vget_low_f32(a45v), vget_low_f32(a67v));
a08v = vcombine_f32(vget_high_f32(a01v), vget_high_f32(a23v));
a12v = vcombine_f32(vget_high_f32(a45v), vget_high_f32(a67v));
vst1q_f32(&a[j + 0], a00v);
vst1q_f32(&a[j + 4], a04v);
vst1q_f32(&a[j + 8], a08v);
vst1q_f32(&a[j + 12], a12v);
}
}
static void cftmdl_128_neon(float* a) {
int j;
const int l = 8;
const float32x4_t vec_swap_sign = vld1q_f32((float32_t*)k_swap_sign);
float32x4_t wk1rv = vld1q_f32(cftmdl_wk1r);
for (j = 0; j < l; j += 2) {
const float32x2_t a_00 = vld1_f32(&a[j + 0]);
const float32x2_t a_08 = vld1_f32(&a[j + 8]);
const float32x2_t a_32 = vld1_f32(&a[j + 32]);
const float32x2_t a_40 = vld1_f32(&a[j + 40]);
const float32x4_t a_00_32 = vcombine_f32(a_00, a_32);
const float32x4_t a_08_40 = vcombine_f32(a_08, a_40);
const float32x4_t x0r0_0i0_0r1_x0i1 = vaddq_f32(a_00_32, a_08_40);
const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40);
const float32x2_t a_16 = vld1_f32(&a[j + 16]);
const float32x2_t a_24 = vld1_f32(&a[j + 24]);
const float32x2_t a_48 = vld1_f32(&a[j + 48]);
const float32x2_t a_56 = vld1_f32(&a[j + 56]);
const float32x4_t a_16_48 = vcombine_f32(a_16, a_48);
const float32x4_t a_24_56 = vcombine_f32(a_24, a_56);
const float32x4_t x2r0_2i0_2r1_x2i1 = vaddq_f32(a_16_48, a_24_56);
const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56);
const float32x4_t xx0 = vaddq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const float32x4_t x3i0_3r0_3i1_x3r1 = vrev64q_f32(x3r0_3i0_3r1_x3i1);
const float32x4_t x1_x3_add =
vmlaq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1);
const float32x4_t x1_x3_sub =
vmlsq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1);
const float32x2_t yy0_a = vdup_lane_f32(vget_high_f32(x1_x3_add), 0);
const float32x2_t yy0_s = vdup_lane_f32(vget_high_f32(x1_x3_sub), 0);
const float32x4_t yy0_as = vcombine_f32(yy0_a, yy0_s);
const float32x2_t yy1_a = vdup_lane_f32(vget_high_f32(x1_x3_add), 1);
const float32x2_t yy1_s = vdup_lane_f32(vget_high_f32(x1_x3_sub), 1);
const float32x4_t yy1_as = vcombine_f32(yy1_a, yy1_s);
const float32x4_t yy0 = vmlaq_f32(yy0_as, vec_swap_sign, yy1_as);
const float32x4_t yy4 = vmulq_f32(wk1rv, yy0);
const float32x4_t xx1_rev = vrev64q_f32(xx1);
const float32x4_t yy4_rev = vrev64q_f32(yy4);
vst1_f32(&a[j + 0], vget_low_f32(xx0));
vst1_f32(&a[j + 32], vget_high_f32(xx0));
vst1_f32(&a[j + 16], vget_low_f32(xx1));
vst1_f32(&a[j + 48], vget_high_f32(xx1_rev));
a[j + 48] = -a[j + 48];
vst1_f32(&a[j + 8], vget_low_f32(x1_x3_add));
vst1_f32(&a[j + 24], vget_low_f32(x1_x3_sub));
vst1_f32(&a[j + 40], vget_low_f32(yy4));
vst1_f32(&a[j + 56], vget_high_f32(yy4_rev));
}
{
const int k = 64;
const int k1 = 2;
const int k2 = 2 * k1;
const float32x4_t wk2rv = vld1q_f32(&rdft_wk2r[k2 + 0]);
const float32x4_t wk2iv = vld1q_f32(&rdft_wk2i[k2 + 0]);
const float32x4_t wk1iv = vld1q_f32(&rdft_wk1i[k2 + 0]);
const float32x4_t wk3rv = vld1q_f32(&rdft_wk3r[k2 + 0]);
const float32x4_t wk3iv = vld1q_f32(&rdft_wk3i[k2 + 0]);
wk1rv = vld1q_f32(&rdft_wk1r[k2 + 0]);
for (j = k; j < l + k; j += 2) {
const float32x2_t a_00 = vld1_f32(&a[j + 0]);
const float32x2_t a_08 = vld1_f32(&a[j + 8]);
const float32x2_t a_32 = vld1_f32(&a[j + 32]);
const float32x2_t a_40 = vld1_f32(&a[j + 40]);
const float32x4_t a_00_32 = vcombine_f32(a_00, a_32);
const float32x4_t a_08_40 = vcombine_f32(a_08, a_40);
const float32x4_t x0r0_0i0_0r1_x0i1 = vaddq_f32(a_00_32, a_08_40);
const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40);
const float32x2_t a_16 = vld1_f32(&a[j + 16]);
const float32x2_t a_24 = vld1_f32(&a[j + 24]);
const float32x2_t a_48 = vld1_f32(&a[j + 48]);
const float32x2_t a_56 = vld1_f32(&a[j + 56]);
const float32x4_t a_16_48 = vcombine_f32(a_16, a_48);
const float32x4_t a_24_56 = vcombine_f32(a_24, a_56);
const float32x4_t x2r0_2i0_2r1_x2i1 = vaddq_f32(a_16_48, a_24_56);
const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56);
const float32x4_t xx = vaddq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const float32x4_t x3i0_3r0_3i1_x3r1 = vrev64q_f32(x3r0_3i0_3r1_x3i1);
const float32x4_t x1_x3_add =
vmlaq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1);
const float32x4_t x1_x3_sub =
vmlsq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1);
float32x4_t xx4 = vmulq_f32(wk2rv, xx1);
float32x4_t xx12 = vmulq_f32(wk1rv, x1_x3_add);
float32x4_t xx22 = vmulq_f32(wk3rv, x1_x3_sub);
xx4 = vmlaq_f32(xx4, wk2iv, vrev64q_f32(xx1));
xx12 = vmlaq_f32(xx12, wk1iv, vrev64q_f32(x1_x3_add));
xx22 = vmlaq_f32(xx22, wk3iv, vrev64q_f32(x1_x3_sub));
vst1_f32(&a[j + 0], vget_low_f32(xx));
vst1_f32(&a[j + 32], vget_high_f32(xx));
vst1_f32(&a[j + 16], vget_low_f32(xx4));
vst1_f32(&a[j + 48], vget_high_f32(xx4));
vst1_f32(&a[j + 8], vget_low_f32(xx12));
vst1_f32(&a[j + 40], vget_high_f32(xx12));
vst1_f32(&a[j + 24], vget_low_f32(xx22));
vst1_f32(&a[j + 56], vget_high_f32(xx22));
}
}
}
__inline static float32x4_t reverse_order_f32x4(float32x4_t in) {
// A B C D -> C D A B
const float32x4_t rev = vcombine_f32(vget_high_f32(in), vget_low_f32(in));
// C D A B -> D C B A
return vrev64q_f32(rev);
}
static void rftfsub_128_neon(float* a) {
const float* c = rdft_w + 32;
int j1, j2;
const float32x4_t mm_half = vdupq_n_f32(0.5f);
// Vectorized code (four at once).
// Note: commented number are indexes for the first iteration of the loop.
for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) {
// Load 'wk'.
const float32x4_t c_j1 = vld1q_f32(&c[j1]); // 1, 2, 3, 4,
const float32x4_t c_k1 = vld1q_f32(&c[29 - j1]); // 28, 29, 30, 31,
const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31,
const float32x4_t wkr_ = reverse_order_f32x4(wkrt); // 31, 30, 29, 28,
const float32x4_t wki_ = c_j1; // 1, 2, 3, 4,
// Load and shuffle 'a'.
// 2, 4, 6, 8, 3, 5, 7, 9
float32x4x2_t a_j2_p = vld2q_f32(&a[0 + j2]);
// 120, 122, 124, 126, 121, 123, 125, 127,
const float32x4x2_t k2_0_4 = vld2q_f32(&a[122 - j2]);
// 126, 124, 122, 120
const float32x4_t a_k2_p0 = reverse_order_f32x4(k2_0_4.val[0]);
// 127, 125, 123, 121
const float32x4_t a_k2_p1 = reverse_order_f32x4(k2_0_4.val[1]);
// Calculate 'x'.
const float32x4_t xr_ = vsubq_f32(a_j2_p.val[0], a_k2_p0);
// 2-126, 4-124, 6-122, 8-120,
const float32x4_t xi_ = vaddq_f32(a_j2_p.val[1], a_k2_p1);
// 3-127, 5-125, 7-123, 9-121,
// Calculate product into 'y'.
// yr = wkr * xr - wki * xi;
// yi = wkr * xi + wki * xr;
const float32x4_t a_ = vmulq_f32(wkr_, xr_);
const float32x4_t b_ = vmulq_f32(wki_, xi_);
const float32x4_t c_ = vmulq_f32(wkr_, xi_);
const float32x4_t d_ = vmulq_f32(wki_, xr_);
const float32x4_t yr_ = vsubq_f32(a_, b_); // 2-126, 4-124, 6-122, 8-120,
const float32x4_t yi_ = vaddq_f32(c_, d_); // 3-127, 5-125, 7-123, 9-121,
// Update 'a'.
// a[j2 + 0] -= yr;
// a[j2 + 1] -= yi;
// a[k2 + 0] += yr;
// a[k2 + 1] -= yi;
// 126, 124, 122, 120,
const float32x4_t a_k2_p0n = vaddq_f32(a_k2_p0, yr_);
// 127, 125, 123, 121,
const float32x4_t a_k2_p1n = vsubq_f32(a_k2_p1, yi_);
// Shuffle in right order and store.
const float32x4_t a_k2_p0nr = vrev64q_f32(a_k2_p0n);
const float32x4_t a_k2_p1nr = vrev64q_f32(a_k2_p1n);
// 124, 125, 126, 127, 120, 121, 122, 123
const float32x4x2_t a_k2_n = vzipq_f32(a_k2_p0nr, a_k2_p1nr);
// 2, 4, 6, 8,
a_j2_p.val[0] = vsubq_f32(a_j2_p.val[0], yr_);
// 3, 5, 7, 9,
a_j2_p.val[1] = vsubq_f32(a_j2_p.val[1], yi_);
// 2, 3, 4, 5, 6, 7, 8, 9,
vst2q_f32(&a[0 + j2], a_j2_p);
vst1q_f32(&a[122 - j2], a_k2_n.val[1]);
vst1q_f32(&a[126 - j2], a_k2_n.val[0]);
}
// Scalar code for the remaining items.
for (; j2 < 64; j1 += 1, j2 += 2) {
const int k2 = 128 - j2;
const int k1 = 32 - j1;
const float wkr = 0.5f - c[k1];
const float wki = c[j1];
const float xr = a[j2 + 0] - a[k2 + 0];
const float xi = a[j2 + 1] + a[k2 + 1];
const float yr = wkr * xr - wki * xi;
const float yi = wkr * xi + wki * xr;
a[j2 + 0] -= yr;
a[j2 + 1] -= yi;
a[k2 + 0] += yr;
a[k2 + 1] -= yi;
}
}
static void rftbsub_128_neon(float* a) {
const float* c = rdft_w + 32;
int j1, j2;
const float32x4_t mm_half = vdupq_n_f32(0.5f);
a[1] = -a[1];
// Vectorized code (four at once).
// Note: commented number are indexes for the first iteration of the loop.
for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) {
// Load 'wk'.
const float32x4_t c_j1 = vld1q_f32(&c[j1]); // 1, 2, 3, 4,
const float32x4_t c_k1 = vld1q_f32(&c[29 - j1]); // 28, 29, 30, 31,
const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31,
const float32x4_t wkr_ = reverse_order_f32x4(wkrt); // 31, 30, 29, 28,
const float32x4_t wki_ = c_j1; // 1, 2, 3, 4,
// Load and shuffle 'a'.
// 2, 4, 6, 8, 3, 5, 7, 9
float32x4x2_t a_j2_p = vld2q_f32(&a[0 + j2]);
// 120, 122, 124, 126, 121, 123, 125, 127,
const float32x4x2_t k2_0_4 = vld2q_f32(&a[122 - j2]);
// 126, 124, 122, 120
const float32x4_t a_k2_p0 = reverse_order_f32x4(k2_0_4.val[0]);
// 127, 125, 123, 121
const float32x4_t a_k2_p1 = reverse_order_f32x4(k2_0_4.val[1]);
// Calculate 'x'.
const float32x4_t xr_ = vsubq_f32(a_j2_p.val[0], a_k2_p0);
// 2-126, 4-124, 6-122, 8-120,
const float32x4_t xi_ = vaddq_f32(a_j2_p.val[1], a_k2_p1);
// 3-127, 5-125, 7-123, 9-121,
// Calculate product into 'y'.
// yr = wkr * xr - wki * xi;
// yi = wkr * xi + wki * xr;
const float32x4_t a_ = vmulq_f32(wkr_, xr_);
const float32x4_t b_ = vmulq_f32(wki_, xi_);
const float32x4_t c_ = vmulq_f32(wkr_, xi_);
const float32x4_t d_ = vmulq_f32(wki_, xr_);
const float32x4_t yr_ = vaddq_f32(a_, b_); // 2-126, 4-124, 6-122, 8-120,
const float32x4_t yi_ = vsubq_f32(c_, d_); // 3-127, 5-125, 7-123, 9-121,
// Update 'a'.
// a[j2 + 0] -= yr;
// a[j2 + 1] -= yi;
// a[k2 + 0] += yr;
// a[k2 + 1] -= yi;
// 126, 124, 122, 120,
const float32x4_t a_k2_p0n = vaddq_f32(a_k2_p0, yr_);
// 127, 125, 123, 121,
const float32x4_t a_k2_p1n = vsubq_f32(yi_, a_k2_p1);
// Shuffle in right order and store.
// 2, 3, 4, 5, 6, 7, 8, 9,
const float32x4_t a_k2_p0nr = vrev64q_f32(a_k2_p0n);
const float32x4_t a_k2_p1nr = vrev64q_f32(a_k2_p1n);
// 124, 125, 126, 127, 120, 121, 122, 123
const float32x4x2_t a_k2_n = vzipq_f32(a_k2_p0nr, a_k2_p1nr);
// 2, 4, 6, 8,
a_j2_p.val[0] = vsubq_f32(a_j2_p.val[0], yr_);
// 3, 5, 7, 9,
a_j2_p.val[1] = vsubq_f32(yi_, a_j2_p.val[1]);
// 2, 3, 4, 5, 6, 7, 8, 9,
vst2q_f32(&a[0 + j2], a_j2_p);
vst1q_f32(&a[122 - j2], a_k2_n.val[1]);
vst1q_f32(&a[126 - j2], a_k2_n.val[0]);
}
// Scalar code for the remaining items.
for (; j2 < 64; j1 += 1, j2 += 2) {
const int k2 = 128 - j2;
const int k1 = 32 - j1;
const float wkr = 0.5f - c[k1];
const float wki = c[j1];
const float xr = a[j2 + 0] - a[k2 + 0];
const float xi = a[j2 + 1] + a[k2 + 1];
const float yr = wkr * xr + wki * xi;
const float yi = wkr * xi - wki * xr;
a[j2 + 0] = a[j2 + 0] - yr;
a[j2 + 1] = yi - a[j2 + 1];
a[k2 + 0] = yr + a[k2 + 0];
a[k2 + 1] = yi - a[k2 + 1];
}
a[65] = -a[65];
}
void aec_rdft_init_neon(void) {
cft1st_128 = cft1st_128_neon;
cftmdl_128 = cftmdl_128_neon;
rftfsub_128 = rftfsub_128_neon;
rftbsub_128 = rftbsub_128_neon;
}

427
webrtc/modules/audio_processing/aec/aec_rdft_sse2.c
View File

@ -1,427 +0,0 @@
/*
* Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "webrtc/modules/audio_processing/aec/aec_rdft.h"
#include <emmintrin.h>
static const ALIGN16_BEG float ALIGN16_END
k_swap_sign[4] = {-1.f, 1.f, -1.f, 1.f};
static void cft1st_128_SSE2(float* a) {
const __m128 mm_swap_sign = _mm_load_ps(k_swap_sign);
int j, k2;
for (k2 = 0, j = 0; j < 128; j += 16, k2 += 4) {
__m128 a00v = _mm_loadu_ps(&a[j + 0]);
__m128 a04v = _mm_loadu_ps(&a[j + 4]);
__m128 a08v = _mm_loadu_ps(&a[j + 8]);
__m128 a12v = _mm_loadu_ps(&a[j + 12]);
__m128 a01v = _mm_shuffle_ps(a00v, a08v, _MM_SHUFFLE(1, 0, 1, 0));
__m128 a23v = _mm_shuffle_ps(a00v, a08v, _MM_SHUFFLE(3, 2, 3, 2));
__m128 a45v = _mm_shuffle_ps(a04v, a12v, _MM_SHUFFLE(1, 0, 1, 0));
__m128 a67v = _mm_shuffle_ps(a04v, a12v, _MM_SHUFFLE(3, 2, 3, 2));
const __m128 wk1rv = _mm_load_ps(&rdft_wk1r[k2]);
const __m128 wk1iv = _mm_load_ps(&rdft_wk1i[k2]);
const __m128 wk2rv = _mm_load_ps(&rdft_wk2r[k2]);
const __m128 wk2iv = _mm_load_ps(&rdft_wk2i[k2]);
const __m128 wk3rv = _mm_load_ps(&rdft_wk3r[k2]);
const __m128 wk3iv = _mm_load_ps(&rdft_wk3i[k2]);
__m128 x0v = _mm_add_ps(a01v, a23v);
const __m128 x1v = _mm_sub_ps(a01v, a23v);
const __m128 x2v = _mm_add_ps(a45v, a67v);
const __m128 x3v = _mm_sub_ps(a45v, a67v);
__m128 x0w;
a01v = _mm_add_ps(x0v, x2v);
x0v = _mm_sub_ps(x0v, x2v);
x0w = _mm_shuffle_ps(x0v, x0v, _MM_SHUFFLE(2, 3, 0, 1));
{
const __m128 a45_0v = _mm_mul_ps(wk2rv, x0v);
const __m128 a45_1v = _mm_mul_ps(wk2iv, x0w);
a45v = _mm_add_ps(a45_0v, a45_1v);
}
{
__m128 a23_0v, a23_1v;
const __m128 x3w = _mm_shuffle_ps(x3v, x3v, _MM_SHUFFLE(2, 3, 0, 1));
const __m128 x3s = _mm_mul_ps(mm_swap_sign, x3w);
x0v = _mm_add_ps(x1v, x3s);
x0w = _mm_shuffle_ps(x0v, x0v, _MM_SHUFFLE(2, 3, 0, 1));
a23_0v = _mm_mul_ps(wk1rv, x0v);
a23_1v = _mm_mul_ps(wk1iv, x0w);
a23v = _mm_add_ps(a23_0v, a23_1v);
x0v = _mm_sub_ps(x1v, x3s);
x0w = _mm_shuffle_ps(x0v, x0v, _MM_SHUFFLE(2, 3, 0, 1));
}
{
const __m128 a67_0v = _mm_mul_ps(wk3rv, x0v);
const __m128 a67_1v = _mm_mul_ps(wk3iv, x0w);
a67v = _mm_add_ps(a67_0v, a67_1v);
}
a00v = _mm_shuffle_ps(a01v, a23v, _MM_SHUFFLE(1, 0, 1, 0));
a04v = _mm_shuffle_ps(a45v, a67v, _MM_SHUFFLE(1, 0, 1, 0));
a08v = _mm_shuffle_ps(a01v, a23v, _MM_SHUFFLE(3, 2, 3, 2));
a12v = _mm_shuffle_ps(a45v, a67v, _MM_SHUFFLE(3, 2, 3, 2));
_mm_storeu_ps(&a[j + 0], a00v);
_mm_storeu_ps(&a[j + 4], a04v);
_mm_storeu_ps(&a[j + 8], a08v);
_mm_storeu_ps(&a[j + 12], a12v);
}
}
static void cftmdl_128_SSE2(float* a) {
const int l = 8;
const __m128 mm_swap_sign = _mm_load_ps(k_swap_sign);
int j0;
__m128 wk1rv = _mm_load_ps(cftmdl_wk1r);
for (j0 = 0; j0 < l; j0 += 2) {
const __m128i a_00 = _mm_loadl_epi64((__m128i*)&a[j0 + 0]);
const __m128i a_08 = _mm_loadl_epi64((__m128i*)&a[j0 + 8]);
const __m128i a_32 = _mm_loadl_epi64((__m128i*)&a[j0 + 32]);
const __m128i a_40 = _mm_loadl_epi64((__m128i*)&a[j0 + 40]);
const __m128 a_00_32 = _mm_shuffle_ps(_mm_castsi128_ps(a_00),
_mm_castsi128_ps(a_32),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 a_08_40 = _mm_shuffle_ps(_mm_castsi128_ps(a_08),
_mm_castsi128_ps(a_40),
_MM_SHUFFLE(1, 0, 1, 0));
__m128 x0r0_0i0_0r1_x0i1 = _mm_add_ps(a_00_32, a_08_40);
const __m128 x1r0_1i0_1r1_x1i1 = _mm_sub_ps(a_00_32, a_08_40);
const __m128i a_16 = _mm_loadl_epi64((__m128i*)&a[j0 + 16]);
const __m128i a_24 = _mm_loadl_epi64((__m128i*)&a[j0 + 24]);
const __m128i a_48 = _mm_loadl_epi64((__m128i*)&a[j0 + 48]);
const __m128i a_56 = _mm_loadl_epi64((__m128i*)&a[j0 + 56]);
const __m128 a_16_48 = _mm_shuffle_ps(_mm_castsi128_ps(a_16),
_mm_castsi128_ps(a_48),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 a_24_56 = _mm_shuffle_ps(_mm_castsi128_ps(a_24),
_mm_castsi128_ps(a_56),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 x2r0_2i0_2r1_x2i1 = _mm_add_ps(a_16_48, a_24_56);
const __m128 x3r0_3i0_3r1_x3i1 = _mm_sub_ps(a_16_48, a_24_56);
const __m128 xx0 = _mm_add_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const __m128 xx1 = _mm_sub_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const __m128 x3i0_3r0_3i1_x3r1 = _mm_castsi128_ps(_mm_shuffle_epi32(
_mm_castps_si128(x3r0_3i0_3r1_x3i1), _MM_SHUFFLE(2, 3, 0, 1)));
const __m128 x3_swapped = _mm_mul_ps(mm_swap_sign, x3i0_3r0_3i1_x3r1);
const __m128 x1_x3_add = _mm_add_ps(x1r0_1i0_1r1_x1i1, x3_swapped);
const __m128 x1_x3_sub = _mm_sub_ps(x1r0_1i0_1r1_x1i1, x3_swapped);
const __m128 yy0 =
_mm_shuffle_ps(x1_x3_add, x1_x3_sub, _MM_SHUFFLE(2, 2, 2, 2));
const __m128 yy1 =
_mm_shuffle_ps(x1_x3_add, x1_x3_sub, _MM_SHUFFLE(3, 3, 3, 3));
const __m128 yy2 = _mm_mul_ps(mm_swap_sign, yy1);
const __m128 yy3 = _mm_add_ps(yy0, yy2);
const __m128 yy4 = _mm_mul_ps(wk1rv, yy3);
_mm_storel_epi64((__m128i*)&a[j0 + 0], _mm_castps_si128(xx0));
_mm_storel_epi64(
(__m128i*)&a[j0 + 32],
_mm_shuffle_epi32(_mm_castps_si128(xx0), _MM_SHUFFLE(3, 2, 3, 2)));
_mm_storel_epi64((__m128i*)&a[j0 + 16], _mm_castps_si128(xx1));
_mm_storel_epi64(
(__m128i*)&a[j0 + 48],
_mm_shuffle_epi32(_mm_castps_si128(xx1), _MM_SHUFFLE(2, 3, 2, 3)));
a[j0 + 48] = -a[j0 + 48];
_mm_storel_epi64((__m128i*)&a[j0 + 8], _mm_castps_si128(x1_x3_add));
_mm_storel_epi64((__m128i*)&a[j0 + 24], _mm_castps_si128(x1_x3_sub));
_mm_storel_epi64((__m128i*)&a[j0 + 40], _mm_castps_si128(yy4));
_mm_storel_epi64(
(__m128i*)&a[j0 + 56],
_mm_shuffle_epi32(_mm_castps_si128(yy4), _MM_SHUFFLE(2, 3, 2, 3)));
}
{
int k = 64;
int k1 = 2;
int k2 = 2 * k1;
const __m128 wk2rv = _mm_load_ps(&rdft_wk2r[k2 + 0]);
const __m128 wk2iv = _mm_load_ps(&rdft_wk2i[k2 + 0]);
const __m128 wk1iv = _mm_load_ps(&rdft_wk1i[k2 + 0]);
const __m128 wk3rv = _mm_load_ps(&rdft_wk3r[k2 + 0]);
const __m128 wk3iv = _mm_load_ps(&rdft_wk3i[k2 + 0]);
wk1rv = _mm_load_ps(&rdft_wk1r[k2 + 0]);
for (j0 = k; j0 < l + k; j0 += 2) {
const __m128i a_00 = _mm_loadl_epi64((__m128i*)&a[j0 + 0]);
const __m128i a_08 = _mm_loadl_epi64((__m128i*)&a[j0 + 8]);
const __m128i a_32 = _mm_loadl_epi64((__m128i*)&a[j0 + 32]);
const __m128i a_40 = _mm_loadl_epi64((__m128i*)&a[j0 + 40]);
const __m128 a_00_32 = _mm_shuffle_ps(_mm_castsi128_ps(a_00),
_mm_castsi128_ps(a_32),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 a_08_40 = _mm_shuffle_ps(_mm_castsi128_ps(a_08),
_mm_castsi128_ps(a_40),
_MM_SHUFFLE(1, 0, 1, 0));
__m128 x0r0_0i0_0r1_x0i1 = _mm_add_ps(a_00_32, a_08_40);
const __m128 x1r0_1i0_1r1_x1i1 = _mm_sub_ps(a_00_32, a_08_40);
const __m128i a_16 = _mm_loadl_epi64((__m128i*)&a[j0 + 16]);
const __m128i a_24 = _mm_loadl_epi64((__m128i*)&a[j0 + 24]);
const __m128i a_48 = _mm_loadl_epi64((__m128i*)&a[j0 + 48]);
const __m128i a_56 = _mm_loadl_epi64((__m128i*)&a[j0 + 56]);
const __m128 a_16_48 = _mm_shuffle_ps(_mm_castsi128_ps(a_16),
_mm_castsi128_ps(a_48),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 a_24_56 = _mm_shuffle_ps(_mm_castsi128_ps(a_24),
_mm_castsi128_ps(a_56),
_MM_SHUFFLE(1, 0, 1, 0));
const __m128 x2r0_2i0_2r1_x2i1 = _mm_add_ps(a_16_48, a_24_56);
const __m128 x3r0_3i0_3r1_x3i1 = _mm_sub_ps(a_16_48, a_24_56);
const __m128 xx = _mm_add_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const __m128 xx1 = _mm_sub_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const __m128 xx2 = _mm_mul_ps(xx1, wk2rv);
const __m128 xx3 =
_mm_mul_ps(wk2iv,
_mm_castsi128_ps(_mm_shuffle_epi32(
_mm_castps_si128(xx1), _MM_SHUFFLE(2, 3, 0, 1))));
const __m128 xx4 = _mm_add_ps(xx2, xx3);
const __m128 x3i0_3r0_3i1_x3r1 = _mm_castsi128_ps(_mm_shuffle_epi32(
_mm_castps_si128(x3r0_3i0_3r1_x3i1), _MM_SHUFFLE(2, 3, 0, 1)));
const __m128 x3_swapped = _mm_mul_ps(mm_swap_sign, x3i0_3r0_3i1_x3r1);
const __m128 x1_x3_add = _mm_add_ps(x1r0_1i0_1r1_x1i1, x3_swapped);
const __m128 x1_x3_sub = _mm_sub_ps(x1r0_1i0_1r1_x1i1, x3_swapped);
const __m128 xx10 = _mm_mul_ps(x1_x3_add, wk1rv);
const __m128 xx11 = _mm_mul_ps(
wk1iv,
_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(x1_x3_add),
_MM_SHUFFLE(2, 3, 0, 1))));
const __m128 xx12 = _mm_add_ps(xx10, xx11);
const __m128 xx20 = _mm_mul_ps(x1_x3_sub, wk3rv);
const __m128 xx21 = _mm_mul_ps(
wk3iv,
_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(x1_x3_sub),
_MM_SHUFFLE(2, 3, 0, 1))));
const __m128 xx22 = _mm_add_ps(xx20, xx21);
_mm_storel_epi64((__m128i*)&a[j0 + 0], _mm_castps_si128(xx));
_mm_storel_epi64(
(__m128i*)&a[j0 + 32],
_mm_shuffle_epi32(_mm_castps_si128(xx), _MM_SHUFFLE(3, 2, 3, 2)));
_mm_storel_epi64((__m128i*)&a[j0 + 16], _mm_castps_si128(xx4));
_mm_storel_epi64(
(__m128i*)&a[j0 + 48],
_mm_shuffle_epi32(_mm_castps_si128(xx4), _MM_SHUFFLE(3, 2, 3, 2)));
_mm_storel_epi64((__m128i*)&a[j0 + 8], _mm_castps_si128(xx12));
_mm_storel_epi64(
(__m128i*)&a[j0 + 40],
_mm_shuffle_epi32(_mm_castps_si128(xx12), _MM_SHUFFLE(3, 2, 3, 2)));
_mm_storel_epi64((__m128i*)&a[j0 + 24], _mm_castps_si128(xx22));
_mm_storel_epi64(
(__m128i*)&a[j0 + 56],
_mm_shuffle_epi32(_mm_castps_si128(xx22), _MM_SHUFFLE(3, 2, 3, 2)));
}
}
}
static void rftfsub_128_SSE2(float* a) {
const float* c = rdft_w + 32;
int j1, j2, k1, k2;
float wkr, wki, xr, xi, yr, yi;
static const ALIGN16_BEG float ALIGN16_END
k_half[4] = {0.5f, 0.5f, 0.5f, 0.5f};
const __m128 mm_half = _mm_load_ps(k_half);
// Vectorized code (four at once).
// Note: commented number are indexes for the first iteration of the loop.
for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) {
// Load 'wk'.
const __m128 c_j1 = _mm_loadu_ps(&c[j1]); // 1, 2, 3, 4,
const __m128 c_k1 = _mm_loadu_ps(&c[29 - j1]); // 28, 29, 30, 31,
const __m128 wkrt = _mm_sub_ps(mm_half, c_k1); // 28, 29, 30, 31,
const __m128 wkr_ =
_mm_shuffle_ps(wkrt, wkrt, _MM_SHUFFLE(0, 1, 2, 3)); // 31, 30, 29, 28,
const __m128 wki_ = c_j1; // 1, 2, 3, 4,
// Load and shuffle 'a'.
const __m128 a_j2_0 = _mm_loadu_ps(&a[0 + j2]); // 2, 3, 4, 5,
const __m128 a_j2_4 = _mm_loadu_ps(&a[4 + j2]); // 6, 7, 8, 9,
const __m128 a_k2_0 = _mm_loadu_ps(&a[122 - j2]); // 120, 121, 122, 123,
const __m128 a_k2_4 = _mm_loadu_ps(&a[126 - j2]); // 124, 125, 126, 127,
const __m128 a_j2_p0 = _mm_shuffle_ps(
a_j2_0, a_j2_4, _MM_SHUFFLE(2, 0, 2, 0)); // 2, 4, 6, 8,
const __m128 a_j2_p1 = _mm_shuffle_ps(
a_j2_0, a_j2_4, _MM_SHUFFLE(3, 1, 3, 1)); // 3, 5, 7, 9,
const __m128 a_k2_p0 = _mm_shuffle_ps(
a_k2_4, a_k2_0, _MM_SHUFFLE(0, 2, 0, 2)); // 126, 124, 122, 120,
const __m128 a_k2_p1 = _mm_shuffle_ps(
a_k2_4, a_k2_0, _MM_SHUFFLE(1, 3, 1, 3)); // 127, 125, 123, 121,
// Calculate 'x'.
const __m128 xr_ = _mm_sub_ps(a_j2_p0, a_k2_p0);
// 2-126, 4-124, 6-122, 8-120,
const __m128 xi_ = _mm_add_ps(a_j2_p1, a_k2_p1);
// 3-127, 5-125, 7-123, 9-121,
// Calculate product into 'y'.
// yr = wkr * xr - wki * xi;
// yi = wkr * xi + wki * xr;
const __m128 a_ = _mm_mul_ps(wkr_, xr_);
const __m128 b_ = _mm_mul_ps(wki_, xi_);
const __m128 c_ = _mm_mul_ps(wkr_, xi_);
const __m128 d_ = _mm_mul_ps(wki_, xr_);
const __m128 yr_ = _mm_sub_ps(a_, b_); // 2-126, 4-124, 6-122, 8-120,
const __m128 yi_ = _mm_add_ps(c_, d_); // 3-127, 5-125, 7-123, 9-121,
// Update 'a'.
// a[j2 + 0] -= yr;
// a[j2 + 1] -= yi;
// a[k2 + 0] += yr;
// a[k2 + 1] -= yi;
const __m128 a_j2_p0n = _mm_sub_ps(a_j2_p0, yr_); // 2, 4, 6, 8,
const __m128 a_j2_p1n = _mm_sub_ps(a_j2_p1, yi_); // 3, 5, 7, 9,
const __m128 a_k2_p0n = _mm_add_ps(a_k2_p0, yr_); // 126, 124, 122, 120,
const __m128 a_k2_p1n = _mm_sub_ps(a_k2_p1, yi_); // 127, 125, 123, 121,
// Shuffle in right order and store.
const __m128 a_j2_0n = _mm_unpacklo_ps(a_j2_p0n, a_j2_p1n);
// 2, 3, 4, 5,
const __m128 a_j2_4n = _mm_unpackhi_ps(a_j2_p0n, a_j2_p1n);
// 6, 7, 8, 9,
const __m128 a_k2_0nt = _mm_unpackhi_ps(a_k2_p0n, a_k2_p1n);
// 122, 123, 120, 121,
const __m128 a_k2_4nt = _mm_unpacklo_ps(a_k2_p0n, a_k2_p1n);
// 126, 127, 124, 125,
const __m128 a_k2_0n = _mm_shuffle_ps(
a_k2_0nt, a_k2_0nt, _MM_SHUFFLE(1, 0, 3, 2)); // 120, 121, 122, 123,
const __m128 a_k2_4n = _mm_shuffle_ps(
a_k2_4nt, a_k2_4nt, _MM_SHUFFLE(1, 0, 3, 2)); // 124, 125, 126, 127,
_mm_storeu_ps(&a[0 + j2], a_j2_0n);
_mm_storeu_ps(&a[4 + j2], a_j2_4n);
_mm_storeu_ps(&a[122 - j2], a_k2_0n);
_mm_storeu_ps(&a[126 - j2], a_k2_4n);
}
// Scalar code for the remaining items.
for (; j2 < 64; j1 += 1, j2 += 2) {
k2 = 128 - j2;
k1 = 32 - j1;
wkr = 0.5f - c[k1];
wki = c[j1];
xr = a[j2 + 0] - a[k2 + 0];
xi = a[j2 + 1] + a[k2 + 1];
yr = wkr * xr - wki * xi;
yi = wkr * xi + wki * xr;
a[j2 + 0] -= yr;
a[j2 + 1] -= yi;
a[k2 + 0] += yr;
a[k2 + 1] -= yi;
}
}
static void rftbsub_128_SSE2(float* a) {
const float* c = rdft_w + 32;
int j1, j2, k1, k2;
float wkr, wki, xr, xi, yr, yi;
static const ALIGN16_BEG float ALIGN16_END
k_half[4] = {0.5f, 0.5f, 0.5f, 0.5f};
const __m128 mm_half = _mm_load_ps(k_half);
a[1] = -a[1];
// Vectorized code (four at once).
// Note: commented number are indexes for the first iteration of the loop.
for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) {
// Load 'wk'.
const __m128 c_j1 = _mm_loadu_ps(&c[j1]); // 1, 2, 3, 4,
const __m128 c_k1 = _mm_loadu_ps(&c[29 - j1]); // 28, 29, 30, 31,
const __m128 wkrt = _mm_sub_ps(mm_half, c_k1); // 28, 29, 30, 31,
const __m128 wkr_ =
_mm_shuffle_ps(wkrt, wkrt, _MM_SHUFFLE(0, 1, 2, 3)); // 31, 30, 29, 28,
const __m128 wki_ = c_j1; // 1, 2, 3, 4,
// Load and shuffle 'a'.
const __m128 a_j2_0 = _mm_loadu_ps(&a[0 + j2]); // 2, 3, 4, 5,
const __m128 a_j2_4 = _mm_loadu_ps(&a[4 + j2]); // 6, 7, 8, 9,
const __m128 a_k2_0 = _mm_loadu_ps(&a[122 - j2]); // 120, 121, 122, 123,
const __m128 a_k2_4 = _mm_loadu_ps(&a[126 - j2]); // 124, 125, 126, 127,
const __m128 a_j2_p0 = _mm_shuffle_ps(
a_j2_0, a_j2_4, _MM_SHUFFLE(2, 0, 2, 0)); // 2, 4, 6, 8,
const __m128 a_j2_p1 = _mm_shuffle_ps(
a_j2_0, a_j2_4, _MM_SHUFFLE(3, 1, 3, 1)); // 3, 5, 7, 9,
const __m128 a_k2_p0 = _mm_shuffle_ps(
a_k2_4, a_k2_0, _MM_SHUFFLE(0, 2, 0, 2)); // 126, 124, 122, 120,
const __m128 a_k2_p1 = _mm_shuffle_ps(
a_k2_4, a_k2_0, _MM_SHUFFLE(1, 3, 1, 3)); // 127, 125, 123, 121,
// Calculate 'x'.
const __m128 xr_ = _mm_sub_ps(a_j2_p0, a_k2_p0);
// 2-126, 4-124, 6-122, 8-120,
const __m128 xi_ = _mm_add_ps(a_j2_p1, a_k2_p1);
// 3-127, 5-125, 7-123, 9-121,
// Calculate product into 'y'.
// yr = wkr * xr + wki * xi;
// yi = wkr * xi - wki * xr;
const __m128 a_ = _mm_mul_ps(wkr_, xr_);
const __m128 b_ = _mm_mul_ps(wki_, xi_);
const __m128 c_ = _mm_mul_ps(wkr_, xi_);
const __m128 d_ = _mm_mul_ps(wki_, xr_);
const __m128 yr_ = _mm_add_ps(a_, b_); // 2-126, 4-124, 6-122, 8-120,
const __m128 yi_ = _mm_sub_ps(c_, d_); // 3-127, 5-125, 7-123, 9-121,
// Update 'a'.
// a[j2 + 0] = a[j2 + 0] - yr;
// a[j2 + 1] = yi - a[j2 + 1];
// a[k2 + 0] = yr + a[k2 + 0];
// a[k2 + 1] = yi - a[k2 + 1];
const __m128 a_j2_p0n = _mm_sub_ps(a_j2_p0, yr_); // 2, 4, 6, 8,
const __m128 a_j2_p1n = _mm_sub_ps(yi_, a_j2_p1); // 3, 5, 7, 9,
const __m128 a_k2_p0n = _mm_add_ps(a_k2_p0, yr_); // 126, 124, 122, 120,
const __m128 a_k2_p1n = _mm_sub_ps(yi_, a_k2_p1); // 127, 125, 123, 121,
// Shuffle in right order and store.
const __m128 a_j2_0n = _mm_unpacklo_ps(a_j2_p0n, a_j2_p1n);
// 2, 3, 4, 5,
const __m128 a_j2_4n = _mm_unpackhi_ps(a_j2_p0n, a_j2_p1n);
// 6, 7, 8, 9,
const __m128 a_k2_0nt = _mm_unpackhi_ps(a_k2_p0n, a_k2_p1n);
// 122, 123, 120, 121,
const __m128 a_k2_4nt = _mm_unpacklo_ps(a_k2_p0n, a_k2_p1n);
// 126, 127, 124, 125,
const __m128 a_k2_0n = _mm_shuffle_ps(
a_k2_0nt, a_k2_0nt, _MM_SHUFFLE(1, 0, 3, 2)); // 120, 121, 122, 123,
const __m128 a_k2_4n = _mm_shuffle_ps(
a_k2_4nt, a_k2_4nt, _MM_SHUFFLE(1, 0, 3, 2)); // 124, 125, 126, 127,
_mm_storeu_ps(&a[0 + j2], a_j2_0n);
_mm_storeu_ps(&a[4 + j2], a_j2_4n);
_mm_storeu_ps(&a[122 - j2], a_k2_0n);
_mm_storeu_ps(&a[126 - j2], a_k2_4n);
}
// Scalar code for the remaining items.
for (; j2 < 64; j1 += 1, j2 += 2) {
k2 = 128 - j2;
k1 = 32 - j1;
wkr = 0.5f - c[k1];
wki = c[j1];
xr = a[j2 + 0] - a[k2 + 0];
xi = a[j2 + 1] + a[k2 + 1];
yr = wkr * xr + wki * xi;
yi = wkr * xi - wki * xr;
a[j2 + 0] = a[j2 + 0] - yr;
a[j2 + 1] = yi - a[j2 + 1];
a[k2 + 0] = yr + a[k2 + 0];
a[k2 + 1] = yi - a[k2 + 1];
}
a[65] = -a[65];
}
void aec_rdft_init_sse2(void) {
cft1st_128 = cft1st_128_SSE2;
cftmdl_128 = cftmdl_128_SSE2;
rftfsub_128 = rftfsub_128_SSE2;
rftbsub_128 = rftbsub_128_SSE2;
}

209
webrtc/modules/audio_processing/aec/aec_resampler.c
View File

@ -1,209 +0,0 @@
/*
* Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
/* Resamples a signal to an arbitrary rate. Used by the AEC to compensate for
* clock skew by resampling the farend signal.
*/
#include "webrtc/modules/audio_processing/aec/aec_resampler.h"
#include <assert.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>
#include "webrtc/modules/audio_processing/aec/aec_core.h"
enum {
kEstimateLengthFrames = 400
};
typedef struct {
float buffer[kResamplerBufferSize];
float position;
int deviceSampleRateHz;
int skewData[kEstimateLengthFrames];
int skewDataIndex;
float skewEstimate;
} AecResampler;
static int EstimateSkew(const int* rawSkew,
int size,
int absLimit,
float* skewEst);
void* WebRtcAec_CreateResampler() {
return malloc(sizeof(AecResampler));
}
int WebRtcAec_InitResampler(void* resampInst, int deviceSampleRateHz) {
AecResampler* obj = (AecResampler*)resampInst;
memset(obj->buffer, 0, sizeof(obj->buffer));
obj->position = 0.0;
obj->deviceSampleRateHz = deviceSampleRateHz;
memset(obj->skewData, 0, sizeof(obj->skewData));
obj->skewDataIndex = 0;
obj->skewEstimate = 0.0;
return 0;
}
void WebRtcAec_FreeResampler(void* resampInst) {
AecResampler* obj = (AecResampler*)resampInst;
free(obj);
}
void WebRtcAec_ResampleLinear(void* resampInst,
const float* inspeech,
size_t size,
float skew,
float* outspeech,
size_t* size_out) {
AecResampler* obj = (AecResampler*)resampInst;
float* y;
float be, tnew;
size_t tn, mm;
assert(size <= 2 * FRAME_LEN);
assert(resampInst != NULL);
assert(inspeech != NULL);
assert(outspeech != NULL);
assert(size_out != NULL);
// Add new frame data in lookahead
memcpy(&obj->buffer[FRAME_LEN + kResamplingDelay],
inspeech,
size * sizeof(inspeech[0]));
// Sample rate ratio
be = 1 + skew;
// Loop over input frame
mm = 0;
y = &obj->buffer[FRAME_LEN]; // Point at current frame
tnew = be * mm + obj->position;
tn = (size_t)tnew;
while (tn < size) {
// Interpolation
outspeech[mm] = y[tn] + (tnew - tn) * (y[tn + 1] - y[tn]);
mm++;
tnew = be * mm + obj->position;
tn = (int)tnew;
}
*size_out = mm;
obj->position += (*size_out) * be - size;
// Shift buffer
memmove(obj->buffer,
&obj->buffer[size],
(kResamplerBufferSize - size) * sizeof(obj->buffer[0]));
}
int WebRtcAec_GetSkew(void* resampInst, int rawSkew, float* skewEst) {
AecResampler* obj = (AecResampler*)resampInst;
int err = 0;
if (obj->skewDataIndex < kEstimateLengthFrames) {
obj->skewData[obj->skewDataIndex] = rawSkew;
obj->skewDataIndex++;
} else if (obj->skewDataIndex == kEstimateLengthFrames) {
err = EstimateSkew(
obj->skewData, kEstimateLengthFrames, obj->deviceSampleRateHz, skewEst);
obj->skewEstimate = *skewEst;
obj->skewDataIndex++;
} else {
*skewEst = obj->skewEstimate;
}
return err;
}
int EstimateSkew(const int* rawSkew,
int size,
int deviceSampleRateHz,
float* skewEst) {
const int absLimitOuter = (int)(0.04f * deviceSampleRateHz);
const int absLimitInner = (int)(0.0025f * deviceSampleRateHz);
int i = 0;
int n = 0;
float rawAvg = 0;
float err = 0;
float rawAbsDev = 0;
int upperLimit = 0;
int lowerLimit = 0;
float cumSum = 0;
float x = 0;
float x2 = 0;
float y = 0;
float xy = 0;
float xAvg = 0;
float denom = 0;
float skew = 0;
*skewEst = 0; // Set in case of error below.
for (i = 0; i < size; i++) {
if ((rawSkew[i] < absLimitOuter && rawSkew[i] > -absLimitOuter)) {
n++;
rawAvg += rawSkew[i];
}
}
if (n == 0) {
return -1;
}
assert(n > 0);
rawAvg /= n;
for (i = 0; i < size; i++) {
if ((rawSkew[i] < absLimitOuter && rawSkew[i] > -absLimitOuter)) {
err = rawSkew[i] - rawAvg;
rawAbsDev += err >= 0 ? err : -err;
}
}
assert(n > 0);
rawAbsDev /= n;
upperLimit = (int)(rawAvg + 5 * rawAbsDev + 1); // +1 for ceiling.
lowerLimit = (int)(rawAvg - 5 * rawAbsDev - 1); // -1 for floor.
n = 0;
for (i = 0; i < size; i++) {
if ((rawSkew[i] < absLimitInner && rawSkew[i] > -absLimitInner) ||
(rawSkew[i] < upperLimit && rawSkew[i] > lowerLimit)) {
n++;
cumSum += rawSkew[i];
x += n;
x2 += n * n;
y += cumSum;
xy += n * cumSum;
}
}
if (n == 0) {
return -1;
}
assert(n > 0);
xAvg = x / n;
denom = x2 - xAvg * x;
if (denom != 0) {
skew = (xy - xAvg * y) / denom;
}
*skewEst = skew;
return 0;
}

39
webrtc/modules/audio_processing/aec/aec_resampler.h
View File

@ -1,39 +0,0 @@
/*
* Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_RESAMPLER_H_
#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_RESAMPLER_H_
#include "webrtc/modules/audio_processing/aec/aec_core.h"
enum {
kResamplingDelay = 1
};
enum {
kResamplerBufferSize = FRAME_LEN * 4
};
// Unless otherwise specified, functions return 0 on success and -1 on error.
void* WebRtcAec_CreateResampler(); // Returns NULL on error.
int WebRtcAec_InitResampler(void* resampInst, int deviceSampleRateHz);
void WebRtcAec_FreeResampler(void* resampInst);
// Estimates skew from raw measurement.
int WebRtcAec_GetSkew(void* resampInst, int rawSkew, float* skewEst);
// Resamples input using linear interpolation.
void WebRtcAec_ResampleLinear(void* resampInst,
const float* inspeech,
size_t size,
float skew,
float* outspeech,
size_t* size_out);
#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_AEC_RESAMPLER_H_

923
webrtc/modules/audio_processing/aec/echo_cancellation.c
View File

@ -1,923 +0,0 @@
/*
* Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
/*
* Contains the API functions for the AEC.
*/
#include "webrtc/modules/audio_processing/aec/include/echo_cancellation.h"
#include <math.h>
#ifdef WEBRTC_AEC_DEBUG_DUMP
#include <stdio.h>
#endif
#include <stdlib.h>
#include <string.h>
#include "webrtc/common_audio/ring_buffer.h"
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include "webrtc/modules/audio_processing/aec/aec_core.h"
#include "webrtc/modules/audio_processing/aec/aec_resampler.h"
#include "webrtc/modules/audio_processing/aec/echo_cancellation_internal.h"
#include "webrtc/typedefs.h"
// Measured delays [ms]
// Device Chrome GTP
// MacBook Air 10
// MacBook Retina 10 100
// MacPro 30?
//
// Win7 Desktop 70 80?
// Win7 T430s 110
// Win8 T420s 70
//
// Daisy 50
// Pixel (w/ preproc?) 240
// Pixel (w/o preproc?) 110 110
// The extended filter mode gives us the flexibility to ignore the system's
// reported delays. We do this for platforms which we believe provide results
// which are incompatible with the AEC's expectations. Based on measurements
// (some provided above) we set a conservative (i.e. lower than measured)
// fixed delay.
//
// WEBRTC_UNTRUSTED_DELAY will only have an impact when |extended_filter_mode|
// is enabled. See the note along with |DelayCorrection| in
// echo_cancellation_impl.h for more details on the mode.
//
// Justification:
// Chromium/Mac: Here, the true latency is so low (~10-20 ms), that it plays
// havoc with the AEC's buffering. To avoid this, we set a fixed delay of 20 ms
// and then compensate by rewinding by 10 ms (in wideband) through
// kDelayDiffOffsetSamples. This trick does not seem to work for larger rewind
// values, but fortunately this is sufficient.
//
// Chromium/Linux(ChromeOS): The values we get on this platform don't correspond
// well to reality. The variance doesn't match the AEC's buffer changes, and the
// bulk values tend to be too low. However, the range across different hardware
// appears to be too large to choose a single value.
//
// GTP/Linux(ChromeOS): TBD, but for the moment we will trust the values.
#if defined(WEBRTC_CHROMIUM_BUILD) && defined(WEBRTC_MAC)
#define WEBRTC_UNTRUSTED_DELAY
#endif
#if defined(WEBRTC_UNTRUSTED_DELAY) && defined(WEBRTC_MAC)
static const int kDelayDiffOffsetSamples = -160;
#else
// Not enabled for now.
static const int kDelayDiffOffsetSamples = 0;
#endif
#if defined(WEBRTC_MAC)
static const int kFixedDelayMs = 20;
#else
static const int kFixedDelayMs = 50;
#endif
#if !defined(WEBRTC_UNTRUSTED_DELAY)
static const int kMinTrustedDelayMs = 20;
#endif
static const int kMaxTrustedDelayMs = 500;
// Maximum length of resampled signal. Must be an integer multiple of frames
// (ceil(1/(1 + MIN_SKEW)*2) + 1)*FRAME_LEN
// The factor of 2 handles wb, and the + 1 is as a safety margin
// TODO(bjornv): Replace with kResamplerBufferSize
#define MAX_RESAMP_LEN (5 * FRAME_LEN)
static const int kMaxBufSizeStart = 62; // In partitions
static const int sampMsNb = 8; // samples per ms in nb
static const int initCheck = 42;
#ifdef WEBRTC_AEC_DEBUG_DUMP
int webrtc_aec_instance_count = 0;
#endif
// Estimates delay to set the position of the far-end buffer read pointer
// (controlled by knownDelay)
static void EstBufDelayNormal(Aec* aecInst);
static void EstBufDelayExtended(Aec* aecInst);
static int ProcessNormal(Aec* self,
const float* const* near,
size_t num_bands,
float* const* out,
size_t num_samples,
int16_t reported_delay_ms,
int32_t skew);
static void ProcessExtended(Aec* self,
const float* const* near,
size_t num_bands,
float* const* out,
size_t num_samples,
int16_t reported_delay_ms,
int32_t skew);
void* WebRtcAec_Create() {
Aec* aecpc = malloc(sizeof(Aec));
if (!aecpc) {
return NULL;
}
aecpc->aec = WebRtcAec_CreateAec();
if (!aecpc->aec) {
WebRtcAec_Free(aecpc);
return NULL;
}
aecpc->resampler = WebRtcAec_CreateResampler();
if (!aecpc->resampler) {
WebRtcAec_Free(aecpc);
return NULL;
}
// Create far-end pre-buffer. The buffer size has to be large enough for
// largest possible drift compensation (kResamplerBufferSize) + "almost" an
// FFT buffer (PART_LEN2 - 1).
aecpc->far_pre_buf =
WebRtc_CreateBuffer(PART_LEN2 + kResamplerBufferSize, sizeof(float));
if (!aecpc->far_pre_buf) {
WebRtcAec_Free(aecpc);
return NULL;
}
aecpc->initFlag = 0;
aecpc->lastError = 0;
#ifdef WEBRTC_AEC_DEBUG_DUMP
{
char filename[64];
sprintf(filename, "aec_buf%d.dat", webrtc_aec_instance_count);
aecpc->bufFile = fopen(filename, "wb");
sprintf(filename, "aec_skew%d.dat", webrtc_aec_instance_count);
aecpc->skewFile = fopen(filename, "wb");
sprintf(filename, "aec_delay%d.dat", webrtc_aec_instance_count);
aecpc->delayFile = fopen(filename, "wb");
webrtc_aec_instance_count++;
}
#endif
return aecpc;
}
void WebRtcAec_Free(void* aecInst) {
Aec* aecpc = aecInst;
if (aecpc == NULL) {
return;
}
WebRtc_FreeBuffer(aecpc->far_pre_buf);
#ifdef WEBRTC_AEC_DEBUG_DUMP
fclose(aecpc->bufFile);
fclose(aecpc->skewFile);
fclose(aecpc->delayFile);
#endif
WebRtcAec_FreeAec(aecpc->aec);
WebRtcAec_FreeResampler(aecpc->resampler);
free(aecpc);
}
int32_t WebRtcAec_Init(void* aecInst, int32_t sampFreq, int32_t scSampFreq) {
Aec* aecpc = aecInst;
AecConfig aecConfig;
if (sampFreq != 8000 &&
sampFreq != 16000 &&
sampFreq != 32000 &&
sampFreq != 48000) {
aecpc->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
aecpc->sampFreq = sampFreq;
if (scSampFreq < 1 || scSampFreq > 96000) {
aecpc->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
aecpc->scSampFreq = scSampFreq;
// Initialize echo canceller core
if (WebRtcAec_InitAec(aecpc->aec, aecpc->sampFreq) == -1) {
aecpc->lastError = AEC_UNSPECIFIED_ERROR;
return -1;
}
if (WebRtcAec_InitResampler(aecpc->resampler, aecpc->scSampFreq) == -1) {
aecpc->lastError = AEC_UNSPECIFIED_ERROR;
return -1;
}
WebRtc_InitBuffer(aecpc->far_pre_buf);
WebRtc_MoveReadPtr(aecpc->far_pre_buf, -PART_LEN); // Start overlap.
aecpc->initFlag = initCheck; // indicates that initialization has been done
if (aecpc->sampFreq == 32000 || aecpc->sampFreq == 48000) {
aecpc->splitSampFreq = 16000;
} else {
aecpc->splitSampFreq = sampFreq;
}
aecpc->delayCtr = 0;
aecpc->sampFactor = (aecpc->scSampFreq * 1.0f) / aecpc->splitSampFreq;
// Sampling frequency multiplier (SWB is processed as 160 frame size).
aecpc->rate_factor = aecpc->splitSampFreq / 8000;
aecpc->sum = 0;
aecpc->counter = 0;
aecpc->checkBuffSize = 1;
aecpc->firstVal = 0;
// We skip the startup_phase completely (setting to 0) if DA-AEC is enabled,
// but not extended_filter mode.
aecpc->startup_phase = WebRtcAec_extended_filter_enabled(aecpc->aec) ||
!WebRtcAec_delay_agnostic_enabled(aecpc->aec);
aecpc->bufSizeStart = 0;
aecpc->checkBufSizeCtr = 0;
aecpc->msInSndCardBuf = 0;
aecpc->filtDelay = -1; // -1 indicates an initialized state.
aecpc->timeForDelayChange = 0;
aecpc->knownDelay = 0;
aecpc->lastDelayDiff = 0;
aecpc->skewFrCtr = 0;
aecpc->resample = kAecFalse;
aecpc->highSkewCtr = 0;
aecpc->skew = 0;
aecpc->farend_started = 0;
// Default settings.
aecConfig.nlpMode = kAecNlpModerate;
aecConfig.skewMode = kAecFalse;
aecConfig.metricsMode = kAecFalse;
aecConfig.delay_logging = kAecFalse;
if (WebRtcAec_set_config(aecpc, aecConfig) == -1) {
aecpc->lastError = AEC_UNSPECIFIED_ERROR;
return -1;
}
return 0;
}
// only buffer L band for farend
int32_t WebRtcAec_BufferFarend(void* aecInst,
const float* farend,
size_t nrOfSamples) {
Aec* aecpc = aecInst;
size_t newNrOfSamples = nrOfSamples;
float new_farend[MAX_RESAMP_LEN];
const float* farend_ptr = farend;
if (farend == NULL) {
aecpc->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (aecpc->initFlag != initCheck) {
aecpc->lastError = AEC_UNINITIALIZED_ERROR;
return -1;
}
// number of samples == 160 for SWB input
if (nrOfSamples != 80 && nrOfSamples != 160) {
aecpc->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
if (aecpc->skewMode == kAecTrue && aecpc->resample == kAecTrue) {
// Resample and get a new number of samples
WebRtcAec_ResampleLinear(aecpc->resampler,
farend,
nrOfSamples,
aecpc->skew,
new_farend,
&newNrOfSamples);
farend_ptr = new_farend;
}
aecpc->farend_started = 1;
WebRtcAec_SetSystemDelay(
aecpc->aec, WebRtcAec_system_delay(aecpc->aec) + (int)newNrOfSamples);
// Write the time-domain data to |far_pre_buf|.
WebRtc_WriteBuffer(aecpc->far_pre_buf, farend_ptr, newNrOfSamples);
// Transform to frequency domain if we have enough data.
while (WebRtc_available_read(aecpc->far_pre_buf) >= PART_LEN2) {
// We have enough data to pass to the FFT, hence read PART_LEN2 samples.
{
float* ptmp = NULL;
float tmp[PART_LEN2];
WebRtc_ReadBuffer(aecpc->far_pre_buf, (void**)&ptmp, tmp, PART_LEN2);
WebRtcAec_BufferFarendPartition(aecpc->aec, ptmp);
#ifdef WEBRTC_AEC_DEBUG_DUMP
WebRtc_WriteBuffer(
WebRtcAec_far_time_buf(aecpc->aec), &ptmp[PART_LEN], 1);
#endif
}
// Rewind |far_pre_buf| PART_LEN samples for overlap before continuing.
WebRtc_MoveReadPtr(aecpc->far_pre_buf, -PART_LEN);
}
return 0;
}
int32_t WebRtcAec_Process(void* aecInst,
const float* const* nearend,
size_t num_bands,
float* const* out,
size_t nrOfSamples,
int16_t msInSndCardBuf,
int32_t skew) {
Aec* aecpc = aecInst;
int32_t retVal = 0;
if (out == NULL) {
aecpc->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (aecpc->initFlag != initCheck) {
aecpc->lastError = AEC_UNINITIALIZED_ERROR;
return -1;
}
// number of samples == 160 for SWB input
if (nrOfSamples != 80 && nrOfSamples != 160) {
aecpc->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
if (msInSndCardBuf < 0) {
msInSndCardBuf = 0;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
retVal = -1;
} else if (msInSndCardBuf > kMaxTrustedDelayMs) {
// The clamping is now done in ProcessExtended/Normal().
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
retVal = -1;
}
// This returns the value of aec->extended_filter_enabled.
if (WebRtcAec_extended_filter_enabled(aecpc->aec)) {
ProcessExtended(aecpc,
nearend,
num_bands,
out,
nrOfSamples,
msInSndCardBuf,
skew);
} else {
if (ProcessNormal(aecpc,
nearend,
num_bands,
out,
nrOfSamples,
msInSndCardBuf,
skew) != 0) {
retVal = -1;
}
}
#ifdef WEBRTC_AEC_DEBUG_DUMP
{
int16_t far_buf_size_ms = (int16_t)(WebRtcAec_system_delay(aecpc->aec) /
(sampMsNb * aecpc->rate_factor));
(void)fwrite(&far_buf_size_ms, 2, 1, aecpc->bufFile);
(void)fwrite(
&aecpc->knownDelay, sizeof(aecpc->knownDelay), 1, aecpc->delayFile);
}
#endif
return retVal;
}
int WebRtcAec_set_config(void* handle, AecConfig config) {
Aec* self = (Aec*)handle;
if (self->initFlag != initCheck) {
self->lastError = AEC_UNINITIALIZED_ERROR;
return -1;
}
if (config.skewMode != kAecFalse && config.skewMode != kAecTrue) {
self->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
self->skewMode = config.skewMode;
if (config.nlpMode != kAecNlpConservative &&
config.nlpMode != kAecNlpModerate &&
config.nlpMode != kAecNlpAggressive) {
self->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
if (config.metricsMode != kAecFalse && config.metricsMode != kAecTrue) {
self->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
if (config.delay_logging != kAecFalse && config.delay_logging != kAecTrue) {
self->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
WebRtcAec_SetConfigCore(
self->aec, config.nlpMode, config.metricsMode, config.delay_logging);
return 0;
}
int WebRtcAec_get_echo_status(void* handle, int* status) {
Aec* self = (Aec*)handle;
if (status == NULL) {
self->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (self->initFlag != initCheck) {
self->lastError = AEC_UNINITIALIZED_ERROR;
return -1;
}
*status = WebRtcAec_echo_state(self->aec);
return 0;
}
int WebRtcAec_GetMetrics(void* handle, AecMetrics* metrics) {
const float kUpWeight = 0.7f;
float dtmp;
int stmp;
Aec* self = (Aec*)handle;
Stats erl;
Stats erle;
Stats a_nlp;
if (handle == NULL) {
return -1;
}
if (metrics == NULL) {
self->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (self->initFlag != initCheck) {
self->lastError = AEC_UNINITIALIZED_ERROR;
return -1;
}
WebRtcAec_GetEchoStats(self->aec, &erl, &erle, &a_nlp);
// ERL
metrics->erl.instant = (int)erl.instant;
if ((erl.himean > kOffsetLevel) && (erl.average > kOffsetLevel)) {
// Use a mix between regular average and upper part average.
dtmp = kUpWeight * erl.himean + (1 - kUpWeight) * erl.average;
metrics->erl.average = (int)dtmp;
} else {
metrics->erl.average = kOffsetLevel;
}
metrics->erl.max = (int)erl.max;
if (erl.min < (kOffsetLevel * (-1))) {
metrics->erl.min = (int)erl.min;
} else {
metrics->erl.min = kOffsetLevel;
}
// ERLE
metrics->erle.instant = (int)erle.instant;
if ((erle.himean > kOffsetLevel) && (erle.average > kOffsetLevel)) {
// Use a mix between regular average and upper part average.
dtmp = kUpWeight * erle.himean + (1 - kUpWeight) * erle.average;
metrics->erle.average = (int)dtmp;
} else {
metrics->erle.average = kOffsetLevel;
}
metrics->erle.max = (int)erle.max;
if (erle.min < (kOffsetLevel * (-1))) {
metrics->erle.min = (int)erle.min;
} else {
metrics->erle.min = kOffsetLevel;
}
// RERL
if ((metrics->erl.average > kOffsetLevel) &&
(metrics->erle.average > kOffsetLevel)) {
stmp = metrics->erl.average + metrics->erle.average;
} else {
stmp = kOffsetLevel;
}
metrics->rerl.average = stmp;
// No other statistics needed, but returned for completeness.
metrics->rerl.instant = stmp;
metrics->rerl.max = stmp;
metrics->rerl.min = stmp;
// A_NLP
metrics->aNlp.instant = (int)a_nlp.instant;
if ((a_nlp.himean > kOffsetLevel) && (a_nlp.average > kOffsetLevel)) {
// Use a mix between regular average and upper part average.
dtmp = kUpWeight * a_nlp.himean + (1 - kUpWeight) * a_nlp.average;
metrics->aNlp.average = (int)dtmp;
} else {
metrics->aNlp.average = kOffsetLevel;
}
metrics->aNlp.max = (int)a_nlp.max;
if (a_nlp.min < (kOffsetLevel * (-1))) {
metrics->aNlp.min = (int)a_nlp.min;
} else {
metrics->aNlp.min = kOffsetLevel;
}
return 0;
}
int WebRtcAec_GetDelayMetrics(void* handle,
int* median,
int* std,
float* fraction_poor_delays) {
Aec* self = handle;
if (median == NULL) {
self->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (std == NULL) {
self->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (self->initFlag != initCheck) {
self->lastError = AEC_UNINITIALIZED_ERROR;
return -1;
}
if (WebRtcAec_GetDelayMetricsCore(self->aec, median, std,
fraction_poor_delays) ==
-1) {
// Logging disabled.
self->lastError = AEC_UNSUPPORTED_FUNCTION_ERROR;
return -1;
}
return 0;
}
int32_t WebRtcAec_get_error_code(void* aecInst) {
Aec* aecpc = aecInst;
return aecpc->lastError;
}
AecCore* WebRtcAec_aec_core(void* handle) {
if (!handle) {
return NULL;
}
return ((Aec*)handle)->aec;
}
static int ProcessNormal(Aec* aecpc,
const float* const* nearend,
size_t num_bands,
float* const* out,
size_t nrOfSamples,
int16_t msInSndCardBuf,
int32_t skew) {
int retVal = 0;
size_t i;
size_t nBlocks10ms;
// Limit resampling to doubling/halving of signal
const float minSkewEst = -0.5f;
const float maxSkewEst = 1.0f;
msInSndCardBuf =
msInSndCardBuf > kMaxTrustedDelayMs ? kMaxTrustedDelayMs : msInSndCardBuf;
// TODO(andrew): we need to investigate if this +10 is really wanted.
msInSndCardBuf += 10;
aecpc->msInSndCardBuf = msInSndCardBuf;
if (aecpc->skewMode == kAecTrue) {
if (aecpc->skewFrCtr < 25) {
aecpc->skewFrCtr++;
} else {
retVal = WebRtcAec_GetSkew(aecpc->resampler, skew, &aecpc->skew);
if (retVal == -1) {
aecpc->skew = 0;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
}
aecpc->skew /= aecpc->sampFactor * nrOfSamples;
if (aecpc->skew < 1.0e-3 && aecpc->skew > -1.0e-3) {
aecpc->resample = kAecFalse;
} else {
aecpc->resample = kAecTrue;
}
if (aecpc->skew < minSkewEst) {
aecpc->skew = minSkewEst;
} else if (aecpc->skew > maxSkewEst) {
aecpc->skew = maxSkewEst;
}
#ifdef WEBRTC_AEC_DEBUG_DUMP
(void)fwrite(&aecpc->skew, sizeof(aecpc->skew), 1, aecpc->skewFile);
#endif
}
}
nBlocks10ms = nrOfSamples / (FRAME_LEN * aecpc->rate_factor);
if (aecpc->startup_phase) {
for (i = 0; i < num_bands; ++i) {
// Only needed if they don't already point to the same place.
if (nearend[i] != out[i]) {
memcpy(out[i], nearend[i], sizeof(nearend[i][0]) * nrOfSamples);
}
}
// The AEC is in the start up mode
// AEC is disabled until the system delay is OK
// Mechanism to ensure that the system delay is reasonably stable.
if (aecpc->checkBuffSize) {
aecpc->checkBufSizeCtr++;
// Before we fill up the far-end buffer we require the system delay
// to be stable (+/-8 ms) compared to the first value. This
// comparison is made during the following 6 consecutive 10 ms
// blocks. If it seems to be stable then we start to fill up the
// far-end buffer.
if (aecpc->counter == 0) {
aecpc->firstVal = aecpc->msInSndCardBuf;
aecpc->sum = 0;
}
if (abs(aecpc->firstVal - aecpc->msInSndCardBuf) <
WEBRTC_SPL_MAX(0.2 * aecpc->msInSndCardBuf, sampMsNb)) {
aecpc->sum += aecpc->msInSndCardBuf;
aecpc->counter++;
} else {
aecpc->counter = 0;
}
if (aecpc->counter * nBlocks10ms >= 6) {
// The far-end buffer size is determined in partitions of
// PART_LEN samples. Use 75% of the average value of the system
// delay as buffer size to start with.
aecpc->bufSizeStart =
WEBRTC_SPL_MIN((3 * aecpc->sum * aecpc->rate_factor * 8) /
(4 * aecpc->counter * PART_LEN),
kMaxBufSizeStart);
// Buffer size has now been determined.
aecpc->checkBuffSize = 0;
}
if (aecpc->checkBufSizeCtr * nBlocks10ms > 50) {
// For really bad systems, don't disable the echo canceller for
// more than 0.5 sec.
aecpc->bufSizeStart = WEBRTC_SPL_MIN(
(aecpc->msInSndCardBuf * aecpc->rate_factor * 3) / 40,
kMaxBufSizeStart);
aecpc->checkBuffSize = 0;
}
}
// If |checkBuffSize| changed in the if-statement above.
if (!aecpc->checkBuffSize) {
// The system delay is now reasonably stable (or has been unstable
// for too long). When the far-end buffer is filled with
// approximately the same amount of data as reported by the system
// we end the startup phase.
int overhead_elements =
WebRtcAec_system_delay(aecpc->aec) / PART_LEN - aecpc->bufSizeStart;
if (overhead_elements == 0) {
// Enable the AEC
aecpc->startup_phase = 0;
} else if (overhead_elements > 0) {
// TODO(bjornv): Do we need a check on how much we actually
// moved the read pointer? It should always be possible to move
// the pointer |overhead_elements| since we have only added data
// to the buffer and no delay compensation nor AEC processing
// has been done.
WebRtcAec_MoveFarReadPtr(aecpc->aec, overhead_elements);
// Enable the AEC
aecpc->startup_phase = 0;
}
}
} else {
// AEC is enabled.
EstBufDelayNormal(aecpc);
// Call the AEC.
// TODO(bjornv): Re-structure such that we don't have to pass
// |aecpc->knownDelay| as input. Change name to something like
// |system_buffer_diff|.
WebRtcAec_ProcessFrames(aecpc->aec,
nearend,
num_bands,
nrOfSamples,
aecpc->knownDelay,
out);
}
return retVal;
}
static void ProcessExtended(Aec* self,
const float* const* near,
size_t num_bands,
float* const* out,
size_t num_samples,
int16_t reported_delay_ms,
int32_t skew) {
size_t i;
const int delay_diff_offset = kDelayDiffOffsetSamples;
#if defined(WEBRTC_UNTRUSTED_DELAY)
reported_delay_ms = kFixedDelayMs;
#else
// This is the usual mode where we trust the reported system delay values.
// Due to the longer filter, we no longer add 10 ms to the reported delay
// to reduce chance of non-causality. Instead we apply a minimum here to avoid
// issues with the read pointer jumping around needlessly.
reported_delay_ms = reported_delay_ms < kMinTrustedDelayMs
? kMinTrustedDelayMs
: reported_delay_ms;
// If the reported delay appears to be bogus, we attempt to recover by using
// the measured fixed delay values. We use >= here because higher layers
// may already clamp to this maximum value, and we would otherwise not
// detect it here.
reported_delay_ms = reported_delay_ms >= kMaxTrustedDelayMs
? kFixedDelayMs
: reported_delay_ms;
#endif
self->msInSndCardBuf = reported_delay_ms;
if (!self->farend_started) {
for (i = 0; i < num_bands; ++i) {
// Only needed if they don't already point to the same place.
if (near[i] != out[i]) {
memcpy(out[i], near[i], sizeof(near[i][0]) * num_samples);
}
}
return;
}
if (self->startup_phase) {
// In the extended mode, there isn't a startup "phase", just a special
// action on the first frame. In the trusted delay case, we'll take the
// current reported delay, unless it's less then our conservative
// measurement.
int startup_size_ms =
reported_delay_ms < kFixedDelayMs ? kFixedDelayMs : reported_delay_ms;
#if defined(WEBRTC_ANDROID)
int target_delay = startup_size_ms * self->rate_factor * 8;
#else
// To avoid putting the AEC in a non-causal state we're being slightly
// conservative and scale by 2. On Android we use a fixed delay and
// therefore there is no need to scale the target_delay.
int target_delay = startup_size_ms * self->rate_factor * 8 / 2;
#endif
int overhead_elements =
(WebRtcAec_system_delay(self->aec) - target_delay) / PART_LEN;
WebRtcAec_MoveFarReadPtr(self->aec, overhead_elements);
self->startup_phase = 0;
}
EstBufDelayExtended(self);
{
// |delay_diff_offset| gives us the option to manually rewind the delay on
// very low delay platforms which can't be expressed purely through
// |reported_delay_ms|.
const int adjusted_known_delay =
WEBRTC_SPL_MAX(0, self->knownDelay + delay_diff_offset);
WebRtcAec_ProcessFrames(self->aec,
near,
num_bands,
num_samples,
adjusted_known_delay,
out);
}
}
static void EstBufDelayNormal(Aec* aecpc) {
int nSampSndCard = aecpc->msInSndCardBuf * sampMsNb * aecpc->rate_factor;
int current_delay = nSampSndCard - WebRtcAec_system_delay(aecpc->aec);
int delay_difference = 0;
// Before we proceed with the delay estimate filtering we:
// 1) Compensate for the frame that will be read.
// 2) Compensate for drift resampling.
// 3) Compensate for non-causality if needed, since the estimated delay can't
// be negative.
// 1) Compensating for the frame(s) that will be read/processed.
current_delay += FRAME_LEN * aecpc->rate_factor;
// 2) Account for resampling frame delay.
if (aecpc->skewMode == kAecTrue && aecpc->resample == kAecTrue) {
current_delay -= kResamplingDelay;
}
// 3) Compensate for non-causality, if needed, by flushing one block.
if (current_delay < PART_LEN) {
current_delay += WebRtcAec_MoveFarReadPtr(aecpc->aec, 1) * PART_LEN;
}
// We use -1 to signal an initialized state in the "extended" implementation;
// compensate for that.
aecpc->filtDelay = aecpc->filtDelay < 0 ? 0 : aecpc->filtDelay;
aecpc->filtDelay =
WEBRTC_SPL_MAX(0, (short)(0.8 * aecpc->filtDelay + 0.2 * current_delay));
delay_difference = aecpc->filtDelay - aecpc->knownDelay;
if (delay_difference > 224) {
if (aecpc->lastDelayDiff < 96) {
aecpc->timeForDelayChange = 0;
} else {
aecpc->timeForDelayChange++;
}
} else if (delay_difference < 96 && aecpc->knownDelay > 0) {
if (aecpc->lastDelayDiff > 224) {
aecpc->timeForDelayChange = 0;
} else {
aecpc->timeForDelayChange++;
}
} else {
aecpc->timeForDelayChange = 0;
}
aecpc->lastDelayDiff = delay_difference;
if (aecpc->timeForDelayChange > 25) {
aecpc->knownDelay = WEBRTC_SPL_MAX((int)aecpc->filtDelay - 160, 0);
}
}
static void EstBufDelayExtended(Aec* self) {
int reported_delay = self->msInSndCardBuf * sampMsNb * self->rate_factor;
int current_delay = reported_delay - WebRtcAec_system_delay(self->aec);
int delay_difference = 0;
// Before we proceed with the delay estimate filtering we:
// 1) Compensate for the frame that will be read.
// 2) Compensate for drift resampling.
// 3) Compensate for non-causality if needed, since the estimated delay can't
// be negative.
// 1) Compensating for the frame(s) that will be read/processed.
current_delay += FRAME_LEN * self->rate_factor;
// 2) Account for resampling frame delay.
if (self->skewMode == kAecTrue && self->resample == kAecTrue) {
current_delay -= kResamplingDelay;
}
// 3) Compensate for non-causality, if needed, by flushing two blocks.
if (current_delay < PART_LEN) {
current_delay += WebRtcAec_MoveFarReadPtr(self->aec, 2) * PART_LEN;
}
if (self->filtDelay == -1) {
self->filtDelay = WEBRTC_SPL_MAX(0, 0.5 * current_delay);
} else {
self->filtDelay = WEBRTC_SPL_MAX(
0, (short)(0.95 * self->filtDelay + 0.05 * current_delay));
}
delay_difference = self->filtDelay - self->knownDelay;
if (delay_difference > 384) {
if (self->lastDelayDiff < 128) {
self->timeForDelayChange = 0;
} else {
self->timeForDelayChange++;
}
} else if (delay_difference < 128 && self->knownDelay > 0) {
if (self->lastDelayDiff > 384) {
self->timeForDelayChange = 0;
} else {
self->timeForDelayChange++;
}
} else {
self->timeForDelayChange = 0;
}
self->lastDelayDiff = delay_difference;
if (self->timeForDelayChange > 25) {
self->knownDelay = WEBRTC_SPL_MAX((int)self->filtDelay - 256, 0);
}
}

67
webrtc/modules/audio_processing/aec/echo_cancellation_internal.h
View File

@ -1,67 +0,0 @@
/*
* Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC_ECHO_CANCELLATION_INTERNAL_H_
#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_ECHO_CANCELLATION_INTERNAL_H_
#include "webrtc/common_audio/ring_buffer.h"
#include "webrtc/modules/audio_processing/aec/aec_core.h"
typedef struct {
int delayCtr;
int sampFreq;
int splitSampFreq;
int scSampFreq;
float sampFactor; // scSampRate / sampFreq
short skewMode;
int bufSizeStart;
int knownDelay;
int rate_factor;
short initFlag; // indicates if AEC has been initialized
// Variables used for averaging far end buffer size
short counter;
int sum;
short firstVal;
short checkBufSizeCtr;
// Variables used for delay shifts
short msInSndCardBuf;
short filtDelay; // Filtered delay estimate.
int timeForDelayChange;
int startup_phase;
int checkBuffSize;
short lastDelayDiff;
#ifdef WEBRTC_AEC_DEBUG_DUMP
FILE* bufFile;
FILE* delayFile;
FILE* skewFile;
#endif
// Structures
void* resampler;
int skewFrCtr;
int resample; // if the skew is small enough we don't resample
int highSkewCtr;
float skew;
RingBuffer* far_pre_buf; // Time domain far-end pre-buffer.
int lastError;
int farend_started;
AecCore* aec;
} Aec;
#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_ECHO_CANCELLATION_INTERNAL_H_

245
webrtc/modules/audio_processing/aec/include/echo_cancellation.h
View File

@ -1,245 +0,0 @@
/*
* Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC_INCLUDE_ECHO_CANCELLATION_H_
#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_INCLUDE_ECHO_CANCELLATION_H_
#include <stddef.h>
#include "webrtc/typedefs.h"
// Errors
#define AEC_UNSPECIFIED_ERROR 12000
#define AEC_UNSUPPORTED_FUNCTION_ERROR 12001
#define AEC_UNINITIALIZED_ERROR 12002
#define AEC_NULL_POINTER_ERROR 12003
#define AEC_BAD_PARAMETER_ERROR 12004
// Warnings
#define AEC_BAD_PARAMETER_WARNING 12050
enum {
kAecNlpConservative = 0,
kAecNlpModerate,
kAecNlpAggressive
};
enum {
kAecFalse = 0,
kAecTrue
};
typedef struct {
int16_t nlpMode; // default kAecNlpModerate
int16_t skewMode; // default kAecFalse
int16_t metricsMode; // default kAecFalse
int delay_logging; // default kAecFalse
// float realSkew;
} AecConfig;
typedef struct {
int instant;
int average;
int max;
int min;
} AecLevel;
typedef struct {
AecLevel rerl;
AecLevel erl;
AecLevel erle;
AecLevel aNlp;
} AecMetrics;
struct AecCore;
#ifdef __cplusplus
extern "C" {
#endif
/*
* Allocates the memory needed by the AEC. The memory needs to be initialized
* separately using the WebRtcAec_Init() function. Returns a pointer to the
* object or NULL on error.
*/
void* WebRtcAec_Create();
/*
* This function releases the memory allocated by WebRtcAec_Create().
*
* Inputs Description
* -------------------------------------------------------------------
* void* aecInst Pointer to the AEC instance
*/
void WebRtcAec_Free(void* aecInst);
/*
* Initializes an AEC instance.
*
* Inputs Description
* -------------------------------------------------------------------
* void* aecInst Pointer to the AEC instance
* int32_t sampFreq Sampling frequency of data
* int32_t scSampFreq Soundcard sampling frequency
*
* Outputs Description
* -------------------------------------------------------------------
* int32_t return 0: OK
* -1: error
*/
int32_t WebRtcAec_Init(void* aecInst, int32_t sampFreq, int32_t scSampFreq);
/*
* Inserts an 80 or 160 sample block of data into the farend buffer.
*
* Inputs Description
* -------------------------------------------------------------------
* void* aecInst Pointer to the AEC instance
* const float* farend In buffer containing one frame of
* farend signal for L band
* int16_t nrOfSamples Number of samples in farend buffer
*
* Outputs Description
* -------------------------------------------------------------------
* int32_t return 0: OK
* -1: error
*/
int32_t WebRtcAec_BufferFarend(void* aecInst,
const float* farend,
size_t nrOfSamples);
/*
* Runs the echo canceller on an 80 or 160 sample blocks of data.
*
* Inputs Description
* -------------------------------------------------------------------
* void* aecInst Pointer to the AEC instance
* float* const* nearend In buffer containing one frame of
* nearend+echo signal for each band
* int num_bands Number of bands in nearend buffer
* int16_t nrOfSamples Number of samples in nearend buffer
* int16_t msInSndCardBuf Delay estimate for sound card and
* system buffers
* int16_t skew Difference between number of samples played
* and recorded at the soundcard (for clock skew
* compensation)
*
* Outputs Description
* -------------------------------------------------------------------
* float* const* out Out buffer, one frame of processed nearend
* for each band
* int32_t return 0: OK
* -1: error
*/
int32_t WebRtcAec_Process(void* aecInst,
const float* const* nearend,
size_t num_bands,
float* const* out,
size_t nrOfSamples,
int16_t msInSndCardBuf,
int32_t skew);
/*
* This function enables the user to set certain parameters on-the-fly.
*
* Inputs Description
* -------------------------------------------------------------------
* void* handle Pointer to the AEC instance
* AecConfig config Config instance that contains all
* properties to be set
*
* Outputs Description
* -------------------------------------------------------------------
* int return 0: OK
* -1: error
*/
int WebRtcAec_set_config(void* handle, AecConfig config);
/*
* Gets the current echo status of the nearend signal.
*
* Inputs Description
* -------------------------------------------------------------------
* void* handle Pointer to the AEC instance
*
* Outputs Description
* -------------------------------------------------------------------
* int* status 0: Almost certainly nearend single-talk
* 1: Might not be neared single-talk
* int return 0: OK
* -1: error
*/
int WebRtcAec_get_echo_status(void* handle, int* status);
/*
* Gets the current echo metrics for the session.
*
* Inputs Description
* -------------------------------------------------------------------
* void* handle Pointer to the AEC instance
*
* Outputs Description
* -------------------------------------------------------------------
* AecMetrics* metrics Struct which will be filled out with the
* current echo metrics.
* int return 0: OK
* -1: error
*/
int WebRtcAec_GetMetrics(void* handle, AecMetrics* metrics);
/*
* Gets the current delay metrics for the session.
*
* Inputs Description
* -------------------------------------------------------------------
* void* handle Pointer to the AEC instance
*
* Outputs Description
* -------------------------------------------------------------------
* int* median Delay median value.
* int* std Delay standard deviation.
* float* fraction_poor_delays Fraction of the delay estimates that may
* cause the AEC to perform poorly.
*
* int return 0: OK
* -1: error
*/
int WebRtcAec_GetDelayMetrics(void* handle,
int* median,
int* std,
float* fraction_poor_delays);
/*
* Gets the last error code.
*
* Inputs Description
* -------------------------------------------------------------------
* void* aecInst Pointer to the AEC instance
*
* Outputs Description
* -------------------------------------------------------------------
* int32_t return 11000-11100: error code
*/
int32_t WebRtcAec_get_error_code(void* aecInst);
// Returns a pointer to the low level AEC handle.
//
// Input:
// - handle : Pointer to the AEC instance.
//
// Return value:
// - AecCore pointer : NULL for error.
//
struct AecCore* WebRtcAec_aec_core(void* handle);
#ifdef __cplusplus
}
#endif
#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_INCLUDE_ECHO_CANCELLATION_H_

367
webrtc/modules/audio_processing/aec3/BUILD.gn Normal file
View File

@ -0,0 +1,367 @@
# Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
#
# Use of this source code is governed by a BSD-style license
# that can be found in the LICENSE file in the root of the source
# tree. An additional intellectual property rights grant can be found
# in the file PATENTS. All contributing project authors may
# be found in the AUTHORS file in the root of the source tree.
import("../../../webrtc.gni")
rtc_library("aec3") {
visibility = [ "*" ]
configs += [ "..:apm_debug_dump" ]
sources = [
"adaptive_fir_filter.cc",
"adaptive_fir_filter_erl.cc",
"aec3_common.cc",
"aec3_fft.cc",
"aec_state.cc",
"aec_state.h",
"alignment_mixer.cc",
"alignment_mixer.h",
"api_call_jitter_metrics.cc",
"api_call_jitter_metrics.h",
"block_buffer.cc",
"block_delay_buffer.cc",
"block_delay_buffer.h",
"block_framer.cc",
"block_framer.h",
"block_processor.cc",
"block_processor.h",
"block_processor_metrics.cc",
"block_processor_metrics.h",
"clockdrift_detector.cc",
"clockdrift_detector.h",
"coarse_filter_update_gain.cc",
"coarse_filter_update_gain.h",
"comfort_noise_generator.cc",
"comfort_noise_generator.h",
"decimator.cc",
"decimator.h",
"delay_estimate.h",
"dominant_nearend_detector.cc",
"dominant_nearend_detector.h",
"downsampled_render_buffer.cc",
"downsampled_render_buffer.h",
"echo_audibility.cc",
"echo_audibility.h",
"echo_canceller3.cc",
"echo_canceller3.h",
"echo_path_delay_estimator.cc",
"echo_path_delay_estimator.h",
"echo_path_variability.cc",
"echo_path_variability.h",
"echo_remover.cc",
"echo_remover.h",
"echo_remover_metrics.cc",
"echo_remover_metrics.h",
"erl_estimator.cc",
"erl_estimator.h",
"erle_estimator.cc",
"erle_estimator.h",
"fft_buffer.cc",
"filter_analyzer.cc",
"filter_analyzer.h",
"frame_blocker.cc",
"frame_blocker.h",
"fullband_erle_estimator.cc",
"fullband_erle_estimator.h",
"matched_filter.cc",
"matched_filter_lag_aggregator.cc",
"matched_filter_lag_aggregator.h",
"moving_average.cc",
"moving_average.h",
"nearend_detector.h",
"refined_filter_update_gain.cc",
"refined_filter_update_gain.h",
"render_buffer.cc",
"render_delay_buffer.cc",
"render_delay_buffer.h",
"render_delay_controller.cc",
"render_delay_controller.h",
"render_delay_controller_metrics.cc",
"render_delay_controller_metrics.h",
"render_signal_analyzer.cc",
"render_signal_analyzer.h",
"residual_echo_estimator.cc",
"residual_echo_estimator.h",
"reverb_decay_estimator.cc",
"reverb_decay_estimator.h",
"reverb_frequency_response.cc",
"reverb_frequency_response.h",
"reverb_model.cc",
"reverb_model.h",
"reverb_model_estimator.cc",
"reverb_model_estimator.h",
"signal_dependent_erle_estimator.cc",
"signal_dependent_erle_estimator.h",
"spectrum_buffer.cc",
"stationarity_estimator.cc",
"stationarity_estimator.h",
"subband_erle_estimator.cc",
"subband_erle_estimator.h",
"subband_nearend_detector.cc",
"subband_nearend_detector.h",
"subtractor.cc",
"subtractor.h",
"subtractor_output.cc",
"subtractor_output.h",
"subtractor_output_analyzer.cc",
"subtractor_output_analyzer.h",
"suppression_filter.cc",
"suppression_filter.h",
"suppression_gain.cc",
"suppression_gain.h",
"transparent_mode.cc",
"transparent_mode.h",
]
defines = []
if (rtc_build_with_neon && current_cpu != "arm64") {
suppressed_configs += [ "//build/config/compiler:compiler_arm_fpu" ]
cflags = [ "-mfpu=neon" ]
}
deps = [
":adaptive_fir_filter",
":adaptive_fir_filter_erl",
":aec3_common",
":aec3_fft",
":fft_data",
":matched_filter",
":render_buffer",
":vector_math",
"..:apm_logging",
"..:audio_buffer",
"..:high_pass_filter",
"../../../api:array_view",
"../../../api/audio:aec3_config",
"../../../api/audio:echo_control",
"../../../common_audio:common_audio_c",
"../../../rtc_base:checks",
"../../../rtc_base:rtc_base_approved",
"../../../rtc_base:safe_minmax",
"../../../rtc_base/experiments:field_trial_parser",
"../../../rtc_base/system:arch",
"../../../system_wrappers",
"../../../system_wrappers:field_trial",
"../../../system_wrappers:metrics",
"../utility:cascaded_biquad_filter",
]
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
if (current_cpu == "x86" || current_cpu == "x64") {
deps += [ ":aec3_avx2" ]
}
}
rtc_source_set("aec3_common") {
sources = [ "aec3_common.h" ]
}
rtc_source_set("aec3_fft") {
sources = [ "aec3_fft.h" ]
deps = [
":aec3_common",
":fft_data",
"../../../api:array_view",
"../../../common_audio/third_party/ooura:fft_size_128",
"../../../rtc_base:checks",
"../../../rtc_base:rtc_base_approved",
"../../../rtc_base/system:arch",
]
}
rtc_source_set("render_buffer") {
sources = [
"block_buffer.h",
"fft_buffer.h",
"render_buffer.h",
"spectrum_buffer.h",
]
deps = [
":aec3_common",
":fft_data",
"../../../api:array_view",
"../../../rtc_base:checks",
"../../../rtc_base:rtc_base_approved",
"../../../rtc_base/system:arch",
]
}
rtc_source_set("adaptive_fir_filter") {
sources = [ "adaptive_fir_filter.h" ]
deps = [
":aec3_common",
":aec3_fft",
":fft_data",
":render_buffer",
"..:apm_logging",
"../../../api:array_view",
"../../../rtc_base/system:arch",
]
}
rtc_source_set("adaptive_fir_filter_erl") {
sources = [ "adaptive_fir_filter_erl.h" ]
deps = [
":aec3_common",
"../../../api:array_view",
"../../../rtc_base/system:arch",
]
}
rtc_source_set("matched_filter") {
sources = [ "matched_filter.h" ]
deps = [
":aec3_common",
"../../../api:array_view",
"../../../rtc_base:rtc_base_approved",
"../../../rtc_base/system:arch",
]
}
rtc_source_set("vector_math") {
sources = [ "vector_math.h" ]
deps = [
":aec3_common",
"../../../api:array_view",
"../../../rtc_base:checks",
"../../../rtc_base/system:arch",
]
}
rtc_source_set("fft_data") {
sources = [ "fft_data.h" ]
deps = [
":aec3_common",
"../../../api:array_view",
"../../../rtc_base/system:arch",
]
}
if (current_cpu == "x86" || current_cpu == "x64") {
rtc_library("aec3_avx2") {
configs += [ "..:apm_debug_dump" ]
sources = [
"adaptive_fir_filter_avx2.cc",
"adaptive_fir_filter_erl_avx2.cc",
"fft_data_avx2.cc",
"matched_filter_avx2.cc",
"vector_math_avx2.cc",
]
if (is_win) {
cflags = [ "/arch:AVX2" ]
} else {
cflags = [
"-mavx2",
"-mfma",
]
}
deps = [
":adaptive_fir_filter",
":adaptive_fir_filter_erl",
":fft_data",
":matched_filter",
":vector_math",
"../../../api:array_view",
"../../../rtc_base:checks",
]
}
}
if (rtc_include_tests) {
rtc_library("aec3_unittests") {
testonly = true
configs += [ "..:apm_debug_dump" ]
sources = [
"mock/mock_block_processor.cc",
"mock/mock_block_processor.h",
"mock/mock_echo_remover.cc",
"mock/mock_echo_remover.h",
"mock/mock_render_delay_buffer.cc",
"mock/mock_render_delay_buffer.h",
"mock/mock_render_delay_controller.cc",
"mock/mock_render_delay_controller.h",
]
deps = [
":adaptive_fir_filter",
":adaptive_fir_filter_erl",
":aec3",
":aec3_common",
":aec3_fft",
":fft_data",
":matched_filter",
":render_buffer",
":vector_math",
"..:apm_logging",
"..:audio_buffer",
"..:audio_processing",
"..:audio_processing_unittests",
"..:high_pass_filter",
"../../../api:array_view",
"../../../api/audio:aec3_config",
"../../../rtc_base:checks",
"../../../rtc_base:rtc_base_approved",
"../../../rtc_base:safe_minmax",
"../../../rtc_base/system:arch",
"../../../system_wrappers",
"../../../test:field_trial",
"../../../test:test_support",
"../utility:cascaded_biquad_filter",
]
absl_deps = [ "//third_party/abseil-cpp/absl/types:optional" ]
defines = []
if (rtc_enable_protobuf) {
sources += [
"adaptive_fir_filter_erl_unittest.cc",
"adaptive_fir_filter_unittest.cc",
"aec3_fft_unittest.cc",
"aec_state_unittest.cc",
"alignment_mixer_unittest.cc",
"api_call_jitter_metrics_unittest.cc",
"block_delay_buffer_unittest.cc",
"block_framer_unittest.cc",
"block_processor_metrics_unittest.cc",
"block_processor_unittest.cc",
"clockdrift_detector_unittest.cc",
"coarse_filter_update_gain_unittest.cc",
"comfort_noise_generator_unittest.cc",
"decimator_unittest.cc",
"echo_canceller3_unittest.cc",
"echo_path_delay_estimator_unittest.cc",
"echo_path_variability_unittest.cc",
"echo_remover_metrics_unittest.cc",
"echo_remover_unittest.cc",
"erl_estimator_unittest.cc",
"erle_estimator_unittest.cc",
"fft_data_unittest.cc",
"filter_analyzer_unittest.cc",
"frame_blocker_unittest.cc",
"matched_filter_lag_aggregator_unittest.cc",
"matched_filter_unittest.cc",
"moving_average_unittest.cc",
"refined_filter_update_gain_unittest.cc",
"render_buffer_unittest.cc",
"render_delay_buffer_unittest.cc",
"render_delay_controller_metrics_unittest.cc",
"render_delay_controller_unittest.cc",
"render_signal_analyzer_unittest.cc",
"residual_echo_estimator_unittest.cc",
"reverb_model_estimator_unittest.cc",
"signal_dependent_erle_estimator_unittest.cc",
"subtractor_unittest.cc",
"suppression_filter_unittest.cc",
"suppression_gain_unittest.cc",
"vector_math_unittest.cc",
]
}
}
}

740
webrtc/modules/audio_processing/aec3/adaptive_fir_filter.cc Normal file
View File

@ -0,0 +1,740 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/adaptive_fir_filter.h"
// Defines WEBRTC_ARCH_X86_FAMILY, used below.
#include "rtc_base/system/arch.h"
#if defined(WEBRTC_HAS_NEON)
#include <arm_neon.h>
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
#include <emmintrin.h>
#endif
#include <math.h>
#include <algorithm>
#include <functional>
#include "modules/audio_processing/aec3/fft_data.h"
#include "rtc_base/checks.h"
namespace webrtc {
namespace aec3 {
// Computes and stores the frequency response of the filter.
void ComputeFrequencyResponse(
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) {
for (auto& H2_ch : *H2) {
H2_ch.fill(0.f);
}
const size_t num_render_channels = H[0].size();
RTC_DCHECK_EQ(H.size(), H2->capacity());
for (size_t p = 0; p < num_partitions; ++p) {
RTC_DCHECK_EQ(kFftLengthBy2Plus1, (*H2)[p].size());
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t j = 0; j < kFftLengthBy2Plus1; ++j) {
float tmp =
H[p][ch].re[j] * H[p][ch].re[j] + H[p][ch].im[j] * H[p][ch].im[j];
(*H2)[p][j] = std::max((*H2)[p][j], tmp);
}
}
}
}
#if defined(WEBRTC_HAS_NEON)
// Computes and stores the frequency response of the filter.
void ComputeFrequencyResponse_Neon(
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) {
for (auto& H2_ch : *H2) {
H2_ch.fill(0.f);
}
const size_t num_render_channels = H[0].size();
RTC_DCHECK_EQ(H.size(), H2->capacity());
for (size_t p = 0; p < num_partitions; ++p) {
RTC_DCHECK_EQ(kFftLengthBy2Plus1, (*H2)[p].size());
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t j = 0; j < kFftLengthBy2; j += 4) {
const float32x4_t re = vld1q_f32(&H[p][ch].re[j]);
const float32x4_t im = vld1q_f32(&H[p][ch].im[j]);
float32x4_t H2_new = vmulq_f32(re, re);
H2_new = vmlaq_f32(H2_new, im, im);
float32x4_t H2_p_j = vld1q_f32(&(*H2)[p][j]);
H2_p_j = vmaxq_f32(H2_p_j, H2_new);
vst1q_f32(&(*H2)[p][j], H2_p_j);
}
float H2_new = H[p][ch].re[kFftLengthBy2] * H[p][ch].re[kFftLengthBy2] +
H[p][ch].im[kFftLengthBy2] * H[p][ch].im[kFftLengthBy2];
(*H2)[p][kFftLengthBy2] = std::max((*H2)[p][kFftLengthBy2], H2_new);
}
}
}
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
// Computes and stores the frequency response of the filter.
void ComputeFrequencyResponse_Sse2(
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) {
for (auto& H2_ch : *H2) {
H2_ch.fill(0.f);
}
const size_t num_render_channels = H[0].size();
RTC_DCHECK_EQ(H.size(), H2->capacity());
// constexpr __mmmask8 kMaxMask = static_cast<__mmmask8>(256u);
for (size_t p = 0; p < num_partitions; ++p) {
RTC_DCHECK_EQ(kFftLengthBy2Plus1, (*H2)[p].size());
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t j = 0; j < kFftLengthBy2; j += 4) {
const __m128 re = _mm_loadu_ps(&H[p][ch].re[j]);
const __m128 re2 = _mm_mul_ps(re, re);
const __m128 im = _mm_loadu_ps(&H[p][ch].im[j]);
const __m128 im2 = _mm_mul_ps(im, im);
const __m128 H2_new = _mm_add_ps(re2, im2);
__m128 H2_k_j = _mm_loadu_ps(&(*H2)[p][j]);
H2_k_j = _mm_max_ps(H2_k_j, H2_new);
_mm_storeu_ps(&(*H2)[p][j], H2_k_j);
}
float H2_new = H[p][ch].re[kFftLengthBy2] * H[p][ch].re[kFftLengthBy2] +
H[p][ch].im[kFftLengthBy2] * H[p][ch].im[kFftLengthBy2];
(*H2)[p][kFftLengthBy2] = std::max((*H2)[p][kFftLengthBy2], H2_new);
}
}
}
#endif
// Adapts the filter partitions as H(t+1)=H(t)+G(t)*conj(X(t)).
void AdaptPartitions(const RenderBuffer& render_buffer,
const FftData& G,
size_t num_partitions,
std::vector<std::vector<FftData>>* H) {
rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
render_buffer.GetFftBuffer();
size_t index = render_buffer.Position();
const size_t num_render_channels = render_buffer_data[index].size();
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
const FftData& X_p_ch = render_buffer_data[index][ch];
FftData& H_p_ch = (*H)[p][ch];
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
H_p_ch.re[k] += X_p_ch.re[k] * G.re[k] + X_p_ch.im[k] * G.im[k];
H_p_ch.im[k] += X_p_ch.re[k] * G.im[k] - X_p_ch.im[k] * G.re[k];
}
}
index = index < (render_buffer_data.size() - 1) ? index + 1 : 0;
}
}
#if defined(WEBRTC_HAS_NEON)
// Adapts the filter partitions. (Neon variant)
void AdaptPartitions_Neon(const RenderBuffer& render_buffer,
const FftData& G,
size_t num_partitions,
std::vector<std::vector<FftData>>* H) {
rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
render_buffer.GetFftBuffer();
const size_t num_render_channels = render_buffer_data[0].size();
const size_t lim1 = std::min(
render_buffer_data.size() - render_buffer.Position(), num_partitions);
const size_t lim2 = num_partitions;
constexpr size_t kNumFourBinBands = kFftLengthBy2 / 4;
size_t X_partition = render_buffer.Position();
size_t limit = lim1;
size_t p = 0;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
FftData& H_p_ch = (*H)[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
for (size_t k = 0, n = 0; n < kNumFourBinBands; ++n, k += 4) {
const float32x4_t G_re = vld1q_f32(&G.re[k]);
const float32x4_t G_im = vld1q_f32(&G.im[k]);
const float32x4_t X_re = vld1q_f32(&X.re[k]);
const float32x4_t X_im = vld1q_f32(&X.im[k]);
const float32x4_t H_re = vld1q_f32(&H_p_ch.re[k]);
const float32x4_t H_im = vld1q_f32(&H_p_ch.im[k]);
const float32x4_t a = vmulq_f32(X_re, G_re);
const float32x4_t e = vmlaq_f32(a, X_im, G_im);
const float32x4_t c = vmulq_f32(X_re, G_im);
const float32x4_t f = vmlsq_f32(c, X_im, G_re);
const float32x4_t g = vaddq_f32(H_re, e);
const float32x4_t h = vaddq_f32(H_im, f);
vst1q_f32(&H_p_ch.re[k], g);
vst1q_f32(&H_p_ch.im[k], h);
}
}
}
X_partition = 0;
limit = lim2;
} while (p < lim2);
X_partition = render_buffer.Position();
limit = lim1;
p = 0;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
FftData& H_p_ch = (*H)[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
H_p_ch.re[kFftLengthBy2] += X.re[kFftLengthBy2] * G.re[kFftLengthBy2] +
X.im[kFftLengthBy2] * G.im[kFftLengthBy2];
H_p_ch.im[kFftLengthBy2] += X.re[kFftLengthBy2] * G.im[kFftLengthBy2] -
X.im[kFftLengthBy2] * G.re[kFftLengthBy2];
}
}
X_partition = 0;
limit = lim2;
} while (p < lim2);
}
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
// Adapts the filter partitions. (SSE2 variant)
void AdaptPartitions_Sse2(const RenderBuffer& render_buffer,
const FftData& G,
size_t num_partitions,
std::vector<std::vector<FftData>>* H) {
rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
render_buffer.GetFftBuffer();
const size_t num_render_channels = render_buffer_data[0].size();
const size_t lim1 = std::min(
render_buffer_data.size() - render_buffer.Position(), num_partitions);
const size_t lim2 = num_partitions;
constexpr size_t kNumFourBinBands = kFftLengthBy2 / 4;
size_t X_partition = render_buffer.Position();
size_t limit = lim1;
size_t p = 0;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
FftData& H_p_ch = (*H)[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
for (size_t k = 0, n = 0; n < kNumFourBinBands; ++n, k += 4) {
const __m128 G_re = _mm_loadu_ps(&G.re[k]);
const __m128 G_im = _mm_loadu_ps(&G.im[k]);
const __m128 X_re = _mm_loadu_ps(&X.re[k]);
const __m128 X_im = _mm_loadu_ps(&X.im[k]);
const __m128 H_re = _mm_loadu_ps(&H_p_ch.re[k]);
const __m128 H_im = _mm_loadu_ps(&H_p_ch.im[k]);
const __m128 a = _mm_mul_ps(X_re, G_re);
const __m128 b = _mm_mul_ps(X_im, G_im);
const __m128 c = _mm_mul_ps(X_re, G_im);
const __m128 d = _mm_mul_ps(X_im, G_re);
const __m128 e = _mm_add_ps(a, b);
const __m128 f = _mm_sub_ps(c, d);
const __m128 g = _mm_add_ps(H_re, e);
const __m128 h = _mm_add_ps(H_im, f);
_mm_storeu_ps(&H_p_ch.re[k], g);
_mm_storeu_ps(&H_p_ch.im[k], h);
}
}
}
X_partition = 0;
limit = lim2;
} while (p < lim2);
X_partition = render_buffer.Position();
limit = lim1;
p = 0;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
FftData& H_p_ch = (*H)[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
H_p_ch.re[kFftLengthBy2] += X.re[kFftLengthBy2] * G.re[kFftLengthBy2] +
X.im[kFftLengthBy2] * G.im[kFftLengthBy2];
H_p_ch.im[kFftLengthBy2] += X.re[kFftLengthBy2] * G.im[kFftLengthBy2] -
X.im[kFftLengthBy2] * G.re[kFftLengthBy2];
}
}
X_partition = 0;
limit = lim2;
} while (p < lim2);
}
#endif
// Produces the filter output.
void ApplyFilter(const RenderBuffer& render_buffer,
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
FftData* S) {
S->re.fill(0.f);
S->im.fill(0.f);
rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
render_buffer.GetFftBuffer();
size_t index = render_buffer.Position();
const size_t num_render_channels = render_buffer_data[index].size();
for (size_t p = 0; p < num_partitions; ++p) {
RTC_DCHECK_EQ(num_render_channels, H[p].size());
for (size_t ch = 0; ch < num_render_channels; ++ch) {
const FftData& X_p_ch = render_buffer_data[index][ch];
const FftData& H_p_ch = H[p][ch];
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
S->re[k] += X_p_ch.re[k] * H_p_ch.re[k] - X_p_ch.im[k] * H_p_ch.im[k];
S->im[k] += X_p_ch.re[k] * H_p_ch.im[k] + X_p_ch.im[k] * H_p_ch.re[k];
}
}
index = index < (render_buffer_data.size() - 1) ? index + 1 : 0;
}
}
#if defined(WEBRTC_HAS_NEON)
// Produces the filter output (Neon variant).
void ApplyFilter_Neon(const RenderBuffer& render_buffer,
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
FftData* S) {
// const RenderBuffer& render_buffer,
// rtc::ArrayView<const FftData> H,
// FftData* S) {
RTC_DCHECK_GE(H.size(), H.size() - 1);
S->Clear();
rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
render_buffer.GetFftBuffer();
const size_t num_render_channels = render_buffer_data[0].size();
const size_t lim1 = std::min(
render_buffer_data.size() - render_buffer.Position(), num_partitions);
const size_t lim2 = num_partitions;
constexpr size_t kNumFourBinBands = kFftLengthBy2 / 4;
size_t X_partition = render_buffer.Position();
size_t p = 0;
size_t limit = lim1;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
const FftData& H_p_ch = H[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
for (size_t k = 0, n = 0; n < kNumFourBinBands; ++n, k += 4) {
const float32x4_t X_re = vld1q_f32(&X.re[k]);
const float32x4_t X_im = vld1q_f32(&X.im[k]);
const float32x4_t H_re = vld1q_f32(&H_p_ch.re[k]);
const float32x4_t H_im = vld1q_f32(&H_p_ch.im[k]);
const float32x4_t S_re = vld1q_f32(&S->re[k]);
const float32x4_t S_im = vld1q_f32(&S->im[k]);
const float32x4_t a = vmulq_f32(X_re, H_re);
const float32x4_t e = vmlsq_f32(a, X_im, H_im);
const float32x4_t c = vmulq_f32(X_re, H_im);
const float32x4_t f = vmlaq_f32(c, X_im, H_re);
const float32x4_t g = vaddq_f32(S_re, e);
const float32x4_t h = vaddq_f32(S_im, f);
vst1q_f32(&S->re[k], g);
vst1q_f32(&S->im[k], h);
}
}
}
limit = lim2;
X_partition = 0;
} while (p < lim2);
X_partition = render_buffer.Position();
p = 0;
limit = lim1;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
const FftData& H_p_ch = H[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
S->re[kFftLengthBy2] += X.re[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2] -
X.im[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2];
S->im[kFftLengthBy2] += X.re[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2] +
X.im[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2];
}
}
limit = lim2;
X_partition = 0;
} while (p < lim2);
}
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
// Produces the filter output (SSE2 variant).
void ApplyFilter_Sse2(const RenderBuffer& render_buffer,
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
FftData* S) {
// const RenderBuffer& render_buffer,
// rtc::ArrayView<const FftData> H,
// FftData* S) {
RTC_DCHECK_GE(H.size(), H.size() - 1);
S->re.fill(0.f);
S->im.fill(0.f);
rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
render_buffer.GetFftBuffer();
const size_t num_render_channels = render_buffer_data[0].size();
const size_t lim1 = std::min(
render_buffer_data.size() - render_buffer.Position(), num_partitions);
const size_t lim2 = num_partitions;
constexpr size_t kNumFourBinBands = kFftLengthBy2 / 4;
size_t X_partition = render_buffer.Position();
size_t p = 0;
size_t limit = lim1;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
const FftData& H_p_ch = H[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
for (size_t k = 0, n = 0; n < kNumFourBinBands; ++n, k += 4) {
const __m128 X_re = _mm_loadu_ps(&X.re[k]);
const __m128 X_im = _mm_loadu_ps(&X.im[k]);
const __m128 H_re = _mm_loadu_ps(&H_p_ch.re[k]);
const __m128 H_im = _mm_loadu_ps(&H_p_ch.im[k]);
const __m128 S_re = _mm_loadu_ps(&S->re[k]);
const __m128 S_im = _mm_loadu_ps(&S->im[k]);
const __m128 a = _mm_mul_ps(X_re, H_re);
const __m128 b = _mm_mul_ps(X_im, H_im);
const __m128 c = _mm_mul_ps(X_re, H_im);
const __m128 d = _mm_mul_ps(X_im, H_re);
const __m128 e = _mm_sub_ps(a, b);
const __m128 f = _mm_add_ps(c, d);
const __m128 g = _mm_add_ps(S_re, e);
const __m128 h = _mm_add_ps(S_im, f);
_mm_storeu_ps(&S->re[k], g);
_mm_storeu_ps(&S->im[k], h);
}
}
}
limit = lim2;
X_partition = 0;
} while (p < lim2);
X_partition = render_buffer.Position();
p = 0;
limit = lim1;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
const FftData& H_p_ch = H[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
S->re[kFftLengthBy2] += X.re[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2] -
X.im[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2];
S->im[kFftLengthBy2] += X.re[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2] +
X.im[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2];
}
}
limit = lim2;
X_partition = 0;
} while (p < lim2);
}
#endif
} // namespace aec3
namespace {
// Ensures that the newly added filter partitions after a size increase are set
// to zero.
void ZeroFilter(size_t old_size,
size_t new_size,
std::vector<std::vector<FftData>>* H) {
RTC_DCHECK_GE(H->size(), old_size);
RTC_DCHECK_GE(H->size(), new_size);
for (size_t p = old_size; p < new_size; ++p) {
RTC_DCHECK_EQ((*H)[p].size(), (*H)[0].size());
for (size_t ch = 0; ch < (*H)[0].size(); ++ch) {
(*H)[p][ch].Clear();
}
}
}
} // namespace
AdaptiveFirFilter::AdaptiveFirFilter(size_t max_size_partitions,
size_t initial_size_partitions,
size_t size_change_duration_blocks,
size_t num_render_channels,
Aec3Optimization optimization,
ApmDataDumper* data_dumper)
: data_dumper_(data_dumper),
fft_(),
optimization_(optimization),
num_render_channels_(num_render_channels),
max_size_partitions_(max_size_partitions),
size_change_duration_blocks_(
static_cast<int>(size_change_duration_blocks)),
current_size_partitions_(initial_size_partitions),
target_size_partitions_(initial_size_partitions),
old_target_size_partitions_(initial_size_partitions),
H_(max_size_partitions_, std::vector<FftData>(num_render_channels_)) {
RTC_DCHECK(data_dumper_);
RTC_DCHECK_GE(max_size_partitions, initial_size_partitions);
RTC_DCHECK_LT(0, size_change_duration_blocks_);
one_by_size_change_duration_blocks_ = 1.f / size_change_duration_blocks_;
ZeroFilter(0, max_size_partitions_, &H_);
SetSizePartitions(current_size_partitions_, true);
}
AdaptiveFirFilter::~AdaptiveFirFilter() = default;
void AdaptiveFirFilter::HandleEchoPathChange() {
// TODO(peah): Check the value and purpose of the code below.
ZeroFilter(current_size_partitions_, max_size_partitions_, &H_);
}
void AdaptiveFirFilter::SetSizePartitions(size_t size, bool immediate_effect) {
RTC_DCHECK_EQ(max_size_partitions_, H_.capacity());
RTC_DCHECK_LE(size, max_size_partitions_);
target_size_partitions_ = std::min(max_size_partitions_, size);
if (immediate_effect) {
size_t old_size_partitions_ = current_size_partitions_;
current_size_partitions_ = old_target_size_partitions_ =
target_size_partitions_;
ZeroFilter(old_size_partitions_, current_size_partitions_, &H_);
partition_to_constrain_ =
std::min(partition_to_constrain_, current_size_partitions_ - 1);
size_change_counter_ = 0;
} else {
size_change_counter_ = size_change_duration_blocks_;
}
}
void AdaptiveFirFilter::UpdateSize() {
RTC_DCHECK_GE(size_change_duration_blocks_, size_change_counter_);
size_t old_size_partitions_ = current_size_partitions_;
if (size_change_counter_ > 0) {
--size_change_counter_;
auto average = [](float from, float to, float from_weight) {
return from * from_weight + to * (1.f - from_weight);
};
float change_factor =
size_change_counter_ * one_by_size_change_duration_blocks_;
current_size_partitions_ = average(old_target_size_partitions_,
target_size_partitions_, change_factor);
partition_to_constrain_ =
std::min(partition_to_constrain_, current_size_partitions_ - 1);
} else {
current_size_partitions_ = old_target_size_partitions_ =
target_size_partitions_;
}
ZeroFilter(old_size_partitions_, current_size_partitions_, &H_);
RTC_DCHECK_LE(0, size_change_counter_);
}
void AdaptiveFirFilter::Filter(const RenderBuffer& render_buffer,
FftData* S) const {
RTC_DCHECK(S);
switch (optimization_) {
#if defined(WEBRTC_ARCH_X86_FAMILY)
case Aec3Optimization::kSse2:
aec3::ApplyFilter_Sse2(render_buffer, current_size_partitions_, H_, S);
break;
case Aec3Optimization::kAvx2:
aec3::ApplyFilter_Avx2(render_buffer, current_size_partitions_, H_, S);
break;
#endif
#if defined(WEBRTC_HAS_NEON)
case Aec3Optimization::kNeon:
aec3::ApplyFilter_Neon(render_buffer, current_size_partitions_, H_, S);
break;
#endif
default:
aec3::ApplyFilter(render_buffer, current_size_partitions_, H_, S);
}
}
void AdaptiveFirFilter::Adapt(const RenderBuffer& render_buffer,
const FftData& G) {
// Adapt the filter and update the filter size.
AdaptAndUpdateSize(render_buffer, G);
// Constrain the filter partitions in a cyclic manner.
Constrain();
}
void AdaptiveFirFilter::Adapt(const RenderBuffer& render_buffer,
const FftData& G,
std::vector<float>* impulse_response) {
// Adapt the filter and update the filter size.
AdaptAndUpdateSize(render_buffer, G);
// Constrain the filter partitions in a cyclic manner.
ConstrainAndUpdateImpulseResponse(impulse_response);
}
void AdaptiveFirFilter::ComputeFrequencyResponse(
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) const {
RTC_DCHECK_GE(max_size_partitions_, H2->capacity());
H2->resize(current_size_partitions_);
switch (optimization_) {
#if defined(WEBRTC_ARCH_X86_FAMILY)
case Aec3Optimization::kSse2:
aec3::ComputeFrequencyResponse_Sse2(current_size_partitions_, H_, H2);
break;
case Aec3Optimization::kAvx2:
aec3::ComputeFrequencyResponse_Avx2(current_size_partitions_, H_, H2);
break;
#endif
#if defined(WEBRTC_HAS_NEON)
case Aec3Optimization::kNeon:
aec3::ComputeFrequencyResponse_Neon(current_size_partitions_, H_, H2);
break;
#endif
default:
aec3::ComputeFrequencyResponse(current_size_partitions_, H_, H2);
}
}
void AdaptiveFirFilter::AdaptAndUpdateSize(const RenderBuffer& render_buffer,
const FftData& G) {
// Update the filter size if needed.
UpdateSize();
// Adapt the filter.
switch (optimization_) {
#if defined(WEBRTC_ARCH_X86_FAMILY)
case Aec3Optimization::kSse2:
aec3::AdaptPartitions_Sse2(render_buffer, G, current_size_partitions_,
&H_);
break;
case Aec3Optimization::kAvx2:
aec3::AdaptPartitions_Avx2(render_buffer, G, current_size_partitions_,
&H_);
break;
#endif
#if defined(WEBRTC_HAS_NEON)
case Aec3Optimization::kNeon:
aec3::AdaptPartitions_Neon(render_buffer, G, current_size_partitions_,
&H_);
break;
#endif
default:
aec3::AdaptPartitions(render_buffer, G, current_size_partitions_, &H_);
}
}
// Constrains the partition of the frequency domain filter to be limited in
// time via setting the relevant time-domain coefficients to zero and updates
// the corresponding values in an externally stored impulse response estimate.
void AdaptiveFirFilter::ConstrainAndUpdateImpulseResponse(
std::vector<float>* impulse_response) {
RTC_DCHECK_EQ(GetTimeDomainLength(max_size_partitions_),
impulse_response->capacity());
impulse_response->resize(GetTimeDomainLength(current_size_partitions_));
std::array<float, kFftLength> h;
impulse_response->resize(GetTimeDomainLength(current_size_partitions_));
std::fill(
impulse_response->begin() + partition_to_constrain_ * kFftLengthBy2,
impulse_response->begin() + (partition_to_constrain_ + 1) * kFftLengthBy2,
0.f);
for (size_t ch = 0; ch < num_render_channels_; ++ch) {
fft_.Ifft(H_[partition_to_constrain_][ch], &h);
static constexpr float kScale = 1.0f / kFftLengthBy2;
std::for_each(h.begin(), h.begin() + kFftLengthBy2,
[](float& a) { a *= kScale; });
std::fill(h.begin() + kFftLengthBy2, h.end(), 0.f);
if (ch == 0) {
std::copy(
h.begin(), h.begin() + kFftLengthBy2,
impulse_response->begin() + partition_to_constrain_ * kFftLengthBy2);
} else {
for (size_t k = 0, j = partition_to_constrain_ * kFftLengthBy2;
k < kFftLengthBy2; ++k, ++j) {
if (fabsf((*impulse_response)[j]) < fabsf(h[k])) {
(*impulse_response)[j] = h[k];
}
}
}
fft_.Fft(&h, &H_[partition_to_constrain_][ch]);
}
partition_to_constrain_ =
partition_to_constrain_ < (current_size_partitions_ - 1)
? partition_to_constrain_ + 1
: 0;
}
// Constrains the a partiton of the frequency domain filter to be limited in
// time via setting the relevant time-domain coefficients to zero.
void AdaptiveFirFilter::Constrain() {
std::array<float, kFftLength> h;
for (size_t ch = 0; ch < num_render_channels_; ++ch) {
fft_.Ifft(H_[partition_to_constrain_][ch], &h);
static constexpr float kScale = 1.0f / kFftLengthBy2;
std::for_each(h.begin(), h.begin() + kFftLengthBy2,
[](float& a) { a *= kScale; });
std::fill(h.begin() + kFftLengthBy2, h.end(), 0.f);
fft_.Fft(&h, &H_[partition_to_constrain_][ch]);
}
partition_to_constrain_ =
partition_to_constrain_ < (current_size_partitions_ - 1)
? partition_to_constrain_ + 1
: 0;
}
void AdaptiveFirFilter::ScaleFilter(float factor) {
for (auto& H_p : H_) {
for (auto& H_p_ch : H_p) {
for (auto& re : H_p_ch.re) {
re *= factor;
}
for (auto& im : H_p_ch.im) {
im *= factor;
}
}
}
}
// Set the filter coefficients.
void AdaptiveFirFilter::SetFilter(size_t num_partitions,
const std::vector<std::vector<FftData>>& H) {
const size_t min_num_partitions =
std::min(current_size_partitions_, num_partitions);
for (size_t p = 0; p < min_num_partitions; ++p) {
RTC_DCHECK_EQ(H_[p].size(), H[p].size());
RTC_DCHECK_EQ(num_render_channels_, H_[p].size());
for (size_t ch = 0; ch < num_render_channels_; ++ch) {
std::copy(H[p][ch].re.begin(), H[p][ch].re.end(), H_[p][ch].re.begin());
std::copy(H[p][ch].im.begin(), H[p][ch].im.end(), H_[p][ch].im.begin());
}
}
}
} // namespace webrtc

191
webrtc/modules/audio_processing/aec3/adaptive_fir_filter.h Normal file
View File

@ -0,0 +1,191 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ADAPTIVE_FIR_FILTER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ADAPTIVE_FIR_FILTER_H_
#include <stddef.h>
#include <array>
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/aec3_fft.h"
#include "modules/audio_processing/aec3/fft_data.h"
#include "modules/audio_processing/aec3/render_buffer.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/system/arch.h"
namespace webrtc {
namespace aec3 {
// Computes and stores the frequency response of the filter.
void ComputeFrequencyResponse(
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2);
#if defined(WEBRTC_HAS_NEON)
void ComputeFrequencyResponse_Neon(
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2);
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
void ComputeFrequencyResponse_Sse2(
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2);
void ComputeFrequencyResponse_Avx2(
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2);
#endif
// Adapts the filter partitions.
void AdaptPartitions(const RenderBuffer& render_buffer,
const FftData& G,
size_t num_partitions,
std::vector<std::vector<FftData>>* H);
#if defined(WEBRTC_HAS_NEON)
void AdaptPartitions_Neon(const RenderBuffer& render_buffer,
const FftData& G,
size_t num_partitions,
std::vector<std::vector<FftData>>* H);
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
void AdaptPartitions_Sse2(const RenderBuffer& render_buffer,
const FftData& G,
size_t num_partitions,
std::vector<std::vector<FftData>>* H);
void AdaptPartitions_Avx2(const RenderBuffer& render_buffer,
const FftData& G,
size_t num_partitions,
std::vector<std::vector<FftData>>* H);
#endif
// Produces the filter output.
void ApplyFilter(const RenderBuffer& render_buffer,
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
FftData* S);
#if defined(WEBRTC_HAS_NEON)
void ApplyFilter_Neon(const RenderBuffer& render_buffer,
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
FftData* S);
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
void ApplyFilter_Sse2(const RenderBuffer& render_buffer,
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
FftData* S);
void ApplyFilter_Avx2(const RenderBuffer& render_buffer,
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
FftData* S);
#endif
} // namespace aec3
// Provides a frequency domain adaptive filter functionality.
class AdaptiveFirFilter {
public:
AdaptiveFirFilter(size_t max_size_partitions,
size_t initial_size_partitions,
size_t size_change_duration_blocks,
size_t num_render_channels,
Aec3Optimization optimization,
ApmDataDumper* data_dumper);
~AdaptiveFirFilter();
AdaptiveFirFilter(const AdaptiveFirFilter&) = delete;
AdaptiveFirFilter& operator=(const AdaptiveFirFilter&) = delete;
// Produces the output of the filter.
void Filter(const RenderBuffer& render_buffer, FftData* S) const;
// Adapts the filter and updates an externally stored impulse response
// estimate.
void Adapt(const RenderBuffer& render_buffer,
const FftData& G,
std::vector<float>* impulse_response);
// Adapts the filter.
void Adapt(const RenderBuffer& render_buffer, const FftData& G);
// Receives reports that known echo path changes have occured and adjusts
// the filter adaptation accordingly.
void HandleEchoPathChange();
// Returns the filter size.
size_t SizePartitions() const { return current_size_partitions_; }
// Sets the filter size.
void SetSizePartitions(size_t size, bool immediate_effect);
// Computes the frequency responses for the filter partitions.
void ComputeFrequencyResponse(
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) const;
// Returns the maximum number of partitions for the filter.
size_t max_filter_size_partitions() const { return max_size_partitions_; }
void DumpFilter(const char* name_frequency_domain) {
for (size_t p = 0; p < max_size_partitions_; ++p) {
data_dumper_->DumpRaw(name_frequency_domain, H_[p][0].re);
data_dumper_->DumpRaw(name_frequency_domain, H_[p][0].im);
}
}
// Scale the filter impulse response and spectrum by a factor.
void ScaleFilter(float factor);
// Set the filter coefficients.
void SetFilter(size_t num_partitions,
const std::vector<std::vector<FftData>>& H);
// Gets the filter coefficients.
const std::vector<std::vector<FftData>>& GetFilter() const { return H_; }
private:
// Adapts the filter and updates the filter size.
void AdaptAndUpdateSize(const RenderBuffer& render_buffer, const FftData& G);
// Constrain the filter partitions in a cyclic manner.
void Constrain();
// Constrains the filter in a cyclic manner and updates the corresponding
// values in the supplied impulse response.
void ConstrainAndUpdateImpulseResponse(std::vector<float>* impulse_response);
// Gradually Updates the current filter size towards the target size.
void UpdateSize();
ApmDataDumper* const data_dumper_;
const Aec3Fft fft_;
const Aec3Optimization optimization_;
const size_t num_render_channels_;
const size_t max_size_partitions_;
const int size_change_duration_blocks_;
float one_by_size_change_duration_blocks_;
size_t current_size_partitions_;
size_t target_size_partitions_;
size_t old_target_size_partitions_;
int size_change_counter_ = 0;
std::vector<std::vector<FftData>> H_;
size_t partition_to_constrain_ = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ADAPTIVE_FIR_FILTER_H_

187
webrtc/modules/audio_processing/aec3/adaptive_fir_filter_avx2.cc Normal file
View File

@ -0,0 +1,187 @@
/*
* Copyright (c) 2020 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/adaptive_fir_filter.h"
#include <immintrin.h>
#include "rtc_base/checks.h"
namespace webrtc {
namespace aec3 {
// Computes and stores the frequency response of the filter.
void ComputeFrequencyResponse_Avx2(
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) {
for (auto& H2_ch : *H2) {
H2_ch.fill(0.f);
}
const size_t num_render_channels = H[0].size();
RTC_DCHECK_EQ(H.size(), H2->capacity());
for (size_t p = 0; p < num_partitions; ++p) {
RTC_DCHECK_EQ(kFftLengthBy2Plus1, (*H2)[p].size());
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t j = 0; j < kFftLengthBy2; j += 8) {
__m256 re = _mm256_loadu_ps(&H[p][ch].re[j]);
__m256 re2 = _mm256_mul_ps(re, re);
__m256 im = _mm256_loadu_ps(&H[p][ch].im[j]);
re2 = _mm256_fmadd_ps(im, im, re2);
__m256 H2_k_j = _mm256_loadu_ps(&(*H2)[p][j]);
H2_k_j = _mm256_max_ps(H2_k_j, re2);
_mm256_storeu_ps(&(*H2)[p][j], H2_k_j);
}
float H2_new = H[p][ch].re[kFftLengthBy2] * H[p][ch].re[kFftLengthBy2] +
H[p][ch].im[kFftLengthBy2] * H[p][ch].im[kFftLengthBy2];
(*H2)[p][kFftLengthBy2] = std::max((*H2)[p][kFftLengthBy2], H2_new);
}
}
}
// Adapts the filter partitions.
void AdaptPartitions_Avx2(const RenderBuffer& render_buffer,
const FftData& G,
size_t num_partitions,
std::vector<std::vector<FftData>>* H) {
rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
render_buffer.GetFftBuffer();
const size_t num_render_channels = render_buffer_data[0].size();
const size_t lim1 = std::min(
render_buffer_data.size() - render_buffer.Position(), num_partitions);
const size_t lim2 = num_partitions;
constexpr size_t kNumEightBinBands = kFftLengthBy2 / 8;
size_t X_partition = render_buffer.Position();
size_t limit = lim1;
size_t p = 0;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
FftData& H_p_ch = (*H)[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
for (size_t k = 0, n = 0; n < kNumEightBinBands; ++n, k += 8) {
const __m256 G_re = _mm256_loadu_ps(&G.re[k]);
const __m256 G_im = _mm256_loadu_ps(&G.im[k]);
const __m256 X_re = _mm256_loadu_ps(&X.re[k]);
const __m256 X_im = _mm256_loadu_ps(&X.im[k]);
const __m256 H_re = _mm256_loadu_ps(&H_p_ch.re[k]);
const __m256 H_im = _mm256_loadu_ps(&H_p_ch.im[k]);
const __m256 a = _mm256_mul_ps(X_re, G_re);
const __m256 b = _mm256_mul_ps(X_im, G_im);
const __m256 c = _mm256_mul_ps(X_re, G_im);
const __m256 d = _mm256_mul_ps(X_im, G_re);
const __m256 e = _mm256_add_ps(a, b);
const __m256 f = _mm256_sub_ps(c, d);
const __m256 g = _mm256_add_ps(H_re, e);
const __m256 h = _mm256_add_ps(H_im, f);
_mm256_storeu_ps(&H_p_ch.re[k], g);
_mm256_storeu_ps(&H_p_ch.im[k], h);
}
}
}
X_partition = 0;
limit = lim2;
} while (p < lim2);
X_partition = render_buffer.Position();
limit = lim1;
p = 0;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
FftData& H_p_ch = (*H)[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
H_p_ch.re[kFftLengthBy2] += X.re[kFftLengthBy2] * G.re[kFftLengthBy2] +
X.im[kFftLengthBy2] * G.im[kFftLengthBy2];
H_p_ch.im[kFftLengthBy2] += X.re[kFftLengthBy2] * G.im[kFftLengthBy2] -
X.im[kFftLengthBy2] * G.re[kFftLengthBy2];
}
}
X_partition = 0;
limit = lim2;
} while (p < lim2);
}
// Produces the filter output (AVX2 variant).
void ApplyFilter_Avx2(const RenderBuffer& render_buffer,
size_t num_partitions,
const std::vector<std::vector<FftData>>& H,
FftData* S) {
RTC_DCHECK_GE(H.size(), H.size() - 1);
S->re.fill(0.f);
S->im.fill(0.f);
rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
render_buffer.GetFftBuffer();
const size_t num_render_channels = render_buffer_data[0].size();
const size_t lim1 = std::min(
render_buffer_data.size() - render_buffer.Position(), num_partitions);
const size_t lim2 = num_partitions;
constexpr size_t kNumEightBinBands = kFftLengthBy2 / 8;
size_t X_partition = render_buffer.Position();
size_t p = 0;
size_t limit = lim1;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
const FftData& H_p_ch = H[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
for (size_t k = 0, n = 0; n < kNumEightBinBands; ++n, k += 8) {
const __m256 X_re = _mm256_loadu_ps(&X.re[k]);
const __m256 X_im = _mm256_loadu_ps(&X.im[k]);
const __m256 H_re = _mm256_loadu_ps(&H_p_ch.re[k]);
const __m256 H_im = _mm256_loadu_ps(&H_p_ch.im[k]);
const __m256 S_re = _mm256_loadu_ps(&S->re[k]);
const __m256 S_im = _mm256_loadu_ps(&S->im[k]);
const __m256 a = _mm256_mul_ps(X_re, H_re);
const __m256 b = _mm256_mul_ps(X_im, H_im);
const __m256 c = _mm256_mul_ps(X_re, H_im);
const __m256 d = _mm256_mul_ps(X_im, H_re);
const __m256 e = _mm256_sub_ps(a, b);
const __m256 f = _mm256_add_ps(c, d);
const __m256 g = _mm256_add_ps(S_re, e);
const __m256 h = _mm256_add_ps(S_im, f);
_mm256_storeu_ps(&S->re[k], g);
_mm256_storeu_ps(&S->im[k], h);
}
}
}
limit = lim2;
X_partition = 0;
} while (p < lim2);
X_partition = render_buffer.Position();
p = 0;
limit = lim1;
do {
for (; p < limit; ++p, ++X_partition) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
const FftData& H_p_ch = H[p][ch];
const FftData& X = render_buffer_data[X_partition][ch];
S->re[kFftLengthBy2] += X.re[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2] -
X.im[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2];
S->im[kFftLengthBy2] += X.re[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2] +
X.im[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2];
}
}
limit = lim2;
X_partition = 0;
} while (p < lim2);
}
} // namespace aec3
} // namespace webrtc

102
webrtc/modules/audio_processing/aec3/adaptive_fir_filter_erl.cc Normal file
View File

@ -0,0 +1,102 @@
/*
* Copyright (c) 2019 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/adaptive_fir_filter_erl.h"
#include <algorithm>
#include <functional>
#if defined(WEBRTC_HAS_NEON)
#include <arm_neon.h>
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
#include <emmintrin.h>
#endif
namespace webrtc {
namespace aec3 {
// Computes and stores the echo return loss estimate of the filter, which is the
// sum of the partition frequency responses.
void ErlComputer(const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl) {
std::fill(erl.begin(), erl.end(), 0.f);
for (auto& H2_j : H2) {
std::transform(H2_j.begin(), H2_j.end(), erl.begin(), erl.begin(),
std::plus<float>());
}
}
#if defined(WEBRTC_HAS_NEON)
// Computes and stores the echo return loss estimate of the filter, which is the
// sum of the partition frequency responses.
void ErlComputer_NEON(
const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl) {
std::fill(erl.begin(), erl.end(), 0.f);
for (auto& H2_j : H2) {
for (size_t k = 0; k < kFftLengthBy2; k += 4) {
const float32x4_t H2_j_k = vld1q_f32(&H2_j[k]);
float32x4_t erl_k = vld1q_f32(&erl[k]);
erl_k = vaddq_f32(erl_k, H2_j_k);
vst1q_f32(&erl[k], erl_k);
}
erl[kFftLengthBy2] += H2_j[kFftLengthBy2];
}
}
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
// Computes and stores the echo return loss estimate of the filter, which is the
// sum of the partition frequency responses.
void ErlComputer_SSE2(
const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl) {
std::fill(erl.begin(), erl.end(), 0.f);
for (auto& H2_j : H2) {
for (size_t k = 0; k < kFftLengthBy2; k += 4) {
const __m128 H2_j_k = _mm_loadu_ps(&H2_j[k]);
__m128 erl_k = _mm_loadu_ps(&erl[k]);
erl_k = _mm_add_ps(erl_k, H2_j_k);
_mm_storeu_ps(&erl[k], erl_k);
}
erl[kFftLengthBy2] += H2_j[kFftLengthBy2];
}
}
#endif
} // namespace aec3
void ComputeErl(const Aec3Optimization& optimization,
const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl) {
RTC_DCHECK_EQ(kFftLengthBy2Plus1, erl.size());
// Update the frequency response and echo return loss for the filter.
switch (optimization) {
#if defined(WEBRTC_ARCH_X86_FAMILY)
case Aec3Optimization::kSse2:
aec3::ErlComputer_SSE2(H2, erl);
break;
case Aec3Optimization::kAvx2:
aec3::ErlComputer_AVX2(H2, erl);
break;
#endif
#if defined(WEBRTC_HAS_NEON)
case Aec3Optimization::kNeon:
aec3::ErlComputer_NEON(H2, erl);
break;
#endif
default:
aec3::ErlComputer(H2, erl);
}
}
} // namespace webrtc

54
webrtc/modules/audio_processing/aec3/adaptive_fir_filter_erl.h Normal file
View File

@ -0,0 +1,54 @@
/*
* Copyright (c) 2019 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ADAPTIVE_FIR_FILTER_ERL_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ADAPTIVE_FIR_FILTER_ERL_H_
#include <stddef.h>
#include <array>
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/system/arch.h"
namespace webrtc {
namespace aec3 {
// Computes and stores the echo return loss estimate of the filter, which is the
// sum of the partition frequency responses.
void ErlComputer(const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl);
#if defined(WEBRTC_HAS_NEON)
void ErlComputer_NEON(
const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl);
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
void ErlComputer_SSE2(
const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl);
void ErlComputer_AVX2(
const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl);
#endif
} // namespace aec3
// Computes the echo return loss based on a frequency response.
void ComputeErl(const Aec3Optimization& optimization,
const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl);
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ADAPTIVE_FIR_FILTER_ERL_H_

37
webrtc/modules/audio_processing/aec3/adaptive_fir_filter_erl_avx2.cc Normal file
View File

@ -0,0 +1,37 @@
/*
* Copyright (c) 2020 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/adaptive_fir_filter_erl.h"
#include <immintrin.h>
namespace webrtc {
namespace aec3 {
// Computes and stores the echo return loss estimate of the filter, which is the
// sum of the partition frequency responses.
void ErlComputer_AVX2(
const std::vector<std::array<float, kFftLengthBy2Plus1>>& H2,
rtc::ArrayView<float> erl) {
std::fill(erl.begin(), erl.end(), 0.f);
for (auto& H2_j : H2) {
for (size_t k = 0; k < kFftLengthBy2; k += 8) {
const __m256 H2_j_k = _mm256_loadu_ps(&H2_j[k]);
__m256 erl_k = _mm256_loadu_ps(&erl[k]);
erl_k = _mm256_add_ps(erl_k, H2_j_k);
_mm256_storeu_ps(&erl[k], erl_k);
}
erl[kFftLengthBy2] += H2_j[kFftLengthBy2];
}
}
} // namespace aec3
} // namespace webrtc

58
webrtc/modules/audio_processing/aec3/aec3_common.cc Normal file
View File

@ -0,0 +1,58 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/aec3_common.h"
#include <stdint.h>
#include "rtc_base/checks.h"
#include "rtc_base/system/arch.h"
#include "system_wrappers/include/cpu_features_wrapper.h"
namespace webrtc {
Aec3Optimization DetectOptimization() {
#if defined(WEBRTC_ARCH_X86_FAMILY)
if (GetCPUInfo(kAVX2) != 0) {
return Aec3Optimization::kAvx2;
} else if (GetCPUInfo(kSSE2) != 0) {
return Aec3Optimization::kSse2;
}
#endif
#if defined(WEBRTC_HAS_NEON)
return Aec3Optimization::kNeon;
#endif
return Aec3Optimization::kNone;
}
float FastApproxLog2f(const float in) {
RTC_DCHECK_GT(in, .0f);
// Read and interpret float as uint32_t and then cast to float.
// This is done to extract the exponent (bits 30 - 23).
// "Right shift" of the exponent is then performed by multiplying
// with the constant (1/2^23). Finally, we subtract a constant to
// remove the bias (https://en.wikipedia.org/wiki/Exponent_bias).
union {
float dummy;
uint32_t a;
} x = {in};
float out = x.a;
out *= 1.1920929e-7f; // 1/2^23
out -= 126.942695f; // Remove bias.
return out;
}
float Log2TodB(const float in_log2) {
return 3.0102999566398121 * in_log2;
}
} // namespace webrtc

114
webrtc/modules/audio_processing/aec3/aec3_common.h Normal file
View File

@ -0,0 +1,114 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_AEC3_COMMON_H_
#define MODULES_AUDIO_PROCESSING_AEC3_AEC3_COMMON_H_
#include <stddef.h>
namespace webrtc {
#ifdef _MSC_VER /* visual c++ */
#define ALIGN16_BEG __declspec(align(16))
#define ALIGN16_END
#else /* gcc or icc */
#define ALIGN16_BEG
#define ALIGN16_END __attribute__((aligned(16)))
#endif
enum class Aec3Optimization { kNone, kSse2, kAvx2, kNeon };
constexpr int kNumBlocksPerSecond = 250;
constexpr int kMetricsReportingIntervalBlocks = 10 * kNumBlocksPerSecond;
constexpr int kMetricsComputationBlocks = 7;
constexpr int kMetricsCollectionBlocks =
kMetricsReportingIntervalBlocks - kMetricsComputationBlocks;
constexpr size_t kFftLengthBy2 = 64;
constexpr size_t kFftLengthBy2Plus1 = kFftLengthBy2 + 1;
constexpr size_t kFftLengthBy2Minus1 = kFftLengthBy2 - 1;
constexpr size_t kFftLength = 2 * kFftLengthBy2;
constexpr size_t kFftLengthBy2Log2 = 6;
constexpr int kRenderTransferQueueSizeFrames = 100;
constexpr size_t kMaxNumBands = 3;
constexpr size_t kFrameSize = 160;
constexpr size_t kSubFrameLength = kFrameSize / 2;
constexpr size_t kBlockSize = kFftLengthBy2;
constexpr size_t kBlockSizeLog2 = kFftLengthBy2Log2;
constexpr size_t kExtendedBlockSize = 2 * kFftLengthBy2;
constexpr size_t kMatchedFilterWindowSizeSubBlocks = 32;
constexpr size_t kMatchedFilterAlignmentShiftSizeSubBlocks =
kMatchedFilterWindowSizeSubBlocks * 3 / 4;
// TODO(peah): Integrate this with how it is done inside audio_processing_impl.
constexpr size_t NumBandsForRate(int sample_rate_hz) {
return static_cast<size_t>(sample_rate_hz / 16000);
}
constexpr bool ValidFullBandRate(int sample_rate_hz) {
return sample_rate_hz == 16000 || sample_rate_hz == 32000 ||
sample_rate_hz == 48000;
}
constexpr int GetTimeDomainLength(int filter_length_blocks) {
return filter_length_blocks * kFftLengthBy2;
}
constexpr size_t GetDownSampledBufferSize(size_t down_sampling_factor,
size_t num_matched_filters) {
return kBlockSize / down_sampling_factor *
(kMatchedFilterAlignmentShiftSizeSubBlocks * num_matched_filters +
kMatchedFilterWindowSizeSubBlocks + 1);
}
constexpr size_t GetRenderDelayBufferSize(size_t down_sampling_factor,
size_t num_matched_filters,
size_t filter_length_blocks) {
return GetDownSampledBufferSize(down_sampling_factor, num_matched_filters) /
(kBlockSize / down_sampling_factor) +
filter_length_blocks + 1;
}
// Detects what kind of optimizations to use for the code.
Aec3Optimization DetectOptimization();
// Computes the log2 of the input in a fast an approximate manner.
float FastApproxLog2f(const float in);
// Returns dB from a power quantity expressed in log2.
float Log2TodB(const float in_log2);
static_assert(1 << kBlockSizeLog2 == kBlockSize,
"Proper number of shifts for blocksize");
static_assert(1 << kFftLengthBy2Log2 == kFftLengthBy2,
"Proper number of shifts for the fft length");
static_assert(1 == NumBandsForRate(16000), "Number of bands for 16 kHz");
static_assert(2 == NumBandsForRate(32000), "Number of bands for 32 kHz");
static_assert(3 == NumBandsForRate(48000), "Number of bands for 48 kHz");
static_assert(ValidFullBandRate(16000),
"Test that 16 kHz is a valid sample rate");
static_assert(ValidFullBandRate(32000),
"Test that 32 kHz is a valid sample rate");
static_assert(ValidFullBandRate(48000),
"Test that 48 kHz is a valid sample rate");
static_assert(!ValidFullBandRate(8001),
"Test that 8001 Hz is not a valid sample rate");
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_AEC3_COMMON_H_

144
webrtc/modules/audio_processing/aec3/aec3_fft.cc Normal file
View File

@ -0,0 +1,144 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/aec3_fft.h"
#include <algorithm>
#include <functional>
#include <iterator>
#include "rtc_base/checks.h"
#include "system_wrappers/include/cpu_features_wrapper.h"
namespace webrtc {
namespace {
const float kHanning64[kFftLengthBy2] = {
0.f, 0.00248461f, 0.00991376f, 0.0222136f, 0.03926189f,
0.06088921f, 0.08688061f, 0.11697778f, 0.15088159f, 0.1882551f,
0.22872687f, 0.27189467f, 0.31732949f, 0.36457977f, 0.41317591f,
0.46263495f, 0.51246535f, 0.56217185f, 0.61126047f, 0.65924333f,
0.70564355f, 0.75f, 0.79187184f, 0.83084292f, 0.86652594f,
0.89856625f, 0.92664544f, 0.95048443f, 0.96984631f, 0.98453864f,
0.99441541f, 0.99937846f, 0.99937846f, 0.99441541f, 0.98453864f,
0.96984631f, 0.95048443f, 0.92664544f, 0.89856625f, 0.86652594f,
0.83084292f, 0.79187184f, 0.75f, 0.70564355f, 0.65924333f,
0.61126047f, 0.56217185f, 0.51246535f, 0.46263495f, 0.41317591f,
0.36457977f, 0.31732949f, 0.27189467f, 0.22872687f, 0.1882551f,
0.15088159f, 0.11697778f, 0.08688061f, 0.06088921f, 0.03926189f,
0.0222136f, 0.00991376f, 0.00248461f, 0.f};
// Hanning window from Matlab command win = sqrt(hanning(128)).
const float kSqrtHanning128[kFftLength] = {
0.00000000000000f, 0.02454122852291f, 0.04906767432742f, 0.07356456359967f,
0.09801714032956f, 0.12241067519922f, 0.14673047445536f, 0.17096188876030f,
0.19509032201613f, 0.21910124015687f, 0.24298017990326f, 0.26671275747490f,
0.29028467725446f, 0.31368174039889f, 0.33688985339222f, 0.35989503653499f,
0.38268343236509f, 0.40524131400499f, 0.42755509343028f, 0.44961132965461f,
0.47139673682600f, 0.49289819222978f, 0.51410274419322f, 0.53499761988710f,
0.55557023301960f, 0.57580819141785f, 0.59569930449243f, 0.61523159058063f,
0.63439328416365f, 0.65317284295378f, 0.67155895484702f, 0.68954054473707f,
0.70710678118655f, 0.72424708295147f, 0.74095112535496f, 0.75720884650648f,
0.77301045336274f, 0.78834642762661f, 0.80320753148064f, 0.81758481315158f,
0.83146961230255f, 0.84485356524971f, 0.85772861000027f, 0.87008699110871f,
0.88192126434835f, 0.89322430119552f, 0.90398929312344f, 0.91420975570353f,
0.92387953251129f, 0.93299279883474f, 0.94154406518302f, 0.94952818059304f,
0.95694033573221f, 0.96377606579544f, 0.97003125319454f, 0.97570213003853f,
0.98078528040323f, 0.98527764238894f, 0.98917650996478f, 0.99247953459871f,
0.99518472667220f, 0.99729045667869f, 0.99879545620517f, 0.99969881869620f,
1.00000000000000f, 0.99969881869620f, 0.99879545620517f, 0.99729045667869f,
0.99518472667220f, 0.99247953459871f, 0.98917650996478f, 0.98527764238894f,
0.98078528040323f, 0.97570213003853f, 0.97003125319454f, 0.96377606579544f,
0.95694033573221f, 0.94952818059304f, 0.94154406518302f, 0.93299279883474f,
0.92387953251129f, 0.91420975570353f, 0.90398929312344f, 0.89322430119552f,
0.88192126434835f, 0.87008699110871f, 0.85772861000027f, 0.84485356524971f,
0.83146961230255f, 0.81758481315158f, 0.80320753148064f, 0.78834642762661f,
0.77301045336274f, 0.75720884650648f, 0.74095112535496f, 0.72424708295147f,
0.70710678118655f, 0.68954054473707f, 0.67155895484702f, 0.65317284295378f,
0.63439328416365f, 0.61523159058063f, 0.59569930449243f, 0.57580819141785f,
0.55557023301960f, 0.53499761988710f, 0.51410274419322f, 0.49289819222978f,
0.47139673682600f, 0.44961132965461f, 0.42755509343028f, 0.40524131400499f,
0.38268343236509f, 0.35989503653499f, 0.33688985339222f, 0.31368174039889f,
0.29028467725446f, 0.26671275747490f, 0.24298017990326f, 0.21910124015687f,
0.19509032201613f, 0.17096188876030f, 0.14673047445536f, 0.12241067519922f,
0.09801714032956f, 0.07356456359967f, 0.04906767432742f, 0.02454122852291f};
bool IsSse2Available() {
#if defined(WEBRTC_ARCH_X86_FAMILY)
return GetCPUInfo(kSSE2) != 0;
#else
return false;
#endif
}
} // namespace
Aec3Fft::Aec3Fft() : ooura_fft_(IsSse2Available()) {}
// TODO(peah): Change x to be std::array once the rest of the code allows this.
void Aec3Fft::ZeroPaddedFft(rtc::ArrayView<const float> x,
Window window,
FftData* X) const {
RTC_DCHECK(X);
RTC_DCHECK_EQ(kFftLengthBy2, x.size());
std::array<float, kFftLength> fft;
std::fill(fft.begin(), fft.begin() + kFftLengthBy2, 0.f);
switch (window) {
case Window::kRectangular:
std::copy(x.begin(), x.end(), fft.begin() + kFftLengthBy2);
break;
case Window::kHanning:
std::transform(x.begin(), x.end(), std::begin(kHanning64),
fft.begin() + kFftLengthBy2,
[](float a, float b) { return a * b; });
break;
case Window::kSqrtHanning:
RTC_NOTREACHED();
break;
default:
RTC_NOTREACHED();
}
Fft(&fft, X);
}
void Aec3Fft::PaddedFft(rtc::ArrayView<const float> x,
rtc::ArrayView<const float> x_old,
Window window,
FftData* X) const {
RTC_DCHECK(X);
RTC_DCHECK_EQ(kFftLengthBy2, x.size());
RTC_DCHECK_EQ(kFftLengthBy2, x_old.size());
std::array<float, kFftLength> fft;
switch (window) {
case Window::kRectangular:
std::copy(x_old.begin(), x_old.end(), fft.begin());
std::copy(x.begin(), x.end(), fft.begin() + x_old.size());
break;
case Window::kHanning:
RTC_NOTREACHED();
break;
case Window::kSqrtHanning:
std::transform(x_old.begin(), x_old.end(), std::begin(kSqrtHanning128),
fft.begin(), std::multiplies<float>());
std::transform(x.begin(), x.end(),
std::begin(kSqrtHanning128) + x_old.size(),
fft.begin() + x_old.size(), std::multiplies<float>());
break;
default:
RTC_NOTREACHED();
}
Fft(&fft, X);
}
} // namespace webrtc

75
webrtc/modules/audio_processing/aec3/aec3_fft.h Normal file
View File

@ -0,0 +1,75 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_AEC3_FFT_H_
#define MODULES_AUDIO_PROCESSING_AEC3_AEC3_FFT_H_
#include <array>
#include "api/array_view.h"
#include "common_audio/third_party/ooura/fft_size_128/ooura_fft.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/fft_data.h"
#include "rtc_base/checks.h"
#include "rtc_base/constructor_magic.h"
namespace webrtc {
// Wrapper class that provides 128 point real valued FFT functionality with the
// FftData type.
class Aec3Fft {
public:
enum class Window { kRectangular, kHanning, kSqrtHanning };
Aec3Fft();
// Computes the FFT. Note that both the input and output are modified.
void Fft(std::array<float, kFftLength>* x, FftData* X) const {
RTC_DCHECK(x);
RTC_DCHECK(X);
ooura_fft_.Fft(x->data());
X->CopyFromPackedArray(*x);
}
// Computes the inverse Fft.
void Ifft(const FftData& X, std::array<float, kFftLength>* x) const {
RTC_DCHECK(x);
X.CopyToPackedArray(x);
ooura_fft_.InverseFft(x->data());
}
// Windows the input using a Hanning window, and then adds padding of
// kFftLengthBy2 initial zeros before computing the Fft.
void ZeroPaddedFft(rtc::ArrayView<const float> x,
Window window,
FftData* X) const;
// Concatenates the kFftLengthBy2 values long x and x_old before computing the
// Fft. After that, x is copied to x_old.
void PaddedFft(rtc::ArrayView<const float> x,
rtc::ArrayView<const float> x_old,
FftData* X) const {
PaddedFft(x, x_old, Window::kRectangular, X);
}
// Padded Fft using a time-domain window.
void PaddedFft(rtc::ArrayView<const float> x,
rtc::ArrayView<const float> x_old,
Window window,
FftData* X) const;
private:
const OouraFft ooura_fft_;
RTC_DISALLOW_COPY_AND_ASSIGN(Aec3Fft);
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_AEC3_FFT_H_

477
webrtc/modules/audio_processing/aec3/aec_state.cc Normal file
View File

@ -0,0 +1,477 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/aec_state.h"
#include <math.h>
#include <algorithm>
#include <numeric>
#include <vector>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomic_ops.h"
#include "rtc_base/checks.h"
#include "system_wrappers/include/field_trial.h"
namespace webrtc {
namespace {
bool DeactivateInitialStateResetAtEchoPathChange() {
return field_trial::IsEnabled(
"WebRTC-Aec3DeactivateInitialStateResetKillSwitch");
}
bool FullResetAtEchoPathChange() {
return !field_trial::IsEnabled("WebRTC-Aec3AecStateFullResetKillSwitch");
}
bool SubtractorAnalyzerResetAtEchoPathChange() {
return !field_trial::IsEnabled(
"WebRTC-Aec3AecStateSubtractorAnalyzerResetKillSwitch");
}
void ComputeAvgRenderReverb(
const SpectrumBuffer& spectrum_buffer,
int delay_blocks,
float reverb_decay,
ReverbModel* reverb_model,
rtc::ArrayView<float, kFftLengthBy2Plus1> reverb_power_spectrum) {
RTC_DCHECK(reverb_model);
const size_t num_render_channels = spectrum_buffer.buffer[0].size();
int idx_at_delay =
spectrum_buffer.OffsetIndex(spectrum_buffer.read, delay_blocks);
int idx_past = spectrum_buffer.IncIndex(idx_at_delay);
std::array<float, kFftLengthBy2Plus1> X2_data;
rtc::ArrayView<const float> X2;
if (num_render_channels > 1) {
auto average_channels =
[](size_t num_render_channels,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
spectrum_band_0,
rtc::ArrayView<float, kFftLengthBy2Plus1> render_power) {
std::fill(render_power.begin(), render_power.end(), 0.f);
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
render_power[k] += spectrum_band_0[ch][k];
}
}
const float normalizer = 1.f / num_render_channels;
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
render_power[k] *= normalizer;
}
};
average_channels(num_render_channels, spectrum_buffer.buffer[idx_past],
X2_data);
reverb_model->UpdateReverbNoFreqShaping(
X2_data, /*power_spectrum_scaling=*/1.0f, reverb_decay);
average_channels(num_render_channels, spectrum_buffer.buffer[idx_at_delay],
X2_data);
X2 = X2_data;
} else {
reverb_model->UpdateReverbNoFreqShaping(
spectrum_buffer.buffer[idx_past][/*channel=*/0],
/*power_spectrum_scaling=*/1.0f, reverb_decay);
X2 = spectrum_buffer.buffer[idx_at_delay][/*channel=*/0];
}
rtc::ArrayView<const float, kFftLengthBy2Plus1> reverb_power =
reverb_model->reverb();
for (size_t k = 0; k < X2.size(); ++k) {
reverb_power_spectrum[k] = X2[k] + reverb_power[k];
}
}
} // namespace
int AecState::instance_count_ = 0;
void AecState::GetResidualEchoScaling(
rtc::ArrayView<float> residual_scaling) const {
bool filter_has_had_time_to_converge;
if (config_.filter.conservative_initial_phase) {
filter_has_had_time_to_converge =
strong_not_saturated_render_blocks_ >= 1.5f * kNumBlocksPerSecond;
} else {
filter_has_had_time_to_converge =
strong_not_saturated_render_blocks_ >= 0.8f * kNumBlocksPerSecond;
}
echo_audibility_.GetResidualEchoScaling(filter_has_had_time_to_converge,
residual_scaling);
}
absl::optional<float> AecState::ErleUncertainty() const {
if (SaturatedEcho()) {
return 1.f;
}
return absl::nullopt;
}
AecState::AecState(const EchoCanceller3Config& config,
size_t num_capture_channels)
: data_dumper_(
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
config_(config),
num_capture_channels_(num_capture_channels),
deactivate_initial_state_reset_at_echo_path_change_(
DeactivateInitialStateResetAtEchoPathChange()),
full_reset_at_echo_path_change_(FullResetAtEchoPathChange()),
subtractor_analyzer_reset_at_echo_path_change_(
SubtractorAnalyzerResetAtEchoPathChange()),
initial_state_(config_),
delay_state_(config_, num_capture_channels_),
transparent_state_(TransparentMode::Create(config_)),
filter_quality_state_(config_, num_capture_channels_),
erl_estimator_(2 * kNumBlocksPerSecond),
erle_estimator_(2 * kNumBlocksPerSecond, config_, num_capture_channels_),
filter_analyzer_(config_, num_capture_channels_),
echo_audibility_(
config_.echo_audibility.use_stationarity_properties_at_init),
reverb_model_estimator_(config_, num_capture_channels_),
subtractor_output_analyzer_(num_capture_channels_) {}
AecState::~AecState() = default;
void AecState::HandleEchoPathChange(
const EchoPathVariability& echo_path_variability) {
const auto full_reset = [&]() {
filter_analyzer_.Reset();
capture_signal_saturation_ = false;
strong_not_saturated_render_blocks_ = 0;
blocks_with_active_render_ = 0;
if (!deactivate_initial_state_reset_at_echo_path_change_) {
initial_state_.Reset();
}
if (transparent_state_) {
transparent_state_->Reset();
}
erle_estimator_.Reset(true);
erl_estimator_.Reset();
filter_quality_state_.Reset();
};
// TODO(peah): Refine the reset scheme according to the type of gain and
// delay adjustment.
if (full_reset_at_echo_path_change_ &&
echo_path_variability.delay_change !=
EchoPathVariability::DelayAdjustment::kNone) {
full_reset();
} else if (echo_path_variability.gain_change) {
erle_estimator_.Reset(false);
}
if (subtractor_analyzer_reset_at_echo_path_change_) {
subtractor_output_analyzer_.HandleEchoPathChange();
}
}
void AecState::Update(
const absl::optional<DelayEstimate>& external_delay,
rtc::ArrayView<const std::vector<std::array<float, kFftLengthBy2Plus1>>>
adaptive_filter_frequency_responses,
rtc::ArrayView<const std::vector<float>> adaptive_filter_impulse_responses,
const RenderBuffer& render_buffer,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> E2_refined,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> Y2,
rtc::ArrayView<const SubtractorOutput> subtractor_output) {
RTC_DCHECK_EQ(num_capture_channels_, Y2.size());
RTC_DCHECK_EQ(num_capture_channels_, subtractor_output.size());
RTC_DCHECK_EQ(num_capture_channels_,
adaptive_filter_frequency_responses.size());
RTC_DCHECK_EQ(num_capture_channels_,
adaptive_filter_impulse_responses.size());
// Analyze the filter outputs and filters.
bool any_filter_converged;
bool all_filters_diverged;
subtractor_output_analyzer_.Update(subtractor_output, &any_filter_converged,
&all_filters_diverged);
bool any_filter_consistent;
float max_echo_path_gain;
filter_analyzer_.Update(adaptive_filter_impulse_responses, render_buffer,
&any_filter_consistent, &max_echo_path_gain);
// Estimate the direct path delay of the filter.
if (config_.filter.use_linear_filter) {
delay_state_.Update(filter_analyzer_.FilterDelaysBlocks(), external_delay,
strong_not_saturated_render_blocks_);
}
const std::vector<std::vector<float>>& aligned_render_block =
render_buffer.Block(-delay_state_.MinDirectPathFilterDelay())[0];
// Update render counters.
bool active_render = false;
for (size_t ch = 0; ch < aligned_render_block.size(); ++ch) {
const float render_energy = std::inner_product(
aligned_render_block[ch].begin(), aligned_render_block[ch].end(),
aligned_render_block[ch].begin(), 0.f);
if (render_energy > (config_.render_levels.active_render_limit *
config_.render_levels.active_render_limit) *
kFftLengthBy2) {
active_render = true;
break;
}
}
blocks_with_active_render_ += active_render ? 1 : 0;
strong_not_saturated_render_blocks_ +=
active_render && !SaturatedCapture() ? 1 : 0;
std::array<float, kFftLengthBy2Plus1> avg_render_spectrum_with_reverb;
ComputeAvgRenderReverb(render_buffer.GetSpectrumBuffer(),
delay_state_.MinDirectPathFilterDelay(), ReverbDecay(),
&avg_render_reverb_, avg_render_spectrum_with_reverb);
if (config_.echo_audibility.use_stationarity_properties) {
// Update the echo audibility evaluator.
echo_audibility_.Update(render_buffer, avg_render_reverb_.reverb(),
delay_state_.MinDirectPathFilterDelay(),
delay_state_.ExternalDelayReported());
}
// Update the ERL and ERLE measures.
if (initial_state_.TransitionTriggered()) {
erle_estimator_.Reset(false);
}
erle_estimator_.Update(render_buffer, adaptive_filter_frequency_responses,
avg_render_spectrum_with_reverb, Y2, E2_refined,
subtractor_output_analyzer_.ConvergedFilters());
erl_estimator_.Update(
subtractor_output_analyzer_.ConvergedFilters(),
render_buffer.Spectrum(delay_state_.MinDirectPathFilterDelay()), Y2);
// Detect and flag echo saturation.
if (config_.ep_strength.echo_can_saturate) {
saturation_detector_.Update(aligned_render_block, SaturatedCapture(),
UsableLinearEstimate(), subtractor_output,
max_echo_path_gain);
} else {
RTC_DCHECK(!saturation_detector_.SaturatedEcho());
}
// Update the decision on whether to use the initial state parameter set.
initial_state_.Update(active_render, SaturatedCapture());
// Detect whether the transparent mode should be activated.
if (transparent_state_) {
transparent_state_->Update(delay_state_.MinDirectPathFilterDelay(),
any_filter_consistent, any_filter_converged,
all_filters_diverged, active_render,
SaturatedCapture());
}
// Analyze the quality of the filter.
filter_quality_state_.Update(active_render, TransparentModeActive(),
SaturatedCapture(), external_delay,
any_filter_converged);
// Update the reverb estimate.
const bool stationary_block =
config_.echo_audibility.use_stationarity_properties &&
echo_audibility_.IsBlockStationary();
reverb_model_estimator_.Update(
filter_analyzer_.GetAdjustedFilters(),
adaptive_filter_frequency_responses,
erle_estimator_.GetInstLinearQualityEstimates(),
delay_state_.DirectPathFilterDelays(),
filter_quality_state_.UsableLinearFilterOutputs(), stationary_block);
erle_estimator_.Dump(data_dumper_);
reverb_model_estimator_.Dump(data_dumper_.get());
data_dumper_->DumpRaw("aec3_active_render", active_render);
data_dumper_->DumpRaw("aec3_erl", Erl());
data_dumper_->DumpRaw("aec3_erl_time_domain", ErlTimeDomain());
data_dumper_->DumpRaw("aec3_erle", Erle()[0]);
data_dumper_->DumpRaw("aec3_usable_linear_estimate", UsableLinearEstimate());
data_dumper_->DumpRaw("aec3_transparent_mode", TransparentModeActive());
data_dumper_->DumpRaw("aec3_filter_delay",
filter_analyzer_.MinFilterDelayBlocks());
data_dumper_->DumpRaw("aec3_any_filter_consistent", any_filter_consistent);
data_dumper_->DumpRaw("aec3_initial_state",
initial_state_.InitialStateActive());
data_dumper_->DumpRaw("aec3_capture_saturation", SaturatedCapture());
data_dumper_->DumpRaw("aec3_echo_saturation", SaturatedEcho());
data_dumper_->DumpRaw("aec3_any_filter_converged", any_filter_converged);
data_dumper_->DumpRaw("aec3_all_filters_diverged", all_filters_diverged);
data_dumper_->DumpRaw("aec3_external_delay_avaliable",
external_delay ? 1 : 0);
data_dumper_->DumpRaw("aec3_filter_tail_freq_resp_est",
GetReverbFrequencyResponse());
}
AecState::InitialState::InitialState(const EchoCanceller3Config& config)
: conservative_initial_phase_(config.filter.conservative_initial_phase),
initial_state_seconds_(config.filter.initial_state_seconds) {
Reset();
}
void AecState::InitialState::InitialState::Reset() {
initial_state_ = true;
strong_not_saturated_render_blocks_ = 0;
}
void AecState::InitialState::InitialState::Update(bool active_render,
bool saturated_capture) {
strong_not_saturated_render_blocks_ +=
active_render && !saturated_capture ? 1 : 0;
// Flag whether the initial state is still active.
bool prev_initial_state = initial_state_;
if (conservative_initial_phase_) {
initial_state_ =
strong_not_saturated_render_blocks_ < 5 * kNumBlocksPerSecond;
} else {
initial_state_ = strong_not_saturated_render_blocks_ <
initial_state_seconds_ * kNumBlocksPerSecond;
}
// Flag whether the transition from the initial state has started.
transition_triggered_ = !initial_state_ && prev_initial_state;
}
AecState::FilterDelay::FilterDelay(const EchoCanceller3Config& config,
size_t num_capture_channels)
: delay_headroom_blocks_(config.delay.delay_headroom_samples / kBlockSize),
filter_delays_blocks_(num_capture_channels, delay_headroom_blocks_),
min_filter_delay_(delay_headroom_blocks_) {}
void AecState::FilterDelay::Update(
rtc::ArrayView<const int> analyzer_filter_delay_estimates_blocks,
const absl::optional<DelayEstimate>& external_delay,
size_t blocks_with_proper_filter_adaptation) {
// Update the delay based on the external delay.
if (external_delay &&
(!external_delay_ || external_delay_->delay != external_delay->delay)) {
external_delay_ = external_delay;
external_delay_reported_ = true;
}
// Override the estimated delay if it is not certain that the filter has had
// time to converge.
const bool delay_estimator_may_not_have_converged =
blocks_with_proper_filter_adaptation < 2 * kNumBlocksPerSecond;
if (delay_estimator_may_not_have_converged && external_delay_) {
const int delay_guess = delay_headroom_blocks_;
std::fill(filter_delays_blocks_.begin(), filter_delays_blocks_.end(),
delay_guess);
} else {
RTC_DCHECK_EQ(filter_delays_blocks_.size(),
analyzer_filter_delay_estimates_blocks.size());
std::copy(analyzer_filter_delay_estimates_blocks.begin(),
analyzer_filter_delay_estimates_blocks.end(),
filter_delays_blocks_.begin());
}
min_filter_delay_ = *std::min_element(filter_delays_blocks_.begin(),
filter_delays_blocks_.end());
}
AecState::FilteringQualityAnalyzer::FilteringQualityAnalyzer(
const EchoCanceller3Config& config,
size_t num_capture_channels)
: use_linear_filter_(config.filter.use_linear_filter),
usable_linear_filter_estimates_(num_capture_channels, false) {}
void AecState::FilteringQualityAnalyzer::Reset() {
std::fill(usable_linear_filter_estimates_.begin(),
usable_linear_filter_estimates_.end(), false);
overall_usable_linear_estimates_ = false;
filter_update_blocks_since_reset_ = 0;
}
void AecState::FilteringQualityAnalyzer::Update(
bool active_render,
bool transparent_mode,
bool saturated_capture,
const absl::optional<DelayEstimate>& external_delay,
bool any_filter_converged) {
// Update blocks counter.
const bool filter_update = active_render && !saturated_capture;
filter_update_blocks_since_reset_ += filter_update ? 1 : 0;
filter_update_blocks_since_start_ += filter_update ? 1 : 0;
// Store convergence flag when observed.
convergence_seen_ = convergence_seen_ || any_filter_converged;
// Verify requirements for achieving a decent filter. The requirements for
// filter adaptation at call startup are more restrictive than after an
// in-call reset.
const bool sufficient_data_to_converge_at_startup =
filter_update_blocks_since_start_ > kNumBlocksPerSecond * 0.4f;
const bool sufficient_data_to_converge_at_reset =
sufficient_data_to_converge_at_startup &&
filter_update_blocks_since_reset_ > kNumBlocksPerSecond * 0.2f;
// The linear filter can only be used if it has had time to converge.
overall_usable_linear_estimates_ = sufficient_data_to_converge_at_startup &&
sufficient_data_to_converge_at_reset;
// The linear filter can only be used if an external delay or convergence have
// been identified
overall_usable_linear_estimates_ =
overall_usable_linear_estimates_ && (external_delay || convergence_seen_);
// If transparent mode is on, deactivate usign the linear filter.
overall_usable_linear_estimates_ =
overall_usable_linear_estimates_ && !transparent_mode;
if (use_linear_filter_) {
std::fill(usable_linear_filter_estimates_.begin(),
usable_linear_filter_estimates_.end(),
overall_usable_linear_estimates_);
}
}
void AecState::SaturationDetector::Update(
rtc::ArrayView<const std::vector<float>> x,
bool saturated_capture,
bool usable_linear_estimate,
rtc::ArrayView<const SubtractorOutput> subtractor_output,
float echo_path_gain) {
saturated_echo_ = false;
if (!saturated_capture) {
return;
}
if (usable_linear_estimate) {
constexpr float kSaturationThreshold = 20000.f;
for (size_t ch = 0; ch < subtractor_output.size(); ++ch) {
saturated_echo_ =
saturated_echo_ ||
(subtractor_output[ch].s_refined_max_abs > kSaturationThreshold ||
subtractor_output[ch].s_coarse_max_abs > kSaturationThreshold);
}
} else {
float max_sample = 0.f;
for (auto& channel : x) {
for (float sample : channel) {
max_sample = std::max(max_sample, fabsf(sample));
}
}
const float kMargin = 10.f;
float peak_echo_amplitude = max_sample * echo_path_gain * kMargin;
saturated_echo_ = saturated_echo_ || peak_echo_amplitude > 32000;
}
}
} // namespace webrtc

294
webrtc/modules/audio_processing/aec3/aec_state.h Normal file
View File

@ -0,0 +1,294 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_AEC_STATE_H_
#define MODULES_AUDIO_PROCESSING_AEC3_AEC_STATE_H_
#include <stddef.h>
#include <array>
#include <memory>
#include <vector>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/delay_estimate.h"
#include "modules/audio_processing/aec3/echo_audibility.h"
#include "modules/audio_processing/aec3/echo_path_variability.h"
#include "modules/audio_processing/aec3/erl_estimator.h"
#include "modules/audio_processing/aec3/erle_estimator.h"
#include "modules/audio_processing/aec3/filter_analyzer.h"
#include "modules/audio_processing/aec3/render_buffer.h"
#include "modules/audio_processing/aec3/reverb_model_estimator.h"
#include "modules/audio_processing/aec3/subtractor_output.h"
#include "modules/audio_processing/aec3/subtractor_output_analyzer.h"
#include "modules/audio_processing/aec3/transparent_mode.h"
namespace webrtc {
class ApmDataDumper;
// Handles the state and the conditions for the echo removal functionality.
class AecState {
public:
AecState(const EchoCanceller3Config& config, size_t num_capture_channels);
~AecState();
// Returns whether the echo subtractor can be used to determine the residual
// echo.
bool UsableLinearEstimate() const {
return filter_quality_state_.LinearFilterUsable() &&
config_.filter.use_linear_filter;
}
// Returns whether the echo subtractor output should be used as output.
bool UseLinearFilterOutput() const {
return filter_quality_state_.LinearFilterUsable() &&
config_.filter.use_linear_filter;
}
// Returns whether the render signal is currently active.
bool ActiveRender() const { return blocks_with_active_render_ > 200; }
// Returns the appropriate scaling of the residual echo to match the
// audibility.
void GetResidualEchoScaling(rtc::ArrayView<float> residual_scaling) const;
// Returns whether the stationary properties of the signals are used in the
// aec.
bool UseStationarityProperties() const {
return config_.echo_audibility.use_stationarity_properties;
}
// Returns the ERLE.
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> Erle() const {
return erle_estimator_.Erle();
}
// Returns an offset to apply to the estimation of the residual echo
// computation. Returning nullopt means that no offset should be used, while
// any other value will be applied as a multiplier to the estimated residual
// echo.
absl::optional<float> ErleUncertainty() const;
// Returns the fullband ERLE estimate in log2 units.
float FullBandErleLog2() const { return erle_estimator_.FullbandErleLog2(); }
// Returns the ERL.
const std::array<float, kFftLengthBy2Plus1>& Erl() const {
return erl_estimator_.Erl();
}
// Returns the time-domain ERL.
float ErlTimeDomain() const { return erl_estimator_.ErlTimeDomain(); }
// Returns the delay estimate based on the linear filter.
int MinDirectPathFilterDelay() const {
return delay_state_.MinDirectPathFilterDelay();
}
// Returns whether the capture signal is saturated.
bool SaturatedCapture() const { return capture_signal_saturation_; }
// Returns whether the echo signal is saturated.
bool SaturatedEcho() const { return saturation_detector_.SaturatedEcho(); }
// Updates the capture signal saturation.
void UpdateCaptureSaturation(bool capture_signal_saturation) {
capture_signal_saturation_ = capture_signal_saturation;
}
// Returns whether the transparent mode is active
bool TransparentModeActive() const {
return transparent_state_ && transparent_state_->Active();
}
// Takes appropriate action at an echo path change.
void HandleEchoPathChange(const EchoPathVariability& echo_path_variability);
// Returns the decay factor for the echo reverberation.
float ReverbDecay() const { return reverb_model_estimator_.ReverbDecay(); }
// Return the frequency response of the reverberant echo.
rtc::ArrayView<const float> GetReverbFrequencyResponse() const {
return reverb_model_estimator_.GetReverbFrequencyResponse();
}
// Returns whether the transition for going out of the initial stated has
// been triggered.
bool TransitionTriggered() const {
return initial_state_.TransitionTriggered();
}
// Updates the aec state.
// TODO(bugs.webrtc.org/10913): Compute multi-channel ERL.
void Update(
const absl::optional<DelayEstimate>& external_delay,
rtc::ArrayView<const std::vector<std::array<float, kFftLengthBy2Plus1>>>
adaptive_filter_frequency_responses,
rtc::ArrayView<const std::vector<float>>
adaptive_filter_impulse_responses,
const RenderBuffer& render_buffer,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> E2_refined,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> Y2,
rtc::ArrayView<const SubtractorOutput> subtractor_output);
// Returns filter length in blocks.
int FilterLengthBlocks() const {
// All filters have the same length, so arbitrarily return channel 0 length.
return filter_analyzer_.FilterLengthBlocks();
}
private:
static int instance_count_;
std::unique_ptr<ApmDataDumper> data_dumper_;
const EchoCanceller3Config config_;
const size_t num_capture_channels_;
const bool deactivate_initial_state_reset_at_echo_path_change_;
const bool full_reset_at_echo_path_change_;
const bool subtractor_analyzer_reset_at_echo_path_change_;
// Class for controlling the transition from the intial state, which in turn
// controls when the filter parameters for the initial state should be used.
class InitialState {
public:
explicit InitialState(const EchoCanceller3Config& config);
// Resets the state to again begin in the initial state.
void Reset();
// Updates the state based on new data.
void Update(bool active_render, bool saturated_capture);
// Returns whether the initial state is active or not.
bool InitialStateActive() const { return initial_state_; }
// Returns that the transition from the initial state has was started.
bool TransitionTriggered() const { return transition_triggered_; }
private:
const bool conservative_initial_phase_;
const float initial_state_seconds_;
bool transition_triggered_ = false;
bool initial_state_ = true;
size_t strong_not_saturated_render_blocks_ = 0;
} initial_state_;
// Class for choosing the direct-path delay relative to the beginning of the
// filter, as well as any other data related to the delay used within
// AecState.
class FilterDelay {
public:
FilterDelay(const EchoCanceller3Config& config,
size_t num_capture_channels);
// Returns whether an external delay has been reported to the AecState (from
// the delay estimator).
bool ExternalDelayReported() const { return external_delay_reported_; }
// Returns the delay in blocks relative to the beginning of the filter that
// corresponds to the direct path of the echo.
rtc::ArrayView<const int> DirectPathFilterDelays() const {
return filter_delays_blocks_;
}
// Returns the minimum delay among the direct path delays relative to the
// beginning of the filter
int MinDirectPathFilterDelay() const { return min_filter_delay_; }
// Updates the delay estimates based on new data.
void Update(
rtc::ArrayView<const int> analyzer_filter_delay_estimates_blocks,
const absl::optional<DelayEstimate>& external_delay,
size_t blocks_with_proper_filter_adaptation);
private:
const int delay_headroom_blocks_;
bool external_delay_reported_ = false;
std::vector<int> filter_delays_blocks_;
int min_filter_delay_;
absl::optional<DelayEstimate> external_delay_;
} delay_state_;
// Classifier for toggling transparent mode when there is no echo.
std::unique_ptr<TransparentMode> transparent_state_;
// Class for analyzing how well the linear filter is, and can be expected to,
// perform on the current signals. The purpose of this is for using to
// select the echo suppression functionality as well as the input to the echo
// suppressor.
class FilteringQualityAnalyzer {
public:
FilteringQualityAnalyzer(const EchoCanceller3Config& config,
size_t num_capture_channels);
// Returns whether the linear filter can be used for the echo
// canceller output.
bool LinearFilterUsable() const { return overall_usable_linear_estimates_; }
// Returns whether an individual filter output can be used for the echo
// canceller output.
const std::vector<bool>& UsableLinearFilterOutputs() const {
return usable_linear_filter_estimates_;
}
// Resets the state of the analyzer.
void Reset();
// Updates the analysis based on new data.
void Update(bool active_render,
bool transparent_mode,
bool saturated_capture,
const absl::optional<DelayEstimate>& external_delay,
bool any_filter_converged);
private:
const bool use_linear_filter_;
bool overall_usable_linear_estimates_ = false;
size_t filter_update_blocks_since_reset_ = 0;
size_t filter_update_blocks_since_start_ = 0;
bool convergence_seen_ = false;
std::vector<bool> usable_linear_filter_estimates_;
} filter_quality_state_;
// Class for detecting whether the echo is to be considered to be
// saturated.
class SaturationDetector {
public:
// Returns whether the echo is to be considered saturated.
bool SaturatedEcho() const { return saturated_echo_; }
// Updates the detection decision based on new data.
void Update(rtc::ArrayView<const std::vector<float>> x,
bool saturated_capture,
bool usable_linear_estimate,
rtc::ArrayView<const SubtractorOutput> subtractor_output,
float echo_path_gain);
private:
bool saturated_echo_ = false;
} saturation_detector_;
ErlEstimator erl_estimator_;
ErleEstimator erle_estimator_;
size_t strong_not_saturated_render_blocks_ = 0;
size_t blocks_with_active_render_ = 0;
bool capture_signal_saturation_ = false;
FilterAnalyzer filter_analyzer_;
EchoAudibility echo_audibility_;
ReverbModelEstimator reverb_model_estimator_;
ReverbModel avg_render_reverb_;
SubtractorOutputAnalyzer subtractor_output_analyzer_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_AEC_STATE_H_

160
webrtc/modules/audio_processing/aec3/alignment_mixer.cc Normal file
View File

@ -0,0 +1,160 @@
/*
* Copyright (c) 2019 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/alignment_mixer.h"
#include <algorithm>
#include "rtc_base/checks.h"
namespace webrtc {
namespace {
AlignmentMixer::MixingVariant ChooseMixingVariant(bool downmix,
bool adaptive_selection,
int num_channels) {
RTC_DCHECK(!(adaptive_selection && downmix));
RTC_DCHECK_LT(0, num_channels);
if (num_channels == 1) {
return AlignmentMixer::MixingVariant::kFixed;
}
if (downmix) {
return AlignmentMixer::MixingVariant::kDownmix;
}
if (adaptive_selection) {
return AlignmentMixer::MixingVariant::kAdaptive;
}
return AlignmentMixer::MixingVariant::kFixed;
}
} // namespace
AlignmentMixer::AlignmentMixer(
size_t num_channels,
const EchoCanceller3Config::Delay::AlignmentMixing& config)
: AlignmentMixer(num_channels,
config.downmix,
config.adaptive_selection,
config.activity_power_threshold,
config.prefer_first_two_channels) {}
AlignmentMixer::AlignmentMixer(size_t num_channels,
bool downmix,
bool adaptive_selection,
float activity_power_threshold,
bool prefer_first_two_channels)
: num_channels_(num_channels),
one_by_num_channels_(1.f / num_channels_),
excitation_energy_threshold_(kBlockSize * activity_power_threshold),
prefer_first_two_channels_(prefer_first_two_channels),
selection_variant_(
ChooseMixingVariant(downmix, adaptive_selection, num_channels_)) {
if (selection_variant_ == MixingVariant::kAdaptive) {
std::fill(strong_block_counters_.begin(), strong_block_counters_.end(), 0);
cumulative_energies_.resize(num_channels_);
std::fill(cumulative_energies_.begin(), cumulative_energies_.end(), 0.f);
}
}
void AlignmentMixer::ProduceOutput(rtc::ArrayView<const std::vector<float>> x,
rtc::ArrayView<float, kBlockSize> y) {
RTC_DCHECK_EQ(x.size(), num_channels_);
if (selection_variant_ == MixingVariant::kDownmix) {
Downmix(x, y);
return;
}
int ch = selection_variant_ == MixingVariant::kFixed ? 0 : SelectChannel(x);
RTC_DCHECK_GE(x.size(), ch);
std::copy(x[ch].begin(), x[ch].end(), y.begin());
}
void AlignmentMixer::Downmix(rtc::ArrayView<const std::vector<float>> x,
rtc::ArrayView<float, kBlockSize> y) const {
RTC_DCHECK_EQ(x.size(), num_channels_);
RTC_DCHECK_GE(num_channels_, 2);
std::copy(x[0].begin(), x[0].end(), y.begin());
for (size_t ch = 1; ch < num_channels_; ++ch) {
for (size_t i = 0; i < kBlockSize; ++i) {
y[i] += x[ch][i];
}
}
for (size_t i = 0; i < kBlockSize; ++i) {
y[i] *= one_by_num_channels_;
}
}
int AlignmentMixer::SelectChannel(rtc::ArrayView<const std::vector<float>> x) {
RTC_DCHECK_EQ(x.size(), num_channels_);
RTC_DCHECK_GE(num_channels_, 2);
RTC_DCHECK_EQ(cumulative_energies_.size(), num_channels_);
constexpr size_t kBlocksToChooseLeftOrRight =
static_cast<size_t>(0.5f * kNumBlocksPerSecond);
const bool good_signal_in_left_or_right =
prefer_first_two_channels_ &&
(strong_block_counters_[0] > kBlocksToChooseLeftOrRight ||
strong_block_counters_[1] > kBlocksToChooseLeftOrRight);
const int num_ch_to_analyze =
good_signal_in_left_or_right ? 2 : num_channels_;
constexpr int kNumBlocksBeforeEnergySmoothing = 60 * kNumBlocksPerSecond;
++block_counter_;
for (int ch = 0; ch < num_ch_to_analyze; ++ch) {
RTC_DCHECK_EQ(x[ch].size(), kBlockSize);
float x2_sum = 0.f;
for (size_t i = 0; i < kBlockSize; ++i) {
x2_sum += x[ch][i] * x[ch][i];
}
if (ch < 2 && x2_sum > excitation_energy_threshold_) {
++strong_block_counters_[ch];
}
if (block_counter_ <= kNumBlocksBeforeEnergySmoothing) {
cumulative_energies_[ch] += x2_sum;
} else {
constexpr float kSmoothing = 1.f / (10 * kNumBlocksPerSecond);
cumulative_energies_[ch] +=
kSmoothing * (x2_sum - cumulative_energies_[ch]);
}
}
// Normalize the energies to allow the energy computations to from now be
// based on smoothing.
if (block_counter_ == kNumBlocksBeforeEnergySmoothing) {
constexpr float kOneByNumBlocksBeforeEnergySmoothing =
1.f / kNumBlocksBeforeEnergySmoothing;
for (int ch = 0; ch < num_ch_to_analyze; ++ch) {
cumulative_energies_[ch] *= kOneByNumBlocksBeforeEnergySmoothing;
}
}
int strongest_ch = 0;
for (int ch = 0; ch < num_ch_to_analyze; ++ch) {
if (cumulative_energies_[ch] > cumulative_energies_[strongest_ch]) {
strongest_ch = ch;
}
}
if ((good_signal_in_left_or_right && selected_channel_ > 1) ||
cumulative_energies_[strongest_ch] >
2.f * cumulative_energies_[selected_channel_]) {
selected_channel_ = strongest_ch;
}
return selected_channel_;
}
} // namespace webrtc

58
webrtc/modules/audio_processing/aec3/alignment_mixer.h Normal file
View File

@ -0,0 +1,58 @@
/*
* Copyright (c) 2019 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ALIGNMENT_MIXER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ALIGNMENT_MIXER_H_
#include <vector>
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
namespace webrtc {
// Performs channel conversion to mono for the purpose of providing a decent
// mono input for the delay estimation. This is achieved by analyzing all
// incoming channels and produce one single channel output.
class AlignmentMixer {
public:
AlignmentMixer(size_t num_channels,
const EchoCanceller3Config::Delay::AlignmentMixing& config);
AlignmentMixer(size_t num_channels,
bool downmix,
bool adaptive_selection,
float excitation_limit,
bool prefer_first_two_channels);
void ProduceOutput(rtc::ArrayView<const std::vector<float>> x,
rtc::ArrayView<float, kBlockSize> y);
enum class MixingVariant { kDownmix, kAdaptive, kFixed };
private:
const size_t num_channels_;
const float one_by_num_channels_;
const float excitation_energy_threshold_;
const bool prefer_first_two_channels_;
const MixingVariant selection_variant_;
std::array<size_t, 2> strong_block_counters_;
std::vector<float> cumulative_energies_;
int selected_channel_ = 0;
size_t block_counter_ = 0;
void Downmix(const rtc::ArrayView<const std::vector<float>> x,
rtc::ArrayView<float, kBlockSize> y) const;
int SelectChannel(rtc::ArrayView<const std::vector<float>> x);
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ALIGNMENT_MIXER_H_

121
webrtc/modules/audio_processing/aec3/api_call_jitter_metrics.cc Normal file
View File

@ -0,0 +1,121 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/api_call_jitter_metrics.h"
#include <algorithm>
#include <limits>
#include "modules/audio_processing/aec3/aec3_common.h"
#include "system_wrappers/include/metrics.h"
namespace webrtc {
namespace {
bool TimeToReportMetrics(int frames_since_last_report) {
constexpr int kNumFramesPerSecond = 100;
constexpr int kReportingIntervalFrames = 10 * kNumFramesPerSecond;
return frames_since_last_report == kReportingIntervalFrames;
}
} // namespace
ApiCallJitterMetrics::Jitter::Jitter()
: max_(0), min_(std::numeric_limits<int>::max()) {}
void ApiCallJitterMetrics::Jitter::Update(int num_api_calls_in_a_row) {
min_ = std::min(min_, num_api_calls_in_a_row);
max_ = std::max(max_, num_api_calls_in_a_row);
}
void ApiCallJitterMetrics::Jitter::Reset() {
min_ = std::numeric_limits<int>::max();
max_ = 0;
}
void ApiCallJitterMetrics::Reset() {
render_jitter_.Reset();
capture_jitter_.Reset();
num_api_calls_in_a_row_ = 0;
frames_since_last_report_ = 0;
last_call_was_render_ = false;
proper_call_observed_ = false;
}
void ApiCallJitterMetrics::ReportRenderCall() {
if (!last_call_was_render_) {
// If the previous call was a capture and a proper call has been observed
// (containing both render and capture data), storing the last number of
// capture calls into the metrics.
if (proper_call_observed_) {
capture_jitter_.Update(num_api_calls_in_a_row_);
}
// Reset the call counter to start counting render calls.
num_api_calls_in_a_row_ = 0;
}
++num_api_calls_in_a_row_;
last_call_was_render_ = true;
}
void ApiCallJitterMetrics::ReportCaptureCall() {
if (last_call_was_render_) {
// If the previous call was a render and a proper call has been observed
// (containing both render and capture data), storing the last number of
// render calls into the metrics.
if (proper_call_observed_) {
render_jitter_.Update(num_api_calls_in_a_row_);
}
// Reset the call counter to start counting capture calls.
num_api_calls_in_a_row_ = 0;
// If this statement is reached, at least one render and one capture call
// have been observed.
proper_call_observed_ = true;
}
++num_api_calls_in_a_row_;
last_call_was_render_ = false;
// Only report and update jitter metrics for when a proper call, containing
// both render and capture data, has been observed.
if (proper_call_observed_ &&
TimeToReportMetrics(++frames_since_last_report_)) {
// Report jitter, where the base basic unit is frames.
constexpr int kMaxJitterToReport = 50;
// Report max and min jitter for render and capture, in units of 20 ms.
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.MaxRenderJitter",
std::min(kMaxJitterToReport, render_jitter().max()), 1,
kMaxJitterToReport, kMaxJitterToReport);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.MinRenderJitter",
std::min(kMaxJitterToReport, render_jitter().min()), 1,
kMaxJitterToReport, kMaxJitterToReport);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.MaxCaptureJitter",
std::min(kMaxJitterToReport, capture_jitter().max()), 1,
kMaxJitterToReport, kMaxJitterToReport);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.MinCaptureJitter",
std::min(kMaxJitterToReport, capture_jitter().min()), 1,
kMaxJitterToReport, kMaxJitterToReport);
frames_since_last_report_ = 0;
Reset();
}
}
bool ApiCallJitterMetrics::WillReportMetricsAtNextCapture() const {
return TimeToReportMetrics(frames_since_last_report_ + 1);
}
} // namespace webrtc

60
webrtc/modules/audio_processing/aec3/api_call_jitter_metrics.h Normal file
View File

@ -0,0 +1,60 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_API_CALL_JITTER_METRICS_H_
#define MODULES_AUDIO_PROCESSING_AEC3_API_CALL_JITTER_METRICS_H_
namespace webrtc {
// Stores data for reporting metrics on the API call jitter.
class ApiCallJitterMetrics {
public:
class Jitter {
public:
Jitter();
void Update(int num_api_calls_in_a_row);
void Reset();
int min() const { return min_; }
int max() const { return max_; }
private:
int max_;
int min_;
};
ApiCallJitterMetrics() { Reset(); }
// Update metrics for render API call.
void ReportRenderCall();
// Update and periodically report metrics for capture API call.
void ReportCaptureCall();
// Methods used only for testing.
const Jitter& render_jitter() const { return render_jitter_; }
const Jitter& capture_jitter() const { return capture_jitter_; }
bool WillReportMetricsAtNextCapture() const;
private:
void Reset();
Jitter render_jitter_;
Jitter capture_jitter_;
int num_api_calls_in_a_row_ = 0;
int frames_since_last_report_ = 0;
bool last_call_was_render_ = false;
bool proper_call_observed_ = false;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_API_CALL_JITTER_METRICS_H_

39
webrtc/modules/audio_processing/aec3/block_buffer.cc Normal file
View File

@ -0,0 +1,39 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/block_buffer.h"
#include <algorithm>
namespace webrtc {
BlockBuffer::BlockBuffer(size_t size,
size_t num_bands,
size_t num_channels,
size_t frame_length)
: size(static_cast<int>(size)),
buffer(size,
std::vector<std::vector<std::vector<float>>>(
num_bands,
std::vector<std::vector<float>>(
num_channels,
std::vector<float>(frame_length, 0.f)))) {
for (auto& block : buffer) {
for (auto& band : block) {
for (auto& channel : band) {
std::fill(channel.begin(), channel.end(), 0.f);
}
}
}
}
BlockBuffer::~BlockBuffer() = default;
} // namespace webrtc

62
webrtc/modules/audio_processing/aec3/block_buffer.h Normal file
View File

@ -0,0 +1,62 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_BLOCK_BUFFER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_BLOCK_BUFFER_H_
#include <stddef.h>
#include <vector>
#include "rtc_base/checks.h"
namespace webrtc {
// Struct for bundling a circular buffer of two dimensional vector objects
// together with the read and write indices.
struct BlockBuffer {
BlockBuffer(size_t size,
size_t num_bands,
size_t num_channels,
size_t frame_length);
~BlockBuffer();
int IncIndex(int index) const {
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
return index < size - 1 ? index + 1 : 0;
}
int DecIndex(int index) const {
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
return index > 0 ? index - 1 : size - 1;
}
int OffsetIndex(int index, int offset) const {
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
RTC_DCHECK_GE(size, offset);
return (size + index + offset) % size;
}
void UpdateWriteIndex(int offset) { write = OffsetIndex(write, offset); }
void IncWriteIndex() { write = IncIndex(write); }
void DecWriteIndex() { write = DecIndex(write); }
void UpdateReadIndex(int offset) { read = OffsetIndex(read, offset); }
void IncReadIndex() { read = IncIndex(read); }
void DecReadIndex() { read = DecIndex(read); }
const int size;
std::vector<std::vector<std::vector<std::vector<float>>>> buffer;
int write = 0;
int read = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_BLOCK_BUFFER_H_

62
webrtc/modules/audio_processing/aec3/block_delay_buffer.cc Normal file
View File

@ -0,0 +1,62 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/block_delay_buffer.h"
#include "api/array_view.h"
#include "rtc_base/checks.h"
namespace webrtc {
BlockDelayBuffer::BlockDelayBuffer(size_t num_channels,
size_t num_bands,
size_t frame_length,
size_t delay_samples)
: frame_length_(frame_length),
delay_(delay_samples),
buf_(num_channels,
std::vector<std::vector<float>>(num_bands,
std::vector<float>(delay_, 0.f))) {}
BlockDelayBuffer::~BlockDelayBuffer() = default;
void BlockDelayBuffer::DelaySignal(AudioBuffer* frame) {
RTC_DCHECK_EQ(buf_.size(), frame->num_channels());
if (delay_ == 0) {
return;
}
const size_t num_bands = buf_[0].size();
const size_t num_channels = buf_.size();
const size_t i_start = last_insert_;
size_t i = 0;
for (size_t ch = 0; ch < num_channels; ++ch) {
RTC_DCHECK_EQ(buf_[ch].size(), frame->num_bands());
RTC_DCHECK_EQ(buf_[ch].size(), num_bands);
rtc::ArrayView<float* const> frame_ch(frame->split_bands(ch), num_bands);
for (size_t band = 0; band < num_bands; ++band) {
RTC_DCHECK_EQ(delay_, buf_[ch][band].size());
i = i_start;
for (size_t k = 0; k < frame_length_; ++k) {
const float tmp = buf_[ch][band][i];
buf_[ch][band][i] = frame_ch[band][k];
frame_ch[band][k] = tmp;
i = i < delay_ - 1 ? i + 1 : 0;
}
}
}
last_insert_ = i;
}
} // namespace webrtc

43
webrtc/modules/audio_processing/aec3/block_delay_buffer.h Normal file
View File

@ -0,0 +1,43 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_BLOCK_DELAY_BUFFER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_BLOCK_DELAY_BUFFER_H_
#include <stddef.h>
#include <vector>
#include "modules/audio_processing/audio_buffer.h"
namespace webrtc {
// Class for applying a fixed delay to the samples in a signal partitioned using
// the audiobuffer band-splitting scheme.
class BlockDelayBuffer {
public:
BlockDelayBuffer(size_t num_channels,
size_t num_bands,
size_t frame_length,
size_t delay_samples);
~BlockDelayBuffer();
// Delays the samples by the specified delay.
void DelaySignal(AudioBuffer* frame);
private:
const size_t frame_length_;
const size_t delay_;
std::vector<std::vector<std::vector<float>>> buf_;
size_t last_insert_ = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_BLOCK_DELAY_BUFFER_H_

86
webrtc/modules/audio_processing/aec3/block_framer.cc Normal file
View File

@ -0,0 +1,86 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/block_framer.h"
#include <algorithm>
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/checks.h"
namespace webrtc {
BlockFramer::BlockFramer(size_t num_bands, size_t num_channels)
: num_bands_(num_bands),
num_channels_(num_channels),
buffer_(num_bands_,
std::vector<std::vector<float>>(
num_channels,
std::vector<float>(kBlockSize, 0.f))) {
RTC_DCHECK_LT(0, num_bands);
RTC_DCHECK_LT(0, num_channels);
}
BlockFramer::~BlockFramer() = default;
// All the constants are chosen so that the buffer is either empty or has enough
// samples for InsertBlockAndExtractSubFrame to produce a frame. In order to
// achieve this, the InsertBlockAndExtractSubFrame and InsertBlock methods need
// to be called in the correct order.
void BlockFramer::InsertBlock(
const std::vector<std::vector<std::vector<float>>>& block) {
RTC_DCHECK_EQ(num_bands_, block.size());
for (size_t band = 0; band < num_bands_; ++band) {
RTC_DCHECK_EQ(num_channels_, block[band].size());
for (size_t channel = 0; channel < num_channels_; ++channel) {
RTC_DCHECK_EQ(kBlockSize, block[band][channel].size());
RTC_DCHECK_EQ(0, buffer_[band][channel].size());
buffer_[band][channel].insert(buffer_[band][channel].begin(),
block[band][channel].begin(),
block[band][channel].end());
}
}
}
void BlockFramer::InsertBlockAndExtractSubFrame(
const std::vector<std::vector<std::vector<float>>>& block,
std::vector<std::vector<rtc::ArrayView<float>>>* sub_frame) {
RTC_DCHECK(sub_frame);
RTC_DCHECK_EQ(num_bands_, block.size());
RTC_DCHECK_EQ(num_bands_, sub_frame->size());
for (size_t band = 0; band < num_bands_; ++band) {
RTC_DCHECK_EQ(num_channels_, block[band].size());
RTC_DCHECK_EQ(num_channels_, (*sub_frame)[0].size());
for (size_t channel = 0; channel < num_channels_; ++channel) {
RTC_DCHECK_LE(kSubFrameLength,
buffer_[band][channel].size() + kBlockSize);
RTC_DCHECK_EQ(kBlockSize, block[band][channel].size());
RTC_DCHECK_GE(kBlockSize, buffer_[band][channel].size());
RTC_DCHECK_EQ(kSubFrameLength, (*sub_frame)[band][channel].size());
const int samples_to_frame =
kSubFrameLength - buffer_[band][channel].size();
std::copy(buffer_[band][channel].begin(), buffer_[band][channel].end(),
(*sub_frame)[band][channel].begin());
std::copy(
block[band][channel].begin(),
block[band][channel].begin() + samples_to_frame,
(*sub_frame)[band][channel].begin() + buffer_[band][channel].size());
buffer_[band][channel].clear();
buffer_[band][channel].insert(
buffer_[band][channel].begin(),
block[band][channel].begin() + samples_to_frame,
block[band][channel].end());
}
}
}
} // namespace webrtc

48
webrtc/modules/audio_processing/aec3/block_framer.h Normal file
View File

@ -0,0 +1,48 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_BLOCK_FRAMER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_BLOCK_FRAMER_H_
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
namespace webrtc {
// Class for producing frames consisting of 2 subframes of 80 samples each
// from 64 sample blocks. The class is designed to work together with the
// FrameBlocker class which performs the reverse conversion. Used together with
// that, this class produces output frames are the same rate as frames are
// received by the FrameBlocker class. Note that the internal buffers will
// overrun if any other rate of packets insertion is used.
class BlockFramer {
public:
BlockFramer(size_t num_bands, size_t num_channels);
~BlockFramer();
BlockFramer(const BlockFramer&) = delete;
BlockFramer& operator=(const BlockFramer&) = delete;
// Adds a 64 sample block into the data that will form the next output frame.
void InsertBlock(const std::vector<std::vector<std::vector<float>>>& block);
// Adds a 64 sample block and extracts an 80 sample subframe.
void InsertBlockAndExtractSubFrame(
const std::vector<std::vector<std::vector<float>>>& block,
std::vector<std::vector<rtc::ArrayView<float>>>* sub_frame);
private:
const size_t num_bands_;
const size_t num_channels_;
std::vector<std::vector<std::vector<float>>> buffer_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_BLOCK_FRAMER_H_

292
webrtc/modules/audio_processing/aec3/block_processor.cc Normal file
View File

@ -0,0 +1,292 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/block_processor.h"
#include <stddef.h>
#include <memory>
#include <utility>
#include <vector>
#include "absl/types/optional.h"
#include "api/audio/echo_canceller3_config.h"
#include "api/audio/echo_control.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/block_processor_metrics.h"
#include "modules/audio_processing/aec3/delay_estimate.h"
#include "modules/audio_processing/aec3/echo_path_variability.h"
#include "modules/audio_processing/aec3/echo_remover.h"
#include "modules/audio_processing/aec3/render_delay_buffer.h"
#include "modules/audio_processing/aec3/render_delay_controller.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomic_ops.h"
#include "rtc_base/checks.h"
#include "rtc_base/logging.h"
namespace webrtc {
namespace {
enum class BlockProcessorApiCall { kCapture, kRender };
class BlockProcessorImpl final : public BlockProcessor {
public:
BlockProcessorImpl(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels,
std::unique_ptr<RenderDelayBuffer> render_buffer,
std::unique_ptr<RenderDelayController> delay_controller,
std::unique_ptr<EchoRemover> echo_remover);
BlockProcessorImpl() = delete;
~BlockProcessorImpl() override;
void ProcessCapture(
bool echo_path_gain_change,
bool capture_signal_saturation,
std::vector<std::vector<std::vector<float>>>* linear_output,
std::vector<std::vector<std::vector<float>>>* capture_block) override;
void BufferRender(
const std::vector<std::vector<std::vector<float>>>& block) override;
void UpdateEchoLeakageStatus(bool leakage_detected) override;
void GetMetrics(EchoControl::Metrics* metrics) const override;
void SetAudioBufferDelay(int delay_ms) override;
private:
static int instance_count_;
std::unique_ptr<ApmDataDumper> data_dumper_;
const EchoCanceller3Config config_;
bool capture_properly_started_ = false;
bool render_properly_started_ = false;
const size_t sample_rate_hz_;
std::unique_ptr<RenderDelayBuffer> render_buffer_;
std::unique_ptr<RenderDelayController> delay_controller_;
std::unique_ptr<EchoRemover> echo_remover_;
BlockProcessorMetrics metrics_;
RenderDelayBuffer::BufferingEvent render_event_;
size_t capture_call_counter_ = 0;
absl::optional<DelayEstimate> estimated_delay_;
};
int BlockProcessorImpl::instance_count_ = 0;
BlockProcessorImpl::BlockProcessorImpl(
const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels,
std::unique_ptr<RenderDelayBuffer> render_buffer,
std::unique_ptr<RenderDelayController> delay_controller,
std::unique_ptr<EchoRemover> echo_remover)
: data_dumper_(
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
config_(config),
sample_rate_hz_(sample_rate_hz),
render_buffer_(std::move(render_buffer)),
delay_controller_(std::move(delay_controller)),
echo_remover_(std::move(echo_remover)),
render_event_(RenderDelayBuffer::BufferingEvent::kNone) {
RTC_DCHECK(ValidFullBandRate(sample_rate_hz_));
}
BlockProcessorImpl::~BlockProcessorImpl() = default;
void BlockProcessorImpl::ProcessCapture(
bool echo_path_gain_change,
bool capture_signal_saturation,
std::vector<std::vector<std::vector<float>>>* linear_output,
std::vector<std::vector<std::vector<float>>>* capture_block) {
RTC_DCHECK(capture_block);
RTC_DCHECK_EQ(NumBandsForRate(sample_rate_hz_), capture_block->size());
RTC_DCHECK_EQ(kBlockSize, (*capture_block)[0][0].size());
capture_call_counter_++;
data_dumper_->DumpRaw("aec3_processblock_call_order",
static_cast<int>(BlockProcessorApiCall::kCapture));
data_dumper_->DumpWav("aec3_processblock_capture_input", kBlockSize,
&(*capture_block)[0][0][0], 16000, 1);
if (render_properly_started_) {
if (!capture_properly_started_) {
capture_properly_started_ = true;
render_buffer_->Reset();
if (delay_controller_)
delay_controller_->Reset(true);
}
} else {
// If no render data has yet arrived, do not process the capture signal.
render_buffer_->HandleSkippedCaptureProcessing();
return;
}
EchoPathVariability echo_path_variability(
echo_path_gain_change, EchoPathVariability::DelayAdjustment::kNone,
false);
if (render_event_ == RenderDelayBuffer::BufferingEvent::kRenderOverrun &&
render_properly_started_) {
echo_path_variability.delay_change =
EchoPathVariability::DelayAdjustment::kBufferFlush;
if (delay_controller_)
delay_controller_->Reset(true);
RTC_LOG(LS_WARNING) << "Reset due to render buffer overrun at block "
<< capture_call_counter_;
}
render_event_ = RenderDelayBuffer::BufferingEvent::kNone;
// Update the render buffers with any newly arrived render blocks and prepare
// the render buffers for reading the render data corresponding to the current
// capture block.
RenderDelayBuffer::BufferingEvent buffer_event =
render_buffer_->PrepareCaptureProcessing();
// Reset the delay controller at render buffer underrun.
if (buffer_event == RenderDelayBuffer::BufferingEvent::kRenderUnderrun) {
if (delay_controller_)
delay_controller_->Reset(false);
}
data_dumper_->DumpWav("aec3_processblock_capture_input2", kBlockSize,
&(*capture_block)[0][0][0], 16000, 1);
bool has_delay_estimator = !config_.delay.use_external_delay_estimator;
if (has_delay_estimator) {
RTC_DCHECK(delay_controller_);
// Compute and apply the render delay required to achieve proper signal
// alignment.
estimated_delay_ = delay_controller_->GetDelay(
render_buffer_->GetDownsampledRenderBuffer(), render_buffer_->Delay(),
(*capture_block)[0]);
if (estimated_delay_) {
bool delay_change =
render_buffer_->AlignFromDelay(estimated_delay_->delay);
if (delay_change) {
rtc::LoggingSeverity log_level =
config_.delay.log_warning_on_delay_changes ? rtc::LS_WARNING
: rtc::LS_INFO;
RTC_LOG_V(log_level) << "Delay changed to " << estimated_delay_->delay
<< " at block " << capture_call_counter_;
echo_path_variability.delay_change =
EchoPathVariability::DelayAdjustment::kNewDetectedDelay;
}
}
echo_path_variability.clock_drift = delay_controller_->HasClockdrift();
} else {
render_buffer_->AlignFromExternalDelay();
}
// Remove the echo from the capture signal.
if (has_delay_estimator || render_buffer_->HasReceivedBufferDelay()) {
echo_remover_->ProcessCapture(
echo_path_variability, capture_signal_saturation, estimated_delay_,
render_buffer_->GetRenderBuffer(), linear_output, capture_block);
}
// Update the metrics.
metrics_.UpdateCapture(false);
}
void BlockProcessorImpl::BufferRender(
const std::vector<std::vector<std::vector<float>>>& block) {
RTC_DCHECK_EQ(NumBandsForRate(sample_rate_hz_), block.size());
RTC_DCHECK_EQ(kBlockSize, block[0][0].size());
data_dumper_->DumpRaw("aec3_processblock_call_order",
static_cast<int>(BlockProcessorApiCall::kRender));
data_dumper_->DumpWav("aec3_processblock_render_input", kBlockSize,
&block[0][0][0], 16000, 1);
data_dumper_->DumpWav("aec3_processblock_render_input2", kBlockSize,
&block[0][0][0], 16000, 1);
render_event_ = render_buffer_->Insert(block);
metrics_.UpdateRender(render_event_ !=
RenderDelayBuffer::BufferingEvent::kNone);
render_properly_started_ = true;
if (delay_controller_)
delay_controller_->LogRenderCall();
}
void BlockProcessorImpl::UpdateEchoLeakageStatus(bool leakage_detected) {
echo_remover_->UpdateEchoLeakageStatus(leakage_detected);
}
void BlockProcessorImpl::GetMetrics(EchoControl::Metrics* metrics) const {
echo_remover_->GetMetrics(metrics);
constexpr int block_size_ms = 4;
absl::optional<size_t> delay = render_buffer_->Delay();
metrics->delay_ms = delay ? static_cast<int>(*delay) * block_size_ms : 0;
}
void BlockProcessorImpl::SetAudioBufferDelay(int delay_ms) {
render_buffer_->SetAudioBufferDelay(delay_ms);
}
} // namespace
BlockProcessor* BlockProcessor::Create(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels) {
std::unique_ptr<RenderDelayBuffer> render_buffer(
RenderDelayBuffer::Create(config, sample_rate_hz, num_render_channels));
std::unique_ptr<RenderDelayController> delay_controller;
if (!config.delay.use_external_delay_estimator) {
delay_controller.reset(RenderDelayController::Create(config, sample_rate_hz,
num_capture_channels));
}
std::unique_ptr<EchoRemover> echo_remover(EchoRemover::Create(
config, sample_rate_hz, num_render_channels, num_capture_channels));
return Create(config, sample_rate_hz, num_render_channels,
num_capture_channels, std::move(render_buffer),
std::move(delay_controller), std::move(echo_remover));
}
BlockProcessor* BlockProcessor::Create(
const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels,
std::unique_ptr<RenderDelayBuffer> render_buffer) {
std::unique_ptr<RenderDelayController> delay_controller;
if (!config.delay.use_external_delay_estimator) {
delay_controller.reset(RenderDelayController::Create(config, sample_rate_hz,
num_capture_channels));
}
std::unique_ptr<EchoRemover> echo_remover(EchoRemover::Create(
config, sample_rate_hz, num_render_channels, num_capture_channels));
return Create(config, sample_rate_hz, num_render_channels,
num_capture_channels, std::move(render_buffer),
std::move(delay_controller), std::move(echo_remover));
}
BlockProcessor* BlockProcessor::Create(
const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels,
std::unique_ptr<RenderDelayBuffer> render_buffer,
std::unique_ptr<RenderDelayController> delay_controller,
std::unique_ptr<EchoRemover> echo_remover) {
return new BlockProcessorImpl(config, sample_rate_hz, num_render_channels,
num_capture_channels, std::move(render_buffer),
std::move(delay_controller),
std::move(echo_remover));
}
} // namespace webrtc

76
webrtc/modules/audio_processing/aec3/block_processor.h Normal file
View File

@ -0,0 +1,76 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_BLOCK_PROCESSOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_BLOCK_PROCESSOR_H_
#include <stddef.h>
#include <memory>
#include <vector>
#include "api/audio/echo_canceller3_config.h"
#include "api/audio/echo_control.h"
#include "modules/audio_processing/aec3/echo_remover.h"
#include "modules/audio_processing/aec3/render_delay_buffer.h"
#include "modules/audio_processing/aec3/render_delay_controller.h"
namespace webrtc {
// Class for performing echo cancellation on 64 sample blocks of audio data.
class BlockProcessor {
public:
static BlockProcessor* Create(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels);
// Only used for testing purposes.
static BlockProcessor* Create(
const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels,
std::unique_ptr<RenderDelayBuffer> render_buffer);
static BlockProcessor* Create(
const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels,
std::unique_ptr<RenderDelayBuffer> render_buffer,
std::unique_ptr<RenderDelayController> delay_controller,
std::unique_ptr<EchoRemover> echo_remover);
virtual ~BlockProcessor() = default;
// Get current metrics.
virtual void GetMetrics(EchoControl::Metrics* metrics) const = 0;
// Provides an optional external estimate of the audio buffer delay.
virtual void SetAudioBufferDelay(int delay_ms) = 0;
// Processes a block of capture data.
virtual void ProcessCapture(
bool echo_path_gain_change,
bool capture_signal_saturation,
std::vector<std::vector<std::vector<float>>>* linear_output,
std::vector<std::vector<std::vector<float>>>* capture_block) = 0;
// Buffers a block of render data supplied by a FrameBlocker object.
virtual void BufferRender(
const std::vector<std::vector<std::vector<float>>>& render_block) = 0;
// Reports whether echo leakage has been detected in the echo canceller
// output.
virtual void UpdateEchoLeakageStatus(bool leakage_detected) = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_BLOCK_PROCESSOR_H_

104
webrtc/modules/audio_processing/aec3/block_processor_metrics.cc Normal file
View File

@ -0,0 +1,104 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/block_processor_metrics.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/checks.h"
#include "system_wrappers/include/metrics.h"
namespace webrtc {
namespace {
enum class RenderUnderrunCategory {
kNone,
kFew,
kSeveral,
kMany,
kConstant,
kNumCategories
};
enum class RenderOverrunCategory {
kNone,
kFew,
kSeveral,
kMany,
kConstant,
kNumCategories
};
} // namespace
void BlockProcessorMetrics::UpdateCapture(bool underrun) {
++capture_block_counter_;
if (underrun) {
++render_buffer_underruns_;
}
if (capture_block_counter_ == kMetricsReportingIntervalBlocks) {
metrics_reported_ = true;
RenderUnderrunCategory underrun_category;
if (render_buffer_underruns_ == 0) {
underrun_category = RenderUnderrunCategory::kNone;
} else if (render_buffer_underruns_ > (capture_block_counter_ >> 1)) {
underrun_category = RenderUnderrunCategory::kConstant;
} else if (render_buffer_underruns_ > 100) {
underrun_category = RenderUnderrunCategory::kMany;
} else if (render_buffer_underruns_ > 10) {
underrun_category = RenderUnderrunCategory::kSeveral;
} else {
underrun_category = RenderUnderrunCategory::kFew;
}
RTC_HISTOGRAM_ENUMERATION(
"WebRTC.Audio.EchoCanceller.RenderUnderruns",
static_cast<int>(underrun_category),
static_cast<int>(RenderUnderrunCategory::kNumCategories));
RenderOverrunCategory overrun_category;
if (render_buffer_overruns_ == 0) {
overrun_category = RenderOverrunCategory::kNone;
} else if (render_buffer_overruns_ > (buffer_render_calls_ >> 1)) {
overrun_category = RenderOverrunCategory::kConstant;
} else if (render_buffer_overruns_ > 100) {
overrun_category = RenderOverrunCategory::kMany;
} else if (render_buffer_overruns_ > 10) {
overrun_category = RenderOverrunCategory::kSeveral;
} else {
overrun_category = RenderOverrunCategory::kFew;
}
RTC_HISTOGRAM_ENUMERATION(
"WebRTC.Audio.EchoCanceller.RenderOverruns",
static_cast<int>(overrun_category),
static_cast<int>(RenderOverrunCategory::kNumCategories));
ResetMetrics();
capture_block_counter_ = 0;
} else {
metrics_reported_ = false;
}
}
void BlockProcessorMetrics::UpdateRender(bool overrun) {
++buffer_render_calls_;
if (overrun) {
++render_buffer_overruns_;
}
}
void BlockProcessorMetrics::ResetMetrics() {
render_buffer_underruns_ = 0;
render_buffer_overruns_ = 0;
buffer_render_calls_ = 0;
}
} // namespace webrtc

47
webrtc/modules/audio_processing/aec3/block_processor_metrics.h Normal file
View File

@ -0,0 +1,47 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_BLOCK_PROCESSOR_METRICS_H_
#define MODULES_AUDIO_PROCESSING_AEC3_BLOCK_PROCESSOR_METRICS_H_
#include "rtc_base/constructor_magic.h"
namespace webrtc {
// Handles the reporting of metrics for the block_processor.
class BlockProcessorMetrics {
public:
BlockProcessorMetrics() = default;
// Updates the metric with new capture data.
void UpdateCapture(bool underrun);
// Updates the metric with new render data.
void UpdateRender(bool overrun);
// Returns true if the metrics have just been reported, otherwise false.
bool MetricsReported() { return metrics_reported_; }
private:
// Resets the metrics.
void ResetMetrics();
int capture_block_counter_ = 0;
bool metrics_reported_ = false;
int render_buffer_underruns_ = 0;
int render_buffer_overruns_ = 0;
int buffer_render_calls_ = 0;
RTC_DISALLOW_COPY_AND_ASSIGN(BlockProcessorMetrics);
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_BLOCK_PROCESSOR_METRICS_H_

61
webrtc/modules/audio_processing/aec3/clockdrift_detector.cc Normal file
View File

@ -0,0 +1,61 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/clockdrift_detector.h"
namespace webrtc {
ClockdriftDetector::ClockdriftDetector()
: level_(Level::kNone), stability_counter_(0) {
delay_history_.fill(0);
}
ClockdriftDetector::~ClockdriftDetector() = default;
void ClockdriftDetector::Update(int delay_estimate) {
if (delay_estimate == delay_history_[0]) {
// Reset clockdrift level if delay estimate is stable for 7500 blocks (30
// seconds).
if (++stability_counter_ > 7500)
level_ = Level::kNone;
return;
}
stability_counter_ = 0;
const int d1 = delay_history_[0] - delay_estimate;
const int d2 = delay_history_[1] - delay_estimate;
const int d3 = delay_history_[2] - delay_estimate;
// Patterns recognized as positive clockdrift:
// [x-3], x-2, x-1, x.
// [x-3], x-1, x-2, x.
const bool probable_drift_up =
(d1 == -1 && d2 == -2) || (d1 == -2 && d2 == -1);
const bool drift_up = probable_drift_up && d3 == -3;
// Patterns recognized as negative clockdrift:
// [x+3], x+2, x+1, x.
// [x+3], x+1, x+2, x.
const bool probable_drift_down = (d1 == 1 && d2 == 2) || (d1 == 2 && d2 == 1);
const bool drift_down = probable_drift_down && d3 == 3;
// Set clockdrift level.
if (drift_up || drift_down) {
level_ = Level::kVerified;
} else if ((probable_drift_up || probable_drift_down) &&
level_ == Level::kNone) {
level_ = Level::kProbable;
}
// Shift delay history one step.
delay_history_[2] = delay_history_[1];
delay_history_[1] = delay_history_[0];
delay_history_[0] = delay_estimate;
}
} // namespace webrtc

40
webrtc/modules/audio_processing/aec3/clockdrift_detector.h Normal file
View File

@ -0,0 +1,40 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_CLOCKDRIFT_DETECTOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_CLOCKDRIFT_DETECTOR_H_
#include <stddef.h>
#include <array>
namespace webrtc {
class ApmDataDumper;
struct DownsampledRenderBuffer;
struct EchoCanceller3Config;
// Detects clockdrift by analyzing the estimated delay.
class ClockdriftDetector {
public:
enum class Level { kNone, kProbable, kVerified, kNumCategories };
ClockdriftDetector();
~ClockdriftDetector();
void Update(int delay_estimate);
Level ClockdriftLevel() const { return level_; }
private:
std::array<int, 3> delay_history_;
Level level_;
size_t stability_counter_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_CLOCKDRIFT_DETECTOR_H_

103
webrtc/modules/audio_processing/aec3/coarse_filter_update_gain.cc Normal file
View File

@ -0,0 +1,103 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/coarse_filter_update_gain.h"
#include <algorithm>
#include <functional>
#include "rtc_base/checks.h"
namespace webrtc {
CoarseFilterUpdateGain::CoarseFilterUpdateGain(
const EchoCanceller3Config::Filter::CoarseConfiguration& config,
size_t config_change_duration_blocks)
: config_change_duration_blocks_(
static_cast<int>(config_change_duration_blocks)) {
SetConfig(config, true);
RTC_DCHECK_LT(0, config_change_duration_blocks_);
one_by_config_change_duration_blocks_ = 1.f / config_change_duration_blocks_;
}
void CoarseFilterUpdateGain::HandleEchoPathChange() {
poor_signal_excitation_counter_ = 0;
call_counter_ = 0;
}
void CoarseFilterUpdateGain::Compute(
const std::array<float, kFftLengthBy2Plus1>& render_power,
const RenderSignalAnalyzer& render_signal_analyzer,
const FftData& E_coarse,
size_t size_partitions,
bool saturated_capture_signal,
FftData* G) {
RTC_DCHECK(G);
++call_counter_;
UpdateCurrentConfig();
if (render_signal_analyzer.PoorSignalExcitation()) {
poor_signal_excitation_counter_ = 0;
}
// Do not update the filter if the render is not sufficiently excited.
if (++poor_signal_excitation_counter_ < size_partitions ||
saturated_capture_signal || call_counter_ <= size_partitions) {
G->re.fill(0.f);
G->im.fill(0.f);
return;
}
// Compute mu.
std::array<float, kFftLengthBy2Plus1> mu;
const auto& X2 = render_power;
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
if (X2[k] > current_config_.noise_gate) {
mu[k] = current_config_.rate / X2[k];
} else {
mu[k] = 0.f;
}
}
// Avoid updating the filter close to narrow bands in the render signals.
render_signal_analyzer.MaskRegionsAroundNarrowBands(&mu);
// G = mu * E * X2.
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
G->re[k] = mu[k] * E_coarse.re[k];
G->im[k] = mu[k] * E_coarse.im[k];
}
}
void CoarseFilterUpdateGain::UpdateCurrentConfig() {
RTC_DCHECK_GE(config_change_duration_blocks_, config_change_counter_);
if (config_change_counter_ > 0) {
if (--config_change_counter_ > 0) {
auto average = [](float from, float to, float from_weight) {
return from * from_weight + to * (1.f - from_weight);
};
float change_factor =
config_change_counter_ * one_by_config_change_duration_blocks_;
current_config_.rate =
average(old_target_config_.rate, target_config_.rate, change_factor);
current_config_.noise_gate =
average(old_target_config_.noise_gate, target_config_.noise_gate,
change_factor);
} else {
current_config_ = old_target_config_ = target_config_;
}
}
RTC_DCHECK_LE(0, config_change_counter_);
}
} // namespace webrtc

74
webrtc/modules/audio_processing/aec3/coarse_filter_update_gain.h Normal file
View File

@ -0,0 +1,74 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_COARSE_FILTER_UPDATE_GAIN_H_
#define MODULES_AUDIO_PROCESSING_AEC3_COARSE_FILTER_UPDATE_GAIN_H_
#include <stddef.h>
#include <array>
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/fft_data.h"
#include "modules/audio_processing/aec3/render_signal_analyzer.h"
namespace webrtc {
// Provides functionality for computing the fixed gain for the coarse filter.
class CoarseFilterUpdateGain {
public:
explicit CoarseFilterUpdateGain(
const EchoCanceller3Config::Filter::CoarseConfiguration& config,
size_t config_change_duration_blocks);
// Takes action in the case of a known echo path change.
void HandleEchoPathChange();
// Computes the gain.
void Compute(const std::array<float, kFftLengthBy2Plus1>& render_power,
const RenderSignalAnalyzer& render_signal_analyzer,
const FftData& E_coarse,
size_t size_partitions,
bool saturated_capture_signal,
FftData* G);
// Sets a new config.
void SetConfig(
const EchoCanceller3Config::Filter::CoarseConfiguration& config,
bool immediate_effect) {
if (immediate_effect) {
old_target_config_ = current_config_ = target_config_ = config;
config_change_counter_ = 0;
} else {
old_target_config_ = current_config_;
target_config_ = config;
config_change_counter_ = config_change_duration_blocks_;
}
}
private:
EchoCanceller3Config::Filter::CoarseConfiguration current_config_;
EchoCanceller3Config::Filter::CoarseConfiguration target_config_;
EchoCanceller3Config::Filter::CoarseConfiguration old_target_config_;
const int config_change_duration_blocks_;
float one_by_config_change_duration_blocks_;
// TODO(peah): Check whether this counter should instead be initialized to a
// large value.
size_t poor_signal_excitation_counter_ = 0;
size_t call_counter_ = 0;
int config_change_counter_ = 0;
void UpdateCurrentConfig();
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_COARSE_FILTER_UPDATE_GAIN_H_

186
webrtc/modules/audio_processing/aec3/comfort_noise_generator.cc Normal file
View File

@ -0,0 +1,186 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/comfort_noise_generator.h"
// Defines WEBRTC_ARCH_X86_FAMILY, used below.
#include "rtc_base/system/arch.h"
#if defined(WEBRTC_ARCH_X86_FAMILY)
#include <emmintrin.h>
#endif
#include <algorithm>
#include <array>
#include <cmath>
#include <cstdint>
#include <functional>
#include <numeric>
#include "common_audio/signal_processing/include/signal_processing_library.h"
#include "modules/audio_processing/aec3/vector_math.h"
#include "rtc_base/checks.h"
namespace webrtc {
namespace {
// Computes the noise floor value that matches a WGN input of noise_floor_dbfs.
float GetNoiseFloorFactor(float noise_floor_dbfs) {
// kdBfsNormalization = 20.f*log10(32768.f).
constexpr float kdBfsNormalization = 90.30899869919436f;
return 64.f * powf(10.f, (kdBfsNormalization + noise_floor_dbfs) * 0.1f);
}
// Table of sqrt(2) * sin(2*pi*i/32).
constexpr float kSqrt2Sin[32] = {
+0.0000000f, +0.2758994f, +0.5411961f, +0.7856950f, +1.0000000f,
+1.1758756f, +1.3065630f, +1.3870398f, +1.4142136f, +1.3870398f,
+1.3065630f, +1.1758756f, +1.0000000f, +0.7856950f, +0.5411961f,
+0.2758994f, +0.0000000f, -0.2758994f, -0.5411961f, -0.7856950f,
-1.0000000f, -1.1758756f, -1.3065630f, -1.3870398f, -1.4142136f,
-1.3870398f, -1.3065630f, -1.1758756f, -1.0000000f, -0.7856950f,
-0.5411961f, -0.2758994f};
void GenerateComfortNoise(Aec3Optimization optimization,
const std::array<float, kFftLengthBy2Plus1>& N2,
uint32_t* seed,
FftData* lower_band_noise,
FftData* upper_band_noise) {
FftData* N_low = lower_band_noise;
FftData* N_high = upper_band_noise;
// Compute square root spectrum.
std::array<float, kFftLengthBy2Plus1> N;
std::copy(N2.begin(), N2.end(), N.begin());
aec3::VectorMath(optimization).Sqrt(N);
// Compute the noise level for the upper bands.
constexpr float kOneByNumBands = 1.f / (kFftLengthBy2Plus1 / 2 + 1);
constexpr int kFftLengthBy2Plus1By2 = kFftLengthBy2Plus1 / 2;
const float high_band_noise_level =
std::accumulate(N.begin() + kFftLengthBy2Plus1By2, N.end(), 0.f) *
kOneByNumBands;
// The analysis and synthesis windowing cause loss of power when
// cross-fading the noise where frames are completely uncorrelated
// (generated with random phase), hence the factor sqrt(2).
// This is not the case for the speech signal where the input is overlapping
// (strong correlation).
N_low->re[0] = N_low->re[kFftLengthBy2] = N_high->re[0] =
N_high->re[kFftLengthBy2] = 0.f;
for (size_t k = 1; k < kFftLengthBy2; k++) {
constexpr int kIndexMask = 32 - 1;
// Generate a random 31-bit integer.
seed[0] = (seed[0] * 69069 + 1) & (0x80000000 - 1);
// Convert to a 5-bit index.
int i = seed[0] >> 26;
// y = sqrt(2) * sin(a)
const float x = kSqrt2Sin[i];
// x = sqrt(2) * cos(a) = sqrt(2) * sin(a + pi/2)
const float y = kSqrt2Sin[(i + 8) & kIndexMask];
// Form low-frequency noise via spectral shaping.
N_low->re[k] = N[k] * x;
N_low->im[k] = N[k] * y;
// Form the high-frequency noise via simple levelling.
N_high->re[k] = high_band_noise_level * x;
N_high->im[k] = high_band_noise_level * y;
}
}
} // namespace
ComfortNoiseGenerator::ComfortNoiseGenerator(const EchoCanceller3Config& config,
Aec3Optimization optimization,
size_t num_capture_channels)
: optimization_(optimization),
seed_(42),
num_capture_channels_(num_capture_channels),
noise_floor_(GetNoiseFloorFactor(config.comfort_noise.noise_floor_dbfs)),
N2_initial_(
std::make_unique<std::vector<std::array<float, kFftLengthBy2Plus1>>>(
num_capture_channels_)),
Y2_smoothed_(num_capture_channels_),
N2_(num_capture_channels_) {
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
(*N2_initial_)[ch].fill(0.f);
Y2_smoothed_[ch].fill(0.f);
N2_[ch].fill(1.0e6f);
}
}
ComfortNoiseGenerator::~ComfortNoiseGenerator() = default;
void ComfortNoiseGenerator::Compute(
bool saturated_capture,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
capture_spectrum,
rtc::ArrayView<FftData> lower_band_noise,
rtc::ArrayView<FftData> upper_band_noise) {
const auto& Y2 = capture_spectrum;
if (!saturated_capture) {
// Smooth Y2.
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
std::transform(Y2_smoothed_[ch].begin(), Y2_smoothed_[ch].end(),
Y2[ch].begin(), Y2_smoothed_[ch].begin(),
[](float a, float b) { return a + 0.1f * (b - a); });
}
if (N2_counter_ > 50) {
// Update N2 from Y2_smoothed.
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
std::transform(N2_[ch].begin(), N2_[ch].end(), Y2_smoothed_[ch].begin(),
N2_[ch].begin(), [](float a, float b) {
return b < a ? (0.9f * b + 0.1f * a) * 1.0002f
: a * 1.0002f;
});
}
}
if (N2_initial_) {
if (++N2_counter_ == 1000) {
N2_initial_.reset();
} else {
// Compute the N2_initial from N2.
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
std::transform(N2_[ch].begin(), N2_[ch].end(),
(*N2_initial_)[ch].begin(), (*N2_initial_)[ch].begin(),
[](float a, float b) {
return a > b ? b + 0.001f * (a - b) : a;
});
}
}
}
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
for (auto& n : N2_[ch]) {
n = std::max(n, noise_floor_);
}
if (N2_initial_) {
for (auto& n : (*N2_initial_)[ch]) {
n = std::max(n, noise_floor_);
}
}
}
}
// Choose N2 estimate to use.
const auto& N2 = N2_initial_ ? (*N2_initial_) : N2_;
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
GenerateComfortNoise(optimization_, N2[ch], &seed_, &lower_band_noise[ch],
&upper_band_noise[ch]);
}
}
} // namespace webrtc

78
webrtc/modules/audio_processing/aec3/comfort_noise_generator.h Normal file
View File

@ -0,0 +1,78 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_COMFORT_NOISE_GENERATOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_COMFORT_NOISE_GENERATOR_H_
#include <stdint.h>
#include <array>
#include <memory>
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/aec_state.h"
#include "modules/audio_processing/aec3/fft_data.h"
#include "rtc_base/constructor_magic.h"
#include "rtc_base/system/arch.h"
namespace webrtc {
namespace aec3 {
#if defined(WEBRTC_ARCH_X86_FAMILY)
void EstimateComfortNoise_SSE2(const std::array<float, kFftLengthBy2Plus1>& N2,
uint32_t* seed,
FftData* lower_band_noise,
FftData* upper_band_noise);
#endif
void EstimateComfortNoise(const std::array<float, kFftLengthBy2Plus1>& N2,
uint32_t* seed,
FftData* lower_band_noise,
FftData* upper_band_noise);
} // namespace aec3
// Generates the comfort noise.
class ComfortNoiseGenerator {
public:
ComfortNoiseGenerator(const EchoCanceller3Config& config,
Aec3Optimization optimization,
size_t num_capture_channels);
ComfortNoiseGenerator() = delete;
~ComfortNoiseGenerator();
ComfortNoiseGenerator(const ComfortNoiseGenerator&) = delete;
// Computes the comfort noise.
void Compute(bool saturated_capture,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
capture_spectrum,
rtc::ArrayView<FftData> lower_band_noise,
rtc::ArrayView<FftData> upper_band_noise);
// Returns the estimate of the background noise spectrum.
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> NoiseSpectrum()
const {
return N2_;
}
private:
const Aec3Optimization optimization_;
uint32_t seed_;
const size_t num_capture_channels_;
const float noise_floor_;
std::unique_ptr<std::vector<std::array<float, kFftLengthBy2Plus1>>>
N2_initial_;
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2_smoothed_;
std::vector<std::array<float, kFftLengthBy2Plus1>> N2_;
int N2_counter_ = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_COMFORT_NOISE_GENERATOR_H_

91
webrtc/modules/audio_processing/aec3/decimator.cc Normal file
View File

@ -0,0 +1,91 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/decimator.h"
#include <array>
#include <vector>
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/checks.h"
namespace webrtc {
namespace {
// signal.butter(2, 3400/8000.0, 'lowpass', analog=False)
const std::vector<CascadedBiQuadFilter::BiQuadParam> GetLowPassFilterDS2() {
return std::vector<CascadedBiQuadFilter::BiQuadParam>{
{{-1.f, 0.f}, {0.13833231f, 0.40743176f}, 0.22711796393486466f},
{{-1.f, 0.f}, {0.13833231f, 0.40743176f}, 0.22711796393486466f},
{{-1.f, 0.f}, {0.13833231f, 0.40743176f}, 0.22711796393486466f}};
}
// signal.ellip(6, 1, 40, 1800/8000, btype='lowpass', analog=False)
const std::vector<CascadedBiQuadFilter::BiQuadParam> GetLowPassFilterDS4() {
return std::vector<CascadedBiQuadFilter::BiQuadParam>{
{{-0.08873842f, 0.99605496f}, {0.75916227f, 0.23841065f}, 0.26250696827f},
{{0.62273832f, 0.78243018f}, {0.74892112f, 0.5410152f}, 0.26250696827f},
{{0.71107693f, 0.70311421f}, {0.74895534f, 0.63924616f}, 0.26250696827f}};
}
// signal.cheby1(1, 6, [1000/8000, 2000/8000], btype='bandpass', analog=False)
const std::vector<CascadedBiQuadFilter::BiQuadParam> GetBandPassFilterDS8() {
return std::vector<CascadedBiQuadFilter::BiQuadParam>{
{{1.f, 0.f}, {0.7601815f, 0.46423542f}, 0.10330478266505948f, true},
{{1.f, 0.f}, {0.7601815f, 0.46423542f}, 0.10330478266505948f, true},
{{1.f, 0.f}, {0.7601815f, 0.46423542f}, 0.10330478266505948f, true},
{{1.f, 0.f}, {0.7601815f, 0.46423542f}, 0.10330478266505948f, true},
{{1.f, 0.f}, {0.7601815f, 0.46423542f}, 0.10330478266505948f, true}};
}
// signal.butter(2, 1000/8000.0, 'highpass', analog=False)
const std::vector<CascadedBiQuadFilter::BiQuadParam> GetHighPassFilter() {
return std::vector<CascadedBiQuadFilter::BiQuadParam>{
{{1.f, 0.f}, {0.72712179f, 0.21296904f}, 0.7570763753338849f}};
}
const std::vector<CascadedBiQuadFilter::BiQuadParam> GetPassThroughFilter() {
return std::vector<CascadedBiQuadFilter::BiQuadParam>{};
}
} // namespace
Decimator::Decimator(size_t down_sampling_factor)
: down_sampling_factor_(down_sampling_factor),
anti_aliasing_filter_(down_sampling_factor_ == 4
? GetLowPassFilterDS4()
: (down_sampling_factor_ == 8
? GetBandPassFilterDS8()
: GetLowPassFilterDS2())),
noise_reduction_filter_(down_sampling_factor_ == 8
? GetPassThroughFilter()
: GetHighPassFilter()) {
RTC_DCHECK(down_sampling_factor_ == 2 || down_sampling_factor_ == 4 ||
down_sampling_factor_ == 8);
}
void Decimator::Decimate(rtc::ArrayView<const float> in,
rtc::ArrayView<float> out) {
RTC_DCHECK_EQ(kBlockSize, in.size());
RTC_DCHECK_EQ(kBlockSize / down_sampling_factor_, out.size());
std::array<float, kBlockSize> x;
// Limit the frequency content of the signal to avoid aliasing.
anti_aliasing_filter_.Process(in, x);
// Reduce the impact of near-end noise.
noise_reduction_filter_.Process(x);
// Downsample the signal.
for (size_t j = 0, k = 0; j < out.size(); ++j, k += down_sampling_factor_) {
RTC_DCHECK_GT(kBlockSize, k);
out[j] = x[k];
}
}
} // namespace webrtc

41
webrtc/modules/audio_processing/aec3/decimator.h Normal file
View File

@ -0,0 +1,41 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_DECIMATOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_DECIMATOR_H_
#include <array>
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/utility/cascaded_biquad_filter.h"
#include "rtc_base/constructor_magic.h"
namespace webrtc {
// Provides functionality for decimating a signal.
class Decimator {
public:
explicit Decimator(size_t down_sampling_factor);
// Downsamples the signal.
void Decimate(rtc::ArrayView<const float> in, rtc::ArrayView<float> out);
private:
const size_t down_sampling_factor_;
CascadedBiQuadFilter anti_aliasing_filter_;
CascadedBiQuadFilter noise_reduction_filter_;
RTC_DISALLOW_COPY_AND_ASSIGN(Decimator);
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_DECIMATOR_H_

31
webrtc/modules/audio_processing/aec3/delay_estimate.h Normal file
View File

@ -0,0 +1,31 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_DELAY_ESTIMATE_H_
#define MODULES_AUDIO_PROCESSING_AEC3_DELAY_ESTIMATE_H_
namespace webrtc {
// Stores delay_estimates.
struct DelayEstimate {
enum class Quality { kCoarse, kRefined };
DelayEstimate(Quality quality, size_t delay)
: quality(quality), delay(delay) {}
Quality quality;
size_t delay;
size_t blocks_since_last_change = 0;
size_t blocks_since_last_update = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_DELAY_ESTIMATE_H_

75
webrtc/modules/audio_processing/aec3/dominant_nearend_detector.cc Normal file
View File

@ -0,0 +1,75 @@
/*
* Copyright (c) 2019 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/dominant_nearend_detector.h"
#include <numeric>
namespace webrtc {
DominantNearendDetector::DominantNearendDetector(
const EchoCanceller3Config::Suppressor::DominantNearendDetection& config,
size_t num_capture_channels)
: enr_threshold_(config.enr_threshold),
enr_exit_threshold_(config.enr_exit_threshold),
snr_threshold_(config.snr_threshold),
hold_duration_(config.hold_duration),
trigger_threshold_(config.trigger_threshold),
use_during_initial_phase_(config.use_during_initial_phase),
num_capture_channels_(num_capture_channels),
trigger_counters_(num_capture_channels_),
hold_counters_(num_capture_channels_) {}
void DominantNearendDetector::Update(
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
nearend_spectrum,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
residual_echo_spectrum,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
comfort_noise_spectrum,
bool initial_state) {
nearend_state_ = false;
auto low_frequency_energy = [](rtc::ArrayView<const float> spectrum) {
RTC_DCHECK_LE(16, spectrum.size());
return std::accumulate(spectrum.begin() + 1, spectrum.begin() + 16, 0.f);
};
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
const float ne_sum = low_frequency_energy(nearend_spectrum[ch]);
const float echo_sum = low_frequency_energy(residual_echo_spectrum[ch]);
const float noise_sum = low_frequency_energy(comfort_noise_spectrum[ch]);
// Detect strong active nearend if the nearend is sufficiently stronger than
// the echo and the nearend noise.
if ((!initial_state || use_during_initial_phase_) &&
echo_sum < enr_threshold_ * ne_sum &&
ne_sum > snr_threshold_ * noise_sum) {
if (++trigger_counters_[ch] >= trigger_threshold_) {
// After a period of strong active nearend activity, flag nearend mode.
hold_counters_[ch] = hold_duration_;
trigger_counters_[ch] = trigger_threshold_;
}
} else {
// Forget previously detected strong active nearend activity.
trigger_counters_[ch] = std::max(0, trigger_counters_[ch] - 1);
}
// Exit nearend-state early at strong echo.
if (echo_sum > enr_exit_threshold_ * ne_sum &&
echo_sum > snr_threshold_ * noise_sum) {
hold_counters_[ch] = 0;
}
// Remain in any nearend mode for a certain duration.
hold_counters_[ch] = std::max(0, hold_counters_[ch] - 1);
nearend_state_ = nearend_state_ || hold_counters_[ch] > 0;
}
}
} // namespace webrtc

56
webrtc/modules/audio_processing/aec3/dominant_nearend_detector.h Normal file
View File

@ -0,0 +1,56 @@
/*
* Copyright (c) 2019 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_DOMINANT_NEAREND_DETECTOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_DOMINANT_NEAREND_DETECTOR_H_
#include <vector>
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/nearend_detector.h"
namespace webrtc {
// Class for selecting whether the suppressor is in the nearend or echo state.
class DominantNearendDetector : public NearendDetector {
public:
DominantNearendDetector(
const EchoCanceller3Config::Suppressor::DominantNearendDetection& config,
size_t num_capture_channels);
// Returns whether the current state is the nearend state.
bool IsNearendState() const override { return nearend_state_; }
// Updates the state selection based on latest spectral estimates.
void Update(rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
nearend_spectrum,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
residual_echo_spectrum,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
comfort_noise_spectrum,
bool initial_state) override;
private:
const float enr_threshold_;
const float enr_exit_threshold_;
const float snr_threshold_;
const int hold_duration_;
const int trigger_threshold_;
const bool use_during_initial_phase_;
const size_t num_capture_channels_;
bool nearend_state_ = false;
std::vector<int> trigger_counters_;
std::vector<int> hold_counters_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_DOMINANT_NEAREND_DETECTOR_H_

25
webrtc/modules/audio_processing/aec3/downsampled_render_buffer.cc Normal file
View File

@ -0,0 +1,25 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/downsampled_render_buffer.h"
#include <algorithm>
namespace webrtc {
DownsampledRenderBuffer::DownsampledRenderBuffer(size_t downsampled_buffer_size)
: size(static_cast<int>(downsampled_buffer_size)),
buffer(downsampled_buffer_size, 0.f) {
std::fill(buffer.begin(), buffer.end(), 0.f);
}
DownsampledRenderBuffer::~DownsampledRenderBuffer() = default;
} // namespace webrtc

58
webrtc/modules/audio_processing/aec3/downsampled_render_buffer.h Normal file
View File

@ -0,0 +1,58 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_DOWNSAMPLED_RENDER_BUFFER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_DOWNSAMPLED_RENDER_BUFFER_H_
#include <stddef.h>
#include <vector>
#include "rtc_base/checks.h"
namespace webrtc {
// Holds the circular buffer of the downsampled render data.
struct DownsampledRenderBuffer {
explicit DownsampledRenderBuffer(size_t downsampled_buffer_size);
~DownsampledRenderBuffer();
int IncIndex(int index) const {
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
return index < size - 1 ? index + 1 : 0;
}
int DecIndex(int index) const {
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
return index > 0 ? index - 1 : size - 1;
}
int OffsetIndex(int index, int offset) const {
RTC_DCHECK_GE(buffer.size(), offset);
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
return (size + index + offset) % size;
}
void UpdateWriteIndex(int offset) { write = OffsetIndex(write, offset); }
void IncWriteIndex() { write = IncIndex(write); }
void DecWriteIndex() { write = DecIndex(write); }
void UpdateReadIndex(int offset) { read = OffsetIndex(read, offset); }
void IncReadIndex() { read = IncIndex(read); }
void DecReadIndex() { read = DecIndex(read); }
const int size;
std::vector<float> buffer;
int write = 0;
int read = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_DOWNSAMPLED_RENDER_BUFFER_H_

118
webrtc/modules/audio_processing/aec3/echo_audibility.cc Normal file
View File

@ -0,0 +1,118 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/echo_audibility.h"
#include <algorithm>
#include <cmath>
#include <utility>
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/block_buffer.h"
#include "modules/audio_processing/aec3/spectrum_buffer.h"
#include "modules/audio_processing/aec3/stationarity_estimator.h"
namespace webrtc {
EchoAudibility::EchoAudibility(bool use_render_stationarity_at_init)
: use_render_stationarity_at_init_(use_render_stationarity_at_init) {
Reset();
}
EchoAudibility::~EchoAudibility() = default;
void EchoAudibility::Update(const RenderBuffer& render_buffer,
rtc::ArrayView<const float> average_reverb,
int delay_blocks,
bool external_delay_seen) {
UpdateRenderNoiseEstimator(render_buffer.GetSpectrumBuffer(),
render_buffer.GetBlockBuffer(),
external_delay_seen);
if (external_delay_seen || use_render_stationarity_at_init_) {
UpdateRenderStationarityFlags(render_buffer, average_reverb, delay_blocks);
}
}
void EchoAudibility::Reset() {
render_stationarity_.Reset();
non_zero_render_seen_ = false;
render_spectrum_write_prev_ = absl::nullopt;
}
void EchoAudibility::UpdateRenderStationarityFlags(
const RenderBuffer& render_buffer,
rtc::ArrayView<const float> average_reverb,
int min_channel_delay_blocks) {
const SpectrumBuffer& spectrum_buffer = render_buffer.GetSpectrumBuffer();
int idx_at_delay = spectrum_buffer.OffsetIndex(spectrum_buffer.read,
min_channel_delay_blocks);
int num_lookahead = render_buffer.Headroom() - min_channel_delay_blocks + 1;
num_lookahead = std::max(0, num_lookahead);
render_stationarity_.UpdateStationarityFlags(spectrum_buffer, average_reverb,
idx_at_delay, num_lookahead);
}
void EchoAudibility::UpdateRenderNoiseEstimator(
const SpectrumBuffer& spectrum_buffer,
const BlockBuffer& block_buffer,
bool external_delay_seen) {
if (!render_spectrum_write_prev_) {
render_spectrum_write_prev_ = spectrum_buffer.write;
render_block_write_prev_ = block_buffer.write;
return;
}
int render_spectrum_write_current = spectrum_buffer.write;
if (!non_zero_render_seen_ && !external_delay_seen) {
non_zero_render_seen_ = !IsRenderTooLow(block_buffer);
}
if (non_zero_render_seen_) {
for (int idx = render_spectrum_write_prev_.value();
idx != render_spectrum_write_current;
idx = spectrum_buffer.DecIndex(idx)) {
render_stationarity_.UpdateNoiseEstimator(spectrum_buffer.buffer[idx]);
}
}
render_spectrum_write_prev_ = render_spectrum_write_current;
}
bool EchoAudibility::IsRenderTooLow(const BlockBuffer& block_buffer) {
const int num_render_channels =
static_cast<int>(block_buffer.buffer[0][0].size());
bool too_low = false;
const int render_block_write_current = block_buffer.write;
if (render_block_write_current == render_block_write_prev_) {
too_low = true;
} else {
for (int idx = render_block_write_prev_; idx != render_block_write_current;
idx = block_buffer.IncIndex(idx)) {
float max_abs_over_channels = 0.f;
for (int ch = 0; ch < num_render_channels; ++ch) {
auto block = block_buffer.buffer[idx][0][ch];
auto r = std::minmax_element(block.cbegin(), block.cend());
float max_abs_channel =
std::max(std::fabs(*r.first), std::fabs(*r.second));
max_abs_over_channels =
std::max(max_abs_over_channels, max_abs_channel);
}
if (max_abs_over_channels < 10.f) {
too_low = true; // Discards all blocks if one of them is too low.
break;
}
}
}
render_block_write_prev_ = render_block_write_current;
return too_low;
}
} // namespace webrtc

86
webrtc/modules/audio_processing/aec3/echo_audibility.h Normal file
View File

@ -0,0 +1,86 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ECHO_AUDIBILITY_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ECHO_AUDIBILITY_H_
#include <stddef.h>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "modules/audio_processing/aec3/block_buffer.h"
#include "modules/audio_processing/aec3/render_buffer.h"
#include "modules/audio_processing/aec3/spectrum_buffer.h"
#include "modules/audio_processing/aec3/stationarity_estimator.h"
#include "rtc_base/constructor_magic.h"
namespace webrtc {
class EchoAudibility {
public:
explicit EchoAudibility(bool use_render_stationarity_at_init);
~EchoAudibility();
EchoAudibility(const EchoAudibility&) = delete;
EchoAudibility& operator=(const EchoAudibility&) = delete;
// Feed new render data to the echo audibility estimator.
void Update(const RenderBuffer& render_buffer,
rtc::ArrayView<const float> average_reverb,
int min_channel_delay_blocks,
bool external_delay_seen);
// Get the residual echo scaling.
void GetResidualEchoScaling(bool filter_has_had_time_to_converge,
rtc::ArrayView<float> residual_scaling) const {
for (size_t band = 0; band < residual_scaling.size(); ++band) {
if (render_stationarity_.IsBandStationary(band) &&
(filter_has_had_time_to_converge ||
use_render_stationarity_at_init_)) {
residual_scaling[band] = 0.f;
} else {
residual_scaling[band] = 1.0f;
}
}
}
// Returns true if the current render block is estimated as stationary.
bool IsBlockStationary() const {
return render_stationarity_.IsBlockStationary();
}
private:
// Reset the EchoAudibility class.
void Reset();
// Updates the render stationarity flags for the current frame.
void UpdateRenderStationarityFlags(const RenderBuffer& render_buffer,
rtc::ArrayView<const float> average_reverb,
int delay_blocks);
// Updates the noise estimator with the new render data since the previous
// call to this method.
void UpdateRenderNoiseEstimator(const SpectrumBuffer& spectrum_buffer,
const BlockBuffer& block_buffer,
bool external_delay_seen);
// Returns a bool being true if the render signal contains just close to zero
// values.
bool IsRenderTooLow(const BlockBuffer& block_buffer);
absl::optional<int> render_spectrum_write_prev_;
int render_block_write_prev_;
bool non_zero_render_seen_;
const bool use_render_stationarity_at_init_;
StationarityEstimator render_stationarity_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ECHO_AUDIBILITY_H_

868
webrtc/modules/audio_processing/aec3/echo_canceller3.cc Normal file
View File

@ -0,0 +1,868 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/echo_canceller3.h"
#include <algorithm>
#include <utility>
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/high_pass_filter.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomic_ops.h"
#include "rtc_base/experiments/field_trial_parser.h"
#include "rtc_base/logging.h"
#include "system_wrappers/include/field_trial.h"
namespace webrtc {
namespace {
enum class EchoCanceller3ApiCall { kCapture, kRender };
bool DetectSaturation(rtc::ArrayView<const float> y) {
for (auto y_k : y) {
if (y_k >= 32700.0f || y_k <= -32700.0f) {
return true;
}
}
return false;
}
// Retrieves a value from a field trial if it is available. If no value is
// present, the default value is returned. If the retrieved value is beyond the
// specified limits, the default value is returned instead.
void RetrieveFieldTrialValue(const char* trial_name,
float min,
float max,
float* value_to_update) {
const std::string field_trial_str = field_trial::FindFullName(trial_name);
FieldTrialParameter<double> field_trial_param(/*key=*/"", *value_to_update);
ParseFieldTrial({&field_trial_param}, field_trial_str);
float field_trial_value = static_cast<float>(field_trial_param.Get());
if (field_trial_value >= min && field_trial_value <= max) {
*value_to_update = field_trial_value;
}
}
void RetrieveFieldTrialValue(const char* trial_name,
int min,
int max,
int* value_to_update) {
const std::string field_trial_str = field_trial::FindFullName(trial_name);
FieldTrialParameter<int> field_trial_param(/*key=*/"", *value_to_update);
ParseFieldTrial({&field_trial_param}, field_trial_str);
float field_trial_value = field_trial_param.Get();
if (field_trial_value >= min && field_trial_value <= max) {
*value_to_update = field_trial_value;
}
}
void FillSubFrameView(
AudioBuffer* frame,
size_t sub_frame_index,
std::vector<std::vector<rtc::ArrayView<float>>>* sub_frame_view) {
RTC_DCHECK_GE(1, sub_frame_index);
RTC_DCHECK_LE(0, sub_frame_index);
RTC_DCHECK_EQ(frame->num_bands(), sub_frame_view->size());
RTC_DCHECK_EQ(frame->num_channels(), (*sub_frame_view)[0].size());
for (size_t band = 0; band < sub_frame_view->size(); ++band) {
for (size_t channel = 0; channel < (*sub_frame_view)[0].size(); ++channel) {
(*sub_frame_view)[band][channel] = rtc::ArrayView<float>(
&frame->split_bands(channel)[band][sub_frame_index * kSubFrameLength],
kSubFrameLength);
}
}
}
void FillSubFrameView(
std::vector<std::vector<std::vector<float>>>* frame,
size_t sub_frame_index,
std::vector<std::vector<rtc::ArrayView<float>>>* sub_frame_view) {
RTC_DCHECK_GE(1, sub_frame_index);
RTC_DCHECK_EQ(frame->size(), sub_frame_view->size());
RTC_DCHECK_EQ((*frame)[0].size(), (*sub_frame_view)[0].size());
for (size_t band = 0; band < frame->size(); ++band) {
for (size_t channel = 0; channel < (*frame)[band].size(); ++channel) {
(*sub_frame_view)[band][channel] = rtc::ArrayView<float>(
&(*frame)[band][channel][sub_frame_index * kSubFrameLength],
kSubFrameLength);
}
}
}
void ProcessCaptureFrameContent(
AudioBuffer* linear_output,
AudioBuffer* capture,
bool level_change,
bool saturated_microphone_signal,
size_t sub_frame_index,
FrameBlocker* capture_blocker,
BlockFramer* linear_output_framer,
BlockFramer* output_framer,
BlockProcessor* block_processor,
std::vector<std::vector<std::vector<float>>>* linear_output_block,
std::vector<std::vector<rtc::ArrayView<float>>>*
linear_output_sub_frame_view,
std::vector<std::vector<std::vector<float>>>* capture_block,
std::vector<std::vector<rtc::ArrayView<float>>>* capture_sub_frame_view) {
FillSubFrameView(capture, sub_frame_index, capture_sub_frame_view);
if (linear_output) {
RTC_DCHECK(linear_output_framer);
RTC_DCHECK(linear_output_block);
RTC_DCHECK(linear_output_sub_frame_view);
FillSubFrameView(linear_output, sub_frame_index,
linear_output_sub_frame_view);
}
capture_blocker->InsertSubFrameAndExtractBlock(*capture_sub_frame_view,
capture_block);
block_processor->ProcessCapture(level_change, saturated_microphone_signal,
linear_output_block, capture_block);
output_framer->InsertBlockAndExtractSubFrame(*capture_block,
capture_sub_frame_view);
if (linear_output) {
RTC_DCHECK(linear_output_framer);
linear_output_framer->InsertBlockAndExtractSubFrame(
*linear_output_block, linear_output_sub_frame_view);
}
}
void ProcessRemainingCaptureFrameContent(
bool level_change,
bool saturated_microphone_signal,
FrameBlocker* capture_blocker,
BlockFramer* linear_output_framer,
BlockFramer* output_framer,
BlockProcessor* block_processor,
std::vector<std::vector<std::vector<float>>>* linear_output_block,
std::vector<std::vector<std::vector<float>>>* block) {
if (!capture_blocker->IsBlockAvailable()) {
return;
}
capture_blocker->ExtractBlock(block);
block_processor->ProcessCapture(level_change, saturated_microphone_signal,
linear_output_block, block);
output_framer->InsertBlock(*block);
if (linear_output_framer) {
RTC_DCHECK(linear_output_block);
linear_output_framer->InsertBlock(*linear_output_block);
}
}
void BufferRenderFrameContent(
std::vector<std::vector<std::vector<float>>>* render_frame,
size_t sub_frame_index,
FrameBlocker* render_blocker,
BlockProcessor* block_processor,
std::vector<std::vector<std::vector<float>>>* block,
std::vector<std::vector<rtc::ArrayView<float>>>* sub_frame_view) {
FillSubFrameView(render_frame, sub_frame_index, sub_frame_view);
render_blocker->InsertSubFrameAndExtractBlock(*sub_frame_view, block);
block_processor->BufferRender(*block);
}
void BufferRemainingRenderFrameContent(
FrameBlocker* render_blocker,
BlockProcessor* block_processor,
std::vector<std::vector<std::vector<float>>>* block) {
if (!render_blocker->IsBlockAvailable()) {
return;
}
render_blocker->ExtractBlock(block);
block_processor->BufferRender(*block);
}
void CopyBufferIntoFrame(const AudioBuffer& buffer,
size_t num_bands,
size_t num_channels,
std::vector<std::vector<std::vector<float>>>* frame) {
RTC_DCHECK_EQ(num_bands, frame->size());
RTC_DCHECK_EQ(num_channels, (*frame)[0].size());
RTC_DCHECK_EQ(AudioBuffer::kSplitBandSize, (*frame)[0][0].size());
for (size_t band = 0; band < num_bands; ++band) {
for (size_t channel = 0; channel < num_channels; ++channel) {
rtc::ArrayView<const float> buffer_view(
&buffer.split_bands_const(channel)[band][0],
AudioBuffer::kSplitBandSize);
std::copy(buffer_view.begin(), buffer_view.end(),
(*frame)[band][channel].begin());
}
}
}
} // namespace
// TODO(webrtc:5298): Move this to a separate file.
EchoCanceller3Config AdjustConfig(const EchoCanceller3Config& config) {
EchoCanceller3Config adjusted_cfg = config;
if (field_trial::IsEnabled("WebRTC-Aec3AntiHowlingMinimizationKillSwitch")) {
adjusted_cfg.suppressor.high_bands_suppression
.anti_howling_activation_threshold = 25.f;
adjusted_cfg.suppressor.high_bands_suppression.anti_howling_gain = 0.01f;
}
if (field_trial::IsEnabled("WebRTC-Aec3UseShortConfigChangeDuration")) {
adjusted_cfg.filter.config_change_duration_blocks = 10;
}
if (field_trial::IsEnabled("WebRTC-Aec3UseZeroInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = 0.f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3UseDot1SecondsInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = .1f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3UseDot2SecondsInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = .2f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3UseDot3SecondsInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = .3f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3UseDot6SecondsInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = .6f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3UseDot9SecondsInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = .9f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3Use1Dot2SecondsInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = 1.2f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3Use1Dot6SecondsInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = 1.6f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3Use2Dot0SecondsInitialStateDuration")) {
adjusted_cfg.filter.initial_state_seconds = 2.0f;
}
if (field_trial::IsEnabled("WebRTC-Aec3EchoSaturationDetectionKillSwitch")) {
adjusted_cfg.ep_strength.echo_can_saturate = false;
}
if (field_trial::IsEnabled("WebRTC-Aec3UseDot2ReverbDefaultLen")) {
adjusted_cfg.ep_strength.default_len = 0.2f;
} else if (field_trial::IsEnabled("WebRTC-Aec3UseDot3ReverbDefaultLen")) {
adjusted_cfg.ep_strength.default_len = 0.3f;
} else if (field_trial::IsEnabled("WebRTC-Aec3UseDot4ReverbDefaultLen")) {
adjusted_cfg.ep_strength.default_len = 0.4f;
} else if (field_trial::IsEnabled("WebRTC-Aec3UseDot5ReverbDefaultLen")) {
adjusted_cfg.ep_strength.default_len = 0.5f;
} else if (field_trial::IsEnabled("WebRTC-Aec3UseDot6ReverbDefaultLen")) {
adjusted_cfg.ep_strength.default_len = 0.6f;
} else if (field_trial::IsEnabled("WebRTC-Aec3UseDot7ReverbDefaultLen")) {
adjusted_cfg.ep_strength.default_len = 0.7f;
} else if (field_trial::IsEnabled("WebRTC-Aec3UseDot8ReverbDefaultLen")) {
adjusted_cfg.ep_strength.default_len = 0.8f;
}
if (field_trial::IsEnabled("WebRTC-Aec3ShortHeadroomKillSwitch")) {
// Two blocks headroom.
adjusted_cfg.delay.delay_headroom_samples = kBlockSize * 2;
}
if (field_trial::IsEnabled("WebRTC-Aec3ClampInstQualityToZeroKillSwitch")) {
adjusted_cfg.erle.clamp_quality_estimate_to_zero = false;
}
if (field_trial::IsEnabled("WebRTC-Aec3ClampInstQualityToOneKillSwitch")) {
adjusted_cfg.erle.clamp_quality_estimate_to_one = false;
}
if (field_trial::IsEnabled("WebRTC-Aec3OnsetDetectionKillSwitch")) {
adjusted_cfg.erle.onset_detection = false;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceRenderDelayEstimationDownmixing")) {
adjusted_cfg.delay.render_alignment_mixing.downmix = true;
adjusted_cfg.delay.render_alignment_mixing.adaptive_selection = false;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceCaptureDelayEstimationDownmixing")) {
adjusted_cfg.delay.capture_alignment_mixing.downmix = true;
adjusted_cfg.delay.capture_alignment_mixing.adaptive_selection = false;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceCaptureDelayEstimationLeftRightPrioritization")) {
adjusted_cfg.delay.capture_alignment_mixing.prefer_first_two_channels =
true;
}
if (field_trial::IsEnabled(
"WebRTC-"
"Aec3RenderDelayEstimationLeftRightPrioritizationKillSwitch")) {
adjusted_cfg.delay.capture_alignment_mixing.prefer_first_two_channels =
false;
}
if (field_trial::IsEnabled("WebRTC-Aec3SensitiveDominantNearendActivation")) {
adjusted_cfg.suppressor.dominant_nearend_detection.enr_threshold = 0.5f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3VerySensitiveDominantNearendActivation")) {
adjusted_cfg.suppressor.dominant_nearend_detection.enr_threshold = 0.75f;
}
if (field_trial::IsEnabled("WebRTC-Aec3TransparentAntiHowlingGain")) {
adjusted_cfg.suppressor.high_bands_suppression.anti_howling_gain = 1.f;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceMoreTransparentNormalSuppressorTuning")) {
adjusted_cfg.suppressor.normal_tuning.mask_lf.enr_transparent = 0.4f;
adjusted_cfg.suppressor.normal_tuning.mask_lf.enr_suppress = 0.5f;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceMoreTransparentNearendSuppressorTuning")) {
adjusted_cfg.suppressor.nearend_tuning.mask_lf.enr_transparent = 1.29f;
adjusted_cfg.suppressor.nearend_tuning.mask_lf.enr_suppress = 1.3f;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceMoreTransparentNormalSuppressorHfTuning")) {
adjusted_cfg.suppressor.normal_tuning.mask_hf.enr_transparent = 0.3f;
adjusted_cfg.suppressor.normal_tuning.mask_hf.enr_suppress = 0.4f;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceMoreTransparentNearendSuppressorHfTuning")) {
adjusted_cfg.suppressor.nearend_tuning.mask_hf.enr_transparent = 1.09f;
adjusted_cfg.suppressor.nearend_tuning.mask_hf.enr_suppress = 1.1f;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceRapidlyAdjustingNormalSuppressorTunings")) {
adjusted_cfg.suppressor.normal_tuning.max_inc_factor = 2.5f;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceRapidlyAdjustingNearendSuppressorTunings")) {
adjusted_cfg.suppressor.nearend_tuning.max_inc_factor = 2.5f;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceSlowlyAdjustingNormalSuppressorTunings")) {
adjusted_cfg.suppressor.normal_tuning.max_dec_factor_lf = .2f;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceSlowlyAdjustingNearendSuppressorTunings")) {
adjusted_cfg.suppressor.nearend_tuning.max_dec_factor_lf = .2f;
}
if (field_trial::IsEnabled("WebRTC-Aec3EnforceStationarityProperties")) {
adjusted_cfg.echo_audibility.use_stationarity_properties = true;
}
if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceStationarityPropertiesAtInit")) {
adjusted_cfg.echo_audibility.use_stationarity_properties_at_init = true;
}
if (field_trial::IsEnabled("WebRTC-Aec3EnforceLowActiveRenderLimit")) {
adjusted_cfg.render_levels.active_render_limit = 50.f;
} else if (field_trial::IsEnabled(
"WebRTC-Aec3EnforceVeryLowActiveRenderLimit")) {
adjusted_cfg.render_levels.active_render_limit = 30.f;
}
if (field_trial::IsEnabled("WebRTC-Aec3NonlinearModeReverbKillSwitch")) {
adjusted_cfg.echo_model.model_reverb_in_nonlinear_mode = false;
}
// Field-trial based override for the whole suppressor tuning.
const std::string suppressor_tuning_override_trial_name =
field_trial::FindFullName("WebRTC-Aec3SuppressorTuningOverride");
FieldTrialParameter<double> nearend_tuning_mask_lf_enr_transparent(
"nearend_tuning_mask_lf_enr_transparent",
adjusted_cfg.suppressor.nearend_tuning.mask_lf.enr_transparent);
FieldTrialParameter<double> nearend_tuning_mask_lf_enr_suppress(
"nearend_tuning_mask_lf_enr_suppress",
adjusted_cfg.suppressor.nearend_tuning.mask_lf.enr_suppress);
FieldTrialParameter<double> nearend_tuning_mask_hf_enr_transparent(
"nearend_tuning_mask_hf_enr_transparent",
adjusted_cfg.suppressor.nearend_tuning.mask_hf.enr_transparent);
FieldTrialParameter<double> nearend_tuning_mask_hf_enr_suppress(
"nearend_tuning_mask_hf_enr_suppress",
adjusted_cfg.suppressor.nearend_tuning.mask_hf.enr_suppress);
FieldTrialParameter<double> nearend_tuning_max_inc_factor(
"nearend_tuning_max_inc_factor",
adjusted_cfg.suppressor.nearend_tuning.max_inc_factor);
FieldTrialParameter<double> nearend_tuning_max_dec_factor_lf(
"nearend_tuning_max_dec_factor_lf",
adjusted_cfg.suppressor.nearend_tuning.max_dec_factor_lf);
FieldTrialParameter<double> normal_tuning_mask_lf_enr_transparent(
"normal_tuning_mask_lf_enr_transparent",
adjusted_cfg.suppressor.normal_tuning.mask_lf.enr_transparent);
FieldTrialParameter<double> normal_tuning_mask_lf_enr_suppress(
"normal_tuning_mask_lf_enr_suppress",
adjusted_cfg.suppressor.normal_tuning.mask_lf.enr_suppress);
FieldTrialParameter<double> normal_tuning_mask_hf_enr_transparent(
"normal_tuning_mask_hf_enr_transparent",
adjusted_cfg.suppressor.normal_tuning.mask_hf.enr_transparent);
FieldTrialParameter<double> normal_tuning_mask_hf_enr_suppress(
"normal_tuning_mask_hf_enr_suppress",
adjusted_cfg.suppressor.normal_tuning.mask_hf.enr_suppress);
FieldTrialParameter<double> normal_tuning_max_inc_factor(
"normal_tuning_max_inc_factor",
adjusted_cfg.suppressor.normal_tuning.max_inc_factor);
FieldTrialParameter<double> normal_tuning_max_dec_factor_lf(
"normal_tuning_max_dec_factor_lf",
adjusted_cfg.suppressor.normal_tuning.max_dec_factor_lf);
FieldTrialParameter<double> dominant_nearend_detection_enr_threshold(
"dominant_nearend_detection_enr_threshold",
adjusted_cfg.suppressor.dominant_nearend_detection.enr_threshold);
FieldTrialParameter<double> dominant_nearend_detection_enr_exit_threshold(
"dominant_nearend_detection_enr_exit_threshold",
adjusted_cfg.suppressor.dominant_nearend_detection.enr_exit_threshold);
FieldTrialParameter<double> dominant_nearend_detection_snr_threshold(
"dominant_nearend_detection_snr_threshold",
adjusted_cfg.suppressor.dominant_nearend_detection.snr_threshold);
FieldTrialParameter<int> dominant_nearend_detection_hold_duration(
"dominant_nearend_detection_hold_duration",
adjusted_cfg.suppressor.dominant_nearend_detection.hold_duration);
FieldTrialParameter<int> dominant_nearend_detection_trigger_threshold(
"dominant_nearend_detection_trigger_threshold",
adjusted_cfg.suppressor.dominant_nearend_detection.trigger_threshold);
FieldTrialParameter<double> ep_strength_default_len(
"ep_strength_default_len", adjusted_cfg.ep_strength.default_len);
ParseFieldTrial(
{&nearend_tuning_mask_lf_enr_transparent,
&nearend_tuning_mask_lf_enr_suppress,
&nearend_tuning_mask_hf_enr_transparent,
&nearend_tuning_mask_hf_enr_suppress, &nearend_tuning_max_inc_factor,
&nearend_tuning_max_dec_factor_lf,
&normal_tuning_mask_lf_enr_transparent,
&normal_tuning_mask_lf_enr_suppress,
&normal_tuning_mask_hf_enr_transparent,
&normal_tuning_mask_hf_enr_suppress, &normal_tuning_max_inc_factor,
&normal_tuning_max_dec_factor_lf,
&dominant_nearend_detection_enr_threshold,
&dominant_nearend_detection_enr_exit_threshold,
&dominant_nearend_detection_snr_threshold,
&dominant_nearend_detection_hold_duration,
&dominant_nearend_detection_trigger_threshold, &ep_strength_default_len},
suppressor_tuning_override_trial_name);
adjusted_cfg.suppressor.nearend_tuning.mask_lf.enr_transparent =
static_cast<float>(nearend_tuning_mask_lf_enr_transparent.Get());
adjusted_cfg.suppressor.nearend_tuning.mask_lf.enr_suppress =
static_cast<float>(nearend_tuning_mask_lf_enr_suppress.Get());
adjusted_cfg.suppressor.nearend_tuning.mask_hf.enr_transparent =
static_cast<float>(nearend_tuning_mask_hf_enr_transparent.Get());
adjusted_cfg.suppressor.nearend_tuning.mask_hf.enr_suppress =
static_cast<float>(nearend_tuning_mask_hf_enr_suppress.Get());
adjusted_cfg.suppressor.nearend_tuning.max_inc_factor =
static_cast<float>(nearend_tuning_max_inc_factor.Get());
adjusted_cfg.suppressor.nearend_tuning.max_dec_factor_lf =
static_cast<float>(nearend_tuning_max_dec_factor_lf.Get());
adjusted_cfg.suppressor.normal_tuning.mask_lf.enr_transparent =
static_cast<float>(normal_tuning_mask_lf_enr_transparent.Get());
adjusted_cfg.suppressor.normal_tuning.mask_lf.enr_suppress =
static_cast<float>(normal_tuning_mask_lf_enr_suppress.Get());
adjusted_cfg.suppressor.normal_tuning.mask_hf.enr_transparent =
static_cast<float>(normal_tuning_mask_hf_enr_transparent.Get());
adjusted_cfg.suppressor.normal_tuning.mask_hf.enr_suppress =
static_cast<float>(normal_tuning_mask_hf_enr_suppress.Get());
adjusted_cfg.suppressor.normal_tuning.max_inc_factor =
static_cast<float>(normal_tuning_max_inc_factor.Get());
adjusted_cfg.suppressor.normal_tuning.max_dec_factor_lf =
static_cast<float>(normal_tuning_max_dec_factor_lf.Get());
adjusted_cfg.suppressor.dominant_nearend_detection.enr_threshold =
static_cast<float>(dominant_nearend_detection_enr_threshold.Get());
adjusted_cfg.suppressor.dominant_nearend_detection.enr_exit_threshold =
static_cast<float>(dominant_nearend_detection_enr_exit_threshold.Get());
adjusted_cfg.suppressor.dominant_nearend_detection.snr_threshold =
static_cast<float>(dominant_nearend_detection_snr_threshold.Get());
adjusted_cfg.suppressor.dominant_nearend_detection.hold_duration =
dominant_nearend_detection_hold_duration.Get();
adjusted_cfg.suppressor.dominant_nearend_detection.trigger_threshold =
dominant_nearend_detection_trigger_threshold.Get();
adjusted_cfg.ep_strength.default_len =
static_cast<float>(ep_strength_default_len.Get());
// Field trial-based overrides of individual suppressor parameters.
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNearendLfMaskTransparentOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.nearend_tuning.mask_lf.enr_transparent);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNearendLfMaskSuppressOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.nearend_tuning.mask_lf.enr_suppress);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNearendHfMaskTransparentOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.nearend_tuning.mask_hf.enr_transparent);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNearendHfMaskSuppressOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.nearend_tuning.mask_hf.enr_suppress);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNearendMaxIncFactorOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.nearend_tuning.max_inc_factor);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNearendMaxDecFactorLfOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.nearend_tuning.max_dec_factor_lf);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNormalLfMaskTransparentOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.normal_tuning.mask_lf.enr_transparent);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNormalLfMaskSuppressOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.normal_tuning.mask_lf.enr_suppress);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNormalHfMaskTransparentOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.normal_tuning.mask_hf.enr_transparent);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNormalHfMaskSuppressOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.normal_tuning.mask_hf.enr_suppress);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNormalMaxIncFactorOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.normal_tuning.max_inc_factor);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorNormalMaxDecFactorLfOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.normal_tuning.max_dec_factor_lf);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorDominantNearendEnrThresholdOverride", 0.f, 100.f,
&adjusted_cfg.suppressor.dominant_nearend_detection.enr_threshold);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorDominantNearendEnrExitThresholdOverride", 0.f,
100.f,
&adjusted_cfg.suppressor.dominant_nearend_detection.enr_exit_threshold);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorDominantNearendSnrThresholdOverride", 0.f, 100.f,
&adjusted_cfg.suppressor.dominant_nearend_detection.snr_threshold);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorDominantNearendHoldDurationOverride", 0, 1000,
&adjusted_cfg.suppressor.dominant_nearend_detection.hold_duration);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorDominantNearendTriggerThresholdOverride", 0, 1000,
&adjusted_cfg.suppressor.dominant_nearend_detection.trigger_threshold);
RetrieveFieldTrialValue(
"WebRTC-Aec3SuppressorAntiHowlingGainOverride", 0.f, 10.f,
&adjusted_cfg.suppressor.high_bands_suppression.anti_howling_gain);
RetrieveFieldTrialValue("WebRTC-Aec3SuppressorEpStrengthDefaultLenOverride",
-1.f, 1.f, &adjusted_cfg.ep_strength.default_len);
return adjusted_cfg;
}
class EchoCanceller3::RenderWriter {
public:
RenderWriter(ApmDataDumper* data_dumper,
SwapQueue<std::vector<std::vector<std::vector<float>>>,
Aec3RenderQueueItemVerifier>* render_transfer_queue,
size_t num_bands,
size_t num_channels);
RenderWriter() = delete;
RenderWriter(const RenderWriter&) = delete;
RenderWriter& operator=(const RenderWriter&) = delete;
~RenderWriter();
void Insert(const AudioBuffer& input);
private:
ApmDataDumper* data_dumper_;
const size_t num_bands_;
const size_t num_channels_;
HighPassFilter high_pass_filter_;
std::vector<std::vector<std::vector<float>>> render_queue_input_frame_;
SwapQueue<std::vector<std::vector<std::vector<float>>>,
Aec3RenderQueueItemVerifier>* render_transfer_queue_;
};
EchoCanceller3::RenderWriter::RenderWriter(
ApmDataDumper* data_dumper,
SwapQueue<std::vector<std::vector<std::vector<float>>>,
Aec3RenderQueueItemVerifier>* render_transfer_queue,
size_t num_bands,
size_t num_channels)
: data_dumper_(data_dumper),
num_bands_(num_bands),
num_channels_(num_channels),
high_pass_filter_(16000, num_channels),
render_queue_input_frame_(
num_bands_,
std::vector<std::vector<float>>(
num_channels_,
std::vector<float>(AudioBuffer::kSplitBandSize, 0.f))),
render_transfer_queue_(render_transfer_queue) {
RTC_DCHECK(data_dumper);
}
EchoCanceller3::RenderWriter::~RenderWriter() = default;
void EchoCanceller3::RenderWriter::Insert(const AudioBuffer& input) {
RTC_DCHECK_EQ(AudioBuffer::kSplitBandSize, input.num_frames_per_band());
RTC_DCHECK_EQ(num_bands_, input.num_bands());
RTC_DCHECK_EQ(num_channels_, input.num_channels());
// TODO(bugs.webrtc.org/8759) Temporary work-around.
if (num_bands_ != input.num_bands())
return;
data_dumper_->DumpWav("aec3_render_input", AudioBuffer::kSplitBandSize,
&input.split_bands_const(0)[0][0], 16000, 1);
CopyBufferIntoFrame(input, num_bands_, num_channels_,
&render_queue_input_frame_);
high_pass_filter_.Process(&render_queue_input_frame_[0]);
static_cast<void>(render_transfer_queue_->Insert(&render_queue_input_frame_));
}
int EchoCanceller3::instance_count_ = 0;
EchoCanceller3::EchoCanceller3(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels)
: EchoCanceller3(AdjustConfig(config),
sample_rate_hz,
num_render_channels,
num_capture_channels,
std::unique_ptr<BlockProcessor>(
BlockProcessor::Create(AdjustConfig(config),
sample_rate_hz,
num_render_channels,
num_capture_channels))) {}
EchoCanceller3::EchoCanceller3(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels,
std::unique_ptr<BlockProcessor> block_processor)
: data_dumper_(
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
config_(config),
sample_rate_hz_(sample_rate_hz),
num_bands_(NumBandsForRate(sample_rate_hz_)),
num_render_channels_(num_render_channels),
num_capture_channels_(num_capture_channels),
output_framer_(num_bands_, num_capture_channels_),
capture_blocker_(num_bands_, num_capture_channels_),
render_blocker_(num_bands_, num_render_channels_),
render_transfer_queue_(
kRenderTransferQueueSizeFrames,
std::vector<std::vector<std::vector<float>>>(
num_bands_,
std::vector<std::vector<float>>(
num_render_channels_,
std::vector<float>(AudioBuffer::kSplitBandSize, 0.f))),
Aec3RenderQueueItemVerifier(num_bands_,
num_render_channels_,
AudioBuffer::kSplitBandSize)),
block_processor_(std::move(block_processor)),
render_queue_output_frame_(
num_bands_,
std::vector<std::vector<float>>(
num_render_channels_,
std::vector<float>(AudioBuffer::kSplitBandSize, 0.f))),
render_block_(
num_bands_,
std::vector<std::vector<float>>(num_render_channels_,
std::vector<float>(kBlockSize, 0.f))),
capture_block_(
num_bands_,
std::vector<std::vector<float>>(num_capture_channels_,
std::vector<float>(kBlockSize, 0.f))),
render_sub_frame_view_(
num_bands_,
std::vector<rtc::ArrayView<float>>(num_render_channels_)),
capture_sub_frame_view_(
num_bands_,
std::vector<rtc::ArrayView<float>>(num_capture_channels_)) {
RTC_DCHECK(ValidFullBandRate(sample_rate_hz_));
if (config_.delay.fixed_capture_delay_samples > 0) {
block_delay_buffer_.reset(new BlockDelayBuffer(
num_capture_channels_, num_bands_, AudioBuffer::kSplitBandSize,
config_.delay.fixed_capture_delay_samples));
}
render_writer_.reset(new RenderWriter(data_dumper_.get(),
&render_transfer_queue_, num_bands_,
num_render_channels_));
RTC_DCHECK_EQ(num_bands_, std::max(sample_rate_hz_, 16000) / 16000);
RTC_DCHECK_GE(kMaxNumBands, num_bands_);
if (config_.filter.export_linear_aec_output) {
linear_output_framer_.reset(new BlockFramer(1, num_capture_channels_));
linear_output_block_ =
std::make_unique<std::vector<std::vector<std::vector<float>>>>(
1, std::vector<std::vector<float>>(
num_capture_channels_, std::vector<float>(kBlockSize, 0.f)));
linear_output_sub_frame_view_ =
std::vector<std::vector<rtc::ArrayView<float>>>(
1, std::vector<rtc::ArrayView<float>>(num_capture_channels_));
}
}
EchoCanceller3::~EchoCanceller3() = default;
void EchoCanceller3::AnalyzeRender(const AudioBuffer& render) {
RTC_DCHECK_RUNS_SERIALIZED(&render_race_checker_);
RTC_DCHECK_EQ(render.num_channels(), num_render_channels_);
data_dumper_->DumpRaw("aec3_call_order",
static_cast<int>(EchoCanceller3ApiCall::kRender));
return render_writer_->Insert(render);
}
void EchoCanceller3::AnalyzeCapture(const AudioBuffer& capture) {
RTC_DCHECK_RUNS_SERIALIZED(&capture_race_checker_);
data_dumper_->DumpWav("aec3_capture_analyze_input", capture.num_frames(),
capture.channels_const()[0], sample_rate_hz_, 1);
saturated_microphone_signal_ = false;
for (size_t channel = 0; channel < capture.num_channels(); ++channel) {
saturated_microphone_signal_ |=
DetectSaturation(rtc::ArrayView<const float>(
capture.channels_const()[channel], capture.num_frames()));
if (saturated_microphone_signal_) {
break;
}
}
}
void EchoCanceller3::ProcessCapture(AudioBuffer* capture, bool level_change) {
ProcessCapture(capture, nullptr, level_change);
}
void EchoCanceller3::ProcessCapture(AudioBuffer* capture,
AudioBuffer* linear_output,
bool level_change) {
RTC_DCHECK_RUNS_SERIALIZED(&capture_race_checker_);
RTC_DCHECK(capture);
RTC_DCHECK_EQ(num_bands_, capture->num_bands());
RTC_DCHECK_EQ(AudioBuffer::kSplitBandSize, capture->num_frames_per_band());
RTC_DCHECK_EQ(capture->num_channels(), num_capture_channels_);
data_dumper_->DumpRaw("aec3_call_order",
static_cast<int>(EchoCanceller3ApiCall::kCapture));
if (linear_output && !linear_output_framer_) {
RTC_LOG(LS_ERROR) << "Trying to retrieve the linear AEC output without "
"properly configuring AEC3.";
RTC_NOTREACHED();
}
// Report capture call in the metrics and periodically update API call
// metrics.
api_call_metrics_.ReportCaptureCall();
// Optionally delay the capture signal.
if (config_.delay.fixed_capture_delay_samples > 0) {
RTC_DCHECK(block_delay_buffer_);
block_delay_buffer_->DelaySignal(capture);
}
rtc::ArrayView<float> capture_lower_band = rtc::ArrayView<float>(
&capture->split_bands(0)[0][0], AudioBuffer::kSplitBandSize);
data_dumper_->DumpWav("aec3_capture_input", capture_lower_band, 16000, 1);
EmptyRenderQueue();
ProcessCaptureFrameContent(linear_output, capture, level_change,
saturated_microphone_signal_, 0, &capture_blocker_,
linear_output_framer_.get(), &output_framer_,
block_processor_.get(), linear_output_block_.get(),
&linear_output_sub_frame_view_, &capture_block_,
&capture_sub_frame_view_);
ProcessCaptureFrameContent(linear_output, capture, level_change,
saturated_microphone_signal_, 1, &capture_blocker_,
linear_output_framer_.get(), &output_framer_,
block_processor_.get(), linear_output_block_.get(),
&linear_output_sub_frame_view_, &capture_block_,
&capture_sub_frame_view_);
ProcessRemainingCaptureFrameContent(
level_change, saturated_microphone_signal_, &capture_blocker_,
linear_output_framer_.get(), &output_framer_, block_processor_.get(),
linear_output_block_.get(), &capture_block_);
data_dumper_->DumpWav("aec3_capture_output", AudioBuffer::kSplitBandSize,
&capture->split_bands(0)[0][0], 16000, 1);
}
EchoControl::Metrics EchoCanceller3::GetMetrics() const {
RTC_DCHECK_RUNS_SERIALIZED(&capture_race_checker_);
Metrics metrics;
block_processor_->GetMetrics(&metrics);
return metrics;
}
void EchoCanceller3::SetAudioBufferDelay(int delay_ms) {
RTC_DCHECK_RUNS_SERIALIZED(&capture_race_checker_);
block_processor_->SetAudioBufferDelay(delay_ms);
}
bool EchoCanceller3::ActiveProcessing() const {
return true;
}
EchoCanceller3Config EchoCanceller3::CreateDefaultConfig(
size_t num_render_channels,
size_t num_capture_channels) {
EchoCanceller3Config cfg;
if (num_render_channels > 1) {
// Use shorter and more rapidly adapting coarse filter to compensate for
// thge increased number of total filter parameters to adapt.
cfg.filter.coarse.length_blocks = 11;
cfg.filter.coarse.rate = 0.95f;
cfg.filter.coarse_initial.length_blocks = 11;
cfg.filter.coarse_initial.rate = 0.95f;
// Use more concervative suppressor behavior for non-nearend speech.
cfg.suppressor.normal_tuning.max_dec_factor_lf = 0.35f;
cfg.suppressor.normal_tuning.max_inc_factor = 1.5f;
}
return cfg;
}
void EchoCanceller3::EmptyRenderQueue() {
RTC_DCHECK_RUNS_SERIALIZED(&capture_race_checker_);
bool frame_to_buffer =
render_transfer_queue_.Remove(&render_queue_output_frame_);
while (frame_to_buffer) {
// Report render call in the metrics.
api_call_metrics_.ReportRenderCall();
BufferRenderFrameContent(&render_queue_output_frame_, 0, &render_blocker_,
block_processor_.get(), &render_block_,
&render_sub_frame_view_);
BufferRenderFrameContent(&render_queue_output_frame_, 1, &render_blocker_,
block_processor_.get(), &render_block_,
&render_sub_frame_view_);
BufferRemainingRenderFrameContent(&render_blocker_, block_processor_.get(),
&render_block_);
frame_to_buffer =
render_transfer_queue_.Remove(&render_queue_output_frame_);
}
}
} // namespace webrtc

196
webrtc/modules/audio_processing/aec3/echo_canceller3.h Normal file
View File

@ -0,0 +1,196 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ECHO_CANCELLER3_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ECHO_CANCELLER3_H_
#include <stddef.h>
#include <memory>
#include <vector>
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "api/audio/echo_control.h"
#include "modules/audio_processing/aec3/api_call_jitter_metrics.h"
#include "modules/audio_processing/aec3/block_delay_buffer.h"
#include "modules/audio_processing/aec3/block_framer.h"
#include "modules/audio_processing/aec3/block_processor.h"
#include "modules/audio_processing/aec3/frame_blocker.h"
#include "modules/audio_processing/audio_buffer.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/checks.h"
#include "rtc_base/race_checker.h"
#include "rtc_base/swap_queue.h"
#include "rtc_base/thread_annotations.h"
namespace webrtc {
// Method for adjusting config parameter dependencies.
// Only to be used externally to AEC3 for testing purposes.
// TODO(webrtc:5298): Move this to a separate file.
EchoCanceller3Config AdjustConfig(const EchoCanceller3Config& config);
// Functor for verifying the invariance of the frames being put into the render
// queue.
class Aec3RenderQueueItemVerifier {
public:
Aec3RenderQueueItemVerifier(size_t num_bands,
size_t num_channels,
size_t frame_length)
: num_bands_(num_bands),
num_channels_(num_channels),
frame_length_(frame_length) {}
bool operator()(const std::vector<std::vector<std::vector<float>>>& v) const {
if (v.size() != num_bands_) {
return false;
}
for (const auto& band : v) {
if (band.size() != num_channels_) {
return false;
}
for (const auto& channel : band) {
if (channel.size() != frame_length_) {
return false;
}
}
}
return true;
}
private:
const size_t num_bands_;
const size_t num_channels_;
const size_t frame_length_;
};
// Main class for the echo canceller3.
// It does 4 things:
// -Receives 10 ms frames of band-split audio.
// -Provides the lower level echo canceller functionality with
// blocks of 64 samples of audio data.
// -Partially handles the jitter in the render and capture API
// call sequence.
//
// The class is supposed to be used in a non-concurrent manner apart from the
// AnalyzeRender call which can be called concurrently with the other methods.
class EchoCanceller3 : public EchoControl {
public:
// Normal c-tor to use.
EchoCanceller3(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels);
// Testing c-tor that is used only for testing purposes.
EchoCanceller3(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels,
std::unique_ptr<BlockProcessor> block_processor);
~EchoCanceller3() override;
EchoCanceller3(const EchoCanceller3&) = delete;
EchoCanceller3& operator=(const EchoCanceller3&) = delete;
// Analyzes and stores an internal copy of the split-band domain render
// signal.
void AnalyzeRender(AudioBuffer* render) override { AnalyzeRender(*render); }
// Analyzes the full-band domain capture signal to detect signal saturation.
void AnalyzeCapture(AudioBuffer* capture) override {
AnalyzeCapture(*capture);
}
// Processes the split-band domain capture signal in order to remove any echo
// present in the signal.
void ProcessCapture(AudioBuffer* capture, bool level_change) override;
// As above, but also returns the linear filter output.
void ProcessCapture(AudioBuffer* capture,
AudioBuffer* linear_output,
bool level_change) override;
// Collect current metrics from the echo canceller.
Metrics GetMetrics() const override;
// Provides an optional external estimate of the audio buffer delay.
void SetAudioBufferDelay(int delay_ms) override;
bool ActiveProcessing() const override;
// Signals whether an external detector has detected echo leakage from the
// echo canceller.
// Note that in the case echo leakage has been flagged, it should be unflagged
// once it is no longer occurring.
void UpdateEchoLeakageStatus(bool leakage_detected) {
RTC_DCHECK_RUNS_SERIALIZED(&capture_race_checker_);
block_processor_->UpdateEchoLeakageStatus(leakage_detected);
}
// Produces a default configuration that is suitable for a certain combination
// of render and capture channels.
static EchoCanceller3Config CreateDefaultConfig(size_t num_render_channels,
size_t num_capture_channels);
private:
class RenderWriter;
// Empties the render SwapQueue.
void EmptyRenderQueue();
// Analyzes and stores an internal copy of the split-band domain render
// signal.
void AnalyzeRender(const AudioBuffer& render);
// Analyzes the full-band domain capture signal to detect signal saturation.
void AnalyzeCapture(const AudioBuffer& capture);
rtc::RaceChecker capture_race_checker_;
rtc::RaceChecker render_race_checker_;
// State that is accessed by the AnalyzeRender call.
std::unique_ptr<RenderWriter> render_writer_
RTC_GUARDED_BY(render_race_checker_);
// State that may be accessed by the capture thread.
static int instance_count_;
std::unique_ptr<ApmDataDumper> data_dumper_;
const EchoCanceller3Config config_;
const int sample_rate_hz_;
const int num_bands_;
const size_t num_render_channels_;
const size_t num_capture_channels_;
std::unique_ptr<BlockFramer> linear_output_framer_
RTC_GUARDED_BY(capture_race_checker_);
BlockFramer output_framer_ RTC_GUARDED_BY(capture_race_checker_);
FrameBlocker capture_blocker_ RTC_GUARDED_BY(capture_race_checker_);
FrameBlocker render_blocker_ RTC_GUARDED_BY(capture_race_checker_);
SwapQueue<std::vector<std::vector<std::vector<float>>>,
Aec3RenderQueueItemVerifier>
render_transfer_queue_;
std::unique_ptr<BlockProcessor> block_processor_
RTC_GUARDED_BY(capture_race_checker_);
std::vector<std::vector<std::vector<float>>> render_queue_output_frame_
RTC_GUARDED_BY(capture_race_checker_);
bool saturated_microphone_signal_ RTC_GUARDED_BY(capture_race_checker_) =
false;
std::vector<std::vector<std::vector<float>>> render_block_
RTC_GUARDED_BY(capture_race_checker_);
std::unique_ptr<std::vector<std::vector<std::vector<float>>>>
linear_output_block_ RTC_GUARDED_BY(capture_race_checker_);
std::vector<std::vector<std::vector<float>>> capture_block_
RTC_GUARDED_BY(capture_race_checker_);
std::vector<std::vector<rtc::ArrayView<float>>> render_sub_frame_view_
RTC_GUARDED_BY(capture_race_checker_);
std::vector<std::vector<rtc::ArrayView<float>>> linear_output_sub_frame_view_
RTC_GUARDED_BY(capture_race_checker_);
std::vector<std::vector<rtc::ArrayView<float>>> capture_sub_frame_view_
RTC_GUARDED_BY(capture_race_checker_);
std::unique_ptr<BlockDelayBuffer> block_delay_buffer_
RTC_GUARDED_BY(capture_race_checker_);
ApiCallJitterMetrics api_call_metrics_ RTC_GUARDED_BY(capture_race_checker_);
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ECHO_CANCELLER3_H_

125
webrtc/modules/audio_processing/aec3/echo_path_delay_estimator.cc Normal file
View File

@ -0,0 +1,125 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/echo_path_delay_estimator.h"
#include <array>
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/downsampled_render_buffer.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/checks.h"
namespace webrtc {
EchoPathDelayEstimator::EchoPathDelayEstimator(
ApmDataDumper* data_dumper,
const EchoCanceller3Config& config,
size_t num_capture_channels)
: data_dumper_(data_dumper),
down_sampling_factor_(config.delay.down_sampling_factor),
sub_block_size_(down_sampling_factor_ != 0
? kBlockSize / down_sampling_factor_
: kBlockSize),
capture_mixer_(num_capture_channels,
config.delay.capture_alignment_mixing),
capture_decimator_(down_sampling_factor_),
matched_filter_(
data_dumper_,
DetectOptimization(),
sub_block_size_,
kMatchedFilterWindowSizeSubBlocks,
config.delay.num_filters,
kMatchedFilterAlignmentShiftSizeSubBlocks,
config.delay.down_sampling_factor == 8
? config.render_levels.poor_excitation_render_limit_ds8
: config.render_levels.poor_excitation_render_limit,
config.delay.delay_estimate_smoothing,
config.delay.delay_candidate_detection_threshold),
matched_filter_lag_aggregator_(data_dumper_,
matched_filter_.GetMaxFilterLag(),
config.delay.delay_selection_thresholds) {
RTC_DCHECK(data_dumper);
RTC_DCHECK(down_sampling_factor_ > 0);
}
EchoPathDelayEstimator::~EchoPathDelayEstimator() = default;
void EchoPathDelayEstimator::Reset(bool reset_delay_confidence) {
Reset(true, reset_delay_confidence);
}
absl::optional<DelayEstimate> EchoPathDelayEstimator::EstimateDelay(
const DownsampledRenderBuffer& render_buffer,
const std::vector<std::vector<float>>& capture) {
RTC_DCHECK_EQ(kBlockSize, capture[0].size());
std::array<float, kBlockSize> downsampled_capture_data;
rtc::ArrayView<float> downsampled_capture(downsampled_capture_data.data(),
sub_block_size_);
std::array<float, kBlockSize> downmixed_capture;
capture_mixer_.ProduceOutput(capture, downmixed_capture);
capture_decimator_.Decimate(downmixed_capture, downsampled_capture);
data_dumper_->DumpWav("aec3_capture_decimator_output",
downsampled_capture.size(), downsampled_capture.data(),
16000 / down_sampling_factor_, 1);
matched_filter_.Update(render_buffer, downsampled_capture);
absl::optional<DelayEstimate> aggregated_matched_filter_lag =
matched_filter_lag_aggregator_.Aggregate(
matched_filter_.GetLagEstimates());
// Run clockdrift detection.
if (aggregated_matched_filter_lag &&
(*aggregated_matched_filter_lag).quality ==
DelayEstimate::Quality::kRefined)
clockdrift_detector_.Update((*aggregated_matched_filter_lag).delay);
// TODO(peah): Move this logging outside of this class once EchoCanceller3
// development is done.
data_dumper_->DumpRaw(
"aec3_echo_path_delay_estimator_delay",
aggregated_matched_filter_lag
? static_cast<int>(aggregated_matched_filter_lag->delay *
down_sampling_factor_)
: -1);
// Return the detected delay in samples as the aggregated matched filter lag
// compensated by the down sampling factor for the signal being correlated.
if (aggregated_matched_filter_lag) {
aggregated_matched_filter_lag->delay *= down_sampling_factor_;
}
if (old_aggregated_lag_ && aggregated_matched_filter_lag &&
old_aggregated_lag_->delay == aggregated_matched_filter_lag->delay) {
++consistent_estimate_counter_;
} else {
consistent_estimate_counter_ = 0;
}
old_aggregated_lag_ = aggregated_matched_filter_lag;
constexpr size_t kNumBlocksPerSecondBy2 = kNumBlocksPerSecond / 2;
if (consistent_estimate_counter_ > kNumBlocksPerSecondBy2) {
Reset(false, false);
}
return aggregated_matched_filter_lag;
}
void EchoPathDelayEstimator::Reset(bool reset_lag_aggregator,
bool reset_delay_confidence) {
if (reset_lag_aggregator) {
matched_filter_lag_aggregator_.Reset(reset_delay_confidence);
}
matched_filter_.Reset();
old_aggregated_lag_ = absl::nullopt;
consistent_estimate_counter_ = 0;
}
} // namespace webrtc

79
webrtc/modules/audio_processing/aec3/echo_path_delay_estimator.h Normal file
View File

@ -0,0 +1,79 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ECHO_PATH_DELAY_ESTIMATOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ECHO_PATH_DELAY_ESTIMATOR_H_
#include <stddef.h>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "modules/audio_processing/aec3/alignment_mixer.h"
#include "modules/audio_processing/aec3/clockdrift_detector.h"
#include "modules/audio_processing/aec3/decimator.h"
#include "modules/audio_processing/aec3/delay_estimate.h"
#include "modules/audio_processing/aec3/matched_filter.h"
#include "modules/audio_processing/aec3/matched_filter_lag_aggregator.h"
#include "rtc_base/constructor_magic.h"
namespace webrtc {
class ApmDataDumper;
struct DownsampledRenderBuffer;
struct EchoCanceller3Config;
// Estimates the delay of the echo path.
class EchoPathDelayEstimator {
public:
EchoPathDelayEstimator(ApmDataDumper* data_dumper,
const EchoCanceller3Config& config,
size_t num_capture_channels);
~EchoPathDelayEstimator();
// Resets the estimation. If the delay confidence is reset, the reset behavior
// is as if the call is restarted.
void Reset(bool reset_delay_confidence);
// Produce a delay estimate if such is avaliable.
absl::optional<DelayEstimate> EstimateDelay(
const DownsampledRenderBuffer& render_buffer,
const std::vector<std::vector<float>>& capture);
// Log delay estimator properties.
void LogDelayEstimationProperties(int sample_rate_hz, size_t shift) const {
matched_filter_.LogFilterProperties(sample_rate_hz, shift,
down_sampling_factor_);
}
// Returns the level of detected clockdrift.
ClockdriftDetector::Level Clockdrift() const {
return clockdrift_detector_.ClockdriftLevel();
}
private:
ApmDataDumper* const data_dumper_;
const size_t down_sampling_factor_;
const size_t sub_block_size_;
AlignmentMixer capture_mixer_;
Decimator capture_decimator_;
MatchedFilter matched_filter_;
MatchedFilterLagAggregator matched_filter_lag_aggregator_;
absl::optional<DelayEstimate> old_aggregated_lag_;
size_t consistent_estimate_counter_ = 0;
ClockdriftDetector clockdrift_detector_;
// Internal reset method with more granularity.
void Reset(bool reset_lag_aggregator, bool reset_delay_confidence);
RTC_DISALLOW_COPY_AND_ASSIGN(EchoPathDelayEstimator);
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ECHO_PATH_DELAY_ESTIMATOR_H_

22
webrtc/modules/audio_processing/aec3/echo_path_variability.cc Normal file
View File

@ -0,0 +1,22 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/echo_path_variability.h"
namespace webrtc {
EchoPathVariability::EchoPathVariability(bool gain_change,
DelayAdjustment delay_change,
bool clock_drift)
: gain_change(gain_change),
delay_change(delay_change),
clock_drift(clock_drift) {}
} // namespace webrtc

37
webrtc/modules/audio_processing/aec3/echo_path_variability.h Normal file
View File

@ -0,0 +1,37 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ECHO_PATH_VARIABILITY_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ECHO_PATH_VARIABILITY_H_
namespace webrtc {
struct EchoPathVariability {
enum class DelayAdjustment {
kNone,
kBufferFlush,
kNewDetectedDelay
};
EchoPathVariability(bool gain_change,
DelayAdjustment delay_change,
bool clock_drift);
bool AudioPathChanged() const {
return gain_change || delay_change != DelayAdjustment::kNone;
}
bool gain_change;
DelayAdjustment delay_change;
bool clock_drift;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ECHO_PATH_VARIABILITY_H_

500
webrtc/modules/audio_processing/aec3/echo_remover.cc Normal file
View File

@ -0,0 +1,500 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/echo_remover.h"
#include <math.h>
#include <stddef.h>
#include <algorithm>
#include <array>
#include <cmath>
#include <memory>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/aec3_fft.h"
#include "modules/audio_processing/aec3/aec_state.h"
#include "modules/audio_processing/aec3/comfort_noise_generator.h"
#include "modules/audio_processing/aec3/echo_path_variability.h"
#include "modules/audio_processing/aec3/echo_remover_metrics.h"
#include "modules/audio_processing/aec3/fft_data.h"
#include "modules/audio_processing/aec3/render_buffer.h"
#include "modules/audio_processing/aec3/render_signal_analyzer.h"
#include "modules/audio_processing/aec3/residual_echo_estimator.h"
#include "modules/audio_processing/aec3/subtractor.h"
#include "modules/audio_processing/aec3/subtractor_output.h"
#include "modules/audio_processing/aec3/suppression_filter.h"
#include "modules/audio_processing/aec3/suppression_gain.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomic_ops.h"
#include "rtc_base/checks.h"
#include "rtc_base/logging.h"
namespace webrtc {
namespace {
// Maximum number of channels for which the capture channel data is stored on
// the stack. If the number of channels are larger than this, they are stored
// using scratch memory that is pre-allocated on the heap. The reason for this
// partitioning is not to waste heap space for handling the more common numbers
// of channels, while at the same time not limiting the support for higher
// numbers of channels by enforcing the capture channel data to be stored on the
// stack using a fixed maximum value.
constexpr size_t kMaxNumChannelsOnStack = 2;
// Chooses the number of channels to store on the heap when that is required due
// to the number of capture channels being larger than the pre-defined number
// of channels to store on the stack.
size_t NumChannelsOnHeap(size_t num_capture_channels) {
return num_capture_channels > kMaxNumChannelsOnStack ? num_capture_channels
: 0;
}
void LinearEchoPower(const FftData& E,
const FftData& Y,
std::array<float, kFftLengthBy2Plus1>* S2) {
for (size_t k = 0; k < E.re.size(); ++k) {
(*S2)[k] = (Y.re[k] - E.re[k]) * (Y.re[k] - E.re[k]) +
(Y.im[k] - E.im[k]) * (Y.im[k] - E.im[k]);
}
}
// Fades between two input signals using a fix-sized transition.
void SignalTransition(rtc::ArrayView<const float> from,
rtc::ArrayView<const float> to,
rtc::ArrayView<float> out) {
if (from == to) {
RTC_DCHECK_EQ(to.size(), out.size());
std::copy(to.begin(), to.end(), out.begin());
} else {
constexpr size_t kTransitionSize = 30;
constexpr float kOneByTransitionSizePlusOne = 1.f / (kTransitionSize + 1);
RTC_DCHECK_EQ(from.size(), to.size());
RTC_DCHECK_EQ(from.size(), out.size());
RTC_DCHECK_LE(kTransitionSize, out.size());
for (size_t k = 0; k < kTransitionSize; ++k) {
float a = (k + 1) * kOneByTransitionSizePlusOne;
out[k] = a * to[k] + (1.f - a) * from[k];
}
std::copy(to.begin() + kTransitionSize, to.end(),
out.begin() + kTransitionSize);
}
}
// Computes a windowed (square root Hanning) padded FFT and updates the related
// memory.
void WindowedPaddedFft(const Aec3Fft& fft,
rtc::ArrayView<const float> v,
rtc::ArrayView<float> v_old,
FftData* V) {
fft.PaddedFft(v, v_old, Aec3Fft::Window::kSqrtHanning, V);
std::copy(v.begin(), v.end(), v_old.begin());
}
// Class for removing the echo from the capture signal.
class EchoRemoverImpl final : public EchoRemover {
public:
EchoRemoverImpl(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels);
~EchoRemoverImpl() override;
EchoRemoverImpl(const EchoRemoverImpl&) = delete;
EchoRemoverImpl& operator=(const EchoRemoverImpl&) = delete;
void GetMetrics(EchoControl::Metrics* metrics) const override;
// Removes the echo from a block of samples from the capture signal. The
// supplied render signal is assumed to be pre-aligned with the capture
// signal.
void ProcessCapture(
EchoPathVariability echo_path_variability,
bool capture_signal_saturation,
const absl::optional<DelayEstimate>& external_delay,
RenderBuffer* render_buffer,
std::vector<std::vector<std::vector<float>>>* linear_output,
std::vector<std::vector<std::vector<float>>>* capture) override;
// Updates the status on whether echo leakage is detected in the output of the
// echo remover.
void UpdateEchoLeakageStatus(bool leakage_detected) override {
echo_leakage_detected_ = leakage_detected;
}
private:
// Selects which of the coarse and refined linear filter outputs that is most
// appropriate to pass to the suppressor and forms the linear filter output by
// smoothly transition between those.
void FormLinearFilterOutput(const SubtractorOutput& subtractor_output,
rtc::ArrayView<float> output);
static int instance_count_;
const EchoCanceller3Config config_;
const Aec3Fft fft_;
std::unique_ptr<ApmDataDumper> data_dumper_;
const Aec3Optimization optimization_;
const int sample_rate_hz_;
const size_t num_render_channels_;
const size_t num_capture_channels_;
const bool use_coarse_filter_output_;
Subtractor subtractor_;
SuppressionGain suppression_gain_;
ComfortNoiseGenerator cng_;
SuppressionFilter suppression_filter_;
RenderSignalAnalyzer render_signal_analyzer_;
ResidualEchoEstimator residual_echo_estimator_;
bool echo_leakage_detected_ = false;
AecState aec_state_;
EchoRemoverMetrics metrics_;
std::vector<std::array<float, kFftLengthBy2>> e_old_;
std::vector<std::array<float, kFftLengthBy2>> y_old_;
size_t block_counter_ = 0;
int gain_change_hangover_ = 0;
bool refined_filter_output_last_selected_ = true;
std::vector<std::array<float, kFftLengthBy2>> e_heap_;
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2_heap_;
std::vector<std::array<float, kFftLengthBy2Plus1>> E2_heap_;
std::vector<std::array<float, kFftLengthBy2Plus1>> R2_heap_;
std::vector<std::array<float, kFftLengthBy2Plus1>> S2_linear_heap_;
std::vector<FftData> Y_heap_;
std::vector<FftData> E_heap_;
std::vector<FftData> comfort_noise_heap_;
std::vector<FftData> high_band_comfort_noise_heap_;
std::vector<SubtractorOutput> subtractor_output_heap_;
};
int EchoRemoverImpl::instance_count_ = 0;
EchoRemoverImpl::EchoRemoverImpl(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels)
: config_(config),
fft_(),
data_dumper_(
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
optimization_(DetectOptimization()),
sample_rate_hz_(sample_rate_hz),
num_render_channels_(num_render_channels),
num_capture_channels_(num_capture_channels),
use_coarse_filter_output_(
config_.filter.enable_coarse_filter_output_usage),
subtractor_(config,
num_render_channels_,
num_capture_channels_,
data_dumper_.get(),
optimization_),
suppression_gain_(config_,
optimization_,
sample_rate_hz,
num_capture_channels),
cng_(config_, optimization_, num_capture_channels_),
suppression_filter_(optimization_,
sample_rate_hz_,
num_capture_channels_),
render_signal_analyzer_(config_),
residual_echo_estimator_(config_, num_render_channels),
aec_state_(config_, num_capture_channels_),
e_old_(num_capture_channels_, {0.f}),
y_old_(num_capture_channels_, {0.f}),
e_heap_(NumChannelsOnHeap(num_capture_channels_), {0.f}),
Y2_heap_(NumChannelsOnHeap(num_capture_channels_)),
E2_heap_(NumChannelsOnHeap(num_capture_channels_)),
R2_heap_(NumChannelsOnHeap(num_capture_channels_)),
S2_linear_heap_(NumChannelsOnHeap(num_capture_channels_)),
Y_heap_(NumChannelsOnHeap(num_capture_channels_)),
E_heap_(NumChannelsOnHeap(num_capture_channels_)),
comfort_noise_heap_(NumChannelsOnHeap(num_capture_channels_)),
high_band_comfort_noise_heap_(NumChannelsOnHeap(num_capture_channels_)),
subtractor_output_heap_(NumChannelsOnHeap(num_capture_channels_)) {
RTC_DCHECK(ValidFullBandRate(sample_rate_hz));
}
EchoRemoverImpl::~EchoRemoverImpl() = default;
void EchoRemoverImpl::GetMetrics(EchoControl::Metrics* metrics) const {
// Echo return loss (ERL) is inverted to go from gain to attenuation.
metrics->echo_return_loss = -10.0 * std::log10(aec_state_.ErlTimeDomain());
metrics->echo_return_loss_enhancement =
Log2TodB(aec_state_.FullBandErleLog2());
}
void EchoRemoverImpl::ProcessCapture(
EchoPathVariability echo_path_variability,
bool capture_signal_saturation,
const absl::optional<DelayEstimate>& external_delay,
RenderBuffer* render_buffer,
std::vector<std::vector<std::vector<float>>>* linear_output,
std::vector<std::vector<std::vector<float>>>* capture) {
++block_counter_;
const std::vector<std::vector<std::vector<float>>>& x =
render_buffer->Block(0);
std::vector<std::vector<std::vector<float>>>* y = capture;
RTC_DCHECK(render_buffer);
RTC_DCHECK(y);
RTC_DCHECK_EQ(x.size(), NumBandsForRate(sample_rate_hz_));
RTC_DCHECK_EQ(y->size(), NumBandsForRate(sample_rate_hz_));
RTC_DCHECK_EQ(x[0].size(), num_render_channels_);
RTC_DCHECK_EQ((*y)[0].size(), num_capture_channels_);
RTC_DCHECK_EQ(x[0][0].size(), kBlockSize);
RTC_DCHECK_EQ((*y)[0][0].size(), kBlockSize);
// Stack allocated data to use when the number of channels is low.
std::array<std::array<float, kFftLengthBy2>, kMaxNumChannelsOnStack> e_stack;
std::array<std::array<float, kFftLengthBy2Plus1>, kMaxNumChannelsOnStack>
Y2_stack;
std::array<std::array<float, kFftLengthBy2Plus1>, kMaxNumChannelsOnStack>
E2_stack;
std::array<std::array<float, kFftLengthBy2Plus1>, kMaxNumChannelsOnStack>
R2_stack;
std::array<std::array<float, kFftLengthBy2Plus1>, kMaxNumChannelsOnStack>
S2_linear_stack;
std::array<FftData, kMaxNumChannelsOnStack> Y_stack;
std::array<FftData, kMaxNumChannelsOnStack> E_stack;
std::array<FftData, kMaxNumChannelsOnStack> comfort_noise_stack;
std::array<FftData, kMaxNumChannelsOnStack> high_band_comfort_noise_stack;
std::array<SubtractorOutput, kMaxNumChannelsOnStack> subtractor_output_stack;
rtc::ArrayView<std::array<float, kFftLengthBy2>> e(e_stack.data(),
num_capture_channels_);
rtc::ArrayView<std::array<float, kFftLengthBy2Plus1>> Y2(
Y2_stack.data(), num_capture_channels_);
rtc::ArrayView<std::array<float, kFftLengthBy2Plus1>> E2(
E2_stack.data(), num_capture_channels_);
rtc::ArrayView<std::array<float, kFftLengthBy2Plus1>> R2(
R2_stack.data(), num_capture_channels_);
rtc::ArrayView<std::array<float, kFftLengthBy2Plus1>> S2_linear(
S2_linear_stack.data(), num_capture_channels_);
rtc::ArrayView<FftData> Y(Y_stack.data(), num_capture_channels_);
rtc::ArrayView<FftData> E(E_stack.data(), num_capture_channels_);
rtc::ArrayView<FftData> comfort_noise(comfort_noise_stack.data(),
num_capture_channels_);
rtc::ArrayView<FftData> high_band_comfort_noise(
high_band_comfort_noise_stack.data(), num_capture_channels_);
rtc::ArrayView<SubtractorOutput> subtractor_output(
subtractor_output_stack.data(), num_capture_channels_);
if (NumChannelsOnHeap(num_capture_channels_) > 0) {
// If the stack-allocated space is too small, use the heap for storing the
// microphone data.
e = rtc::ArrayView<std::array<float, kFftLengthBy2>>(e_heap_.data(),
num_capture_channels_);
Y2 = rtc::ArrayView<std::array<float, kFftLengthBy2Plus1>>(
Y2_heap_.data(), num_capture_channels_);
E2 = rtc::ArrayView<std::array<float, kFftLengthBy2Plus1>>(
E2_heap_.data(), num_capture_channels_);
R2 = rtc::ArrayView<std::array<float, kFftLengthBy2Plus1>>(
R2_heap_.data(), num_capture_channels_);
S2_linear = rtc::ArrayView<std::array<float, kFftLengthBy2Plus1>>(
S2_linear_heap_.data(), num_capture_channels_);
Y = rtc::ArrayView<FftData>(Y_heap_.data(), num_capture_channels_);
E = rtc::ArrayView<FftData>(E_heap_.data(), num_capture_channels_);
comfort_noise = rtc::ArrayView<FftData>(comfort_noise_heap_.data(),
num_capture_channels_);
high_band_comfort_noise = rtc::ArrayView<FftData>(
high_band_comfort_noise_heap_.data(), num_capture_channels_);
subtractor_output = rtc::ArrayView<SubtractorOutput>(
subtractor_output_heap_.data(), num_capture_channels_);
}
data_dumper_->DumpWav("aec3_echo_remover_capture_input", kBlockSize,
&(*y)[0][0][0], 16000, 1);
data_dumper_->DumpWav("aec3_echo_remover_render_input", kBlockSize,
&x[0][0][0], 16000, 1);
data_dumper_->DumpRaw("aec3_echo_remover_capture_input", (*y)[0][0]);
data_dumper_->DumpRaw("aec3_echo_remover_render_input", x[0][0]);
aec_state_.UpdateCaptureSaturation(capture_signal_saturation);
if (echo_path_variability.AudioPathChanged()) {
// Ensure that the gain change is only acted on once per frame.
if (echo_path_variability.gain_change) {
if (gain_change_hangover_ == 0) {
constexpr int kMaxBlocksPerFrame = 3;
gain_change_hangover_ = kMaxBlocksPerFrame;
rtc::LoggingSeverity log_level =
config_.delay.log_warning_on_delay_changes ? rtc::LS_WARNING
: rtc::LS_VERBOSE;
RTC_LOG_V(log_level)
<< "Gain change detected at block " << block_counter_;
} else {
echo_path_variability.gain_change = false;
}
}
subtractor_.HandleEchoPathChange(echo_path_variability);
aec_state_.HandleEchoPathChange(echo_path_variability);
if (echo_path_variability.delay_change !=
EchoPathVariability::DelayAdjustment::kNone) {
suppression_gain_.SetInitialState(true);
}
}
if (gain_change_hangover_ > 0) {
--gain_change_hangover_;
}
// Analyze the render signal.
render_signal_analyzer_.Update(*render_buffer,
aec_state_.MinDirectPathFilterDelay());
// State transition.
if (aec_state_.TransitionTriggered()) {
subtractor_.ExitInitialState();
suppression_gain_.SetInitialState(false);
}
// Perform linear echo cancellation.
subtractor_.Process(*render_buffer, (*y)[0], render_signal_analyzer_,
aec_state_, subtractor_output);
// Compute spectra.
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
FormLinearFilterOutput(subtractor_output[ch], e[ch]);
WindowedPaddedFft(fft_, (*y)[0][ch], y_old_[ch], &Y[ch]);
WindowedPaddedFft(fft_, e[ch], e_old_[ch], &E[ch]);
LinearEchoPower(E[ch], Y[ch], &S2_linear[ch]);
Y[ch].Spectrum(optimization_, Y2[ch]);
E[ch].Spectrum(optimization_, E2[ch]);
}
// Optionally return the linear filter output.
if (linear_output) {
RTC_DCHECK_GE(1, linear_output->size());
RTC_DCHECK_EQ(num_capture_channels_, linear_output[0].size());
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
RTC_DCHECK_EQ(kBlockSize, (*linear_output)[0][ch].size());
std::copy(e[ch].begin(), e[ch].end(), (*linear_output)[0][ch].begin());
}
}
// Update the AEC state information.
aec_state_.Update(external_delay, subtractor_.FilterFrequencyResponses(),
subtractor_.FilterImpulseResponses(), *render_buffer, E2,
Y2, subtractor_output);
// Choose the linear output.
const auto& Y_fft = aec_state_.UseLinearFilterOutput() ? E : Y;
data_dumper_->DumpWav("aec3_output_linear", kBlockSize, &(*y)[0][0][0], 16000,
1);
data_dumper_->DumpWav("aec3_output_linear2", kBlockSize, &e[0][0], 16000, 1);
// Estimate the residual echo power.
residual_echo_estimator_.Estimate(aec_state_, *render_buffer, S2_linear, Y2,
R2);
// Estimate the comfort noise.
cng_.Compute(aec_state_.SaturatedCapture(), Y2, comfort_noise,
high_band_comfort_noise);
// Suppressor nearend estimate.
if (aec_state_.UsableLinearEstimate()) {
// E2 is bound by Y2.
for (size_t ch = 0; ch < num_capture_channels_; ++ch) {
std::transform(E2[ch].begin(), E2[ch].end(), Y2[ch].begin(),
E2[ch].begin(),
[](float a, float b) { return std::min(a, b); });
}
}
const auto& nearend_spectrum = aec_state_.UsableLinearEstimate() ? E2 : Y2;
// Suppressor echo estimate.
const auto& echo_spectrum =
aec_state_.UsableLinearEstimate() ? S2_linear : R2;
// Compute preferred gains.
float high_bands_gain;
std::array<float, kFftLengthBy2Plus1> G;
suppression_gain_.GetGain(nearend_spectrum, echo_spectrum, R2,
cng_.NoiseSpectrum(), render_signal_analyzer_,
aec_state_, x, &high_bands_gain, &G);
suppression_filter_.ApplyGain(comfort_noise, high_band_comfort_noise, G,
high_bands_gain, Y_fft, y);
// Update the metrics.
metrics_.Update(aec_state_, cng_.NoiseSpectrum()[0], G);
// Debug outputs for the purpose of development and analysis.
data_dumper_->DumpWav("aec3_echo_estimate", kBlockSize,
&subtractor_output[0].s_refined[0], 16000, 1);
data_dumper_->DumpRaw("aec3_output", (*y)[0][0]);
data_dumper_->DumpRaw("aec3_narrow_render",
render_signal_analyzer_.NarrowPeakBand() ? 1 : 0);
data_dumper_->DumpRaw("aec3_N2", cng_.NoiseSpectrum()[0]);
data_dumper_->DumpRaw("aec3_suppressor_gain", G);
data_dumper_->DumpWav("aec3_output",
rtc::ArrayView<const float>(&(*y)[0][0][0], kBlockSize),
16000, 1);
data_dumper_->DumpRaw("aec3_using_subtractor_output[0]",
aec_state_.UseLinearFilterOutput() ? 1 : 0);
data_dumper_->DumpRaw("aec3_E2", E2[0]);
data_dumper_->DumpRaw("aec3_S2_linear", S2_linear[0]);
data_dumper_->DumpRaw("aec3_Y2", Y2[0]);
data_dumper_->DumpRaw(
"aec3_X2", render_buffer->Spectrum(
aec_state_.MinDirectPathFilterDelay())[/*channel=*/0]);
data_dumper_->DumpRaw("aec3_R2", R2[0]);
data_dumper_->DumpRaw("aec3_filter_delay",
aec_state_.MinDirectPathFilterDelay());
data_dumper_->DumpRaw("aec3_capture_saturation",
aec_state_.SaturatedCapture() ? 1 : 0);
}
void EchoRemoverImpl::FormLinearFilterOutput(
const SubtractorOutput& subtractor_output,
rtc::ArrayView<float> output) {
RTC_DCHECK_EQ(subtractor_output.e_refined.size(), output.size());
RTC_DCHECK_EQ(subtractor_output.e_coarse.size(), output.size());
bool use_refined_output = true;
if (use_coarse_filter_output_) {
// As the output of the refined adaptive filter generally should be better
// than the coarse filter output, add a margin and threshold for when
// choosing the coarse filter output.
if (subtractor_output.e2_coarse < 0.9f * subtractor_output.e2_refined &&
subtractor_output.y2 > 30.f * 30.f * kBlockSize &&
(subtractor_output.s2_refined > 60.f * 60.f * kBlockSize ||
subtractor_output.s2_coarse > 60.f * 60.f * kBlockSize)) {
use_refined_output = false;
} else {
// If the refined filter is diverged, choose the filter output that has
// the lowest power.
if (subtractor_output.e2_coarse < subtractor_output.e2_refined &&
subtractor_output.y2 < subtractor_output.e2_refined) {
use_refined_output = false;
}
}
}
SignalTransition(refined_filter_output_last_selected_
? subtractor_output.e_refined
: subtractor_output.e_coarse,
use_refined_output ? subtractor_output.e_refined
: subtractor_output.e_coarse,
output);
refined_filter_output_last_selected_ = use_refined_output;
}
} // namespace
EchoRemover* EchoRemover::Create(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels) {
return new EchoRemoverImpl(config, sample_rate_hz, num_render_channels,
num_capture_channels);
}
} // namespace webrtc

55
webrtc/modules/audio_processing/aec3/echo_remover.h Normal file
View File

@ -0,0 +1,55 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ECHO_REMOVER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ECHO_REMOVER_H_
#include <vector>
#include "absl/types/optional.h"
#include "api/audio/echo_canceller3_config.h"
#include "api/audio/echo_control.h"
#include "modules/audio_processing/aec3/delay_estimate.h"
#include "modules/audio_processing/aec3/echo_path_variability.h"
#include "modules/audio_processing/aec3/render_buffer.h"
namespace webrtc {
// Class for removing the echo from the capture signal.
class EchoRemover {
public:
static EchoRemover* Create(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels,
size_t num_capture_channels);
virtual ~EchoRemover() = default;
// Get current metrics.
virtual void GetMetrics(EchoControl::Metrics* metrics) const = 0;
// Removes the echo from a block of samples from the capture signal. The
// supplied render signal is assumed to be pre-aligned with the capture
// signal.
virtual void ProcessCapture(
EchoPathVariability echo_path_variability,
bool capture_signal_saturation,
const absl::optional<DelayEstimate>& external_delay,
RenderBuffer* render_buffer,
std::vector<std::vector<std::vector<float>>>* linear_output,
std::vector<std::vector<std::vector<float>>>* capture) = 0;
// Updates the status on whether echo leakage is detected in the output of the
// echo remover.
virtual void UpdateEchoLeakageStatus(bool leakage_detected) = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ECHO_REMOVER_H_

246
webrtc/modules/audio_processing/aec3/echo_remover_metrics.cc Normal file
View File

@ -0,0 +1,246 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/echo_remover_metrics.h"
#include <math.h>
#include <stddef.h>
#include <algorithm>
#include <cmath>
#include <numeric>
#include "rtc_base/checks.h"
#include "rtc_base/numerics/safe_minmax.h"
#include "system_wrappers/include/metrics.h"
namespace webrtc {
namespace {
constexpr float kOneByMetricsCollectionBlocks = 1.f / kMetricsCollectionBlocks;
} // namespace
EchoRemoverMetrics::DbMetric::DbMetric() : DbMetric(0.f, 0.f, 0.f) {}
EchoRemoverMetrics::DbMetric::DbMetric(float sum_value,
float floor_value,
float ceil_value)
: sum_value(sum_value), floor_value(floor_value), ceil_value(ceil_value) {}
void EchoRemoverMetrics::DbMetric::Update(float value) {
sum_value += value;
floor_value = std::min(floor_value, value);
ceil_value = std::max(ceil_value, value);
}
void EchoRemoverMetrics::DbMetric::UpdateInstant(float value) {
sum_value = value;
floor_value = std::min(floor_value, value);
ceil_value = std::max(ceil_value, value);
}
EchoRemoverMetrics::EchoRemoverMetrics() {
ResetMetrics();
}
void EchoRemoverMetrics::ResetMetrics() {
erl_.fill(DbMetric(0.f, 10000.f, 0.000f));
erl_time_domain_ = DbMetric(0.f, 10000.f, 0.000f);
erle_.fill(DbMetric(0.f, 0.f, 1000.f));
erle_time_domain_ = DbMetric(0.f, 0.f, 1000.f);
active_render_count_ = 0;
saturated_capture_ = false;
}
void EchoRemoverMetrics::Update(
const AecState& aec_state,
const std::array<float, kFftLengthBy2Plus1>& comfort_noise_spectrum,
const std::array<float, kFftLengthBy2Plus1>& suppressor_gain) {
metrics_reported_ = false;
if (++block_counter_ <= kMetricsCollectionBlocks) {
aec3::UpdateDbMetric(aec_state.Erl(), &erl_);
erl_time_domain_.UpdateInstant(aec_state.ErlTimeDomain());
aec3::UpdateDbMetric(aec_state.Erle()[0], &erle_);
erle_time_domain_.UpdateInstant(aec_state.FullBandErleLog2());
active_render_count_ += (aec_state.ActiveRender() ? 1 : 0);
saturated_capture_ = saturated_capture_ || aec_state.SaturatedCapture();
} else {
// Report the metrics over several frames in order to lower the impact of
// the logarithms involved on the computational complexity.
constexpr int kMetricsCollectionBlocksBy2 = kMetricsCollectionBlocks / 2;
switch (block_counter_) {
case kMetricsCollectionBlocks + 1:
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErleBand0.Average",
aec3::TransformDbMetricForReporting(true, 0.f, 19.f, 0.f,
kOneByMetricsCollectionBlocks,
erle_[0].sum_value),
0, 19, 20);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErleBand0.Max",
aec3::TransformDbMetricForReporting(true, 0.f, 19.f, 0.f, 1.f,
erle_[0].ceil_value),
0, 19, 20);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErleBand0.Min",
aec3::TransformDbMetricForReporting(true, 0.f, 19.f, 0.f, 1.f,
erle_[0].floor_value),
0, 19, 20);
break;
case kMetricsCollectionBlocks + 2:
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErleBand1.Average",
aec3::TransformDbMetricForReporting(true, 0.f, 19.f, 0.f,
kOneByMetricsCollectionBlocks,
erle_[1].sum_value),
0, 19, 20);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErleBand1.Max",
aec3::TransformDbMetricForReporting(true, 0.f, 19.f, 0.f, 1.f,
erle_[1].ceil_value),
0, 19, 20);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErleBand1.Min",
aec3::TransformDbMetricForReporting(true, 0.f, 19.f, 0.f, 1.f,
erle_[1].floor_value),
0, 19, 20);
break;
case kMetricsCollectionBlocks + 3:
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErlBand0.Average",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f,
kOneByMetricsCollectionBlocks,
erl_[0].sum_value),
0, 59, 30);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErlBand0.Max",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f, 1.f,
erl_[0].ceil_value),
0, 59, 30);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErlBand0.Min",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f, 1.f,
erl_[0].floor_value),
0, 59, 30);
break;
case kMetricsCollectionBlocks + 4:
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErlBand1.Average",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f,
kOneByMetricsCollectionBlocks,
erl_[1].sum_value),
0, 59, 30);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErlBand1.Max",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f, 1.f,
erl_[1].ceil_value),
0, 59, 30);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.ErlBand1.Min",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f, 1.f,
erl_[1].floor_value),
0, 59, 30);
break;
case kMetricsCollectionBlocks + 5:
RTC_HISTOGRAM_BOOLEAN(
"WebRTC.Audio.EchoCanceller.UsableLinearEstimate",
static_cast<int>(aec_state.UsableLinearEstimate() ? 1 : 0));
RTC_HISTOGRAM_BOOLEAN(
"WebRTC.Audio.EchoCanceller.ActiveRender",
static_cast<int>(
active_render_count_ > kMetricsCollectionBlocksBy2 ? 1 : 0));
RTC_HISTOGRAM_COUNTS_LINEAR("WebRTC.Audio.EchoCanceller.FilterDelay",
aec_state.MinDirectPathFilterDelay(), 0, 30,
31);
RTC_HISTOGRAM_BOOLEAN("WebRTC.Audio.EchoCanceller.CaptureSaturation",
static_cast<int>(saturated_capture_ ? 1 : 0));
break;
case kMetricsCollectionBlocks + 6:
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.Erl.Value",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f, 1.f,
erl_time_domain_.sum_value),
0, 59, 30);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.Erl.Max",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f, 1.f,
erl_time_domain_.ceil_value),
0, 59, 30);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.Erl.Min",
aec3::TransformDbMetricForReporting(true, 0.f, 59.f, 30.f, 1.f,
erl_time_domain_.floor_value),
0, 59, 30);
break;
case kMetricsCollectionBlocks + 7:
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.Erle.Value",
aec3::TransformDbMetricForReporting(false, 0.f, 19.f, 0.f, 1.f,
erle_time_domain_.sum_value),
0, 19, 20);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.Erle.Max",
aec3::TransformDbMetricForReporting(false, 0.f, 19.f, 0.f, 1.f,
erle_time_domain_.ceil_value),
0, 19, 20);
RTC_HISTOGRAM_COUNTS_LINEAR(
"WebRTC.Audio.EchoCanceller.Erle.Min",
aec3::TransformDbMetricForReporting(false, 0.f, 19.f, 0.f, 1.f,
erle_time_domain_.floor_value),
0, 19, 20);
metrics_reported_ = true;
RTC_DCHECK_EQ(kMetricsReportingIntervalBlocks, block_counter_);
block_counter_ = 0;
ResetMetrics();
break;
default:
RTC_NOTREACHED();
break;
}
}
}
namespace aec3 {
void UpdateDbMetric(const std::array<float, kFftLengthBy2Plus1>& value,
std::array<EchoRemoverMetrics::DbMetric, 2>* statistic) {
RTC_DCHECK(statistic);
// Truncation is intended in the band width computation.
constexpr int kNumBands = 2;
constexpr int kBandWidth = 65 / kNumBands;
constexpr float kOneByBandWidth = 1.f / kBandWidth;
RTC_DCHECK_EQ(kNumBands, statistic->size());
RTC_DCHECK_EQ(65, value.size());
for (size_t k = 0; k < statistic->size(); ++k) {
float average_band =
std::accumulate(value.begin() + kBandWidth * k,
value.begin() + kBandWidth * (k + 1), 0.f) *
kOneByBandWidth;
(*statistic)[k].Update(average_band);
}
}
int TransformDbMetricForReporting(bool negate,
float min_value,
float max_value,
float offset,
float scaling,
float value) {
float new_value = 10.f * std::log10(value * scaling + 1e-10f) + offset;
if (negate) {
new_value = -new_value;
}
return static_cast<int>(rtc::SafeClamp(new_value, min_value, max_value));
}
} // namespace aec3
} // namespace webrtc

81
webrtc/modules/audio_processing/aec3/echo_remover_metrics.h Normal file
View File

@ -0,0 +1,81 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ECHO_REMOVER_METRICS_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ECHO_REMOVER_METRICS_H_
#include <array>
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/aec_state.h"
#include "rtc_base/constructor_magic.h"
namespace webrtc {
// Handles the reporting of metrics for the echo remover.
class EchoRemoverMetrics {
public:
struct DbMetric {
DbMetric();
DbMetric(float sum_value, float floor_value, float ceil_value);
void Update(float value);
void UpdateInstant(float value);
float sum_value;
float floor_value;
float ceil_value;
};
EchoRemoverMetrics();
// Updates the metric with new data.
void Update(
const AecState& aec_state,
const std::array<float, kFftLengthBy2Plus1>& comfort_noise_spectrum,
const std::array<float, kFftLengthBy2Plus1>& suppressor_gain);
// Returns true if the metrics have just been reported, otherwise false.
bool MetricsReported() { return metrics_reported_; }
private:
// Resets the metrics.
void ResetMetrics();
int block_counter_ = 0;
std::array<DbMetric, 2> erl_;
DbMetric erl_time_domain_;
std::array<DbMetric, 2> erle_;
DbMetric erle_time_domain_;
int active_render_count_ = 0;
bool saturated_capture_ = false;
bool metrics_reported_ = false;
RTC_DISALLOW_COPY_AND_ASSIGN(EchoRemoverMetrics);
};
namespace aec3 {
// Updates a banded metric of type DbMetric with the values in the supplied
// array.
void UpdateDbMetric(const std::array<float, kFftLengthBy2Plus1>& value,
std::array<EchoRemoverMetrics::DbMetric, 2>* statistic);
// Transforms a DbMetric from the linear domain into the logarithmic domain.
int TransformDbMetricForReporting(bool negate,
float min_value,
float max_value,
float offset,
float scaling,
float value);
} // namespace aec3
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ECHO_REMOVER_METRICS_H_

146
webrtc/modules/audio_processing/aec3/erl_estimator.cc Normal file
View File

@ -0,0 +1,146 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/erl_estimator.h"
#include <algorithm>
#include <numeric>
#include "rtc_base/checks.h"
namespace webrtc {
namespace {
constexpr float kMinErl = 0.01f;
constexpr float kMaxErl = 1000.f;
} // namespace
ErlEstimator::ErlEstimator(size_t startup_phase_length_blocks_)
: startup_phase_length_blocks__(startup_phase_length_blocks_) {
erl_.fill(kMaxErl);
hold_counters_.fill(0);
erl_time_domain_ = kMaxErl;
hold_counter_time_domain_ = 0;
}
ErlEstimator::~ErlEstimator() = default;
void ErlEstimator::Reset() {
blocks_since_reset_ = 0;
}
void ErlEstimator::Update(
const std::vector<bool>& converged_filters,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> render_spectra,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
capture_spectra) {
const size_t num_capture_channels = converged_filters.size();
RTC_DCHECK_EQ(capture_spectra.size(), num_capture_channels);
// Corresponds to WGN of power -46 dBFS.
constexpr float kX2Min = 44015068.0f;
const auto first_converged_iter =
std::find(converged_filters.begin(), converged_filters.end(), true);
const bool any_filter_converged =
first_converged_iter != converged_filters.end();
if (++blocks_since_reset_ < startup_phase_length_blocks__ ||
!any_filter_converged) {
return;
}
// Use the maximum spectrum across capture and the maximum across render.
std::array<float, kFftLengthBy2Plus1> max_capture_spectrum_data;
std::array<float, kFftLengthBy2Plus1> max_capture_spectrum =
capture_spectra[/*channel=*/0];
if (num_capture_channels > 1) {
// Initialize using the first channel with a converged filter.
const size_t first_converged =
std::distance(converged_filters.begin(), first_converged_iter);
RTC_DCHECK_GE(first_converged, 0);
RTC_DCHECK_LT(first_converged, num_capture_channels);
max_capture_spectrum_data = capture_spectra[first_converged];
for (size_t ch = first_converged + 1; ch < num_capture_channels; ++ch) {
if (!converged_filters[ch]) {
continue;
}
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
max_capture_spectrum_data[k] =
std::max(max_capture_spectrum_data[k], capture_spectra[ch][k]);
}
}
max_capture_spectrum = max_capture_spectrum_data;
}
const size_t num_render_channels = render_spectra.size();
std::array<float, kFftLengthBy2Plus1> max_render_spectrum_data;
rtc::ArrayView<const float, kFftLengthBy2Plus1> max_render_spectrum =
render_spectra[/*channel=*/0];
if (num_render_channels > 1) {
std::copy(render_spectra[0].begin(), render_spectra[0].end(),
max_render_spectrum_data.begin());
for (size_t ch = 1; ch < num_render_channels; ++ch) {
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
max_render_spectrum_data[k] =
std::max(max_render_spectrum_data[k], render_spectra[ch][k]);
}
}
max_render_spectrum = max_render_spectrum_data;
}
const auto& X2 = max_render_spectrum;
const auto& Y2 = max_capture_spectrum;
// Update the estimates in a maximum statistics manner.
for (size_t k = 1; k < kFftLengthBy2; ++k) {
if (X2[k] > kX2Min) {
const float new_erl = Y2[k] / X2[k];
if (new_erl < erl_[k]) {
hold_counters_[k - 1] = 1000;
erl_[k] += 0.1f * (new_erl - erl_[k]);
erl_[k] = std::max(erl_[k], kMinErl);
}
}
}
std::for_each(hold_counters_.begin(), hold_counters_.end(),
[](int& a) { --a; });
std::transform(hold_counters_.begin(), hold_counters_.end(), erl_.begin() + 1,
erl_.begin() + 1, [](int a, float b) {
return a > 0 ? b : std::min(kMaxErl, 2.f * b);
});
erl_[0] = erl_[1];
erl_[kFftLengthBy2] = erl_[kFftLengthBy2 - 1];
// Compute ERL over all frequency bins.
const float X2_sum = std::accumulate(X2.begin(), X2.end(), 0.0f);
if (X2_sum > kX2Min * X2.size()) {
const float Y2_sum = std::accumulate(Y2.begin(), Y2.end(), 0.0f);
const float new_erl = Y2_sum / X2_sum;
if (new_erl < erl_time_domain_) {
hold_counter_time_domain_ = 1000;
erl_time_domain_ += 0.1f * (new_erl - erl_time_domain_);
erl_time_domain_ = std::max(erl_time_domain_, kMinErl);
}
}
--hold_counter_time_domain_;
erl_time_domain_ = (hold_counter_time_domain_ > 0)
? erl_time_domain_
: std::min(kMaxErl, 2.f * erl_time_domain_);
}
} // namespace webrtc

57
webrtc/modules/audio_processing/aec3/erl_estimator.h Normal file
View File

@ -0,0 +1,57 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ERL_ESTIMATOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ERL_ESTIMATOR_H_
#include <stddef.h>
#include <array>
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/constructor_magic.h"
namespace webrtc {
// Estimates the echo return loss based on the signal spectra.
class ErlEstimator {
public:
explicit ErlEstimator(size_t startup_phase_length_blocks_);
~ErlEstimator();
// Resets the ERL estimation.
void Reset();
// Updates the ERL estimate.
void Update(const std::vector<bool>& converged_filters,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
render_spectra,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
capture_spectra);
// Returns the most recent ERL estimate.
const std::array<float, kFftLengthBy2Plus1>& Erl() const { return erl_; }
float ErlTimeDomain() const { return erl_time_domain_; }
private:
const size_t startup_phase_length_blocks__;
std::array<float, kFftLengthBy2Plus1> erl_;
std::array<int, kFftLengthBy2Minus1> hold_counters_;
float erl_time_domain_;
int hold_counter_time_domain_;
size_t blocks_since_reset_ = 0;
RTC_DISALLOW_COPY_AND_ASSIGN(ErlEstimator);
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ERL_ESTIMATOR_H_

86
webrtc/modules/audio_processing/aec3/erle_estimator.cc Normal file
View File

@ -0,0 +1,86 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/erle_estimator.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/checks.h"
namespace webrtc {
ErleEstimator::ErleEstimator(size_t startup_phase_length_blocks,
const EchoCanceller3Config& config,
size_t num_capture_channels)
: startup_phase_length_blocks_(startup_phase_length_blocks),
fullband_erle_estimator_(config.erle, num_capture_channels),
subband_erle_estimator_(config, num_capture_channels) {
if (config.erle.num_sections > 1) {
signal_dependent_erle_estimator_ =
std::make_unique<SignalDependentErleEstimator>(config,
num_capture_channels);
}
Reset(true);
}
ErleEstimator::~ErleEstimator() = default;
void ErleEstimator::Reset(bool delay_change) {
fullband_erle_estimator_.Reset();
subband_erle_estimator_.Reset();
if (signal_dependent_erle_estimator_) {
signal_dependent_erle_estimator_->Reset();
}
if (delay_change) {
blocks_since_reset_ = 0;
}
}
void ErleEstimator::Update(
const RenderBuffer& render_buffer,
rtc::ArrayView<const std::vector<std::array<float, kFftLengthBy2Plus1>>>
filter_frequency_responses,
rtc::ArrayView<const float, kFftLengthBy2Plus1>
avg_render_spectrum_with_reverb,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> capture_spectra,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
subtractor_spectra,
const std::vector<bool>& converged_filters) {
RTC_DCHECK_EQ(subband_erle_estimator_.Erle().size(), capture_spectra.size());
RTC_DCHECK_EQ(subband_erle_estimator_.Erle().size(),
subtractor_spectra.size());
const auto& X2_reverb = avg_render_spectrum_with_reverb;
const auto& Y2 = capture_spectra;
const auto& E2 = subtractor_spectra;
if (++blocks_since_reset_ < startup_phase_length_blocks_) {
return;
}
subband_erle_estimator_.Update(X2_reverb, Y2, E2, converged_filters);
if (signal_dependent_erle_estimator_) {
signal_dependent_erle_estimator_->Update(
render_buffer, filter_frequency_responses, X2_reverb, Y2, E2,
subband_erle_estimator_.Erle(), converged_filters);
}
fullband_erle_estimator_.Update(X2_reverb, Y2, E2, converged_filters);
}
void ErleEstimator::Dump(
const std::unique_ptr<ApmDataDumper>& data_dumper) const {
fullband_erle_estimator_.Dump(data_dumper);
subband_erle_estimator_.Dump(data_dumper);
if (signal_dependent_erle_estimator_) {
signal_dependent_erle_estimator_->Dump(data_dumper);
}
}
} // namespace webrtc

99
webrtc/modules/audio_processing/aec3/erle_estimator.h Normal file
View File

@ -0,0 +1,99 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_ERLE_ESTIMATOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_ERLE_ESTIMATOR_H_
#include <stddef.h>
#include <array>
#include <memory>
#include <vector>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/fullband_erle_estimator.h"
#include "modules/audio_processing/aec3/render_buffer.h"
#include "modules/audio_processing/aec3/signal_dependent_erle_estimator.h"
#include "modules/audio_processing/aec3/subband_erle_estimator.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
namespace webrtc {
// Estimates the echo return loss enhancement. One estimate is done per subband
// and another one is done using the aggreation of energy over all the subbands.
class ErleEstimator {
public:
ErleEstimator(size_t startup_phase_length_blocks,
const EchoCanceller3Config& config,
size_t num_capture_channels);
~ErleEstimator();
// Resets the fullband ERLE estimator and the subbands ERLE estimators.
void Reset(bool delay_change);
// Updates the ERLE estimates.
void Update(
const RenderBuffer& render_buffer,
rtc::ArrayView<const std::vector<std::array<float, kFftLengthBy2Plus1>>>
filter_frequency_responses,
rtc::ArrayView<const float, kFftLengthBy2Plus1>
avg_render_spectrum_with_reverb,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
capture_spectra,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
subtractor_spectra,
const std::vector<bool>& converged_filters);
// Returns the most recent subband ERLE estimates.
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> Erle() const {
return signal_dependent_erle_estimator_
? signal_dependent_erle_estimator_->Erle()
: subband_erle_estimator_.Erle();
}
// Returns the subband ERLE that are estimated during onsets (only used for
// testing).
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> ErleOnsets()
const {
return subband_erle_estimator_.ErleOnsets();
}
// Returns the fullband ERLE estimate.
float FullbandErleLog2() const {
return fullband_erle_estimator_.FullbandErleLog2();
}
// Returns an estimation of the current linear filter quality based on the
// current and past fullband ERLE estimates. The returned value is a float
// vector with content between 0 and 1 where 1 indicates that, at this current
// time instant, the linear filter is reaching its maximum subtraction
// performance.
rtc::ArrayView<const absl::optional<float>> GetInstLinearQualityEstimates()
const {
return fullband_erle_estimator_.GetInstLinearQualityEstimates();
}
void Dump(const std::unique_ptr<ApmDataDumper>& data_dumper) const;
private:
const size_t startup_phase_length_blocks_;
FullBandErleEstimator fullband_erle_estimator_;
SubbandErleEstimator subband_erle_estimator_;
std::unique_ptr<SignalDependentErleEstimator>
signal_dependent_erle_estimator_;
size_t blocks_since_reset_ = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_ERLE_ESTIMATOR_H_

27
webrtc/modules/audio_processing/aec3/fft_buffer.cc Normal file
View File

@ -0,0 +1,27 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/fft_buffer.h"
namespace webrtc {
FftBuffer::FftBuffer(size_t size, size_t num_channels)
: size(static_cast<int>(size)),
buffer(size, std::vector<FftData>(num_channels)) {
for (auto& block : buffer) {
for (auto& channel_fft_data : block) {
channel_fft_data.Clear();
}
}
}
FftBuffer::~FftBuffer() = default;
} // namespace webrtc

60
webrtc/modules/audio_processing/aec3/fft_buffer.h Normal file
View File

@ -0,0 +1,60 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_FFT_BUFFER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_FFT_BUFFER_H_
#include <stddef.h>
#include <vector>
#include "modules/audio_processing/aec3/fft_data.h"
#include "rtc_base/checks.h"
namespace webrtc {
// Struct for bundling a circular buffer of FftData objects together with the
// read and write indices.
struct FftBuffer {
FftBuffer(size_t size, size_t num_channels);
~FftBuffer();
int IncIndex(int index) const {
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
return index < size - 1 ? index + 1 : 0;
}
int DecIndex(int index) const {
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
return index > 0 ? index - 1 : size - 1;
}
int OffsetIndex(int index, int offset) const {
RTC_DCHECK_GE(buffer.size(), offset);
RTC_DCHECK_EQ(buffer.size(), static_cast<size_t>(size));
return (size + index + offset) % size;
}
void UpdateWriteIndex(int offset) { write = OffsetIndex(write, offset); }
void IncWriteIndex() { write = IncIndex(write); }
void DecWriteIndex() { write = DecIndex(write); }
void UpdateReadIndex(int offset) { read = OffsetIndex(read, offset); }
void IncReadIndex() { read = IncIndex(read); }
void DecReadIndex() { read = DecIndex(read); }
const int size;
std::vector<std::vector<FftData>> buffer;
int write = 0;
int read = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_FFT_BUFFER_H_

104
webrtc/modules/audio_processing/aec3/fft_data.h Normal file
View File

@ -0,0 +1,104 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_FFT_DATA_H_
#define MODULES_AUDIO_PROCESSING_AEC3_FFT_DATA_H_
// Defines WEBRTC_ARCH_X86_FAMILY, used below.
#include "rtc_base/system/arch.h"
#if defined(WEBRTC_ARCH_X86_FAMILY)
#include <emmintrin.h>
#endif
#include <algorithm>
#include <array>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
namespace webrtc {
// Struct that holds imaginary data produced from 128 point real-valued FFTs.
struct FftData {
// Copies the data in src.
void Assign(const FftData& src) {
std::copy(src.re.begin(), src.re.end(), re.begin());
std::copy(src.im.begin(), src.im.end(), im.begin());
im[0] = im[kFftLengthBy2] = 0;
}
// Clears all the imaginary.
void Clear() {
re.fill(0.f);
im.fill(0.f);
}
// Computes the power spectrum of the data.
void SpectrumAVX2(rtc::ArrayView<float> power_spectrum) const;
// Computes the power spectrum of the data.
void Spectrum(Aec3Optimization optimization,
rtc::ArrayView<float> power_spectrum) const {
RTC_DCHECK_EQ(kFftLengthBy2Plus1, power_spectrum.size());
switch (optimization) {
#if defined(WEBRTC_ARCH_X86_FAMILY)
case Aec3Optimization::kSse2: {
constexpr int kNumFourBinBands = kFftLengthBy2 / 4;
constexpr int kLimit = kNumFourBinBands * 4;
for (size_t k = 0; k < kLimit; k += 4) {
const __m128 r = _mm_loadu_ps(&re[k]);
const __m128 i = _mm_loadu_ps(&im[k]);
const __m128 ii = _mm_mul_ps(i, i);
const __m128 rr = _mm_mul_ps(r, r);
const __m128 rrii = _mm_add_ps(rr, ii);
_mm_storeu_ps(&power_spectrum[k], rrii);
}
power_spectrum[kFftLengthBy2] = re[kFftLengthBy2] * re[kFftLengthBy2] +
im[kFftLengthBy2] * im[kFftLengthBy2];
} break;
case Aec3Optimization::kAvx2:
SpectrumAVX2(power_spectrum);
break;
#endif
default:
std::transform(re.begin(), re.end(), im.begin(), power_spectrum.begin(),
[](float a, float b) { return a * a + b * b; });
}
}
// Copy the data from an interleaved array.
void CopyFromPackedArray(const std::array<float, kFftLength>& v) {
re[0] = v[0];
re[kFftLengthBy2] = v[1];
im[0] = im[kFftLengthBy2] = 0;
for (size_t k = 1, j = 2; k < kFftLengthBy2; ++k) {
re[k] = v[j++];
im[k] = v[j++];
}
}
// Copies the data into an interleaved array.
void CopyToPackedArray(std::array<float, kFftLength>* v) const {
RTC_DCHECK(v);
(*v)[0] = re[0];
(*v)[1] = re[kFftLengthBy2];
for (size_t k = 1, j = 2; k < kFftLengthBy2; ++k) {
(*v)[j++] = re[k];
(*v)[j++] = im[k];
}
}
std::array<float, kFftLengthBy2Plus1> re;
std::array<float, kFftLengthBy2Plus1> im;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_FFT_DATA_H_

33
webrtc/modules/audio_processing/aec3/fft_data_avx2.cc Normal file
View File

@ -0,0 +1,33 @@
/*
* Copyright (c) 2020 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/fft_data.h"
#include <immintrin.h>
#include "api/array_view.h"
namespace webrtc {
// Computes the power spectrum of the data.
void FftData::SpectrumAVX2(rtc::ArrayView<float> power_spectrum) const {
RTC_DCHECK_EQ(kFftLengthBy2Plus1, power_spectrum.size());
for (size_t k = 0; k < kFftLengthBy2; k += 8) {
__m256 r = _mm256_loadu_ps(&re[k]);
__m256 i = _mm256_loadu_ps(&im[k]);
__m256 ii = _mm256_mul_ps(i, i);
ii = _mm256_fmadd_ps(r, r, ii);
_mm256_storeu_ps(&power_spectrum[k], ii);
}
power_spectrum[kFftLengthBy2] = re[kFftLengthBy2] * re[kFftLengthBy2] +
im[kFftLengthBy2] * im[kFftLengthBy2];
}
} // namespace webrtc

280
webrtc/modules/audio_processing/aec3/filter_analyzer.cc Normal file
View File

@ -0,0 +1,280 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/filter_analyzer.h"
#include <math.h>
#include <algorithm>
#include <array>
#include <numeric>
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/render_buffer.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomic_ops.h"
#include "rtc_base/checks.h"
namespace webrtc {
namespace {
size_t FindPeakIndex(rtc::ArrayView<const float> filter_time_domain,
size_t peak_index_in,
size_t start_sample,
size_t end_sample) {
size_t peak_index_out = peak_index_in;
float max_h2 =
filter_time_domain[peak_index_out] * filter_time_domain[peak_index_out];
for (size_t k = start_sample; k <= end_sample; ++k) {
float tmp = filter_time_domain[k] * filter_time_domain[k];
if (tmp > max_h2) {
peak_index_out = k;
max_h2 = tmp;
}
}
return peak_index_out;
}
} // namespace
int FilterAnalyzer::instance_count_ = 0;
FilterAnalyzer::FilterAnalyzer(const EchoCanceller3Config& config,
size_t num_capture_channels)
: data_dumper_(
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
bounded_erl_(config.ep_strength.bounded_erl),
default_gain_(config.ep_strength.default_gain),
h_highpass_(num_capture_channels,
std::vector<float>(
GetTimeDomainLength(config.filter.refined.length_blocks),
0.f)),
filter_analysis_states_(num_capture_channels,
FilterAnalysisState(config)),
filter_delays_blocks_(num_capture_channels, 0) {
Reset();
}
FilterAnalyzer::~FilterAnalyzer() = default;
void FilterAnalyzer::Reset() {
blocks_since_reset_ = 0;
ResetRegion();
for (auto& state : filter_analysis_states_) {
state.Reset(default_gain_);
}
std::fill(filter_delays_blocks_.begin(), filter_delays_blocks_.end(), 0);
}
void FilterAnalyzer::Update(
rtc::ArrayView<const std::vector<float>> filters_time_domain,
const RenderBuffer& render_buffer,
bool* any_filter_consistent,
float* max_echo_path_gain) {
RTC_DCHECK(any_filter_consistent);
RTC_DCHECK(max_echo_path_gain);
RTC_DCHECK_EQ(filters_time_domain.size(), filter_analysis_states_.size());
RTC_DCHECK_EQ(filters_time_domain.size(), h_highpass_.size());
++blocks_since_reset_;
SetRegionToAnalyze(filters_time_domain[0].size());
AnalyzeRegion(filters_time_domain, render_buffer);
// Aggregate the results for all capture channels.
auto& st_ch0 = filter_analysis_states_[0];
*any_filter_consistent = st_ch0.consistent_estimate;
*max_echo_path_gain = st_ch0.gain;
min_filter_delay_blocks_ = filter_delays_blocks_[0];
for (size_t ch = 1; ch < filters_time_domain.size(); ++ch) {
auto& st_ch = filter_analysis_states_[ch];
*any_filter_consistent =
*any_filter_consistent || st_ch.consistent_estimate;
*max_echo_path_gain = std::max(*max_echo_path_gain, st_ch.gain);
min_filter_delay_blocks_ =
std::min(min_filter_delay_blocks_, filter_delays_blocks_[ch]);
}
}
void FilterAnalyzer::AnalyzeRegion(
rtc::ArrayView<const std::vector<float>> filters_time_domain,
const RenderBuffer& render_buffer) {
// Preprocess the filter to avoid issues with low-frequency components in the
// filter.
PreProcessFilters(filters_time_domain);
data_dumper_->DumpRaw("aec3_linear_filter_processed_td", h_highpass_[0]);
constexpr float kOneByBlockSize = 1.f / kBlockSize;
for (size_t ch = 0; ch < filters_time_domain.size(); ++ch) {
RTC_DCHECK_LT(region_.start_sample_, filters_time_domain[ch].size());
RTC_DCHECK_LT(region_.end_sample_, filters_time_domain[ch].size());
auto& st_ch = filter_analysis_states_[ch];
RTC_DCHECK_EQ(h_highpass_[ch].size(), filters_time_domain[ch].size());
RTC_DCHECK_GT(h_highpass_[ch].size(), 0);
st_ch.peak_index = std::min(st_ch.peak_index, h_highpass_[ch].size() - 1);
st_ch.peak_index =
FindPeakIndex(h_highpass_[ch], st_ch.peak_index, region_.start_sample_,
region_.end_sample_);
filter_delays_blocks_[ch] = st_ch.peak_index >> kBlockSizeLog2;
UpdateFilterGain(h_highpass_[ch], &st_ch);
st_ch.filter_length_blocks =
filters_time_domain[ch].size() * kOneByBlockSize;
st_ch.consistent_estimate = st_ch.consistent_filter_detector.Detect(
h_highpass_[ch], region_,
render_buffer.Block(-filter_delays_blocks_[ch])[0], st_ch.peak_index,
filter_delays_blocks_[ch]);
}
}
void FilterAnalyzer::UpdateFilterGain(
rtc::ArrayView<const float> filter_time_domain,
FilterAnalysisState* st) {
bool sufficient_time_to_converge =
blocks_since_reset_ > 5 * kNumBlocksPerSecond;
if (sufficient_time_to_converge && st->consistent_estimate) {
st->gain = fabsf(filter_time_domain[st->peak_index]);
} else {
// TODO(peah): Verify whether this check against a float is ok.
if (st->gain) {
st->gain = std::max(st->gain, fabsf(filter_time_domain[st->peak_index]));
}
}
if (bounded_erl_ && st->gain) {
st->gain = std::max(st->gain, 0.01f);
}
}
void FilterAnalyzer::PreProcessFilters(
rtc::ArrayView<const std::vector<float>> filters_time_domain) {
for (size_t ch = 0; ch < filters_time_domain.size(); ++ch) {
RTC_DCHECK_LT(region_.start_sample_, filters_time_domain[ch].size());
RTC_DCHECK_LT(region_.end_sample_, filters_time_domain[ch].size());
RTC_DCHECK_GE(h_highpass_[ch].capacity(), filters_time_domain[ch].size());
h_highpass_[ch].resize(filters_time_domain[ch].size());
// Minimum phase high-pass filter with cutoff frequency at about 600 Hz.
constexpr std::array<float, 3> h = {
{0.7929742f, -0.36072128f, -0.47047766f}};
std::fill(h_highpass_[ch].begin() + region_.start_sample_,
h_highpass_[ch].begin() + region_.end_sample_ + 1, 0.f);
for (size_t k = std::max(h.size() - 1, region_.start_sample_);
k <= region_.end_sample_; ++k) {
for (size_t j = 0; j < h.size(); ++j) {
h_highpass_[ch][k] += filters_time_domain[ch][k - j] * h[j];
}
}
}
}
void FilterAnalyzer::ResetRegion() {
region_.start_sample_ = 0;
region_.end_sample_ = 0;
}
void FilterAnalyzer::SetRegionToAnalyze(size_t filter_size) {
constexpr size_t kNumberBlocksToUpdate = 1;
auto& r = region_;
r.start_sample_ = r.end_sample_ >= filter_size - 1 ? 0 : r.end_sample_ + 1;
r.end_sample_ =
std::min(r.start_sample_ + kNumberBlocksToUpdate * kBlockSize - 1,
filter_size - 1);
// Check range.
RTC_DCHECK_LT(r.start_sample_, filter_size);
RTC_DCHECK_LT(r.end_sample_, filter_size);
RTC_DCHECK_LE(r.start_sample_, r.end_sample_);
}
FilterAnalyzer::ConsistentFilterDetector::ConsistentFilterDetector(
const EchoCanceller3Config& config)
: active_render_threshold_(config.render_levels.active_render_limit *
config.render_levels.active_render_limit *
kFftLengthBy2) {
Reset();
}
void FilterAnalyzer::ConsistentFilterDetector::Reset() {
significant_peak_ = false;
filter_floor_accum_ = 0.f;
filter_secondary_peak_ = 0.f;
filter_floor_low_limit_ = 0;
filter_floor_high_limit_ = 0;
consistent_estimate_counter_ = 0;
consistent_delay_reference_ = -10;
}
bool FilterAnalyzer::ConsistentFilterDetector::Detect(
rtc::ArrayView<const float> filter_to_analyze,
const FilterRegion& region,
rtc::ArrayView<const std::vector<float>> x_block,
size_t peak_index,
int delay_blocks) {
if (region.start_sample_ == 0) {
filter_floor_accum_ = 0.f;
filter_secondary_peak_ = 0.f;
filter_floor_low_limit_ = peak_index < 64 ? 0 : peak_index - 64;
filter_floor_high_limit_ =
peak_index > filter_to_analyze.size() - 129 ? 0 : peak_index + 128;
}
for (size_t k = region.start_sample_;
k < std::min(region.end_sample_ + 1, filter_floor_low_limit_); ++k) {
float abs_h = fabsf(filter_to_analyze[k]);
filter_floor_accum_ += abs_h;
filter_secondary_peak_ = std::max(filter_secondary_peak_, abs_h);
}
for (size_t k = std::max(filter_floor_high_limit_, region.start_sample_);
k <= region.end_sample_; ++k) {
float abs_h = fabsf(filter_to_analyze[k]);
filter_floor_accum_ += abs_h;
filter_secondary_peak_ = std::max(filter_secondary_peak_, abs_h);
}
if (region.end_sample_ == filter_to_analyze.size() - 1) {
float filter_floor = filter_floor_accum_ /
(filter_floor_low_limit_ + filter_to_analyze.size() -
filter_floor_high_limit_);
float abs_peak = fabsf(filter_to_analyze[peak_index]);
significant_peak_ = abs_peak > 10.f * filter_floor &&
abs_peak > 2.f * filter_secondary_peak_;
}
if (significant_peak_) {
bool active_render_block = false;
for (auto& x_channel : x_block) {
const float x_energy = std::inner_product(
x_channel.begin(), x_channel.end(), x_channel.begin(), 0.f);
if (x_energy > active_render_threshold_) {
active_render_block = true;
break;
}
}
if (consistent_delay_reference_ == delay_blocks) {
if (active_render_block) {
++consistent_estimate_counter_;
}
} else {
consistent_estimate_counter_ = 0;
consistent_delay_reference_ = delay_blocks;
}
}
return consistent_estimate_counter_ > 1.5f * kNumBlocksPerSecond;
}
} // namespace webrtc

149
webrtc/modules/audio_processing/aec3/filter_analyzer.h Normal file
View File

@ -0,0 +1,149 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_FILTER_ANALYZER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_FILTER_ANALYZER_H_
#include <stddef.h>
#include <array>
#include <memory>
#include <vector>
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/constructor_magic.h"
namespace webrtc {
class ApmDataDumper;
class RenderBuffer;
// Class for analyzing the properties of an adaptive filter.
class FilterAnalyzer {
public:
FilterAnalyzer(const EchoCanceller3Config& config,
size_t num_capture_channels);
~FilterAnalyzer();
FilterAnalyzer(const FilterAnalyzer&) = delete;
FilterAnalyzer& operator=(const FilterAnalyzer&) = delete;
// Resets the analysis.
void Reset();
// Updates the estimates with new input data.
void Update(rtc::ArrayView<const std::vector<float>> filters_time_domain,
const RenderBuffer& render_buffer,
bool* any_filter_consistent,
float* max_echo_path_gain);
// Returns the delay in blocks for each filter.
rtc::ArrayView<const int> FilterDelaysBlocks() const {
return filter_delays_blocks_;
}
// Returns the minimum delay of all filters in terms of blocks.
int MinFilterDelayBlocks() const { return min_filter_delay_blocks_; }
// Returns the number of blocks for the current used filter.
int FilterLengthBlocks() const {
return filter_analysis_states_[0].filter_length_blocks;
}
// Returns the preprocessed filter.
rtc::ArrayView<const std::vector<float>> GetAdjustedFilters() const {
return h_highpass_;
}
// Public for testing purposes only.
void SetRegionToAnalyze(size_t filter_size);
private:
struct FilterAnalysisState;
void AnalyzeRegion(
rtc::ArrayView<const std::vector<float>> filters_time_domain,
const RenderBuffer& render_buffer);
void UpdateFilterGain(rtc::ArrayView<const float> filters_time_domain,
FilterAnalysisState* st);
void PreProcessFilters(
rtc::ArrayView<const std::vector<float>> filters_time_domain);
void ResetRegion();
struct FilterRegion {
size_t start_sample_;
size_t end_sample_;
};
// This class checks whether the shape of the impulse response has been
// consistent over time.
class ConsistentFilterDetector {
public:
explicit ConsistentFilterDetector(const EchoCanceller3Config& config);
void Reset();
bool Detect(rtc::ArrayView<const float> filter_to_analyze,
const FilterRegion& region,
rtc::ArrayView<const std::vector<float>> x_block,
size_t peak_index,
int delay_blocks);
private:
bool significant_peak_;
float filter_floor_accum_;
float filter_secondary_peak_;
size_t filter_floor_low_limit_;
size_t filter_floor_high_limit_;
const float active_render_threshold_;
size_t consistent_estimate_counter_ = 0;
int consistent_delay_reference_ = -10;
};
struct FilterAnalysisState {
explicit FilterAnalysisState(const EchoCanceller3Config& config)
: filter_length_blocks(config.filter.refined_initial.length_blocks),
consistent_filter_detector(config) {
Reset(config.ep_strength.default_gain);
}
void Reset(float default_gain) {
peak_index = 0;
gain = default_gain;
consistent_filter_detector.Reset();
}
float gain;
size_t peak_index;
int filter_length_blocks;
bool consistent_estimate = false;
ConsistentFilterDetector consistent_filter_detector;
};
static int instance_count_;
std::unique_ptr<ApmDataDumper> data_dumper_;
const bool bounded_erl_;
const float default_gain_;
std::vector<std::vector<float>> h_highpass_;
size_t blocks_since_reset_ = 0;
FilterRegion region_;
std::vector<FilterAnalysisState> filter_analysis_states_;
std::vector<int> filter_delays_blocks_;
int min_filter_delay_blocks_ = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_FILTER_ANALYZER_H_

88
webrtc/modules/audio_processing/aec3/frame_blocker.cc Normal file
View File

@ -0,0 +1,88 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/frame_blocker.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/checks.h"
namespace webrtc {
FrameBlocker::FrameBlocker(size_t num_bands, size_t num_channels)
: num_bands_(num_bands),
num_channels_(num_channels),
buffer_(num_bands_, std::vector<std::vector<float>>(num_channels)) {
RTC_DCHECK_LT(0, num_bands);
RTC_DCHECK_LT(0, num_channels);
for (auto& band : buffer_) {
for (auto& channel : band) {
channel.reserve(kBlockSize);
RTC_DCHECK(channel.empty());
}
}
}
FrameBlocker::~FrameBlocker() = default;
void FrameBlocker::InsertSubFrameAndExtractBlock(
const std::vector<std::vector<rtc::ArrayView<float>>>& sub_frame,
std::vector<std::vector<std::vector<float>>>* block) {
RTC_DCHECK(block);
RTC_DCHECK_EQ(num_bands_, block->size());
RTC_DCHECK_EQ(num_bands_, sub_frame.size());
for (size_t band = 0; band < num_bands_; ++band) {
RTC_DCHECK_EQ(num_channels_, (*block)[band].size());
RTC_DCHECK_EQ(num_channels_, sub_frame[band].size());
for (size_t channel = 0; channel < num_channels_; ++channel) {
RTC_DCHECK_GE(kBlockSize - 16, buffer_[band][channel].size());
RTC_DCHECK_EQ(kBlockSize, (*block)[band][channel].size());
RTC_DCHECK_EQ(kSubFrameLength, sub_frame[band][channel].size());
const int samples_to_block = kBlockSize - buffer_[band][channel].size();
(*block)[band][channel].clear();
(*block)[band][channel].insert((*block)[band][channel].begin(),
buffer_[band][channel].begin(),
buffer_[band][channel].end());
(*block)[band][channel].insert(
(*block)[band][channel].begin() + buffer_[band][channel].size(),
sub_frame[band][channel].begin(),
sub_frame[band][channel].begin() + samples_to_block);
buffer_[band][channel].clear();
buffer_[band][channel].insert(
buffer_[band][channel].begin(),
sub_frame[band][channel].begin() + samples_to_block,
sub_frame[band][channel].end());
}
}
}
bool FrameBlocker::IsBlockAvailable() const {
return kBlockSize == buffer_[0][0].size();
}
void FrameBlocker::ExtractBlock(
std::vector<std::vector<std::vector<float>>>* block) {
RTC_DCHECK(block);
RTC_DCHECK_EQ(num_bands_, block->size());
RTC_DCHECK(IsBlockAvailable());
for (size_t band = 0; band < num_bands_; ++band) {
RTC_DCHECK_EQ(num_channels_, (*block)[band].size());
for (size_t channel = 0; channel < num_channels_; ++channel) {
RTC_DCHECK_EQ(kBlockSize, buffer_[band][channel].size());
RTC_DCHECK_EQ(kBlockSize, (*block)[band][channel].size());
(*block)[band][channel].clear();
(*block)[band][channel].insert((*block)[band][channel].begin(),
buffer_[band][channel].begin(),
buffer_[band][channel].end());
buffer_[band][channel].clear();
}
}
}
} // namespace webrtc

50
webrtc/modules/audio_processing/aec3/frame_blocker.h Normal file
View File

@ -0,0 +1,50 @@
/*
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_FRAME_BLOCKER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_FRAME_BLOCKER_H_
#include <stddef.h>
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
namespace webrtc {
// Class for producing 64 sample multiband blocks from frames consisting of 2
// subframes of 80 samples.
class FrameBlocker {
public:
FrameBlocker(size_t num_bands, size_t num_channels);
~FrameBlocker();
FrameBlocker(const FrameBlocker&) = delete;
FrameBlocker& operator=(const FrameBlocker&) = delete;
// Inserts one 80 sample multiband subframe from the multiband frame and
// extracts one 64 sample multiband block.
void InsertSubFrameAndExtractBlock(
const std::vector<std::vector<rtc::ArrayView<float>>>& sub_frame,
std::vector<std::vector<std::vector<float>>>* block);
// Reports whether a multiband block of 64 samples is available for
// extraction.
bool IsBlockAvailable() const;
// Extracts a multiband block of 64 samples.
void ExtractBlock(std::vector<std::vector<std::vector<float>>>* block);
private:
const size_t num_bands_;
const size_t num_channels_;
std::vector<std::vector<std::vector<float>>> buffer_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_FRAME_BLOCKER_H_

200
webrtc/modules/audio_processing/aec3/fullband_erle_estimator.cc Normal file
View File

@ -0,0 +1,200 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/fullband_erle_estimator.h"
#include <algorithm>
#include <memory>
#include <numeric>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/checks.h"
#include "rtc_base/numerics/safe_minmax.h"
namespace webrtc {
namespace {
constexpr float kEpsilon = 1e-3f;
constexpr float kX2BandEnergyThreshold = 44015068.0f;
constexpr int kBlocksToHoldErle = 100;
constexpr int kPointsToAccumulate = 6;
} // namespace
FullBandErleEstimator::FullBandErleEstimator(
const EchoCanceller3Config::Erle& config,
size_t num_capture_channels)
: min_erle_log2_(FastApproxLog2f(config.min + kEpsilon)),
max_erle_lf_log2(FastApproxLog2f(config.max_l + kEpsilon)),
hold_counters_time_domain_(num_capture_channels, 0),
erle_time_domain_log2_(num_capture_channels, min_erle_log2_),
instantaneous_erle_(num_capture_channels, ErleInstantaneous(config)),
linear_filters_qualities_(num_capture_channels) {
Reset();
}
FullBandErleEstimator::~FullBandErleEstimator() = default;
void FullBandErleEstimator::Reset() {
for (auto& instantaneous_erle_ch : instantaneous_erle_) {
instantaneous_erle_ch.Reset();
}
UpdateQualityEstimates();
std::fill(erle_time_domain_log2_.begin(), erle_time_domain_log2_.end(),
min_erle_log2_);
std::fill(hold_counters_time_domain_.begin(),
hold_counters_time_domain_.end(), 0);
}
void FullBandErleEstimator::Update(
rtc::ArrayView<const float> X2,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> Y2,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> E2,
const std::vector<bool>& converged_filters) {
for (size_t ch = 0; ch < Y2.size(); ++ch) {
if (converged_filters[ch]) {
// Computes the fullband ERLE.
const float X2_sum = std::accumulate(X2.begin(), X2.end(), 0.0f);
if (X2_sum > kX2BandEnergyThreshold * X2.size()) {
const float Y2_sum =
std::accumulate(Y2[ch].begin(), Y2[ch].end(), 0.0f);
const float E2_sum =
std::accumulate(E2[ch].begin(), E2[ch].end(), 0.0f);
if (instantaneous_erle_[ch].Update(Y2_sum, E2_sum)) {
hold_counters_time_domain_[ch] = kBlocksToHoldErle;
erle_time_domain_log2_[ch] +=
0.1f * ((instantaneous_erle_[ch].GetInstErleLog2().value()) -
erle_time_domain_log2_[ch]);
erle_time_domain_log2_[ch] = rtc::SafeClamp(
erle_time_domain_log2_[ch], min_erle_log2_, max_erle_lf_log2);
}
}
}
--hold_counters_time_domain_[ch];
if (hold_counters_time_domain_[ch] <= 0) {
erle_time_domain_log2_[ch] =
std::max(min_erle_log2_, erle_time_domain_log2_[ch] - 0.044f);
}
if (hold_counters_time_domain_[ch] == 0) {
instantaneous_erle_[ch].ResetAccumulators();
}
}
UpdateQualityEstimates();
}
void FullBandErleEstimator::Dump(
const std::unique_ptr<ApmDataDumper>& data_dumper) const {
data_dumper->DumpRaw("aec3_fullband_erle_log2", FullbandErleLog2());
instantaneous_erle_[0].Dump(data_dumper);
}
void FullBandErleEstimator::UpdateQualityEstimates() {
for (size_t ch = 0; ch < instantaneous_erle_.size(); ++ch) {
linear_filters_qualities_[ch] =
instantaneous_erle_[ch].GetQualityEstimate();
}
}
FullBandErleEstimator::ErleInstantaneous::ErleInstantaneous(
const EchoCanceller3Config::Erle& config)
: clamp_inst_quality_to_zero_(config.clamp_quality_estimate_to_zero),
clamp_inst_quality_to_one_(config.clamp_quality_estimate_to_one) {
Reset();
}
FullBandErleEstimator::ErleInstantaneous::~ErleInstantaneous() = default;
bool FullBandErleEstimator::ErleInstantaneous::Update(const float Y2_sum,
const float E2_sum) {
bool update_estimates = false;
E2_acum_ += E2_sum;
Y2_acum_ += Y2_sum;
num_points_++;
if (num_points_ == kPointsToAccumulate) {
if (E2_acum_ > 0.f) {
update_estimates = true;
erle_log2_ = FastApproxLog2f(Y2_acum_ / E2_acum_ + kEpsilon);
}
num_points_ = 0;
E2_acum_ = 0.f;
Y2_acum_ = 0.f;
}
if (update_estimates) {
UpdateMaxMin();
UpdateQualityEstimate();
}
return update_estimates;
}
void FullBandErleEstimator::ErleInstantaneous::Reset() {
ResetAccumulators();
max_erle_log2_ = -10.f; // -30 dB.
min_erle_log2_ = 33.f; // 100 dB.
inst_quality_estimate_ = 0.f;
}
void FullBandErleEstimator::ErleInstantaneous::ResetAccumulators() {
erle_log2_ = absl::nullopt;
inst_quality_estimate_ = 0.f;
num_points_ = 0;
E2_acum_ = 0.f;
Y2_acum_ = 0.f;
}
void FullBandErleEstimator::ErleInstantaneous::Dump(
const std::unique_ptr<ApmDataDumper>& data_dumper) const {
data_dumper->DumpRaw("aec3_fullband_erle_inst_log2",
erle_log2_ ? *erle_log2_ : -10.f);
data_dumper->DumpRaw(
"aec3_erle_instantaneous_quality",
GetQualityEstimate() ? GetQualityEstimate().value() : 0.f);
data_dumper->DumpRaw("aec3_fullband_erle_max_log2", max_erle_log2_);
data_dumper->DumpRaw("aec3_fullband_erle_min_log2", min_erle_log2_);
}
void FullBandErleEstimator::ErleInstantaneous::UpdateMaxMin() {
RTC_DCHECK(erle_log2_);
if (erle_log2_.value() > max_erle_log2_) {
max_erle_log2_ = erle_log2_.value();
} else {
max_erle_log2_ -= 0.0004; // Forget factor, approx 1dB every 3 sec.
}
if (erle_log2_.value() < min_erle_log2_) {
min_erle_log2_ = erle_log2_.value();
} else {
min_erle_log2_ += 0.0004; // Forget factor, approx 1dB every 3 sec.
}
}
void FullBandErleEstimator::ErleInstantaneous::UpdateQualityEstimate() {
const float alpha = 0.07f;
float quality_estimate = 0.f;
RTC_DCHECK(erle_log2_);
// TODO(peah): Currently, the estimate can become be less than 0; this should
// be corrected.
if (max_erle_log2_ > min_erle_log2_) {
quality_estimate = (erle_log2_.value() - min_erle_log2_) /
(max_erle_log2_ - min_erle_log2_);
}
if (quality_estimate > inst_quality_estimate_) {
inst_quality_estimate_ = quality_estimate;
} else {
inst_quality_estimate_ +=
alpha * (quality_estimate - inst_quality_estimate_);
}
}
} // namespace webrtc

118
webrtc/modules/audio_processing/aec3/fullband_erle_estimator.h Normal file
View File

@ -0,0 +1,118 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_FULLBAND_ERLE_ESTIMATOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_FULLBAND_ERLE_ESTIMATOR_H_
#include <memory>
#include <vector>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
namespace webrtc {
// Estimates the echo return loss enhancement using the energy of all the
// freuquency bands.
class FullBandErleEstimator {
public:
FullBandErleEstimator(const EchoCanceller3Config::Erle& config,
size_t num_capture_channels);
~FullBandErleEstimator();
// Resets the ERLE estimator.
void Reset();
// Updates the ERLE estimator.
void Update(rtc::ArrayView<const float> X2,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> Y2,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> E2,
const std::vector<bool>& converged_filters);
// Returns the fullband ERLE estimates in log2 units.
float FullbandErleLog2() const {
float min_erle = erle_time_domain_log2_[0];
for (size_t ch = 1; ch < erle_time_domain_log2_.size(); ++ch) {
min_erle = std::min(min_erle, erle_time_domain_log2_[ch]);
}
return min_erle;
}
// Returns an estimation of the current linear filter quality. It returns a
// float number between 0 and 1 mapping 1 to the highest possible quality.
rtc::ArrayView<const absl::optional<float>> GetInstLinearQualityEstimates()
const {
return linear_filters_qualities_;
}
void Dump(const std::unique_ptr<ApmDataDumper>& data_dumper) const;
private:
void UpdateQualityEstimates();
class ErleInstantaneous {
public:
explicit ErleInstantaneous(const EchoCanceller3Config::Erle& config);
~ErleInstantaneous();
// Updates the estimator with a new point, returns true
// if the instantaneous ERLE was updated due to having enough
// points for performing the estimate.
bool Update(const float Y2_sum, const float E2_sum);
// Resets the instantaneous ERLE estimator to its initial state.
void Reset();
// Resets the members related with an instantaneous estimate.
void ResetAccumulators();
// Returns the instantaneous ERLE in log2 units.
absl::optional<float> GetInstErleLog2() const { return erle_log2_; }
// Gets an indication between 0 and 1 of the performance of the linear
// filter for the current time instant.
absl::optional<float> GetQualityEstimate() const {
if (erle_log2_) {
float value = inst_quality_estimate_;
if (clamp_inst_quality_to_zero_) {
value = std::max(0.f, value);
}
if (clamp_inst_quality_to_one_) {
value = std::min(1.f, value);
}
return absl::optional<float>(value);
}
return absl::nullopt;
}
void Dump(const std::unique_ptr<ApmDataDumper>& data_dumper) const;
private:
void UpdateMaxMin();
void UpdateQualityEstimate();
const bool clamp_inst_quality_to_zero_;
const bool clamp_inst_quality_to_one_;
absl::optional<float> erle_log2_;
float inst_quality_estimate_;
float max_erle_log2_;
float min_erle_log2_;
float Y2_acum_;
float E2_acum_;
int num_points_;
};
const float min_erle_log2_;
const float max_erle_lf_log2;
std::vector<int> hold_counters_time_domain_;
std::vector<float> erle_time_domain_log2_;
std::vector<ErleInstantaneous> instantaneous_erle_;
std::vector<absl::optional<float>> linear_filters_qualities_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_FULLBAND_ERLE_ESTIMATOR_H_

464
webrtc/modules/audio_processing/aec3/matched_filter.cc Normal file
View File

@ -0,0 +1,464 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/matched_filter.h"
// Defines WEBRTC_ARCH_X86_FAMILY, used below.
#include "rtc_base/system/arch.h"
#if defined(WEBRTC_HAS_NEON)
#include <arm_neon.h>
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
#include <emmintrin.h>
#endif
#include <algorithm>
#include <cstddef>
#include <initializer_list>
#include <iterator>
#include <numeric>
#include "modules/audio_processing/aec3/downsampled_render_buffer.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/checks.h"
#include "rtc_base/logging.h"
namespace webrtc {
namespace aec3 {
#if defined(WEBRTC_HAS_NEON)
void MatchedFilterCore_NEON(size_t x_start_index,
float x2_sum_threshold,
float smoothing,
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float> h,
bool* filters_updated,
float* error_sum) {
const int h_size = static_cast<int>(h.size());
const int x_size = static_cast<int>(x.size());
RTC_DCHECK_EQ(0, h_size % 4);
// Process for all samples in the sub-block.
for (size_t i = 0; i < y.size(); ++i) {
// Apply the matched filter as filter * x, and compute x * x.
RTC_DCHECK_GT(x_size, x_start_index);
const float* x_p = &x[x_start_index];
const float* h_p = &h[0];
// Initialize values for the accumulation.
float32x4_t s_128 = vdupq_n_f32(0);
float32x4_t x2_sum_128 = vdupq_n_f32(0);
float x2_sum = 0.f;
float s = 0;
// Compute loop chunk sizes until, and after, the wraparound of the circular
// buffer for x.
const int chunk1 =
std::min(h_size, static_cast<int>(x_size - x_start_index));
// Perform the loop in two chunks.
const int chunk2 = h_size - chunk1;
for (int limit : {chunk1, chunk2}) {
// Perform 128 bit vector operations.
const int limit_by_4 = limit >> 2;
for (int k = limit_by_4; k > 0; --k, h_p += 4, x_p += 4) {
// Load the data into 128 bit vectors.
const float32x4_t x_k = vld1q_f32(x_p);
const float32x4_t h_k = vld1q_f32(h_p);
// Compute and accumulate x * x and h * x.
x2_sum_128 = vmlaq_f32(x2_sum_128, x_k, x_k);
s_128 = vmlaq_f32(s_128, h_k, x_k);
}
// Perform non-vector operations for any remaining items.
for (int k = limit - limit_by_4 * 4; k > 0; --k, ++h_p, ++x_p) {
const float x_k = *x_p;
x2_sum += x_k * x_k;
s += *h_p * x_k;
}
x_p = &x[0];
}
// Combine the accumulated vector and scalar values.
float* v = reinterpret_cast<float*>(&x2_sum_128);
x2_sum += v[0] + v[1] + v[2] + v[3];
v = reinterpret_cast<float*>(&s_128);
s += v[0] + v[1] + v[2] + v[3];
// Compute the matched filter error.
float e = y[i] - s;
const bool saturation = y[i] >= 32000.f || y[i] <= -32000.f;
(*error_sum) += e * e;
// Update the matched filter estimate in an NLMS manner.
if (x2_sum > x2_sum_threshold && !saturation) {
RTC_DCHECK_LT(0.f, x2_sum);
const float alpha = smoothing * e / x2_sum;
const float32x4_t alpha_128 = vmovq_n_f32(alpha);
// filter = filter + smoothing * (y - filter * x) * x / x * x.
float* h_p = &h[0];
x_p = &x[x_start_index];
// Perform the loop in two chunks.
for (int limit : {chunk1, chunk2}) {
// Perform 128 bit vector operations.
const int limit_by_4 = limit >> 2;
for (int k = limit_by_4; k > 0; --k, h_p += 4, x_p += 4) {
// Load the data into 128 bit vectors.
float32x4_t h_k = vld1q_f32(h_p);
const float32x4_t x_k = vld1q_f32(x_p);
// Compute h = h + alpha * x.
h_k = vmlaq_f32(h_k, alpha_128, x_k);
// Store the result.
vst1q_f32(h_p, h_k);
}
// Perform non-vector operations for any remaining items.
for (int k = limit - limit_by_4 * 4; k > 0; --k, ++h_p, ++x_p) {
*h_p += alpha * *x_p;
}
x_p = &x[0];
}
*filters_updated = true;
}
x_start_index = x_start_index > 0 ? x_start_index - 1 : x_size - 1;
}
}
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
void MatchedFilterCore_SSE2(size_t x_start_index,
float x2_sum_threshold,
float smoothing,
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float> h,
bool* filters_updated,
float* error_sum) {
const int h_size = static_cast<int>(h.size());
const int x_size = static_cast<int>(x.size());
RTC_DCHECK_EQ(0, h_size % 4);
// Process for all samples in the sub-block.
for (size_t i = 0; i < y.size(); ++i) {
// Apply the matched filter as filter * x, and compute x * x.
RTC_DCHECK_GT(x_size, x_start_index);
const float* x_p = &x[x_start_index];
const float* h_p = &h[0];
// Initialize values for the accumulation.
__m128 s_128 = _mm_set1_ps(0);
__m128 x2_sum_128 = _mm_set1_ps(0);
float x2_sum = 0.f;
float s = 0;
// Compute loop chunk sizes until, and after, the wraparound of the circular
// buffer for x.
const int chunk1 =
std::min(h_size, static_cast<int>(x_size - x_start_index));
// Perform the loop in two chunks.
const int chunk2 = h_size - chunk1;
for (int limit : {chunk1, chunk2}) {
// Perform 128 bit vector operations.
const int limit_by_4 = limit >> 2;
for (int k = limit_by_4; k > 0; --k, h_p += 4, x_p += 4) {
// Load the data into 128 bit vectors.
const __m128 x_k = _mm_loadu_ps(x_p);
const __m128 h_k = _mm_loadu_ps(h_p);
const __m128 xx = _mm_mul_ps(x_k, x_k);
// Compute and accumulate x * x and h * x.
x2_sum_128 = _mm_add_ps(x2_sum_128, xx);
const __m128 hx = _mm_mul_ps(h_k, x_k);
s_128 = _mm_add_ps(s_128, hx);
}
// Perform non-vector operations for any remaining items.
for (int k = limit - limit_by_4 * 4; k > 0; --k, ++h_p, ++x_p) {
const float x_k = *x_p;
x2_sum += x_k * x_k;
s += *h_p * x_k;
}
x_p = &x[0];
}
// Combine the accumulated vector and scalar values.
float* v = reinterpret_cast<float*>(&x2_sum_128);
x2_sum += v[0] + v[1] + v[2] + v[3];
v = reinterpret_cast<float*>(&s_128);
s += v[0] + v[1] + v[2] + v[3];
// Compute the matched filter error.
float e = y[i] - s;
const bool saturation = y[i] >= 32000.f || y[i] <= -32000.f;
(*error_sum) += e * e;
// Update the matched filter estimate in an NLMS manner.
if (x2_sum > x2_sum_threshold && !saturation) {
RTC_DCHECK_LT(0.f, x2_sum);
const float alpha = smoothing * e / x2_sum;
const __m128 alpha_128 = _mm_set1_ps(alpha);
// filter = filter + smoothing * (y - filter * x) * x / x * x.
float* h_p = &h[0];
x_p = &x[x_start_index];
// Perform the loop in two chunks.
for (int limit : {chunk1, chunk2}) {
// Perform 128 bit vector operations.
const int limit_by_4 = limit >> 2;
for (int k = limit_by_4; k > 0; --k, h_p += 4, x_p += 4) {
// Load the data into 128 bit vectors.
__m128 h_k = _mm_loadu_ps(h_p);
const __m128 x_k = _mm_loadu_ps(x_p);
// Compute h = h + alpha * x.
const __m128 alpha_x = _mm_mul_ps(alpha_128, x_k);
h_k = _mm_add_ps(h_k, alpha_x);
// Store the result.
_mm_storeu_ps(h_p, h_k);
}
// Perform non-vector operations for any remaining items.
for (int k = limit - limit_by_4 * 4; k > 0; --k, ++h_p, ++x_p) {
*h_p += alpha * *x_p;
}
x_p = &x[0];
}
*filters_updated = true;
}
x_start_index = x_start_index > 0 ? x_start_index - 1 : x_size - 1;
}
}
#endif
void MatchedFilterCore(size_t x_start_index,
float x2_sum_threshold,
float smoothing,
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float> h,
bool* filters_updated,
float* error_sum) {
// Process for all samples in the sub-block.
for (size_t i = 0; i < y.size(); ++i) {
// Apply the matched filter as filter * x, and compute x * x.
float x2_sum = 0.f;
float s = 0;
size_t x_index = x_start_index;
for (size_t k = 0; k < h.size(); ++k) {
x2_sum += x[x_index] * x[x_index];
s += h[k] * x[x_index];
x_index = x_index < (x.size() - 1) ? x_index + 1 : 0;
}
// Compute the matched filter error.
float e = y[i] - s;
const bool saturation = y[i] >= 32000.f || y[i] <= -32000.f;
(*error_sum) += e * e;
// Update the matched filter estimate in an NLMS manner.
if (x2_sum > x2_sum_threshold && !saturation) {
RTC_DCHECK_LT(0.f, x2_sum);
const float alpha = smoothing * e / x2_sum;
// filter = filter + smoothing * (y - filter * x) * x / x * x.
size_t x_index = x_start_index;
for (size_t k = 0; k < h.size(); ++k) {
h[k] += alpha * x[x_index];
x_index = x_index < (x.size() - 1) ? x_index + 1 : 0;
}
*filters_updated = true;
}
x_start_index = x_start_index > 0 ? x_start_index - 1 : x.size() - 1;
}
}
} // namespace aec3
MatchedFilter::MatchedFilter(ApmDataDumper* data_dumper,
Aec3Optimization optimization,
size_t sub_block_size,
size_t window_size_sub_blocks,
int num_matched_filters,
size_t alignment_shift_sub_blocks,
float excitation_limit,
float smoothing,
float matching_filter_threshold)
: data_dumper_(data_dumper),
optimization_(optimization),
sub_block_size_(sub_block_size),
filter_intra_lag_shift_(alignment_shift_sub_blocks * sub_block_size_),
filters_(
num_matched_filters,
std::vector<float>(window_size_sub_blocks * sub_block_size_, 0.f)),
lag_estimates_(num_matched_filters),
filters_offsets_(num_matched_filters, 0),
excitation_limit_(excitation_limit),
smoothing_(smoothing),
matching_filter_threshold_(matching_filter_threshold) {
RTC_DCHECK(data_dumper);
RTC_DCHECK_LT(0, window_size_sub_blocks);
RTC_DCHECK((kBlockSize % sub_block_size) == 0);
RTC_DCHECK((sub_block_size % 4) == 0);
}
MatchedFilter::~MatchedFilter() = default;
void MatchedFilter::Reset() {
for (auto& f : filters_) {
std::fill(f.begin(), f.end(), 0.f);
}
for (auto& l : lag_estimates_) {
l = MatchedFilter::LagEstimate();
}
}
void MatchedFilter::Update(const DownsampledRenderBuffer& render_buffer,
rtc::ArrayView<const float> capture) {
RTC_DCHECK_EQ(sub_block_size_, capture.size());
auto& y = capture;
const float x2_sum_threshold =
filters_[0].size() * excitation_limit_ * excitation_limit_;
// Apply all matched filters.
size_t alignment_shift = 0;
for (size_t n = 0; n < filters_.size(); ++n) {
float error_sum = 0.f;
bool filters_updated = false;
size_t x_start_index =
(render_buffer.read + alignment_shift + sub_block_size_ - 1) %
render_buffer.buffer.size();
switch (optimization_) {
#if defined(WEBRTC_ARCH_X86_FAMILY)
case Aec3Optimization::kSse2:
aec3::MatchedFilterCore_SSE2(x_start_index, x2_sum_threshold,
smoothing_, render_buffer.buffer, y,
filters_[n], &filters_updated, &error_sum);
break;
case Aec3Optimization::kAvx2:
aec3::MatchedFilterCore_AVX2(x_start_index, x2_sum_threshold,
smoothing_, render_buffer.buffer, y,
filters_[n], &filters_updated, &error_sum);
break;
#endif
#if defined(WEBRTC_HAS_NEON)
case Aec3Optimization::kNeon:
aec3::MatchedFilterCore_NEON(x_start_index, x2_sum_threshold,
smoothing_, render_buffer.buffer, y,
filters_[n], &filters_updated, &error_sum);
break;
#endif
default:
aec3::MatchedFilterCore(x_start_index, x2_sum_threshold, smoothing_,
render_buffer.buffer, y, filters_[n],
&filters_updated, &error_sum);
}
// Compute anchor for the matched filter error.
const float error_sum_anchor =
std::inner_product(y.begin(), y.end(), y.begin(), 0.f);
// Estimate the lag in the matched filter as the distance to the portion in
// the filter that contributes the most to the matched filter output. This
// is detected as the peak of the matched filter.
const size_t lag_estimate = std::distance(
filters_[n].begin(),
std::max_element(
filters_[n].begin(), filters_[n].end(),
[](float a, float b) -> bool { return a * a < b * b; }));
// Update the lag estimates for the matched filter.
lag_estimates_[n] = LagEstimate(
error_sum_anchor - error_sum,
(lag_estimate > 2 && lag_estimate < (filters_[n].size() - 10) &&
error_sum < matching_filter_threshold_ * error_sum_anchor),
lag_estimate + alignment_shift, filters_updated);
RTC_DCHECK_GE(10, filters_.size());
switch (n) {
case 0:
data_dumper_->DumpRaw("aec3_correlator_0_h", filters_[0]);
break;
case 1:
data_dumper_->DumpRaw("aec3_correlator_1_h", filters_[1]);
break;
case 2:
data_dumper_->DumpRaw("aec3_correlator_2_h", filters_[2]);
break;
case 3:
data_dumper_->DumpRaw("aec3_correlator_3_h", filters_[3]);
break;
case 4:
data_dumper_->DumpRaw("aec3_correlator_4_h", filters_[4]);
break;
case 5:
data_dumper_->DumpRaw("aec3_correlator_5_h", filters_[5]);
break;
case 6:
data_dumper_->DumpRaw("aec3_correlator_6_h", filters_[6]);
break;
case 7:
data_dumper_->DumpRaw("aec3_correlator_7_h", filters_[7]);
break;
case 8:
data_dumper_->DumpRaw("aec3_correlator_8_h", filters_[8]);
break;
case 9:
data_dumper_->DumpRaw("aec3_correlator_9_h", filters_[9]);
break;
default:
RTC_NOTREACHED();
}
alignment_shift += filter_intra_lag_shift_;
}
}
void MatchedFilter::LogFilterProperties(int sample_rate_hz,
size_t shift,
size_t downsampling_factor) const {
size_t alignment_shift = 0;
constexpr int kFsBy1000 = 16;
for (size_t k = 0; k < filters_.size(); ++k) {
int start = static_cast<int>(alignment_shift * downsampling_factor);
int end = static_cast<int>((alignment_shift + filters_[k].size()) *
downsampling_factor);
RTC_LOG(LS_VERBOSE) << "Filter " << k << ": start: "
<< (start - static_cast<int>(shift)) / kFsBy1000
<< " ms, end: "
<< (end - static_cast<int>(shift)) / kFsBy1000
<< " ms.";
alignment_shift += filter_intra_lag_shift_;
}
}
} // namespace webrtc

149
webrtc/modules/audio_processing/aec3/matched_filter.h Normal file
View File

@ -0,0 +1,149 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_MATCHED_FILTER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_MATCHED_FILTER_H_
#include <stddef.h>
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/system/arch.h"
namespace webrtc {
class ApmDataDumper;
struct DownsampledRenderBuffer;
namespace aec3 {
#if defined(WEBRTC_HAS_NEON)
// Filter core for the matched filter that is optimized for NEON.
void MatchedFilterCore_NEON(size_t x_start_index,
float x2_sum_threshold,
float smoothing,
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float> h,
bool* filters_updated,
float* error_sum);
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
// Filter core for the matched filter that is optimized for SSE2.
void MatchedFilterCore_SSE2(size_t x_start_index,
float x2_sum_threshold,
float smoothing,
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float> h,
bool* filters_updated,
float* error_sum);
// Filter core for the matched filter that is optimized for AVX2.
void MatchedFilterCore_AVX2(size_t x_start_index,
float x2_sum_threshold,
float smoothing,
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float> h,
bool* filters_updated,
float* error_sum);
#endif
// Filter core for the matched filter.
void MatchedFilterCore(size_t x_start_index,
float x2_sum_threshold,
float smoothing,
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float> h,
bool* filters_updated,
float* error_sum);
} // namespace aec3
// Produces recursively updated cross-correlation estimates for several signal
// shifts where the intra-shift spacing is uniform.
class MatchedFilter {
public:
// Stores properties for the lag estimate corresponding to a particular signal
// shift.
struct LagEstimate {
LagEstimate() = default;
LagEstimate(float accuracy, bool reliable, size_t lag, bool updated)
: accuracy(accuracy), reliable(reliable), lag(lag), updated(updated) {}
float accuracy = 0.f;
bool reliable = false;
size_t lag = 0;
bool updated = false;
};
MatchedFilter(ApmDataDumper* data_dumper,
Aec3Optimization optimization,
size_t sub_block_size,
size_t window_size_sub_blocks,
int num_matched_filters,
size_t alignment_shift_sub_blocks,
float excitation_limit,
float smoothing,
float matching_filter_threshold);
MatchedFilter() = delete;
MatchedFilter(const MatchedFilter&) = delete;
MatchedFilter& operator=(const MatchedFilter&) = delete;
~MatchedFilter();
// Updates the correlation with the values in the capture buffer.
void Update(const DownsampledRenderBuffer& render_buffer,
rtc::ArrayView<const float> capture);
// Resets the matched filter.
void Reset();
// Returns the current lag estimates.
rtc::ArrayView<const MatchedFilter::LagEstimate> GetLagEstimates() const {
return lag_estimates_;
}
// Returns the maximum filter lag.
size_t GetMaxFilterLag() const {
return filters_.size() * filter_intra_lag_shift_ + filters_[0].size();
}
// Log matched filter properties.
void LogFilterProperties(int sample_rate_hz,
size_t shift,
size_t downsampling_factor) const;
private:
ApmDataDumper* const data_dumper_;
const Aec3Optimization optimization_;
const size_t sub_block_size_;
const size_t filter_intra_lag_shift_;
std::vector<std::vector<float>> filters_;
std::vector<LagEstimate> lag_estimates_;
std::vector<size_t> filters_offsets_;
const float excitation_limit_;
const float smoothing_;
const float matching_filter_threshold_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_MATCHED_FILTER_H_

132
webrtc/modules/audio_processing/aec3/matched_filter_avx2.cc Normal file
View File

@ -0,0 +1,132 @@
/*
* Copyright (c) 2020 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/matched_filter.h"
#include <immintrin.h>
#include "rtc_base/checks.h"
namespace webrtc {
namespace aec3 {
void MatchedFilterCore_AVX2(size_t x_start_index,
float x2_sum_threshold,
float smoothing,
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float> h,
bool* filters_updated,
float* error_sum) {
const int h_size = static_cast<int>(h.size());
const int x_size = static_cast<int>(x.size());
RTC_DCHECK_EQ(0, h_size % 8);
// Process for all samples in the sub-block.
for (size_t i = 0; i < y.size(); ++i) {
// Apply the matched filter as filter * x, and compute x * x.
RTC_DCHECK_GT(x_size, x_start_index);
const float* x_p = &x[x_start_index];
const float* h_p = &h[0];
// Initialize values for the accumulation.
__m256 s_256 = _mm256_set1_ps(0);
__m256 x2_sum_256 = _mm256_set1_ps(0);
float x2_sum = 0.f;
float s = 0;
// Compute loop chunk sizes until, and after, the wraparound of the circular
// buffer for x.
const int chunk1 =
std::min(h_size, static_cast<int>(x_size - x_start_index));
// Perform the loop in two chunks.
const int chunk2 = h_size - chunk1;
for (int limit : {chunk1, chunk2}) {
// Perform 256 bit vector operations.
const int limit_by_8 = limit >> 3;
for (int k = limit_by_8; k > 0; --k, h_p += 8, x_p += 8) {
// Load the data into 256 bit vectors.
__m256 x_k = _mm256_loadu_ps(x_p);
__m256 h_k = _mm256_loadu_ps(h_p);
// Compute and accumulate x * x and h * x.
x2_sum_256 = _mm256_fmadd_ps(x_k, x_k, x2_sum_256);
s_256 = _mm256_fmadd_ps(h_k, x_k, s_256);
}
// Perform non-vector operations for any remaining items.
for (int k = limit - limit_by_8 * 8; k > 0; --k, ++h_p, ++x_p) {
const float x_k = *x_p;
x2_sum += x_k * x_k;
s += *h_p * x_k;
}
x_p = &x[0];
}
// Sum components together.
__m128 x2_sum_128 = _mm_add_ps(_mm256_extractf128_ps(x2_sum_256, 0),
_mm256_extractf128_ps(x2_sum_256, 1));
__m128 s_128 = _mm_add_ps(_mm256_extractf128_ps(s_256, 0),
_mm256_extractf128_ps(s_256, 1));
// Combine the accumulated vector and scalar values.
float* v = reinterpret_cast<float*>(&x2_sum_128);
x2_sum += v[0] + v[1] + v[2] + v[3];
v = reinterpret_cast<float*>(&s_128);
s += v[0] + v[1] + v[2] + v[3];
// Compute the matched filter error.
float e = y[i] - s;
const bool saturation = y[i] >= 32000.f || y[i] <= -32000.f;
(*error_sum) += e * e;
// Update the matched filter estimate in an NLMS manner.
if (x2_sum > x2_sum_threshold && !saturation) {
RTC_DCHECK_LT(0.f, x2_sum);
const float alpha = smoothing * e / x2_sum;
const __m256 alpha_256 = _mm256_set1_ps(alpha);
// filter = filter + smoothing * (y - filter * x) * x / x * x.
float* h_p = &h[0];
x_p = &x[x_start_index];
// Perform the loop in two chunks.
for (int limit : {chunk1, chunk2}) {
// Perform 256 bit vector operations.
const int limit_by_8 = limit >> 3;
for (int k = limit_by_8; k > 0; --k, h_p += 8, x_p += 8) {
// Load the data into 256 bit vectors.
__m256 h_k = _mm256_loadu_ps(h_p);
__m256 x_k = _mm256_loadu_ps(x_p);
// Compute h = h + alpha * x.
h_k = _mm256_fmadd_ps(x_k, alpha_256, h_k);
// Store the result.
_mm256_storeu_ps(h_p, h_k);
}
// Perform non-vector operations for any remaining items.
for (int k = limit - limit_by_8 * 8; k > 0; --k, ++h_p, ++x_p) {
*h_p += alpha * *x_p;
}
x_p = &x[0];
}
*filters_updated = true;
}
x_start_index = x_start_index > 0 ? x_start_index - 1 : x_size - 1;
}
}
} // namespace aec3
} // namespace webrtc

97
webrtc/modules/audio_processing/aec3/matched_filter_lag_aggregator.cc Normal file
View File

@ -0,0 +1,97 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/matched_filter_lag_aggregator.h"
#include <algorithm>
#include <iterator>
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/checks.h"
namespace webrtc {
MatchedFilterLagAggregator::MatchedFilterLagAggregator(
ApmDataDumper* data_dumper,
size_t max_filter_lag,
const EchoCanceller3Config::Delay::DelaySelectionThresholds& thresholds)
: data_dumper_(data_dumper),
histogram_(max_filter_lag + 1, 0),
thresholds_(thresholds) {
RTC_DCHECK(data_dumper);
RTC_DCHECK_LE(thresholds_.initial, thresholds_.converged);
histogram_data_.fill(0);
}
MatchedFilterLagAggregator::~MatchedFilterLagAggregator() = default;
void MatchedFilterLagAggregator::Reset(bool hard_reset) {
std::fill(histogram_.begin(), histogram_.end(), 0);
histogram_data_.fill(0);
histogram_data_index_ = 0;
if (hard_reset) {
significant_candidate_found_ = false;
}
}
absl::optional<DelayEstimate> MatchedFilterLagAggregator::Aggregate(
rtc::ArrayView<const MatchedFilter::LagEstimate> lag_estimates) {
// Choose the strongest lag estimate as the best one.
float best_accuracy = 0.f;
int best_lag_estimate_index = -1;
for (size_t k = 0; k < lag_estimates.size(); ++k) {
if (lag_estimates[k].updated && lag_estimates[k].reliable) {
if (lag_estimates[k].accuracy > best_accuracy) {
best_accuracy = lag_estimates[k].accuracy;
best_lag_estimate_index = static_cast<int>(k);
}
}
}
// TODO(peah): Remove this logging once all development is done.
data_dumper_->DumpRaw("aec3_echo_path_delay_estimator_best_index",
best_lag_estimate_index);
data_dumper_->DumpRaw("aec3_echo_path_delay_estimator_histogram", histogram_);
if (best_lag_estimate_index != -1) {
RTC_DCHECK_GT(histogram_.size(), histogram_data_[histogram_data_index_]);
RTC_DCHECK_LE(0, histogram_data_[histogram_data_index_]);
--histogram_[histogram_data_[histogram_data_index_]];
histogram_data_[histogram_data_index_] =
lag_estimates[best_lag_estimate_index].lag;
RTC_DCHECK_GT(histogram_.size(), histogram_data_[histogram_data_index_]);
RTC_DCHECK_LE(0, histogram_data_[histogram_data_index_]);
++histogram_[histogram_data_[histogram_data_index_]];
histogram_data_index_ =
(histogram_data_index_ + 1) % histogram_data_.size();
const int candidate =
std::distance(histogram_.begin(),
std::max_element(histogram_.begin(), histogram_.end()));
significant_candidate_found_ =
significant_candidate_found_ ||
histogram_[candidate] > thresholds_.converged;
if (histogram_[candidate] > thresholds_.converged ||
(histogram_[candidate] > thresholds_.initial &&
!significant_candidate_found_)) {
DelayEstimate::Quality quality = significant_candidate_found_
? DelayEstimate::Quality::kRefined
: DelayEstimate::Quality::kCoarse;
return DelayEstimate(quality, candidate);
}
}
return absl::nullopt;
}
} // namespace webrtc

58
webrtc/modules/audio_processing/aec3/matched_filter_lag_aggregator.h Normal file
View File

@ -0,0 +1,58 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_MATCHED_FILTER_LAG_AGGREGATOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_MATCHED_FILTER_LAG_AGGREGATOR_H_
#include <vector>
#include "absl/types/optional.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/delay_estimate.h"
#include "modules/audio_processing/aec3/matched_filter.h"
namespace webrtc {
class ApmDataDumper;
// Aggregates lag estimates produced by the MatchedFilter class into a single
// reliable combined lag estimate.
class MatchedFilterLagAggregator {
public:
MatchedFilterLagAggregator(
ApmDataDumper* data_dumper,
size_t max_filter_lag,
const EchoCanceller3Config::Delay::DelaySelectionThresholds& thresholds);
MatchedFilterLagAggregator() = delete;
MatchedFilterLagAggregator(const MatchedFilterLagAggregator&) = delete;
MatchedFilterLagAggregator& operator=(const MatchedFilterLagAggregator&) =
delete;
~MatchedFilterLagAggregator();
// Resets the aggregator.
void Reset(bool hard_reset);
// Aggregates the provided lag estimates.
absl::optional<DelayEstimate> Aggregate(
rtc::ArrayView<const MatchedFilter::LagEstimate> lag_estimates);
private:
ApmDataDumper* const data_dumper_;
std::vector<int> histogram_;
std::array<int, 250> histogram_data_;
int histogram_data_index_ = 0;
bool significant_candidate_found_ = false;
const EchoCanceller3Config::Delay::DelaySelectionThresholds thresholds_;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_MATCHED_FILTER_LAG_AGGREGATOR_H_

60
webrtc/modules/audio_processing/aec3/moving_average.cc Normal file
View File

@ -0,0 +1,60 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/moving_average.h"
#include <algorithm>
#include <functional>
#include "rtc_base/checks.h"
namespace webrtc {
namespace aec3 {
MovingAverage::MovingAverage(size_t num_elem, size_t mem_len)
: num_elem_(num_elem),
mem_len_(mem_len - 1),
scaling_(1.0f / static_cast<float>(mem_len)),
memory_(num_elem * mem_len_, 0.f),
mem_index_(0) {
RTC_DCHECK(num_elem_ > 0);
RTC_DCHECK(mem_len > 0);
}
MovingAverage::~MovingAverage() = default;
void MovingAverage::Average(rtc::ArrayView<const float> input,
rtc::ArrayView<float> output) {
RTC_DCHECK(input.size() == num_elem_);
RTC_DCHECK(output.size() == num_elem_);
// Sum all contributions.
std::copy(input.begin(), input.end(), output.begin());
for (auto i = memory_.begin(); i < memory_.end(); i += num_elem_) {
std::transform(i, i + num_elem_, output.begin(), output.begin(),
std::plus<float>());
}
// Divide by mem_len_.
for (float& o : output) {
o *= scaling_;
}
// Update memory.
if (mem_len_ > 0) {
std::copy(input.begin(), input.end(),
memory_.begin() + mem_index_ * num_elem_);
mem_index_ = (mem_index_ + 1) % mem_len_;
}
}
} // namespace aec3
} // namespace webrtc

45
webrtc/modules/audio_processing/aec3/moving_average.h Normal file
View File

@ -0,0 +1,45 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_MOVING_AVERAGE_H_
#define MODULES_AUDIO_PROCESSING_AEC3_MOVING_AVERAGE_H_
#include <stddef.h>
#include <vector>
#include "api/array_view.h"
namespace webrtc {
namespace aec3 {
class MovingAverage {
public:
// Creates an instance of MovingAverage that accepts inputs of length num_elem
// and averages over mem_len inputs.
MovingAverage(size_t num_elem, size_t mem_len);
~MovingAverage();
// Computes the average of input and mem_len-1 previous inputs and stores the
// result in output.
void Average(rtc::ArrayView<const float> input, rtc::ArrayView<float> output);
private:
const size_t num_elem_;
const size_t mem_len_;
const float scaling_;
std::vector<float> memory_;
size_t mem_index_;
};
} // namespace aec3
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_MOVING_AVERAGE_H_

42
webrtc/modules/audio_processing/aec3/nearend_detector.h Normal file
View File

@ -0,0 +1,42 @@
/*
* Copyright (c) 2019 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_NEAREND_DETECTOR_H_
#define MODULES_AUDIO_PROCESSING_AEC3_NEAREND_DETECTOR_H_
#include <vector>
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
namespace webrtc {
// Class for selecting whether the suppressor is in the nearend or echo state.
class NearendDetector {
public:
virtual ~NearendDetector() {}
// Returns whether the current state is the nearend state.
virtual bool IsNearendState() const = 0;
// Updates the state selection based on latest spectral estimates.
virtual void Update(
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
nearend_spectrum,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
residual_echo_spectrum,
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>>
comfort_noise_spectrum,
bool initial_state) = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_NEAREND_DETECTOR_H_

174
webrtc/modules/audio_processing/aec3/refined_filter_update_gain.cc Normal file
View File

@ -0,0 +1,174 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/refined_filter_update_gain.h"
#include <algorithm>
#include <functional>
#include "modules/audio_processing/aec3/adaptive_fir_filter.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/echo_path_variability.h"
#include "modules/audio_processing/aec3/fft_data.h"
#include "modules/audio_processing/aec3/render_signal_analyzer.h"
#include "modules/audio_processing/aec3/subtractor_output.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomic_ops.h"
#include "rtc_base/checks.h"
namespace webrtc {
namespace {
constexpr float kHErrorInitial = 10000.f;
constexpr int kPoorExcitationCounterInitial = 1000;
} // namespace
int RefinedFilterUpdateGain::instance_count_ = 0;
RefinedFilterUpdateGain::RefinedFilterUpdateGain(
const EchoCanceller3Config::Filter::RefinedConfiguration& config,
size_t config_change_duration_blocks)
: data_dumper_(
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
config_change_duration_blocks_(
static_cast<int>(config_change_duration_blocks)),
poor_excitation_counter_(kPoorExcitationCounterInitial) {
SetConfig(config, true);
H_error_.fill(kHErrorInitial);
RTC_DCHECK_LT(0, config_change_duration_blocks_);
one_by_config_change_duration_blocks_ = 1.f / config_change_duration_blocks_;
}
RefinedFilterUpdateGain::~RefinedFilterUpdateGain() {}
void RefinedFilterUpdateGain::HandleEchoPathChange(
const EchoPathVariability& echo_path_variability) {
if (echo_path_variability.gain_change) {
// TODO(bugs.webrtc.org/9526) Handle gain changes.
}
if (echo_path_variability.delay_change !=
EchoPathVariability::DelayAdjustment::kNone) {
H_error_.fill(kHErrorInitial);
}
if (!echo_path_variability.gain_change) {
poor_excitation_counter_ = kPoorExcitationCounterInitial;
call_counter_ = 0;
}
}
void RefinedFilterUpdateGain::Compute(
const std::array<float, kFftLengthBy2Plus1>& render_power,
const RenderSignalAnalyzer& render_signal_analyzer,
const SubtractorOutput& subtractor_output,
rtc::ArrayView<const float> erl,
size_t size_partitions,
bool saturated_capture_signal,
FftData* gain_fft) {
RTC_DCHECK(gain_fft);
// Introducing shorter notation to improve readability.
const FftData& E_refined = subtractor_output.E_refined;
const auto& E2_refined = subtractor_output.E2_refined;
const auto& E2_coarse = subtractor_output.E2_coarse;
FftData* G = gain_fft;
const auto& X2 = render_power;
++call_counter_;
UpdateCurrentConfig();
if (render_signal_analyzer.PoorSignalExcitation()) {
poor_excitation_counter_ = 0;
}
// Do not update the filter if the render is not sufficiently excited.
if (++poor_excitation_counter_ < size_partitions ||
saturated_capture_signal || call_counter_ <= size_partitions) {
G->re.fill(0.f);
G->im.fill(0.f);
} else {
// Corresponds to WGN of power -39 dBFS.
std::array<float, kFftLengthBy2Plus1> mu;
// mu = H_error / (0.5* H_error* X2 + n * E2).
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
if (X2[k] >= current_config_.noise_gate) {
mu[k] = H_error_[k] /
(0.5f * H_error_[k] * X2[k] + size_partitions * E2_refined[k]);
} else {
mu[k] = 0.f;
}
}
// Avoid updating the filter close to narrow bands in the render signals.
render_signal_analyzer.MaskRegionsAroundNarrowBands(&mu);
// H_error = H_error - 0.5 * mu * X2 * H_error.
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
H_error_[k] -= 0.5f * mu[k] * X2[k] * H_error_[k];
}
// G = mu * E.
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
G->re[k] = mu[k] * E_refined.re[k];
G->im[k] = mu[k] * E_refined.im[k];
}
}
// H_error = H_error + factor * erl.
for (size_t k = 0; k < kFftLengthBy2Plus1; ++k) {
if (E2_coarse[k] >= E2_refined[k]) {
H_error_[k] += current_config_.leakage_converged * erl[k];
} else {
H_error_[k] += current_config_.leakage_diverged * erl[k];
}
H_error_[k] = std::max(H_error_[k], current_config_.error_floor);
H_error_[k] = std::min(H_error_[k], current_config_.error_ceil);
}
data_dumper_->DumpRaw("aec3_refined_gain_H_error", H_error_);
}
void RefinedFilterUpdateGain::UpdateCurrentConfig() {
RTC_DCHECK_GE(config_change_duration_blocks_, config_change_counter_);
if (config_change_counter_ > 0) {
if (--config_change_counter_ > 0) {
auto average = [](float from, float to, float from_weight) {
return from * from_weight + to * (1.f - from_weight);
};
float change_factor =
config_change_counter_ * one_by_config_change_duration_blocks_;
current_config_.leakage_converged =
average(old_target_config_.leakage_converged,
target_config_.leakage_converged, change_factor);
current_config_.leakage_diverged =
average(old_target_config_.leakage_diverged,
target_config_.leakage_diverged, change_factor);
current_config_.error_floor =
average(old_target_config_.error_floor, target_config_.error_floor,
change_factor);
current_config_.error_ceil =
average(old_target_config_.error_ceil, target_config_.error_ceil,
change_factor);
current_config_.noise_gate =
average(old_target_config_.noise_gate, target_config_.noise_gate,
change_factor);
} else {
current_config_ = old_target_config_ = target_config_;
}
}
RTC_DCHECK_LE(0, config_change_counter_);
}
} // namespace webrtc

89
webrtc/modules/audio_processing/aec3/refined_filter_update_gain.h Normal file
View File

@ -0,0 +1,89 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_REFINED_FILTER_UPDATE_GAIN_H_
#define MODULES_AUDIO_PROCESSING_AEC3_REFINED_FILTER_UPDATE_GAIN_H_
#include <stddef.h>
#include <array>
#include <memory>
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
namespace webrtc {
class AdaptiveFirFilter;
class ApmDataDumper;
struct EchoPathVariability;
struct FftData;
class RenderSignalAnalyzer;
struct SubtractorOutput;
// Provides functionality for computing the adaptive gain for the refined
// filter.
class RefinedFilterUpdateGain {
public:
RefinedFilterUpdateGain(
const EchoCanceller3Config::Filter::RefinedConfiguration& config,
size_t config_change_duration_blocks);
~RefinedFilterUpdateGain();
RefinedFilterUpdateGain(const RefinedFilterUpdateGain&) = delete;
RefinedFilterUpdateGain& operator=(const RefinedFilterUpdateGain&) = delete;
// Takes action in the case of a known echo path change.
void HandleEchoPathChange(const EchoPathVariability& echo_path_variability);
// Computes the gain.
void Compute(const std::array<float, kFftLengthBy2Plus1>& render_power,
const RenderSignalAnalyzer& render_signal_analyzer,
const SubtractorOutput& subtractor_output,
rtc::ArrayView<const float> erl,
size_t size_partitions,
bool saturated_capture_signal,
FftData* gain_fft);
// Sets a new config.
void SetConfig(
const EchoCanceller3Config::Filter::RefinedConfiguration& config,
bool immediate_effect) {
if (immediate_effect) {
old_target_config_ = current_config_ = target_config_ = config;
config_change_counter_ = 0;
} else {
old_target_config_ = current_config_;
target_config_ = config;
config_change_counter_ = config_change_duration_blocks_;
}
}
private:
static int instance_count_;
std::unique_ptr<ApmDataDumper> data_dumper_;
const int config_change_duration_blocks_;
float one_by_config_change_duration_blocks_;
EchoCanceller3Config::Filter::RefinedConfiguration current_config_;
EchoCanceller3Config::Filter::RefinedConfiguration target_config_;
EchoCanceller3Config::Filter::RefinedConfiguration old_target_config_;
std::array<float, kFftLengthBy2Plus1> H_error_;
size_t poor_excitation_counter_;
size_t call_counter_ = 0;
int config_change_counter_ = 0;
// Updates the current config towards the target config.
void UpdateCurrentConfig();
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_REFINED_FILTER_UPDATE_GAIN_H_

80
webrtc/modules/audio_processing/aec3/render_buffer.cc Normal file
View File

@ -0,0 +1,80 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/render_buffer.h"
#include <algorithm>
#include <functional>
#include "modules/audio_processing/aec3/aec3_common.h"
#include "rtc_base/checks.h"
namespace webrtc {
RenderBuffer::RenderBuffer(BlockBuffer* block_buffer,
SpectrumBuffer* spectrum_buffer,
FftBuffer* fft_buffer)
: block_buffer_(block_buffer),
spectrum_buffer_(spectrum_buffer),
fft_buffer_(fft_buffer) {
RTC_DCHECK(block_buffer_);
RTC_DCHECK(spectrum_buffer_);
RTC_DCHECK(fft_buffer_);
RTC_DCHECK_EQ(block_buffer_->buffer.size(), fft_buffer_->buffer.size());
RTC_DCHECK_EQ(spectrum_buffer_->buffer.size(), fft_buffer_->buffer.size());
RTC_DCHECK_EQ(spectrum_buffer_->read, fft_buffer_->read);
RTC_DCHECK_EQ(spectrum_buffer_->write, fft_buffer_->write);
}
RenderBuffer::~RenderBuffer() = default;
void RenderBuffer::SpectralSum(
size_t num_spectra,
std::array<float, kFftLengthBy2Plus1>* X2) const {
X2->fill(0.f);
int position = spectrum_buffer_->read;
for (size_t j = 0; j < num_spectra; ++j) {
for (const auto& channel_spectrum : spectrum_buffer_->buffer[position]) {
std::transform(X2->begin(), X2->end(), channel_spectrum.begin(),
X2->begin(), std::plus<float>());
}
position = spectrum_buffer_->IncIndex(position);
}
}
void RenderBuffer::SpectralSums(
size_t num_spectra_shorter,
size_t num_spectra_longer,
std::array<float, kFftLengthBy2Plus1>* X2_shorter,
std::array<float, kFftLengthBy2Plus1>* X2_longer) const {
RTC_DCHECK_LE(num_spectra_shorter, num_spectra_longer);
X2_shorter->fill(0.f);
int position = spectrum_buffer_->read;
size_t j = 0;
for (; j < num_spectra_shorter; ++j) {
for (const auto& channel_spectrum : spectrum_buffer_->buffer[position]) {
std::transform(X2_shorter->begin(), X2_shorter->end(),
channel_spectrum.begin(), X2_shorter->begin(),
std::plus<float>());
}
position = spectrum_buffer_->IncIndex(position);
}
std::copy(X2_shorter->begin(), X2_shorter->end(), X2_longer->begin());
for (; j < num_spectra_longer; ++j) {
for (const auto& channel_spectrum : spectrum_buffer_->buffer[position]) {
std::transform(X2_longer->begin(), X2_longer->end(),
channel_spectrum.begin(), X2_longer->begin(),
std::plus<float>());
}
position = spectrum_buffer_->IncIndex(position);
}
}
} // namespace webrtc

116
webrtc/modules/audio_processing/aec3/render_buffer.h Normal file
View File

@ -0,0 +1,116 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_RENDER_BUFFER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_RENDER_BUFFER_H_
#include <stddef.h>
#include <array>
#include <vector>
#include "api/array_view.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/block_buffer.h"
#include "modules/audio_processing/aec3/fft_buffer.h"
#include "modules/audio_processing/aec3/fft_data.h"
#include "modules/audio_processing/aec3/spectrum_buffer.h"
#include "rtc_base/checks.h"
namespace webrtc {
// Provides a buffer of the render data for the echo remover.
class RenderBuffer {
public:
RenderBuffer(BlockBuffer* block_buffer,
SpectrumBuffer* spectrum_buffer,
FftBuffer* fft_buffer);
RenderBuffer() = delete;
RenderBuffer(const RenderBuffer&) = delete;
RenderBuffer& operator=(const RenderBuffer&) = delete;
~RenderBuffer();
// Get a block.
const std::vector<std::vector<std::vector<float>>>& Block(
int buffer_offset_blocks) const {
int position =
block_buffer_->OffsetIndex(block_buffer_->read, buffer_offset_blocks);
return block_buffer_->buffer[position];
}
// Get the spectrum from one of the FFTs in the buffer.
rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> Spectrum(
int buffer_offset_ffts) const {
int position = spectrum_buffer_->OffsetIndex(spectrum_buffer_->read,
buffer_offset_ffts);
return spectrum_buffer_->buffer[position];
}
// Returns the circular fft buffer.
rtc::ArrayView<const std::vector<FftData>> GetFftBuffer() const {
return fft_buffer_->buffer;
}
// Returns the current position in the circular buffer.
size_t Position() const {
RTC_DCHECK_EQ(spectrum_buffer_->read, fft_buffer_->read);
RTC_DCHECK_EQ(spectrum_buffer_->write, fft_buffer_->write);
return fft_buffer_->read;
}
// Returns the sum of the spectrums for a certain number of FFTs.
void SpectralSum(size_t num_spectra,
std::array<float, kFftLengthBy2Plus1>* X2) const;
// Returns the sums of the spectrums for two numbers of FFTs.
void SpectralSums(size_t num_spectra_shorter,
size_t num_spectra_longer,
std::array<float, kFftLengthBy2Plus1>* X2_shorter,
std::array<float, kFftLengthBy2Plus1>* X2_longer) const;
// Gets the recent activity seen in the render signal.
bool GetRenderActivity() const { return render_activity_; }
// Specifies the recent activity seen in the render signal.
void SetRenderActivity(bool activity) { render_activity_ = activity; }
// Returns the headroom between the write and the read positions in the
// buffer.
int Headroom() const {
// The write and read indices are decreased over time.
int headroom =
fft_buffer_->write < fft_buffer_->read
? fft_buffer_->read - fft_buffer_->write
: fft_buffer_->size - fft_buffer_->write + fft_buffer_->read;
RTC_DCHECK_LE(0, headroom);
RTC_DCHECK_GE(fft_buffer_->size, headroom);
return headroom;
}
// Returns a reference to the spectrum buffer.
const SpectrumBuffer& GetSpectrumBuffer() const { return *spectrum_buffer_; }
// Returns a reference to the block buffer.
const BlockBuffer& GetBlockBuffer() const { return *block_buffer_; }
private:
const BlockBuffer* const block_buffer_;
const SpectrumBuffer* const spectrum_buffer_;
const FftBuffer* const fft_buffer_;
bool render_activity_ = false;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_RENDER_BUFFER_H_

523
webrtc/modules/audio_processing/aec3/render_delay_buffer.cc Normal file
View File

@ -0,0 +1,523 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/render_delay_buffer.h"
#include <string.h>
#include <algorithm>
#include <cmath>
#include <memory>
#include <numeric>
#include <vector>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/aec3_fft.h"
#include "modules/audio_processing/aec3/alignment_mixer.h"
#include "modules/audio_processing/aec3/block_buffer.h"
#include "modules/audio_processing/aec3/decimator.h"
#include "modules/audio_processing/aec3/downsampled_render_buffer.h"
#include "modules/audio_processing/aec3/fft_buffer.h"
#include "modules/audio_processing/aec3/fft_data.h"
#include "modules/audio_processing/aec3/render_buffer.h"
#include "modules/audio_processing/aec3/spectrum_buffer.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomic_ops.h"
#include "rtc_base/checks.h"
#include "rtc_base/logging.h"
#include "system_wrappers/include/field_trial.h"
namespace webrtc {
namespace {
bool UpdateCaptureCallCounterOnSkippedBlocks() {
return !field_trial::IsEnabled(
"WebRTC-Aec3RenderBufferCallCounterUpdateKillSwitch");
}
class RenderDelayBufferImpl final : public RenderDelayBuffer {
public:
RenderDelayBufferImpl(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels);
RenderDelayBufferImpl() = delete;
~RenderDelayBufferImpl() override;
void Reset() override;
BufferingEvent Insert(
const std::vector<std::vector<std::vector<float>>>& block) override;
BufferingEvent PrepareCaptureProcessing() override;
void HandleSkippedCaptureProcessing() override;
bool AlignFromDelay(size_t delay) override;
void AlignFromExternalDelay() override;
size_t Delay() const override { return ComputeDelay(); }
size_t MaxDelay() const override {
return blocks_.buffer.size() - 1 - buffer_headroom_;
}
RenderBuffer* GetRenderBuffer() override { return &echo_remover_buffer_; }
const DownsampledRenderBuffer& GetDownsampledRenderBuffer() const override {
return low_rate_;
}
int BufferLatency() const;
void SetAudioBufferDelay(int delay_ms) override;
bool HasReceivedBufferDelay() override;
private:
static int instance_count_;
std::unique_ptr<ApmDataDumper> data_dumper_;
const Aec3Optimization optimization_;
const EchoCanceller3Config config_;
const bool update_capture_call_counter_on_skipped_blocks_;
const float render_linear_amplitude_gain_;
const rtc::LoggingSeverity delay_log_level_;
size_t down_sampling_factor_;
const int sub_block_size_;
BlockBuffer blocks_;
SpectrumBuffer spectra_;
FftBuffer ffts_;
absl::optional<size_t> delay_;
RenderBuffer echo_remover_buffer_;
DownsampledRenderBuffer low_rate_;
AlignmentMixer render_mixer_;
Decimator render_decimator_;
const Aec3Fft fft_;
std::vector<float> render_ds_;
const int buffer_headroom_;
bool last_call_was_render_ = false;
int num_api_calls_in_a_row_ = 0;
int max_observed_jitter_ = 1;
int64_t capture_call_counter_ = 0;
int64_t render_call_counter_ = 0;
bool render_activity_ = false;
size_t render_activity_counter_ = 0;
absl::optional<int> external_audio_buffer_delay_;
bool external_audio_buffer_delay_verified_after_reset_ = false;
size_t min_latency_blocks_ = 0;
size_t excess_render_detection_counter_ = 0;
int MapDelayToTotalDelay(size_t delay) const;
int ComputeDelay() const;
void ApplyTotalDelay(int delay);
void InsertBlock(const std::vector<std::vector<std::vector<float>>>& block,
int previous_write);
bool DetectActiveRender(rtc::ArrayView<const float> x) const;
bool DetectExcessRenderBlocks();
void IncrementWriteIndices();
void IncrementLowRateReadIndices();
void IncrementReadIndices();
bool RenderOverrun();
bool RenderUnderrun();
};
int RenderDelayBufferImpl::instance_count_ = 0;
RenderDelayBufferImpl::RenderDelayBufferImpl(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels)
: data_dumper_(
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
optimization_(DetectOptimization()),
config_(config),
update_capture_call_counter_on_skipped_blocks_(
UpdateCaptureCallCounterOnSkippedBlocks()),
render_linear_amplitude_gain_(
std::pow(10.0f, config_.render_levels.render_power_gain_db / 20.f)),
delay_log_level_(config_.delay.log_warning_on_delay_changes
? rtc::LS_WARNING
: rtc::LS_VERBOSE),
down_sampling_factor_(config.delay.down_sampling_factor),
sub_block_size_(static_cast<int>(down_sampling_factor_ > 0
? kBlockSize / down_sampling_factor_
: kBlockSize)),
blocks_(GetRenderDelayBufferSize(down_sampling_factor_,
config.delay.num_filters,
config.filter.refined.length_blocks),
NumBandsForRate(sample_rate_hz),
num_render_channels,
kBlockSize),
spectra_(blocks_.buffer.size(), num_render_channels),
ffts_(blocks_.buffer.size(), num_render_channels),
delay_(config_.delay.default_delay),
echo_remover_buffer_(&blocks_, &spectra_, &ffts_),
low_rate_(GetDownSampledBufferSize(down_sampling_factor_,
config.delay.num_filters)),
render_mixer_(num_render_channels, config.delay.render_alignment_mixing),
render_decimator_(down_sampling_factor_),
fft_(),
render_ds_(sub_block_size_, 0.f),
buffer_headroom_(config.filter.refined.length_blocks) {
RTC_DCHECK_EQ(blocks_.buffer.size(), ffts_.buffer.size());
RTC_DCHECK_EQ(spectra_.buffer.size(), ffts_.buffer.size());
for (size_t i = 0; i < blocks_.buffer.size(); ++i) {
RTC_DCHECK_EQ(blocks_.buffer[i][0].size(), ffts_.buffer[i].size());
RTC_DCHECK_EQ(spectra_.buffer[i].size(), ffts_.buffer[i].size());
}
Reset();
}
RenderDelayBufferImpl::~RenderDelayBufferImpl() = default;
// Resets the buffer delays and clears the reported delays.
void RenderDelayBufferImpl::Reset() {
last_call_was_render_ = false;
num_api_calls_in_a_row_ = 1;
min_latency_blocks_ = 0;
excess_render_detection_counter_ = 0;
// Initialize the read index to one sub-block before the write index.
low_rate_.read = low_rate_.OffsetIndex(low_rate_.write, sub_block_size_);
// Check for any external audio buffer delay and whether it is feasible.
if (external_audio_buffer_delay_) {
const int headroom = 2;
size_t audio_buffer_delay_to_set;
// Minimum delay is 1 (like the low-rate render buffer).
if (*external_audio_buffer_delay_ <= headroom) {
audio_buffer_delay_to_set = 1;
} else {
audio_buffer_delay_to_set = *external_audio_buffer_delay_ - headroom;
}
audio_buffer_delay_to_set = std::min(audio_buffer_delay_to_set, MaxDelay());
// When an external delay estimate is available, use that delay as the
// initial render buffer delay.
ApplyTotalDelay(audio_buffer_delay_to_set);
delay_ = ComputeDelay();
external_audio_buffer_delay_verified_after_reset_ = false;
} else {
// If an external delay estimate is not available, use that delay as the
// initial delay. Set the render buffer delays to the default delay.
ApplyTotalDelay(config_.delay.default_delay);
// Unset the delays which are set by AlignFromDelay.
delay_ = absl::nullopt;
}
}
// Inserts a new block into the render buffers.
RenderDelayBuffer::BufferingEvent RenderDelayBufferImpl::Insert(
const std::vector<std::vector<std::vector<float>>>& block) {
++render_call_counter_;
if (delay_) {
if (!last_call_was_render_) {
last_call_was_render_ = true;
num_api_calls_in_a_row_ = 1;
} else {
if (++num_api_calls_in_a_row_ > max_observed_jitter_) {
max_observed_jitter_ = num_api_calls_in_a_row_;
RTC_LOG_V(delay_log_level_)
<< "New max number api jitter observed at render block "
<< render_call_counter_ << ": " << num_api_calls_in_a_row_
<< " blocks";
}
}
}
// Increase the write indices to where the new blocks should be written.
const int previous_write = blocks_.write;
IncrementWriteIndices();
// Allow overrun and do a reset when render overrun occurrs due to more render
// data being inserted than capture data is received.
BufferingEvent event =
RenderOverrun() ? BufferingEvent::kRenderOverrun : BufferingEvent::kNone;
// Detect and update render activity.
if (!render_activity_) {
render_activity_counter_ += DetectActiveRender(block[0][0]) ? 1 : 0;
render_activity_ = render_activity_counter_ >= 20;
}
// Insert the new render block into the specified position.
InsertBlock(block, previous_write);
if (event != BufferingEvent::kNone) {
Reset();
}
return event;
}
void RenderDelayBufferImpl::HandleSkippedCaptureProcessing() {
if (update_capture_call_counter_on_skipped_blocks_) {
++capture_call_counter_;
}
}
// Prepares the render buffers for processing another capture block.
RenderDelayBuffer::BufferingEvent
RenderDelayBufferImpl::PrepareCaptureProcessing() {
RenderDelayBuffer::BufferingEvent event = BufferingEvent::kNone;
++capture_call_counter_;
if (delay_) {
if (last_call_was_render_) {
last_call_was_render_ = false;
num_api_calls_in_a_row_ = 1;
} else {
if (++num_api_calls_in_a_row_ > max_observed_jitter_) {
max_observed_jitter_ = num_api_calls_in_a_row_;
RTC_LOG_V(delay_log_level_)
<< "New max number api jitter observed at capture block "
<< capture_call_counter_ << ": " << num_api_calls_in_a_row_
<< " blocks";
}
}
}
if (DetectExcessRenderBlocks()) {
// Too many render blocks compared to capture blocks. Risk of delay ending
// up before the filter used by the delay estimator.
RTC_LOG_V(delay_log_level_)
<< "Excess render blocks detected at block " << capture_call_counter_;
Reset();
event = BufferingEvent::kRenderOverrun;
} else if (RenderUnderrun()) {
// Don't increment the read indices of the low rate buffer if there is a
// render underrun.
RTC_LOG_V(delay_log_level_)
<< "Render buffer underrun detected at block " << capture_call_counter_;
IncrementReadIndices();
// Incrementing the buffer index without increasing the low rate buffer
// index means that the delay is reduced by one.
if (delay_ && *delay_ > 0)
delay_ = *delay_ - 1;
event = BufferingEvent::kRenderUnderrun;
} else {
// Increment the read indices in the render buffers to point to the most
// recent block to use in the capture processing.
IncrementLowRateReadIndices();
IncrementReadIndices();
}
echo_remover_buffer_.SetRenderActivity(render_activity_);
if (render_activity_) {
render_activity_counter_ = 0;
render_activity_ = false;
}
return event;
}
// Sets the delay and returns a bool indicating whether the delay was changed.
bool RenderDelayBufferImpl::AlignFromDelay(size_t delay) {
RTC_DCHECK(!config_.delay.use_external_delay_estimator);
if (!external_audio_buffer_delay_verified_after_reset_ &&
external_audio_buffer_delay_ && delay_) {
int difference = static_cast<int>(delay) - static_cast<int>(*delay_);
RTC_LOG_V(delay_log_level_)
<< "Mismatch between first estimated delay after reset "
"and externally reported audio buffer delay: "
<< difference << " blocks";
external_audio_buffer_delay_verified_after_reset_ = true;
}
if (delay_ && *delay_ == delay) {
return false;
}
delay_ = delay;
// Compute the total delay and limit the delay to the allowed range.
int total_delay = MapDelayToTotalDelay(*delay_);
total_delay =
std::min(MaxDelay(), static_cast<size_t>(std::max(total_delay, 0)));
// Apply the delay to the buffers.
ApplyTotalDelay(total_delay);
return true;
}
void RenderDelayBufferImpl::SetAudioBufferDelay(int delay_ms) {
if (!external_audio_buffer_delay_) {
RTC_LOG_V(delay_log_level_)
<< "Receiving a first externally reported audio buffer delay of "
<< delay_ms << " ms.";
}
// Convert delay from milliseconds to blocks (rounded down).
external_audio_buffer_delay_ = delay_ms / 4;
}
bool RenderDelayBufferImpl::HasReceivedBufferDelay() {
return external_audio_buffer_delay_.has_value();
}
// Maps the externally computed delay to the delay used internally.
int RenderDelayBufferImpl::MapDelayToTotalDelay(
size_t external_delay_blocks) const {
const int latency_blocks = BufferLatency();
return latency_blocks + static_cast<int>(external_delay_blocks);
}
// Returns the delay (not including call jitter).
int RenderDelayBufferImpl::ComputeDelay() const {
const int latency_blocks = BufferLatency();
int internal_delay = spectra_.read >= spectra_.write
? spectra_.read - spectra_.write
: spectra_.size + spectra_.read - spectra_.write;
return internal_delay - latency_blocks;
}
// Set the read indices according to the delay.
void RenderDelayBufferImpl::ApplyTotalDelay(int delay) {
RTC_LOG_V(delay_log_level_)
<< "Applying total delay of " << delay << " blocks.";
blocks_.read = blocks_.OffsetIndex(blocks_.write, -delay);
spectra_.read = spectra_.OffsetIndex(spectra_.write, delay);
ffts_.read = ffts_.OffsetIndex(ffts_.write, delay);
}
void RenderDelayBufferImpl::AlignFromExternalDelay() {
RTC_DCHECK(config_.delay.use_external_delay_estimator);
if (external_audio_buffer_delay_) {
const int64_t delay = render_call_counter_ - capture_call_counter_ +
*external_audio_buffer_delay_;
const int64_t delay_with_headroom =
delay - config_.delay.delay_headroom_samples / kBlockSize;
ApplyTotalDelay(delay_with_headroom);
}
}
// Inserts a block into the render buffers.
void RenderDelayBufferImpl::InsertBlock(
const std::vector<std::vector<std::vector<float>>>& block,
int previous_write) {
auto& b = blocks_;
auto& lr = low_rate_;
auto& ds = render_ds_;
auto& f = ffts_;
auto& s = spectra_;
const size_t num_bands = b.buffer[b.write].size();
const size_t num_render_channels = b.buffer[b.write][0].size();
RTC_DCHECK_EQ(block.size(), b.buffer[b.write].size());
for (size_t band = 0; band < num_bands; ++band) {
RTC_DCHECK_EQ(block[band].size(), num_render_channels);
RTC_DCHECK_EQ(b.buffer[b.write][band].size(), num_render_channels);
for (size_t ch = 0; ch < num_render_channels; ++ch) {
RTC_DCHECK_EQ(block[band][ch].size(), b.buffer[b.write][band][ch].size());
std::copy(block[band][ch].begin(), block[band][ch].end(),
b.buffer[b.write][band][ch].begin());
}
}
if (render_linear_amplitude_gain_ != 1.f) {
for (size_t band = 0; band < num_bands; ++band) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t k = 0; k < 64; ++k) {
b.buffer[b.write][band][ch][k] *= render_linear_amplitude_gain_;
}
}
}
}
std::array<float, kBlockSize> downmixed_render;
render_mixer_.ProduceOutput(b.buffer[b.write][0], downmixed_render);
render_decimator_.Decimate(downmixed_render, ds);
data_dumper_->DumpWav("aec3_render_decimator_output", ds.size(), ds.data(),
16000 / down_sampling_factor_, 1);
std::copy(ds.rbegin(), ds.rend(), lr.buffer.begin() + lr.write);
for (size_t channel = 0; channel < b.buffer[b.write][0].size(); ++channel) {
fft_.PaddedFft(b.buffer[b.write][0][channel],
b.buffer[previous_write][0][channel],
&f.buffer[f.write][channel]);
f.buffer[f.write][channel].Spectrum(optimization_,
s.buffer[s.write][channel]);
}
}
bool RenderDelayBufferImpl::DetectActiveRender(
rtc::ArrayView<const float> x) const {
const float x_energy = std::inner_product(x.begin(), x.end(), x.begin(), 0.f);
return x_energy > (config_.render_levels.active_render_limit *
config_.render_levels.active_render_limit) *
kFftLengthBy2;
}
bool RenderDelayBufferImpl::DetectExcessRenderBlocks() {
bool excess_render_detected = false;
const size_t latency_blocks = static_cast<size_t>(BufferLatency());
// The recently seen minimum latency in blocks. Should be close to 0.
min_latency_blocks_ = std::min(min_latency_blocks_, latency_blocks);
// After processing a configurable number of blocks the minimum latency is
// checked.
if (++excess_render_detection_counter_ >=
config_.buffering.excess_render_detection_interval_blocks) {
// If the minimum latency is not lower than the threshold there have been
// more render than capture frames.
excess_render_detected = min_latency_blocks_ >
config_.buffering.max_allowed_excess_render_blocks;
// Reset the counter and let the minimum latency be the current latency.
min_latency_blocks_ = latency_blocks;
excess_render_detection_counter_ = 0;
}
data_dumper_->DumpRaw("aec3_latency_blocks", latency_blocks);
data_dumper_->DumpRaw("aec3_min_latency_blocks", min_latency_blocks_);
data_dumper_->DumpRaw("aec3_excess_render_detected", excess_render_detected);
return excess_render_detected;
}
// Computes the latency in the buffer (the number of unread sub-blocks).
int RenderDelayBufferImpl::BufferLatency() const {
const DownsampledRenderBuffer& l = low_rate_;
int latency_samples = (l.buffer.size() + l.read - l.write) % l.buffer.size();
int latency_blocks = latency_samples / sub_block_size_;
return latency_blocks;
}
// Increments the write indices for the render buffers.
void RenderDelayBufferImpl::IncrementWriteIndices() {
low_rate_.UpdateWriteIndex(-sub_block_size_);
blocks_.IncWriteIndex();
spectra_.DecWriteIndex();
ffts_.DecWriteIndex();
}
// Increments the read indices of the low rate render buffers.
void RenderDelayBufferImpl::IncrementLowRateReadIndices() {
low_rate_.UpdateReadIndex(-sub_block_size_);
}
// Increments the read indices for the render buffers.
void RenderDelayBufferImpl::IncrementReadIndices() {
if (blocks_.read != blocks_.write) {
blocks_.IncReadIndex();
spectra_.DecReadIndex();
ffts_.DecReadIndex();
}
}
// Checks for a render buffer overrun.
bool RenderDelayBufferImpl::RenderOverrun() {
return low_rate_.read == low_rate_.write || blocks_.read == blocks_.write;
}
// Checks for a render buffer underrun.
bool RenderDelayBufferImpl::RenderUnderrun() {
return low_rate_.read == low_rate_.write;
}
} // namespace
RenderDelayBuffer* RenderDelayBuffer::Create(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels) {
return new RenderDelayBufferImpl(config, sample_rate_hz, num_render_channels);
}
} // namespace webrtc

86
webrtc/modules/audio_processing/aec3/render_delay_buffer.h Normal file
View File

@ -0,0 +1,86 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_RENDER_DELAY_BUFFER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_RENDER_DELAY_BUFFER_H_
#include <stddef.h>
#include <vector>
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/downsampled_render_buffer.h"
#include "modules/audio_processing/aec3/render_buffer.h"
namespace webrtc {
// Class for buffering the incoming render blocks such that these may be
// extracted with a specified delay.
class RenderDelayBuffer {
public:
enum class BufferingEvent {
kNone,
kRenderUnderrun,
kRenderOverrun,
kApiCallSkew
};
static RenderDelayBuffer* Create(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_render_channels);
virtual ~RenderDelayBuffer() = default;
// Resets the buffer alignment.
virtual void Reset() = 0;
// Inserts a block into the buffer.
virtual BufferingEvent Insert(
const std::vector<std::vector<std::vector<float>>>& block) = 0;
// Updates the buffers one step based on the specified buffer delay. Returns
// an enum indicating whether there was a special event that occurred.
virtual BufferingEvent PrepareCaptureProcessing() = 0;
// Called on capture blocks where PrepareCaptureProcessing is not called.
virtual void HandleSkippedCaptureProcessing() = 0;
// Sets the buffer delay and returns a bool indicating whether the delay
// changed.
virtual bool AlignFromDelay(size_t delay) = 0;
// Sets the buffer delay from the most recently reported external delay.
virtual void AlignFromExternalDelay() = 0;
// Gets the buffer delay.
virtual size_t Delay() const = 0;
// Gets the buffer delay.
virtual size_t MaxDelay() const = 0;
// Returns the render buffer for the echo remover.
virtual RenderBuffer* GetRenderBuffer() = 0;
// Returns the downsampled render buffer.
virtual const DownsampledRenderBuffer& GetDownsampledRenderBuffer() const = 0;
// Returns the maximum non calusal offset that can occur in the delay buffer.
static int DelayEstimatorOffset(const EchoCanceller3Config& config);
// Provides an optional external estimate of the audio buffer delay.
virtual void SetAudioBufferDelay(int delay_ms) = 0;
// Returns whether an external delay estimate has been reported via
// SetAudioBufferDelay.
virtual bool HasReceivedBufferDelay() = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_RENDER_DELAY_BUFFER_H_

196
webrtc/modules/audio_processing/aec3/render_delay_controller.cc Normal file
View File

@ -0,0 +1,196 @@
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/aec3/render_delay_controller.h"
#include <stddef.h>
#include <algorithm>
#include <memory>
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/aec3_common.h"
#include "modules/audio_processing/aec3/delay_estimate.h"
#include "modules/audio_processing/aec3/downsampled_render_buffer.h"
#include "modules/audio_processing/aec3/echo_path_delay_estimator.h"
#include "modules/audio_processing/aec3/render_delay_controller_metrics.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomic_ops.h"
#include "rtc_base/checks.h"
namespace webrtc {
namespace {
class RenderDelayControllerImpl final : public RenderDelayController {
public:
RenderDelayControllerImpl(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_capture_channels);
RenderDelayControllerImpl() = delete;
RenderDelayControllerImpl(const RenderDelayControllerImpl&) = delete;
RenderDelayControllerImpl& operator=(const RenderDelayControllerImpl&) =
delete;
~RenderDelayControllerImpl() override;
void Reset(bool reset_delay_confidence) override;
void LogRenderCall() override;
absl::optional<DelayEstimate> GetDelay(
const DownsampledRenderBuffer& render_buffer,
size_t render_delay_buffer_delay,
const std::vector<std::vector<float>>& capture) override;
bool HasClockdrift() const override;
private:
static int instance_count_;
std::unique_ptr<ApmDataDumper> data_dumper_;
const int hysteresis_limit_blocks_;
const int delay_headroom_samples_;
absl::optional<DelayEstimate> delay_;
EchoPathDelayEstimator delay_estimator_;
RenderDelayControllerMetrics metrics_;
absl::optional<DelayEstimate> delay_samples_;
size_t capture_call_counter_ = 0;
int delay_change_counter_ = 0;
DelayEstimate::Quality last_delay_estimate_quality_;
};
DelayEstimate ComputeBufferDelay(
const absl::optional<DelayEstimate>& current_delay,
int hysteresis_limit_blocks,
int delay_headroom_samples,
DelayEstimate estimated_delay) {
// Subtract delay headroom.
const int delay_with_headroom_samples = std::max(
static_cast<int>(estimated_delay.delay) - delay_headroom_samples, 0);
// Compute the buffer delay increase required to achieve the desired latency.
size_t new_delay_blocks = delay_with_headroom_samples >> kBlockSizeLog2;
// Add hysteresis.
if (current_delay) {
size_t current_delay_blocks = current_delay->delay;
if (new_delay_blocks > current_delay_blocks &&
new_delay_blocks <= current_delay_blocks + hysteresis_limit_blocks) {
new_delay_blocks = current_delay_blocks;
}
}
DelayEstimate new_delay = estimated_delay;
new_delay.delay = new_delay_blocks;
return new_delay;
}
int RenderDelayControllerImpl::instance_count_ = 0;
RenderDelayControllerImpl::RenderDelayControllerImpl(
const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_capture_channels)
: data_dumper_(
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
hysteresis_limit_blocks_(
static_cast<int>(config.delay.hysteresis_limit_blocks)),
delay_headroom_samples_(config.delay.delay_headroom_samples),
delay_estimator_(data_dumper_.get(), config, num_capture_channels),
last_delay_estimate_quality_(DelayEstimate::Quality::kCoarse) {
RTC_DCHECK(ValidFullBandRate(sample_rate_hz));
delay_estimator_.LogDelayEstimationProperties(sample_rate_hz, 0);
}
RenderDelayControllerImpl::~RenderDelayControllerImpl() = default;
void RenderDelayControllerImpl::Reset(bool reset_delay_confidence) {
delay_ = absl::nullopt;
delay_samples_ = absl::nullopt;
delay_estimator_.Reset(reset_delay_confidence);
delay_change_counter_ = 0;
if (reset_delay_confidence) {
last_delay_estimate_quality_ = DelayEstimate::Quality::kCoarse;
}
}
void RenderDelayControllerImpl::LogRenderCall() {}
absl::optional<DelayEstimate> RenderDelayControllerImpl::GetDelay(
const DownsampledRenderBuffer& render_buffer,
size_t render_delay_buffer_delay,
const std::vector<std::vector<float>>& capture) {
RTC_DCHECK_EQ(kBlockSize, capture[0].size());
++capture_call_counter_;
auto delay_samples = delay_estimator_.EstimateDelay(render_buffer, capture);
if (delay_samples) {
if (!delay_samples_ || delay_samples->delay != delay_samples_->delay) {
delay_change_counter_ = 0;
}
if (delay_samples_) {
delay_samples_->blocks_since_last_change =
delay_samples_->delay == delay_samples->delay
? delay_samples_->blocks_since_last_change + 1
: 0;
delay_samples_->blocks_since_last_update = 0;
delay_samples_->delay = delay_samples->delay;
delay_samples_->quality = delay_samples->quality;
} else {
delay_samples_ = delay_samples;
}
} else {
if (delay_samples_) {
++delay_samples_->blocks_since_last_change;
++delay_samples_->blocks_since_last_update;
}
}
if (delay_change_counter_ < 2 * kNumBlocksPerSecond) {
++delay_change_counter_;
}
if (delay_samples_) {
// Compute the render delay buffer delay.
const bool use_hysteresis =
last_delay_estimate_quality_ == DelayEstimate::Quality::kRefined &&
delay_samples_->quality == DelayEstimate::Quality::kRefined;
delay_ = ComputeBufferDelay(delay_,
use_hysteresis ? hysteresis_limit_blocks_ : 0,
delay_headroom_samples_, *delay_samples_);
last_delay_estimate_quality_ = delay_samples_->quality;
}
metrics_.Update(delay_samples_ ? absl::optional<size_t>(delay_samples_->delay)
: absl::nullopt,
delay_ ? delay_->delay : 0, 0, delay_estimator_.Clockdrift());
data_dumper_->DumpRaw("aec3_render_delay_controller_delay",
delay_samples ? delay_samples->delay : 0);
data_dumper_->DumpRaw("aec3_render_delay_controller_buffer_delay",
delay_ ? delay_->delay : 0);
return delay_;
}
bool RenderDelayControllerImpl::HasClockdrift() const {
return delay_estimator_.Clockdrift() != ClockdriftDetector::Level::kNone;
}
} // namespace
RenderDelayController* RenderDelayController::Create(
const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_capture_channels) {
return new RenderDelayControllerImpl(config, sample_rate_hz,
num_capture_channels);
}
} // namespace webrtc

50
webrtc/modules/audio_processing/aec3/render_delay_controller.h Normal file
View File

@ -0,0 +1,50 @@
/*
* Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef MODULES_AUDIO_PROCESSING_AEC3_RENDER_DELAY_CONTROLLER_H_
#define MODULES_AUDIO_PROCESSING_AEC3_RENDER_DELAY_CONTROLLER_H_
#include "absl/types/optional.h"
#include "api/array_view.h"
#include "api/audio/echo_canceller3_config.h"
#include "modules/audio_processing/aec3/delay_estimate.h"
#include "modules/audio_processing/aec3/downsampled_render_buffer.h"
#include "modules/audio_processing/aec3/render_delay_buffer.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
namespace webrtc {
// Class for aligning the render and capture signal using a RenderDelayBuffer.
class RenderDelayController {
public:
static RenderDelayController* Create(const EchoCanceller3Config& config,
int sample_rate_hz,
size_t num_capture_channels);
virtual ~RenderDelayController() = default;
// Resets the delay controller. If the delay confidence is reset, the reset
// behavior is as if the call is restarted.
virtual void Reset(bool reset_delay_confidence) = 0;
// Logs a render call.
virtual void LogRenderCall() = 0;
// Aligns the render buffer content with the capture signal.
virtual absl::optional<DelayEstimate> GetDelay(
const DownsampledRenderBuffer& render_buffer,
size_t render_delay_buffer_delay,
const std::vector<std::vector<float>>& capture) = 0;
// Returns true if clockdrift has been detected.
virtual bool HasClockdrift() const = 0;
};
} // namespace webrtc
#endif // MODULES_AUDIO_PROCESSING_AEC3_RENDER_DELAY_CONTROLLER_H_

Some files were not shown because too many files have changed in this diff Show More