368 lines
18 KiB
C++
368 lines
18 KiB
C++
/*******************************************************************************
|
|
Copyright (c) 2005-2009 David Williams
|
|
|
|
This software is provided 'as-is', without any express or implied
|
|
warranty. In no event will the authors be held liable for any damages
|
|
arising from the use of this software.
|
|
|
|
Permission is granted to anyone to use this software for any purpose,
|
|
including commercial applications, and to alter it and redistribute it
|
|
freely, subject to the following restrictions:
|
|
|
|
1. The origin of this software must not be misrepresented; you must not
|
|
claim that you wrote the original software. If you use this software
|
|
in a product, an acknowledgment in the product documentation would be
|
|
appreciated but is not required.
|
|
|
|
2. Altered source versions must be plainly marked as such, and must not be
|
|
misrepresented as being the original software.
|
|
|
|
3. This notice may not be removed or altered from any source
|
|
distribution.
|
|
*******************************************************************************/
|
|
|
|
#include "PolyVox/Impl/Timer.h"
|
|
|
|
namespace PolyVox
|
|
{
|
|
template< typename Sampler, typename ControllerType>
|
|
Vector3DFloat computeCentralDifferenceGradient(const Sampler& volIter, ControllerType& controller)
|
|
{
|
|
//FIXME - Should actually use DensityType here, both in principle and because the maths may be
|
|
//faster (and to reduce casts). So it would be good to add a way to get DensityType from a voxel.
|
|
//But watch out for when the DensityType is unsigned and the difference could be negative.
|
|
float voxel1nx = static_cast<float>(controller.convertToDensity(volIter.peekVoxel1nx0py0pz()));
|
|
float voxel1px = static_cast<float>(controller.convertToDensity(volIter.peekVoxel1px0py0pz()));
|
|
|
|
float voxel1ny = static_cast<float>(controller.convertToDensity(volIter.peekVoxel0px1ny0pz()));
|
|
float voxel1py = static_cast<float>(controller.convertToDensity(volIter.peekVoxel0px1py0pz()));
|
|
|
|
float voxel1nz = static_cast<float>(controller.convertToDensity(volIter.peekVoxel0px0py1nz()));
|
|
float voxel1pz = static_cast<float>(controller.convertToDensity(volIter.peekVoxel0px0py1pz()));
|
|
|
|
return Vector3DFloat
|
|
(
|
|
voxel1nx - voxel1px,
|
|
voxel1ny - voxel1py,
|
|
voxel1nz - voxel1pz
|
|
);
|
|
}
|
|
|
|
template< typename VolumeType, typename MeshType, typename ControllerType >
|
|
void extractMarchingCubesMeshCustom(VolumeType* volData, Region region, MeshType* result, ControllerType controller)
|
|
{
|
|
// Validate parameters
|
|
POLYVOX_THROW_IF(volData == nullptr, std::invalid_argument, "Provided volume cannot be null");
|
|
POLYVOX_THROW_IF(result == nullptr, std::invalid_argument, "Provided mesh cannot be null");
|
|
|
|
// For profiling this function
|
|
Timer timer;
|
|
|
|
// Performance note: Profiling indicates that simply adding vertices and indices to the std::vector is one
|
|
// of the bottlenecks when generating the mesh. Reserving space in advance helps here but is wasteful in the
|
|
// common case that no/few vertices are generated. Maybe it's worth reserving a couple of thousand or so?
|
|
result->clear();
|
|
|
|
// Store some commonly used values for performance and convienience
|
|
const uint32_t uRegionWidthInVoxels = region.getWidthInVoxels();
|
|
const uint32_t uRegionHeightInVoxels = region.getHeightInVoxels();
|
|
const uint32_t uRegionDepthInVoxels = region.getDepthInVoxels();
|
|
|
|
typename ControllerType::DensityType tThreshold = controller.getThreshold();
|
|
|
|
// A naive implemetation of Marching Cubes might sample the eight corner voxels of every cell to determine the cell index.
|
|
// However, when processing the cells sequentially we cn observe that many of the voxels are shared with previous adjacent
|
|
// cells, and so we can obtain these by careful bit-shifting. These variables keep track of previous cells for this purpose.
|
|
// We don't clear the arrays because the algorithm ensures that we only read from elements we have previously written to.
|
|
uint8_t uPreviousCellIndex = 0;
|
|
Array1DUint8 pPreviousRowCellIndices(uRegionWidthInVoxels);
|
|
Array2DUint8 pPreviousSliceCellIndices(uRegionWidthInVoxels, uRegionHeightInVoxels);
|
|
|
|
// A given vertex may be shared by multiple triangles, so we need to keep track of the indices into the vertex array.
|
|
// We don't clear the arrays because the algorithm ensures that we only read from elements we have previously written to.
|
|
Array<2, Vector3DInt32> pIndices(uRegionWidthInVoxels, uRegionHeightInVoxels);
|
|
Array<2, Vector3DInt32> pPreviousIndices(uRegionWidthInVoxels, uRegionHeightInVoxels);
|
|
|
|
// A sampler pointing at the beginning of the region, which gets incremented to always point at the beginning of a slice.
|
|
typename VolumeType::Sampler startOfSlice(volData);
|
|
startOfSlice.setPosition(region.getLowerX(), region.getLowerY(), region.getLowerZ());
|
|
|
|
for (uint32_t uZRegSpace = 0; uZRegSpace < uRegionDepthInVoxels; uZRegSpace++)
|
|
{
|
|
// A sampler pointing at the beginning of the slice, which gets incremented to always point at the beginning of a row.
|
|
typename VolumeType::Sampler startOfRow = startOfSlice;
|
|
|
|
for (uint32_t uYRegSpace = 0; uYRegSpace < uRegionHeightInVoxels; uYRegSpace++)
|
|
{
|
|
// Copying a sampler which is already pointing at the correct location seems (slightly) faster than
|
|
// calling setPosition(). Therefore we make use of 'startOfRow' and 'startOfSlice' to reset the sampler.
|
|
typename VolumeType::Sampler sampler = startOfRow;
|
|
|
|
for (uint32_t uXRegSpace = 0; uXRegSpace < uRegionWidthInVoxels; uXRegSpace++)
|
|
{
|
|
// Note: In many cases the provided region will be (mostly) empty which means mesh vertices/indices
|
|
// are not generated and the only thing that is done for each cell is the computation of uCellIndex.
|
|
// It appears that retriving the voxel value is not so expensive and that it is the bitwise combining
|
|
// which actually carries the cost.
|
|
//
|
|
// If we really need to speed this up more then it may be possible to pack 4 8-bit cell indices into
|
|
// a single 32-bit value and then perform the bitwise logic on all four of them at the same time.
|
|
// However, this complicates the code and there would still be the cost of packing/unpacking so it's
|
|
// not clear if there is really a benefit. It's something to consider in the future.
|
|
|
|
// Each bit of the cell index specifies whether a given corner of the cell is above or below the threshold.
|
|
uint8_t uCellIndex = 0;
|
|
|
|
// Four bits of our cube index are obtained by looking at the cube index for
|
|
// the previous slice and copying four of those bits into their new positions.
|
|
uint8_t uPreviousCellIndexZ = pPreviousSliceCellIndices(uXRegSpace, uYRegSpace);
|
|
uPreviousCellIndexZ >>= 4;
|
|
uCellIndex |= uPreviousCellIndexZ;
|
|
|
|
// Two bits of our cube index are obtained by looking at the cube index for
|
|
// the previous row and copying two of those bits into their new positions.
|
|
uint8_t uPreviousCellIndexY = pPreviousRowCellIndices(uXRegSpace);
|
|
uPreviousCellIndexY &= 204; //204 = 128+64+8+4
|
|
uPreviousCellIndexY >>= 2;
|
|
uCellIndex |= uPreviousCellIndexY;
|
|
|
|
// One bit of our cube index are obtained by looking at the cube index for
|
|
// the previous cell and copying one of those bits into it's new position.
|
|
uint8_t UPreviousCellIndexX = uPreviousCellIndex;
|
|
UPreviousCellIndexX &= 170; //170 = 128+32+8+2
|
|
UPreviousCellIndexX >>= 1;
|
|
uCellIndex |= UPreviousCellIndexX;
|
|
|
|
// The last bit of our cube index is obtained by looking
|
|
// at the relevant voxel and comparing it to the threshold
|
|
typename VolumeType::VoxelType v111 = sampler.getVoxel();
|
|
if (controller.convertToDensity(v111) < tThreshold) uCellIndex |= 128;
|
|
|
|
// The current value becomes the previous value, ready for the next iteration.
|
|
uPreviousCellIndex = uCellIndex;
|
|
pPreviousRowCellIndices(uXRegSpace) = uCellIndex;
|
|
pPreviousSliceCellIndices(uXRegSpace, uYRegSpace) = uCellIndex;
|
|
|
|
// 12 bits of uEdge determine whether a vertex is placed on each of the 12 edges of the cell.
|
|
uint16_t uEdge = edgeTable[uCellIndex];
|
|
|
|
// Test whether any vertices and indices should be generated for the current cell (i.e. it is occupied).
|
|
// Performance note: This condition is usually false because most cells in a volume are completely above
|
|
// or below the threshold and hence unoccupied. However, even when it is always false (testing on an empty
|
|
// volume) it still incurs significant overhead, probably because the code is large and bloats the for loop
|
|
// which contains it. On my empty volume test case the code as given runs in 34ms, but if I replace the
|
|
// condition with 'false' it runs in 24ms and gives the same output (i.e. none).
|
|
//
|
|
// An improvement is to move the code into a seperate function which does speed things up (30ms), but this
|
|
// is messy as the function needs to be passed about 10 differnt parameters, probably adding some overhead
|
|
// in its self. This does indeed seem to slow down the case when cells are occupied, by about 10-20%.
|
|
//
|
|
// Overall I don't know the right solution, but I'm leaving the code as-is to avoid making it messy. If we
|
|
// can reduce the number of parameters which need to be passed then it might be worth moving it into a
|
|
// function, or otherwise it may simply be worth trying to shorten the code (e.g. adding other function
|
|
// calls). For now we will leave it as-is, until we have more information from real-world profiling.
|
|
if (uEdge != 0)
|
|
{
|
|
auto v111Density = controller.convertToDensity(v111);
|
|
|
|
// Performance note: Computing normals is one of the bottlencks in the mesh generation process. The
|
|
// central difference approach actually samples the same voxel more than once as we call it on two
|
|
// adjacent voxels. Perhaps we could expand this and eliminate dupicates in the future. Alternatively,
|
|
// we could compute vertex normals from adjacent face normals instead of via central differencing,
|
|
// but not for vertices on the edge of the region (as this causes visual discontinities).
|
|
const Vector3DFloat n000 = computeCentralDifferenceGradient(sampler, controller);
|
|
|
|
/* Find the vertices where the surface intersects the cube */
|
|
if ((uEdge & 64) && (uXRegSpace > 0))
|
|
{
|
|
sampler.moveNegativeX();
|
|
typename VolumeType::VoxelType v011 = sampler.getVoxel();
|
|
auto v011Density = controller.convertToDensity(v011);
|
|
const float fInterp = static_cast<float>(tThreshold - v011Density) / static_cast<float>(v111Density - v011Density);
|
|
|
|
// Compute the position
|
|
const Vector3DFloat v3dPosition(static_cast<float>(uXRegSpace - 1) + fInterp, static_cast<float>(uYRegSpace), static_cast<float>(uZRegSpace));
|
|
|
|
// Compute the normal
|
|
const Vector3DFloat n100 = computeCentralDifferenceGradient(sampler, controller);
|
|
Vector3DFloat v3dNormal = (n100*fInterp) + (n000*(1 - fInterp));
|
|
|
|
// The gradient for a voxel can be zero (e.g. solid voxel surrounded by empty ones) and so
|
|
// the interpolated normal can also be zero (e.g. a grid of alternating solid and empty voxels).
|
|
if (v3dNormal.lengthSquared() > 0.000001f)
|
|
{
|
|
v3dNormal.normalise();
|
|
}
|
|
|
|
// Allow the controller to decide how the material should be derived from the voxels.
|
|
const typename VolumeType::VoxelType uMaterial = controller.blendMaterials(v011, v111, fInterp);
|
|
|
|
MarchingCubesVertex<typename VolumeType::VoxelType> surfaceVertex;
|
|
const Vector3DUint16 v3dScaledPosition(static_cast<uint16_t>(v3dPosition.getX() * 256.0f), static_cast<uint16_t>(v3dPosition.getY() * 256.0f), static_cast<uint16_t>(v3dPosition.getZ() * 256.0f));
|
|
surfaceVertex.encodedPosition = v3dScaledPosition;
|
|
surfaceVertex.encodedNormal = encodeNormal(v3dNormal);
|
|
surfaceVertex.data = uMaterial;
|
|
|
|
const uint32_t uLastVertexIndex = result->addVertex(surfaceVertex);
|
|
pIndices(uXRegSpace, uYRegSpace).setX(uLastVertexIndex);
|
|
|
|
sampler.movePositiveX();
|
|
}
|
|
if ((uEdge & 32) && (uYRegSpace > 0))
|
|
{
|
|
sampler.moveNegativeY();
|
|
typename VolumeType::VoxelType v101 = sampler.getVoxel();
|
|
auto v101Density = controller.convertToDensity(v101);
|
|
const float fInterp = static_cast<float>(tThreshold - v101Density) / static_cast<float>(v111Density - v101Density);
|
|
|
|
// Compute the position
|
|
const Vector3DFloat v3dPosition(static_cast<float>(uXRegSpace), static_cast<float>(uYRegSpace - 1) + fInterp, static_cast<float>(uZRegSpace));
|
|
|
|
// Compute the normal
|
|
const Vector3DFloat n010 = computeCentralDifferenceGradient(sampler, controller);
|
|
Vector3DFloat v3dNormal = (n010*fInterp) + (n000*(1 - fInterp));
|
|
|
|
// The gradient for a voxel can be zero (e.g. solid voxel surrounded by empty ones) and so
|
|
// the interpolated normal can also be zero (e.g. a grid of alternating solid and empty voxels).
|
|
if (v3dNormal.lengthSquared() > 0.000001f)
|
|
{
|
|
v3dNormal.normalise();
|
|
}
|
|
|
|
// Allow the controller to decide how the material should be derived from the voxels.
|
|
const typename VolumeType::VoxelType uMaterial = controller.blendMaterials(v101, v111, fInterp);
|
|
|
|
MarchingCubesVertex<typename VolumeType::VoxelType> surfaceVertex;
|
|
const Vector3DUint16 v3dScaledPosition(static_cast<uint16_t>(v3dPosition.getX() * 256.0f), static_cast<uint16_t>(v3dPosition.getY() * 256.0f), static_cast<uint16_t>(v3dPosition.getZ() * 256.0f));
|
|
surfaceVertex.encodedPosition = v3dScaledPosition;
|
|
surfaceVertex.encodedNormal = encodeNormal(v3dNormal);
|
|
surfaceVertex.data = uMaterial;
|
|
|
|
uint32_t uLastVertexIndex = result->addVertex(surfaceVertex);
|
|
pIndices(uXRegSpace, uYRegSpace).setY(uLastVertexIndex);
|
|
|
|
sampler.movePositiveY();
|
|
}
|
|
if ((uEdge & 1024) && (uZRegSpace > 0))
|
|
{
|
|
sampler.moveNegativeZ();
|
|
typename VolumeType::VoxelType v110 = sampler.getVoxel();
|
|
auto v110Density = controller.convertToDensity(v110);
|
|
const float fInterp = static_cast<float>(tThreshold - v110Density) / static_cast<float>(v111Density - v110Density);
|
|
|
|
// Compute the position
|
|
const Vector3DFloat v3dPosition(static_cast<float>(uXRegSpace), static_cast<float>(uYRegSpace), static_cast<float>(uZRegSpace - 1) + fInterp);
|
|
|
|
// Compute the normal
|
|
const Vector3DFloat n001 = computeCentralDifferenceGradient(sampler, controller);
|
|
Vector3DFloat v3dNormal = (n001*fInterp) + (n000*(1 - fInterp));
|
|
|
|
// The gradient for a voxel can be zero (e.g. solid voxel surrounded by empty ones) and so
|
|
// the interpolated normal can also be zero (e.g. a grid of alternating solid and empty voxels).
|
|
if (v3dNormal.lengthSquared() > 0.000001f)
|
|
{
|
|
v3dNormal.normalise();
|
|
}
|
|
|
|
// Allow the controller to decide how the material should be derived from the voxels.
|
|
const typename VolumeType::VoxelType uMaterial = controller.blendMaterials(v110, v111, fInterp);
|
|
|
|
MarchingCubesVertex<typename VolumeType::VoxelType> surfaceVertex;
|
|
const Vector3DUint16 v3dScaledPosition(static_cast<uint16_t>(v3dPosition.getX() * 256.0f), static_cast<uint16_t>(v3dPosition.getY() * 256.0f), static_cast<uint16_t>(v3dPosition.getZ() * 256.0f));
|
|
surfaceVertex.encodedPosition = v3dScaledPosition;
|
|
surfaceVertex.encodedNormal = encodeNormal(v3dNormal);
|
|
surfaceVertex.data = uMaterial;
|
|
|
|
const uint32_t uLastVertexIndex = result->addVertex(surfaceVertex);
|
|
pIndices(uXRegSpace, uYRegSpace).setZ(uLastVertexIndex);
|
|
|
|
sampler.movePositiveZ();
|
|
}
|
|
|
|
// Now output the indices. For the first row, column or slice there aren't
|
|
// any (the region size in cells is one less than the region size in voxels)
|
|
if ((uXRegSpace != 0) && (uYRegSpace != 0) && (uZRegSpace != 0))
|
|
{
|
|
|
|
int32_t indlist[12];
|
|
|
|
/* Find the vertices where the surface intersects the cube */
|
|
if (uEdge & 1)
|
|
{
|
|
indlist[0] = pPreviousIndices(uXRegSpace, uYRegSpace - 1).getX();
|
|
}
|
|
if (uEdge & 2)
|
|
{
|
|
indlist[1] = pPreviousIndices(uXRegSpace, uYRegSpace).getY();
|
|
}
|
|
if (uEdge & 4)
|
|
{
|
|
indlist[2] = pPreviousIndices(uXRegSpace, uYRegSpace).getX();
|
|
}
|
|
if (uEdge & 8)
|
|
{
|
|
indlist[3] = pPreviousIndices(uXRegSpace - 1, uYRegSpace).getY();
|
|
}
|
|
if (uEdge & 16)
|
|
{
|
|
indlist[4] = pIndices(uXRegSpace, uYRegSpace - 1).getX();
|
|
}
|
|
if (uEdge & 32)
|
|
{
|
|
indlist[5] = pIndices(uXRegSpace, uYRegSpace).getY();
|
|
}
|
|
if (uEdge & 64)
|
|
{
|
|
indlist[6] = pIndices(uXRegSpace, uYRegSpace).getX();
|
|
}
|
|
if (uEdge & 128)
|
|
{
|
|
indlist[7] = pIndices(uXRegSpace - 1, uYRegSpace).getY();
|
|
}
|
|
if (uEdge & 256)
|
|
{
|
|
indlist[8] = pIndices(uXRegSpace - 1, uYRegSpace - 1).getZ();
|
|
}
|
|
if (uEdge & 512)
|
|
{
|
|
indlist[9] = pIndices(uXRegSpace, uYRegSpace - 1).getZ();
|
|
}
|
|
if (uEdge & 1024)
|
|
{
|
|
indlist[10] = pIndices(uXRegSpace, uYRegSpace).getZ();
|
|
}
|
|
if (uEdge & 2048)
|
|
{
|
|
indlist[11] = pIndices(uXRegSpace - 1, uYRegSpace).getZ();
|
|
}
|
|
|
|
for (int i = 0; triTable[uCellIndex][i] != -1; i += 3)
|
|
{
|
|
const int32_t ind0 = indlist[triTable[uCellIndex][i]];
|
|
const int32_t ind1 = indlist[triTable[uCellIndex][i + 1]];
|
|
const int32_t ind2 = indlist[triTable[uCellIndex][i + 2]];
|
|
|
|
if ((ind0 != -1) && (ind1 != -1) && (ind2 != -1))
|
|
{
|
|
result->addTriangle(ind0, ind1, ind2);
|
|
}
|
|
} // For each triangle
|
|
}
|
|
} // For each cell
|
|
sampler.movePositiveX();
|
|
} // For X
|
|
startOfRow.movePositiveY();
|
|
} // For Y
|
|
startOfSlice.movePositiveZ();
|
|
|
|
pIndices.swap(pPreviousIndices);
|
|
} // For Z
|
|
|
|
result->setOffset(region.getLowerCorner());
|
|
|
|
POLYVOX_LOG_TRACE("Marching cubes surface extraction took ", timer.elapsedTimeInMilliSeconds(),
|
|
"ms (Region size = ", region.getWidthInVoxels(), "x", region.getHeightInVoxels(),
|
|
"x", region.getDepthInVoxels(), ")");
|
|
}
|
|
}
|