Skip to content

A fast 2D and 3D game engine — TypeScript SDK, C++/WebAssembly core, visual editor. Ship one project to web, desktop, WeChat MiniGames, playable ads, and native Android / iOS.

14.1 KB
Raw Blame
// Copyright (c) 2001-2004 NovodeX AG. All rights reserved.
// Copyright (c) 2004-2008 AGEIA Technologies, Inc. All rights reserved.
// SPDX-FileCopyrightText: Copyright (c) 2008-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

#include "GuCookingSDF.h"
#include "cooking/PxTriangleMeshDesc.h"
#include "GuSDF.h"
#include "GuCooking.h"
#include "PxSDFBuilder.h"

using namespace physx;

namespace
{
	struct MeshData
	{
		MeshData(const PxTriangleMeshDesc& desc)
		{
			m_positions.resize(desc.points.count);
			m_indices.resize(desc.triangles.count * 3);

			immediateCooking::gatherStrided(desc.points.data, &m_positions[0], desc.points.count, sizeof(PxVec3), desc.points.stride);
			immediateCooking::gatherStrided(desc.triangles.data, &m_indices[0], desc.triangles.count, 3 * sizeof(PxU32), desc.triangles.stride);
		}

		void GetBounds(PxVec3& outMinExtents, PxVec3& outMaxExtents) const
		{
			PxVec3 minExtents(FLT_MAX);
			PxVec3 maxExtents(-FLT_MAX);

			for (PxU32 i = 0; i < m_positions.size(); ++i)
			{
				const PxVec3& a = m_positions[i];

				minExtents = a.minimum(minExtents);
				maxExtents = a.maximum(maxExtents);
			}

			outMinExtents = minExtents;
			outMaxExtents = maxExtents;
		}

		PxArray<PxVec3> m_positions;
		PxArray<PxU32> m_indices;
	};
}

static void quantizeSparseSDF(PxSdfBitsPerSubgridPixel::Enum bitsPerSubgridPixel,
	const PxArray<PxReal>& uncompressedSdfDataSubgrids, PxArray<PxU8>& compressedSdfDataSubgrids,
	PxReal subgridsMinSdfValue, PxReal subgridsMaxSdfValue)
{
	PxU32 bytesPerPixel = PxU32(bitsPerSubgridPixel);

	compressedSdfDataSubgrids.resize(uncompressedSdfDataSubgrids.size() * bytesPerPixel);

	PxReal* ptr32 = reinterpret_cast<PxReal*>(compressedSdfDataSubgrids.begin());
	PxU16* ptr16 = reinterpret_cast<PxU16*>(compressedSdfDataSubgrids.begin());
	PxU8* ptr8 = compressedSdfDataSubgrids.begin();

	PxReal s = 1.0f / (subgridsMaxSdfValue - subgridsMinSdfValue);

	for (PxU32 i = 0; i < uncompressedSdfDataSubgrids.size(); ++i)
	{
		PxReal v = uncompressedSdfDataSubgrids[i];
		PxReal vNormalized;
		if (v == FLT_MAX)
			vNormalized = 0.0f; //Not all subgrids in the 3d texture are used, can assign an arbitrary value to unused subgrids
		else
			vNormalized = (v - subgridsMinSdfValue) * s;

		switch (bitsPerSubgridPixel)
		{
		case PxSdfBitsPerSubgridPixel::e8_BIT_PER_PIXEL:
			PX_ASSERT(vNormalized >= 0.0f);
			PX_ASSERT(vNormalized <= 1.0f);
			ptr8[i] = PxU8(255.0f * vNormalized);
			break;
		case PxSdfBitsPerSubgridPixel::e16_BIT_PER_PIXEL:
			PX_ASSERT(vNormalized >= 0.0f);
			PX_ASSERT(vNormalized <= 1.0f);
			ptr16[i] = PxU16(65535.0f * vNormalized);
			break;
		case PxSdfBitsPerSubgridPixel::e32_BIT_PER_PIXEL:
			ptr32[i] = v;
			break;			
		}
	}
}

static PX_FORCE_INLINE PxU32 idxCompact(PxU32 x, PxU32 y, PxU32 z, PxU32 width, PxU32 height)
{
	return z * (width) * (height)+y * (width)+x;
}

static void convert16To32Bits(PxSimpleTriangleMesh mesh, PxArray<PxU32>& indices32)
{
	indices32.resize(3 * mesh.triangles.count);
	if (mesh.flags & PxMeshFlag::e16_BIT_INDICES)
	{
		// conversion; 16 bit index -> 32 bit index & stride
		PxU32* dest = indices32.begin();
		const PxU32* pastLastDest = indices32.begin() + 3 * mesh.triangles.count;
		const PxU8* source = reinterpret_cast<const PxU8*>(mesh.triangles.data);
		while (dest < pastLastDest)
		{
			const PxU16 * trig16 = reinterpret_cast<const PxU16*>(source);
			*dest++ = trig16[0];
			*dest++ = trig16[1];
			*dest++ = trig16[2];
			source += mesh.triangles.stride;
		}
	}
	else
	{
		immediateCooking::gatherStrided(mesh.triangles.data, indices32.begin(), mesh.triangles.count, sizeof(PxU32) * 3, mesh.triangles.stride);
	}
}

static bool createSDFSparse(PxTriangleMeshDesc& desc, PxSDFDesc& sdfDesc, PxArray<PxReal>& sdfCoarse, PxArray<PxU8>& sdfDataSubgrids,
	PxArray<PxU32>& sdfSubgridsStartSlots)
{
	PX_ASSERT(sdfDesc.subgridSize > 0);

	MeshData mesh(desc);

	PxVec3 meshLower, meshUpper;
	if (sdfDesc.sdfBounds.isEmpty())
		mesh.GetBounds(meshLower, meshUpper);
	else
	{
		meshLower = sdfDesc.sdfBounds.minimum;
		meshUpper = sdfDesc.sdfBounds.maximum;
	}

	PxVec3 edges = meshUpper - meshLower;

	const PxReal spacing = sdfDesc.spacing;

	// tweak spacing to avoid edge cases for vertices laying on the boundary
	// just covers the case where an edge is a whole multiple of the spacing.
	PxReal spacingEps = spacing * (1.0f - 1e-4f);

	// make sure to have at least one particle in each dimension
	PxI32 dx, dy, dz;
	dx = spacing > edges.x ? 1 : PxI32(edges.x / spacingEps);
	dy = spacing > edges.y ? 1 : PxI32(edges.y / spacingEps);
	dz = spacing > edges.z ? 1 : PxI32(edges.z / spacingEps);

	dx += 4;
	dy += 4;
	dz += 4;

	//Make sure that dx, dy and dz are multiple of subgridSize
	dx = ((dx + sdfDesc.subgridSize - 1) / sdfDesc.subgridSize) * sdfDesc.subgridSize;
	dy = ((dy + sdfDesc.subgridSize - 1) / sdfDesc.subgridSize) * sdfDesc.subgridSize;
	dz = ((dz + sdfDesc.subgridSize - 1) / sdfDesc.subgridSize) * sdfDesc.subgridSize;

	PX_ASSERT(dx % sdfDesc.subgridSize == 0);
	PX_ASSERT(dy % sdfDesc.subgridSize == 0);
	PX_ASSERT(dz % sdfDesc.subgridSize == 0);

	// we shift the voxelization bounds so that the voxel centers
	// lie symmetrically to the center of the object. this reduces the 
	// chance of missing features, and also better aligns the particles
	// with the mesh
	PxVec3 meshOffset;
	meshOffset.x = 0.5f * (spacing - (edges.x - (dx - 1)*spacing));
	meshOffset.y = 0.5f * (spacing - (edges.y - (dy - 1)*spacing));
	meshOffset.z = 0.5f * (spacing - (edges.z - (dz - 1)*spacing));
	meshLower -= meshOffset;

	sdfDesc.meshLower = meshLower;

	sdfDesc.dims.x = dx;
	sdfDesc.dims.y = dy;
	sdfDesc.dims.z = dz;

	PxArray<PxU32> indices32;
	PxArray<PxVec3> vertices;
	const PxVec3* verticesPtr = NULL;
	bool baseMeshSpecified = sdfDesc.baseMesh.triangles.data && sdfDesc.baseMesh.points.data;
	if (baseMeshSpecified)
	{
		convert16To32Bits(sdfDesc.baseMesh, indices32);

		if (sdfDesc.baseMesh.points.stride != sizeof(PxVec3))
		{
			vertices.resize(sdfDesc.baseMesh.points.count);
			immediateCooking::gatherStrided(sdfDesc.baseMesh.points.data, vertices.begin(), sdfDesc.baseMesh.points.count, sizeof(PxVec3), sdfDesc.baseMesh.points.stride);
			verticesPtr = vertices.begin();
		}
		else
			verticesPtr = reinterpret_cast<const PxVec3*>(sdfDesc.baseMesh.points.data);
	}

	PxReal subgridsMinSdfValue = 0.0f;
	PxReal subgridsMaxSdfValue = 1.0f;
	PxReal narrowBandThickness = sdfDesc.narrowBandThicknessRelativeToSdfBoundsDiagonal * edges.magnitude();

	bool success = true; // catch GPU errors.

	if (sdfDesc.sdfBuilder == NULL) 
	{
		PxArray<PxReal> denseSdf;			
		PxArray<PxReal> sparseSdf;
		Gu::SDFUsingWindingNumbersSparse(
			baseMeshSpecified ? verticesPtr : &mesh.m_positions[0],
			baseMeshSpecified ? indices32.begin() : &mesh.m_indices[0],
			baseMeshSpecified ? indices32.size() : mesh.m_indices.size(),
			dx, dy, dz,
			meshLower, meshLower + PxVec3(static_cast<PxReal>(dx), static_cast<PxReal>(dy), static_cast<PxReal>(dz)) * spacing, narrowBandThickness, sdfDesc.subgridSize,
			sdfCoarse, sdfSubgridsStartSlots, sparseSdf, denseSdf, subgridsMinSdfValue, subgridsMaxSdfValue, 16, sdfDesc.sdfBuilder);

		PxArray<PxReal> uncompressedSdfDataSubgrids;
		Gu::convertSparseSDFTo3DTextureLayout(dx, dy, dz, sdfDesc.subgridSize, sdfSubgridsStartSlots.begin(), sparseSdf.begin(), sparseSdf.size(), uncompressedSdfDataSubgrids,
			sdfDesc.sdfSubgrids3DTexBlockDim.x, sdfDesc.sdfSubgrids3DTexBlockDim.y, sdfDesc.sdfSubgrids3DTexBlockDim.z);

		if (sdfDesc.bitsPerSubgridPixel == 4)
		{
			//32bit values are stored as normal floats while 16bit and 8bit values are scaled to 0...1 range and then scaled back to original range
			subgridsMinSdfValue = 0.0f;
			subgridsMaxSdfValue = 1.0f;
		}

		quantizeSparseSDF(sdfDesc.bitsPerSubgridPixel, uncompressedSdfDataSubgrids, sdfDataSubgrids,
			subgridsMinSdfValue, subgridsMaxSdfValue);			
	}
	else
	{
		PxU32* indices = baseMeshSpecified ? indices32.begin() : &mesh.m_indices[0];
		PxU32 numTriangleIndices = baseMeshSpecified ? indices32.size() : mesh.m_indices.size();
		const PxVec3* verts = baseMeshSpecified ? verticesPtr : &mesh.m_positions[0];
		PxU32 numVertices = baseMeshSpecified ? sdfDesc.baseMesh.points.count : mesh.m_positions.size();

		PxArray<PxU32> repairedIndices;
		//Analyze the mesh to catch and fix some special cases
		//There are meshes where every triangle is present once with cw and once with ccw orientation. Try to filter out only one set
		Gu::analyzeAndFixMesh(verts, indices, numTriangleIndices, repairedIndices);
		const PxU32* ind = repairedIndices.size() > 0 ? repairedIndices.begin() : indices;
		if (repairedIndices.size() > 0)
			numTriangleIndices = repairedIndices.size();

		//The GPU SDF builder does sparse SDF, 3d texture layout and quantization in one go to best utilize the gpu
		success = sdfDesc.sdfBuilder->buildSparseSDF(verts,
			numVertices,
			ind,
			numTriangleIndices,
			dx, dy, dz,
			meshLower, meshLower + PxVec3(static_cast<PxReal>(dx), static_cast<PxReal>(dy), static_cast<PxReal>(dz)) * spacing, narrowBandThickness, sdfDesc.subgridSize, sdfDesc.bitsPerSubgridPixel,
			sdfCoarse, sdfSubgridsStartSlots, sdfDataSubgrids, subgridsMinSdfValue, subgridsMaxSdfValue, 
			sdfDesc.sdfSubgrids3DTexBlockDim.x, sdfDesc.sdfSubgrids3DTexBlockDim.y, sdfDesc.sdfSubgrids3DTexBlockDim.z, 0);
	}

	sdfDesc.sdf.count = sdfCoarse.size();
	sdfDesc.sdf.stride = sizeof(PxReal);
	sdfDesc.sdf.data = sdfCoarse.begin();
	sdfDesc.sdfSubgrids.count = sdfDataSubgrids.size();
	sdfDesc.sdfSubgrids.stride = sizeof(PxU8);
	sdfDesc.sdfSubgrids.data = sdfDataSubgrids.begin();
	sdfDesc.sdfStartSlots.count = sdfSubgridsStartSlots.size();
	sdfDesc.sdfStartSlots.stride = sizeof(PxU32);
	sdfDesc.sdfStartSlots.data = sdfSubgridsStartSlots.begin();

	sdfDesc.subgridsMinSdfValue = subgridsMinSdfValue;
	sdfDesc.subgridsMaxSdfValue = subgridsMaxSdfValue;

	return success; // false if we had GPU errors.
}

static bool createSDF(PxTriangleMeshDesc& desc, PxSDFDesc& sdfDesc, PxArray<PxReal>& sdf, PxArray<PxU8>& sdfDataSubgrids, PxArray<PxU32>& sdfSubgridsStartSlots)
{
	if (sdfDesc.subgridSize > 0)
	{
		return createSDFSparse(desc, sdfDesc, sdf, sdfDataSubgrids, sdfSubgridsStartSlots);
	}

	MeshData mesh(desc);

	PxVec3 meshLower, meshUpper;
	if (sdfDesc.sdfBounds.isEmpty())
		mesh.GetBounds(meshLower, meshUpper);
	else
	{
		meshLower = sdfDesc.sdfBounds.minimum;
		meshUpper = sdfDesc.sdfBounds.maximum;
	}

	PxVec3 edges = meshUpper - meshLower;

	const PxReal spacing = sdfDesc.spacing;

	// tweak spacing to avoid edge cases for vertices laying on the boundary
	// just covers the case where an edge is a whole multiple of the spacing.
	PxReal spacingEps = spacing * (1.0f - 1e-4f);

	// make sure to have at least one particle in each dimension
	PxI32 dx, dy, dz;
	dx = spacing > edges.x ? 1 : PxI32(edges.x / spacingEps);
	dy = spacing > edges.y ? 1 : PxI32(edges.y / spacingEps);
	dz = spacing > edges.z ? 1 : PxI32(edges.z / spacingEps);

	dx += 4;
	dy += 4;
	dz += 4;

	const PxU32 numVoxels = dx * dy * dz;

	// we shift the voxelization bounds so that the voxel centers
	// lie symmetrically to the center of the object. this reduces the 
	// chance of missing features, and also better aligns the particles
	// with the mesh
	PxVec3 meshOffset;
	meshOffset.x = 0.5f * (spacing - (edges.x - (dx - 1)*spacing));
	meshOffset.y = 0.5f * (spacing - (edges.y - (dy - 1)*spacing));
	meshOffset.z = 0.5f * (spacing - (edges.z - (dz - 1)*spacing));
	meshLower -= meshOffset;

	sdfDesc.meshLower = meshLower;

	sdfDesc.dims.x = dx;
	sdfDesc.dims.y = dy;
	sdfDesc.dims.z = dz;

	sdf.resize(numVoxels);

	PxArray<PxU32> indices32;
	PxArray<PxVec3> vertices;
	const PxVec3* verticesPtr = NULL;
	bool baseMeshSpecified = sdfDesc.baseMesh.triangles.data && sdfDesc.baseMesh.points.data;
	if (baseMeshSpecified)
	{
		convert16To32Bits(sdfDesc.baseMesh, indices32);

		if (sdfDesc.baseMesh.points.stride != sizeof(PxVec3))
		{
			vertices.resize(sdfDesc.baseMesh.points.count);
			immediateCooking::gatherStrided(sdfDesc.baseMesh.points.data, vertices.begin(), sdfDesc.baseMesh.points.count, sizeof(PxVec3), sdfDesc.baseMesh.points.stride);
			verticesPtr = vertices.begin();
		}
		else
			verticesPtr = reinterpret_cast<const PxVec3*>(sdfDesc.baseMesh.points.data);
	}

	PxU32* indices = baseMeshSpecified ? indices32.begin() : &mesh.m_indices[0];
	PxU32 numTriangleIndices = baseMeshSpecified ? indices32.size() : mesh.m_indices.size();
	const PxVec3* verts = baseMeshSpecified ? verticesPtr : &mesh.m_positions[0];
	PxU32 numVertices = baseMeshSpecified ? sdfDesc.baseMesh.points.count : mesh.m_positions.size();

	if (sdfDesc.sdfBuilder == NULL) 
	{
		Gu::SDFUsingWindingNumbers(verts, indices, numTriangleIndices, dx, dy, dz, &sdf[0], meshLower,
			meshLower + PxVec3(static_cast<PxReal>(dx), static_cast<PxReal>(dy), static_cast<PxReal>(dz)) * spacing, NULL, true,
			sdfDesc.numThreadsForSdfConstruction, sdfDesc.sdfBuilder);
	}
	else
	{
		PxArray<PxU32> repairedIndices;
		//Analyze the mesh to catch and fix some special cases
		//There are meshes where every triangle is present once with cw and once with ccw orientation. Try to filter out only one set
		Gu::analyzeAndFixMesh(verts, indices, numTriangleIndices, repairedIndices);
		const PxU32* ind = repairedIndices.size() > 0 ? repairedIndices.begin() : indices;
		if (repairedIndices.size() > 0)
			numTriangleIndices = repairedIndices.size();

		bool success = sdfDesc.sdfBuilder->buildSDF(verts, numVertices, ind, numTriangleIndices, dx, dy, dz, meshLower,
			meshLower + PxVec3(static_cast<PxReal>(dx), static_cast<PxReal>(dy), static_cast<PxReal>(dz)) * spacing, true, &sdf[0]);

		// return false if the cooking failed.
		if (!success)
			return false;
	}

	sdfDesc.sdf.count = sdfDesc.dims.x * sdfDesc.dims.y * sdfDesc.dims.z;
	sdfDesc.sdf.stride = sizeof(PxReal);
	sdfDesc.sdf.data = &sdf[0];

	return true;
}

bool physx::buildSDF(PxTriangleMeshDesc& desc, PxArray<PxReal>& sdf, PxArray<PxU8>& sdfDataSubgrids, PxArray<PxU32>& sdfSubgridsStartSlots)
{
	PxSDFDesc& sdfDesc = *desc.sdfDesc;

	if (!sdfDesc.sdf.data && sdfDesc.spacing > 0.f)
	{
		// Calculate signed distance field here if no sdf data provided.
		if (!createSDF(desc, sdfDesc, sdf, sdfDataSubgrids, sdfSubgridsStartSlots))
			return false;

		sdfDesc.sdf.stride = sizeof(PxReal);
	}
	return true;
}