Bart/bullet3
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Accelerate GPU raycaster with PLBVH.

Browse Source
This commit is contained in:
Jackson Lee
2014-02-23 20:40:58 -08:00
parent e955192971
commit e4fbd5332d
10 changed files with 732 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

170
src/Bullet3OpenCL/Raycast/b3GpuRaycast.cpp
View File

@@ -8,6 +8,11 @@
#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
#include "Bullet3OpenCL/ParallelPrimitives/b3FillCL.h"
#include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
#include "Bullet3OpenCL/BroadphaseCollision/b3GpuBroadphaseInterface.h"
#include "Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvh.h"
#include "Bullet3OpenCL/Raycast/kernels/rayCastKernels.h"
@@ -20,7 +25,24 @@ struct b3GpuRaycastInternalData
cl_context m_context;
cl_device_id m_device;
cl_command_queue m_q;
cl_kernel m_raytraceKernel;
cl_kernel m_raytraceKernel;
cl_kernel m_raytracePairsKernel;
cl_kernel m_findRayRigidPairIndexRanges;
b3GpuParallelLinearBvh* m_plbvh;
b3RadixSort32CL* m_radixSorter;
b3FillCL* m_fill;
//1 element per ray
b3OpenCLArray<b3RayInfo>* m_gpuRays;
b3OpenCLArray<b3RayHit>* m_gpuHitResults;
b3OpenCLArray<int>* m_firstRayRigidPairIndexPerRay;
b3OpenCLArray<int>* m_numRayRigidPairsPerRay;
//1 element per (ray index, rigid index) pair
b3OpenCLArray<int>* m_gpuNumRayRigidPairs;
b3OpenCLArray<b3Int2>* m_gpuRayRigidPairs;
int m_test;
};
@@ -31,7 +53,19 @@ b3GpuRaycast::b3GpuRaycast(cl_context ctx,cl_device_id device, cl_command_queue
m_data->m_device = device;
m_data->m_q = q;
m_data->m_raytraceKernel = 0;
m_data->m_raytracePairsKernel = 0;
m_data->m_findRayRigidPairIndexRanges = 0;
m_data->m_plbvh = new b3GpuParallelLinearBvh(ctx, device, q);
m_data->m_radixSorter = new b3RadixSort32CL(ctx, device, q);
m_data->m_fill = new b3FillCL(ctx, device, q);
m_data->m_gpuRays = new b3OpenCLArray<b3RayInfo>(ctx, q);
m_data->m_gpuHitResults = new b3OpenCLArray<b3RayHit>(ctx, q);
m_data->m_firstRayRigidPairIndexPerRay = new b3OpenCLArray<int>(ctx, q);
m_data->m_numRayRigidPairsPerRay = new b3OpenCLArray<int>(ctx, q);
m_data->m_gpuNumRayRigidPairs = new b3OpenCLArray<int>(ctx, q);
m_data->m_gpuRayRigidPairs = new b3OpenCLArray<b3Int2>(ctx, q);
{
cl_int errNum=0;
@@ -39,6 +73,10 @@ b3GpuRaycast::b3GpuRaycast(cl_context ctx,cl_device_id device, cl_command_queue
b3Assert(errNum==CL_SUCCESS);
m_data->m_raytraceKernel = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,rayCastKernelCL, "rayCastKernel",&errNum,prog);
b3Assert(errNum==CL_SUCCESS);
m_data->m_raytracePairsKernel = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,rayCastKernelCL, "rayCastPairsKernel",&errNum,prog);
b3Assert(errNum==CL_SUCCESS);
m_data->m_findRayRigidPairIndexRanges = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,rayCastKernelCL, "findRayRigidPairIndexRanges",&errNum,prog);
b3Assert(errNum==CL_SUCCESS);
clReleaseProgram(prog);
}
@@ -48,6 +86,20 @@ b3GpuRaycast::b3GpuRaycast(cl_context ctx,cl_device_id device, cl_command_queue
b3GpuRaycast::~b3GpuRaycast()
{
clReleaseKernel(m_data->m_raytraceKernel);
clReleaseKernel(m_data->m_raytracePairsKernel);
clReleaseKernel(m_data->m_findRayRigidPairIndexRanges);
delete m_data->m_plbvh;
delete m_data->m_radixSorter;
delete m_data->m_fill;
delete m_data->m_gpuRays;
delete m_data->m_gpuHitResults;
delete m_data->m_firstRayRigidPairIndexPerRay;
delete m_data->m_numRayRigidPairsPerRay;
delete m_data->m_gpuNumRayRigidPairs;
delete m_data->m_gpuRayRigidPairs;
delete m_data;
}
@@ -206,27 +258,32 @@ void b3GpuRaycast::castRaysHost(const b3AlignedObjectArray<b3RayInfo>& rays, b3A
}
///todo: add some acceleration structure (AABBs, tree etc)
void b3GpuRaycast::castRays(const b3AlignedObjectArray<b3RayInfo>& rays, b3AlignedObjectArray<b3RayHit>& hitResults,
int numBodies,const struct b3RigidBodyData* bodies, int numCollidables, const struct b3Collidable* collidables, const struct b3GpuNarrowPhaseInternalData* narrowphaseData)
int numBodies,const struct b3RigidBodyData* bodies, int numCollidables, const struct b3Collidable* collidables,
const struct b3GpuNarrowPhaseInternalData* narrowphaseData, class b3GpuBroadphaseInterface* broadphase)
{
//castRaysHost(rays,hitResults,numBodies,bodies,numCollidables,collidables,narrowphaseData);
B3_PROFILE("castRaysGPU");
b3OpenCLArray<b3RayInfo> gpuRays(m_data->m_context,m_data->m_q);
b3OpenCLArray<b3RayHit> gpuHitResults(m_data->m_context,m_data->m_q);
{
B3_PROFILE("raycast copyFromHost");
gpuRays.copyFromHost(rays);
gpuHitResults.resize(hitResults.size());
gpuHitResults.copyFromHost(hitResults);
m_data->m_gpuRays->copyFromHost(rays);
m_data->m_gpuHitResults->copyFromHost(hitResults);
}
int numRays = hitResults.size();
{
m_data->m_firstRayRigidPairIndexPerRay->resize(numRays);
m_data->m_numRayRigidPairsPerRay->resize(numRays);
m_data->m_gpuNumRayRigidPairs->resize(1);
m_data->m_gpuRayRigidPairs->resize(numRays * 16);
}
//run kernel
const bool USE_BRUTE_FORCE_RAYCAST = false;
if(USE_BRUTE_FORCE_RAYCAST)
{
B3_PROFILE("raycast launch1D");
@@ -234,8 +291,8 @@ void b3GpuRaycast::castRays(const b3AlignedObjectArray<b3RayInfo>& rays, b3Align
int numRays = rays.size();
launcher.setConst(numRays);
launcher.setBuffer(gpuRays.getBufferCL());
launcher.setBuffer(gpuHitResults.getBufferCL());
launcher.setBuffer(m_data->m_gpuRays->getBufferCL());
launcher.setBuffer(m_data->m_gpuHitResults->getBufferCL());
launcher.setConst(numBodies);
launcher.setBuffer(narrowphaseData->m_bodyBufferGPU->getBufferCL());
@@ -246,11 +303,90 @@ void b3GpuRaycast::castRays(const b3AlignedObjectArray<b3RayInfo>& rays, b3Align
launcher.launch1D(numRays);
clFinish(m_data->m_q);
}
else
{
//printf("broadphase->getAllAabbsGPU().size(): %d \n", broadphase->getAllAabbsGPU().size());
m_data->m_plbvh->build( broadphase->getAllAabbsGPU() );
m_data->m_plbvh->testRaysAgainstBvhAabbs(*m_data->m_gpuRays, *m_data->m_gpuNumRayRigidPairs, *m_data->m_gpuRayRigidPairs);
int numRayRigidPairs = -1;
m_data->m_gpuNumRayRigidPairs->copyToHostPointer(&numRayRigidPairs, 1);
if( numRayRigidPairs > m_data->m_gpuRayRigidPairs->size() )
{
numRayRigidPairs = m_data->m_gpuRayRigidPairs->size();
m_data->m_gpuNumRayRigidPairs->copyFromHostPointer(&numRayRigidPairs, 1);
}
m_data->m_gpuRayRigidPairs->resize(numRayRigidPairs); //Radix sort needs b3OpenCLArray::size() to be correct
//Sort ray-rigid pairs by ray index
{
B3_PROFILE("sort ray-rigid pairs");
m_data->m_radixSorter->execute( *reinterpret_cast< b3OpenCLArray<b3SortData>* >(m_data->m_gpuRayRigidPairs) );
}
//detect start,count of each ray pair
{
B3_PROFILE("detect ray-rigid pair index ranges");
{
B3_PROFILE("reset ray-rigid pair index ranges");
m_data->m_fill->execute(*m_data->m_firstRayRigidPairIndexPerRay, numRayRigidPairs, numRays); //atomic_min used to find first index
m_data->m_fill->execute(*m_data->m_numRayRigidPairsPerRay, 0, numRays);
clFinish(m_data->m_q);
}
b3BufferInfoCL bufferInfo[] =
{
b3BufferInfoCL( m_data->m_gpuRayRigidPairs->getBufferCL() ),
b3BufferInfoCL( m_data->m_firstRayRigidPairIndexPerRay->getBufferCL() ),
b3BufferInfoCL( m_data->m_numRayRigidPairsPerRay->getBufferCL() )
};
b3LauncherCL launcher(m_data->m_q, m_data->m_findRayRigidPairIndexRanges, "m_findRayRigidPairIndexRanges");
launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
launcher.setConst(numRayRigidPairs);
launcher.launch1D(numRayRigidPairs);
clFinish(m_data->m_q);
}
{
B3_PROFILE("ray-rigid intersection");
b3BufferInfoCL bufferInfo[] =
{
b3BufferInfoCL( m_data->m_gpuRays->getBufferCL() ),
b3BufferInfoCL( m_data->m_gpuHitResults->getBufferCL() ),
b3BufferInfoCL( m_data->m_firstRayRigidPairIndexPerRay->getBufferCL() ),
b3BufferInfoCL( m_data->m_numRayRigidPairsPerRay->getBufferCL() ),
b3BufferInfoCL( narrowphaseData->m_bodyBufferGPU->getBufferCL() ),
b3BufferInfoCL( narrowphaseData->m_collidablesGPU->getBufferCL() ),
b3BufferInfoCL( narrowphaseData->m_convexFacesGPU->getBufferCL() ),
b3BufferInfoCL( narrowphaseData->m_convexPolyhedraGPU->getBufferCL() ),
b3BufferInfoCL( m_data->m_gpuRayRigidPairs->getBufferCL() )
};
b3LauncherCL launcher(m_data->m_q, m_data->m_raytracePairsKernel, "m_raytracePairsKernel");
launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
launcher.setConst(numRays);
launcher.launch1D(numRays);
clFinish(m_data->m_q);
}
}
//copy results
{
B3_PROFILE("raycast copyToHost");
gpuHitResults.copyToHost(hitResults);
m_data->m_gpuHitResults->copyToHost(hitResults);
}
}

3
src/Bullet3OpenCL/Raycast/b3GpuRaycast.h
View File

@@ -23,8 +23,7 @@ public:
void castRays(const b3AlignedObjectArray<b3RayInfo>& rays, b3AlignedObjectArray<b3RayHit>& hitResults,
int numBodies,const struct b3RigidBodyData* bodies, int numCollidables, const struct b3Collidable* collidables,
const struct b3GpuNarrowPhaseInternalData* narrowphaseData
);
const struct b3GpuNarrowPhaseInternalData* narrowphaseData, class b3GpuBroadphaseInterface* broadphase);

100
src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl
View File

@@ -337,3 +337,103 @@ __kernel void rayCastKernel(
}
}
__kernel void findRayRigidPairIndexRanges(__global int2* rayRigidPairs,
__global int* out_firstRayRigidPairIndexPerRay,
__global int* out_numRayRigidPairsPerRay,
int numRayRigidPairs)
{
int rayRigidPairIndex = get_global_id(0);
if (rayRigidPairIndex >= numRayRigidPairs) return;
int rayIndex = rayRigidPairs[rayRigidPairIndex].x;
atomic_min(&out_firstRayRigidPairIndexPerRay[rayIndex], rayRigidPairIndex);
atomic_inc(&out_numRayRigidPairsPerRay[rayIndex]);
}
__kernel void rayCastPairsKernel(const __global b3RayInfo* rays,
__global b3RayHit* hitResults,
__global int* firstRayRigidPairIndexPerRay,
__global int* numRayRigidPairsPerRay,
__global Body* bodies,
__global Collidable* collidables,
__global const b3GpuFace* faces,
__global const ConvexPolyhedronCL* convexShapes,
__global int2* rayRigidPairs,
int numRays)
{
int i = get_global_id(0);
if (i >= numRays) return;
float4 rayFrom = rays[i].m_from;
float4 rayTo = rays[i].m_to;
hitResults[i].m_hitFraction = 1.f;
float hitFraction = 1.f;
float4 hitPoint;
float4 hitNormal;
int hitBodyIndex = -1;
//
for(int pair = 0; pair < numRayRigidPairsPerRay[i]; ++pair)
{
int rayRigidPairIndex = pair + firstRayRigidPairIndexPerRay[i];
int b = rayRigidPairs[rayRigidPairIndex].y;
if (hitResults[i].m_hitResult2 == b) continue;
Body body = bodies[b];
Collidable rigidCollidable = collidables[body.m_collidableIdx];
float4 pos = body.m_pos;
float4 orn = body.m_quat;
if (rigidCollidable.m_shapeType == SHAPE_CONVEX_HULL)
{
float4 invPos = (float4)(0,0,0,0);
float4 invOrn = (float4)(0,0,0,0);
float4 rayFromLocal = (float4)(0,0,0,0);
float4 rayToLocal = (float4)(0,0,0,0);
invOrn = qtInvert(orn);
invPos = qtRotate(invOrn, -pos);
rayFromLocal = qtRotate( invOrn, rayFrom ) + invPos;
rayToLocal = qtRotate( invOrn, rayTo) + invPos;
rayFromLocal.w = 0.f;
rayToLocal.w = 0.f;
int numFaces = convexShapes[rigidCollidable.m_shapeIndex].m_numFaces;
int faceOffset = convexShapes[rigidCollidable.m_shapeIndex].m_faceOffset;
if (numFaces && rayConvex(rayFromLocal, rayToLocal, numFaces, faceOffset,faces, &hitFraction, &hitNormal))
{
hitBodyIndex = b;
hitPoint = setInterpolate3(rayFrom, rayTo, hitFraction);
}
}
if (rigidCollidable.m_shapeType == SHAPE_SPHERE)
{
float radius = rigidCollidable.m_radius;
if (sphere_intersect(pos, radius, rayFrom, rayTo, &hitFraction))
{
hitBodyIndex = b;
hitPoint = setInterpolate3(rayFrom, rayTo, hitFraction);
hitNormal = (float4) (hitPoint - bodies[b].m_pos);
}
}
}
if (hitBodyIndex >= 0)
{
hitResults[i].m_hitFraction = hitFraction;
hitResults[i].m_hitPoint = hitPoint;
hitResults[i].m_hitNormal = normalize(hitNormal);
hitResults[i].m_hitResult0 = hitBodyIndex;
}
}

97
src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.h
View File

@@ -281,4 +281,101 @@ static const char* rayCastKernelCL= \
" hitResults[i].m_hitResult0 = hitBodyIndex;\n"
" }\n"
"}\n"
"__kernel void findRayRigidPairIndexRanges(__global int2* rayRigidPairs, \n"
" __global int* out_firstRayRigidPairIndexPerRay,\n"
" __global int* out_numRayRigidPairsPerRay,\n"
" int numRayRigidPairs)\n"
"{\n"
" int rayRigidPairIndex = get_global_id(0);\n"
" if (rayRigidPairIndex >= numRayRigidPairs) return;\n"
" \n"
" int rayIndex = rayRigidPairs[rayRigidPairIndex].x;\n"
" \n"
" atomic_min(&out_firstRayRigidPairIndexPerRay[rayIndex], rayRigidPairIndex);\n"
" atomic_inc(&out_numRayRigidPairsPerRay[rayIndex]);\n"
"}\n"
"__kernel void rayCastPairsKernel(const __global b3RayInfo* rays, \n"
" __global b3RayHit* hitResults, \n"
" __global int* firstRayRigidPairIndexPerRay,\n"
" __global int* numRayRigidPairsPerRay,\n"
" \n"
" __global Body* bodies,\n"
" __global Collidable* collidables,\n"
" __global const b3GpuFace* faces,\n"
" __global const ConvexPolyhedronCL* convexShapes,\n"
" \n"
" __global int2* rayRigidPairs,\n"
" int numRays)\n"
"{\n"
" int i = get_global_id(0);\n"
" if (i >= numRays) return;\n"
" \n"
" float4 rayFrom = rays[i].m_from;\n"
" float4 rayTo = rays[i].m_to;\n"
" \n"
" hitResults[i].m_hitFraction = 1.f;\n"
" \n"
" float hitFraction = 1.f;\n"
" float4 hitPoint;\n"
" float4 hitNormal;\n"
" int hitBodyIndex = -1;\n"
" \n"
" //\n"
" for(int pair = 0; pair < numRayRigidPairsPerRay[i]; ++pair)\n"
" {\n"
" int rayRigidPairIndex = pair + firstRayRigidPairIndexPerRay[i];\n"
" int b = rayRigidPairs[rayRigidPairIndex].y;\n"
" \n"
" if (hitResults[i].m_hitResult2 == b) continue;\n"
" \n"
" Body body = bodies[b];\n"
" Collidable rigidCollidable = collidables[body.m_collidableIdx];\n"
" \n"
" float4 pos = body.m_pos;\n"
" float4 orn = body.m_quat;\n"
" \n"
" if (rigidCollidable.m_shapeType == SHAPE_CONVEX_HULL)\n"
" {\n"
" float4 invPos = (float4)(0,0,0,0);\n"
" float4 invOrn = (float4)(0,0,0,0);\n"
" float4 rayFromLocal = (float4)(0,0,0,0);\n"
" float4 rayToLocal = (float4)(0,0,0,0);\n"
" invOrn = qtInvert(orn);\n"
" invPos = qtRotate(invOrn, -pos);\n"
" rayFromLocal = qtRotate( invOrn, rayFrom ) + invPos;\n"
" rayToLocal = qtRotate( invOrn, rayTo) + invPos;\n"
" rayFromLocal.w = 0.f;\n"
" rayToLocal.w = 0.f;\n"
" int numFaces = convexShapes[rigidCollidable.m_shapeIndex].m_numFaces;\n"
" int faceOffset = convexShapes[rigidCollidable.m_shapeIndex].m_faceOffset;\n"
" \n"
" if (numFaces && rayConvex(rayFromLocal, rayToLocal, numFaces, faceOffset,faces, &hitFraction, &hitNormal))\n"
" {\n"
" hitBodyIndex = b;\n"
" hitPoint = setInterpolate3(rayFrom, rayTo, hitFraction);\n"
" }\n"
" }\n"
" \n"
" if (rigidCollidable.m_shapeType == SHAPE_SPHERE)\n"
" {\n"
" float radius = rigidCollidable.m_radius;\n"
" \n"
" if (sphere_intersect(pos, radius, rayFrom, rayTo, &hitFraction))\n"
" {\n"
" hitBodyIndex = b;\n"
" hitPoint = setInterpolate3(rayFrom, rayTo, hitFraction);\n"
" hitNormal = (float4) (hitPoint - bodies[b].m_pos);\n"
" }\n"
" }\n"
" }\n"
" \n"
" if (hitBodyIndex >= 0)\n"
" {\n"
" hitResults[i].m_hitFraction = hitFraction;\n"
" hitResults[i].m_hitPoint = hitPoint;\n"
" hitResults[i].m_hitNormal = normalize(hitNormal);\n"
" hitResults[i].m_hitResult0 = hitBodyIndex;\n"
" }\n"
" \n"
"}\n"
;