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

make bvh kernel work on AMD GPU (and reported an OpenCL compiler bug)

Browse Source 
enable source-level OpenCL debugging of a single kernel
This commit is contained in:
erwin coumans
2013-03-20 13:54:00 -07:00
parent 39884e4665
commit 9a693fb850
6 changed files with 192 additions and 171 deletions
Download Patch File Download Diff File Expand all files Collapse all files

167
opencl/gpu_sat/kernels/bvhTraversal.h
View File

@@ -49,7 +49,7 @@ static const char* bvhTraversalKernelCL= \
" }\n"
"*/\n"
"\n"
"int getTriangleIndex(__global const btQuantizedBvhNode* rootNode)\n"
"int getTriangleIndex(const btQuantizedBvhNode* rootNode)\n"
"{\n"
" unsigned int x=0;\n"
" unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);\n"
@@ -57,13 +57,13 @@ static const char* bvhTraversalKernelCL= \
" return (rootNode->m_escapeIndexOrTriangleIndex&~(y));\n"
"}\n"
"\n"
"bool isLeaf(__global const btQuantizedBvhNode* rootNode)\n"
"int isLeaf(const btQuantizedBvhNode* rootNode)\n"
"{\n"
" //skipindex is negative (internal node), triangleindex >=0 (leafnode)\n"
" return (rootNode->m_escapeIndexOrTriangleIndex >= 0);\n"
" return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;\n"
"}\n"
" \n"
"int getEscapeIndex(__global const btQuantizedBvhNode* rootNode)\n"
"int getEscapeIndex(const btQuantizedBvhNode* rootNode)\n"
"{\n"
" return -rootNode->m_escapeIndexOrTriangleIndex;\n"
"}\n"
@@ -131,13 +131,30 @@ static const char* bvhTraversalKernelCL= \
"} btAabbCL;\n"
"\n"
"\n"
"bool testQuantizedAabbAgainstQuantizedAabb(__private const unsigned short int* aabbMin1,__private const unsigned short int* aabbMax1,__global const unsigned short int* aabbMin2,__global const unsigned short int* aabbMax2)\n"
"int testQuantizedAabbAgainstQuantizedAabb(\n"
" const unsigned short int* aabbMin1,\n"
" const unsigned short int* aabbMax1,\n"
" const unsigned short int* aabbMin2,\n"
" const unsigned short int* aabbMax2)\n"
"{\n"
" bool overlap = true;\n"
" overlap = (aabbMin1[0] > aabbMax2[0] || aabbMax1[0] < aabbMin2[0]) ? false : overlap;\n"
" overlap = (aabbMin1[2] > aabbMax2[2] || aabbMax1[2] < aabbMin2[2]) ? false : overlap;\n"
" overlap = (aabbMin1[1] > aabbMax2[1] || aabbMax1[1] < aabbMin2[1]) ? false : overlap;\n"
" return overlap;\n"
" //int overlap = 1;\n"
" if (aabbMin1[0] > aabbMax2[0])\n"
" return 0;\n"
" if (aabbMax1[0] < aabbMin2[0])\n"
" return 0;\n"
" if (aabbMin1[1] > aabbMax2[1])\n"
" return 0;\n"
" if (aabbMax1[1] < aabbMin2[1])\n"
" return 0;\n"
" if (aabbMin1[2] > aabbMax2[2])\n"
" return 0;\n"
" if (aabbMax1[2] < aabbMin2[2])\n"
" return 0;\n"
" return 1;\n"
" //overlap = ((aabbMin1[0] > aabbMax2[0]) || (aabbMax1[0] < aabbMin2[0])) ? 0 : overlap;\n"
" //overlap = ((aabbMin1[2] > aabbMax2[2]) || (aabbMax1[2] < aabbMin2[2])) ? 0 : overlap;\n"
" //overlap = ((aabbMin1[1] > aabbMax2[1]) || (aabbMax1[1] < aabbMin2[1])) ? 0 : overlap;\n"
" //return overlap;\n"
"}\n"
"\n"
"\n"
@@ -178,88 +195,78 @@ static const char* bvhTraversalKernelCL= \
" int numPairs,\n"
" int maxNumConcavePairsCapacity)\n"
"{\n"
"\n"
" int i = get_global_id(0);\n"
" int id = get_global_id(0);\n"
" if (id>=numPairs)\n"
" return;\n"
" \n"
" if (i<numPairs)\n"
" int bodyIndexA = pairs[id].x;\n"
" int bodyIndexB = pairs[id].y;\n"
" int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
" int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
" \n"
" //once the broadphase avoids static-static pairs, we can remove this test\n"
" if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))\n"
" {\n"
" return;\n"
" }\n"
" \n"
" if (collidables[collidableIndexA].m_shapeType!=SHAPE_CONCAVE_TRIMESH)\n"
" return;\n"
"\n"
" if (collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL)\n"
" return;\n"
"\n"
" unsigned short int quantizedQueryAabbMin[3];\n"
" unsigned short int quantizedQueryAabbMax[3];\n"
" quantizeWithClamp(quantizedQueryAabbMin,aabbs[bodyIndexB].m_min,false,bvhAabbMin, bvhAabbMax,bvhQuantization);\n"
" quantizeWithClamp(quantizedQueryAabbMax,aabbs[bodyIndexB].m_max,true ,bvhAabbMin, bvhAabbMax,bvhQuantization);\n"
" \n"
" int bodyIndexA = pairs[i].x;\n"
" int bodyIndexB = pairs[i].y;\n"
"\n"
" int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
" int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
" \n"
" int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
" int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
" \n"
" \n"
" //once the broadphase avoids static-static pairs, we can remove this test\n"
" if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))\n"
" for (int i=0;i<numSubtreeHeaders;i++)\n"
" {\n"
" btBvhSubtreeInfo subtree = subtreeHeaders[i];\n"
" \n"
" int overlap = testQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,subtree.m_quantizedAabbMin,subtree.m_quantizedAabbMax);\n"
" if (overlap != 0)\n"
" {\n"
" return;\n"
" }\n"
" \n"
" if ((collidables[collidableIndexA].m_shapeType==SHAPE_CONCAVE_TRIMESH))// && (collidables[collidableIndexB].m_shapeType==SHAPE_CONVEX_HULL))\n"
" {\n"
"\n"
" \n"
" unsigned short int quantizedQueryAabbMin[3];\n"
" unsigned short int quantizedQueryAabbMax[3];\n"
" quantizeWithClamp(quantizedQueryAabbMin,aabbs[bodyIndexB].m_min,false,bvhAabbMin, bvhAabbMax,bvhQuantization);\n"
" quantizeWithClamp(quantizedQueryAabbMax,aabbs[bodyIndexB].m_max,true ,bvhAabbMin, bvhAabbMax,bvhQuantization);\n"
"\n"
"\n"
" int i;\n"
" for (i=0;i<numSubtreeHeaders;i++)\n"
" int startNodeIndex = subtree.m_rootNodeIndex;\n"
" int endNodeIndex = subtree.m_rootNodeIndex+subtree.m_subtreeSize;\n"
" int curIndex = startNodeIndex;\n"
" int escapeIndex;\n"
" int isLeafNode;\n"
" int aabbOverlap;\n"
" while (curIndex < endNodeIndex)\n"
" {\n"
" const __global btBvhSubtreeInfo* subtree = &subtreeHeaders[i];\n"
" //PCK: unsigned instead of bool\n"
" unsigned overlap = testQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,subtree->m_quantizedAabbMin,subtree->m_quantizedAabbMax);\n"
" if (overlap != 0)\n"
" btQuantizedBvhNode rootNode = quantizedNodes[curIndex];\n"
" aabbOverlap = testQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,rootNode.m_quantizedAabbMin,rootNode.m_quantizedAabbMax);\n"
" isLeafNode = isLeaf(&rootNode);\n"
" if (aabbOverlap)\n"
" {\n"
" int startNodeIndex = subtree->m_rootNodeIndex;\n"
" int endNodeIndex = subtree->m_rootNodeIndex+subtree->m_subtreeSize;\n"
"\n"
" int curIndex = startNodeIndex;\n"
" int subTreeSize = endNodeIndex - startNodeIndex;\n"
" __global const btQuantizedBvhNode* rootNode = &quantizedNodes[startNodeIndex];\n"
" int escapeIndex;\n"
" bool isLeafNode;\n"
" unsigned aabbOverlap;\n"
" while (curIndex < endNodeIndex)\n"
" if (isLeafNode)\n"
" {\n"
" aabbOverlap = testQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,rootNode->m_quantizedAabbMin,rootNode->m_quantizedAabbMax);\n"
" isLeafNode = isLeaf(rootNode);\n"
" if (isLeafNode && aabbOverlap)\n"
" int triangleIndex = getTriangleIndex(&rootNode);\n"
" \n"
" int pairIdx = atomic_inc(numConcavePairsOut);\n"
" if (pairIdx<maxNumConcavePairsCapacity)\n"
" {\n"
" //do your thing! nodeCallback->processNode(rootNode->getPartId(),rootNode->getTriangleIndex());\n"
" int triangleIndex = getTriangleIndex(rootNode);\n"
" int pairIdx = atomic_inc(numConcavePairsOut);\n"
" if (pairIdx<maxNumConcavePairsCapacity)\n"
" {\n"
" //int4 newPair;\n"
" concavePairsOut[pairIdx].x = bodyIndexA;\n"
" concavePairsOut[pairIdx].y = bodyIndexB;\n"
" concavePairsOut[pairIdx].z = triangleIndex;\n"
" concavePairsOut[pairIdx].w = 3;\n"
" }\n"
" } \n"
" if ((aabbOverlap != 0) || isLeafNode)\n"
" {\n"
" rootNode++;\n"
" curIndex++;\n"
" } else\n"
" {\n"
" escapeIndex = getEscapeIndex(rootNode);\n"
" rootNode += escapeIndex;\n"
" curIndex += escapeIndex;\n"
" int4 newPair = (int4)(bodyIndexA,bodyIndexB,triangleIndex,3);\n"
" concavePairsOut[pairIdx] = newPair;\n"
" }\n"
" } \n"
" curIndex++;\n"
" } else\n"
" {\n"
" if (isLeafNode)\n"
" {\n"
" curIndex++;\n"
" } else\n"
" {\n"
" escapeIndex = getEscapeIndex(&rootNode);\n"
" curIndex += escapeIndex;\n"
" }\n"
" }\n"
" }\n"
" }//SHAPE_CONCAVE_TRIMESH\n"
" }//i<numpairs\n"
" }\n"
" }\n"
"\n"
"}\n"
;