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

Fix PLBVH reduction kernels, simplify nodes per level calculation.

Browse Source 
Also calculate index ranges for each internal node.
This commit is contained in:
Jackson Lee
2014-02-19 21:49:30 -08:00
parent 7f0e361fa0
commit e955192971
4 changed files with 193 additions and 175 deletions
Download Patch File Download Diff File Expand all files Collapse all files

114
src/Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvhKernels.h
View File

@@ -75,59 +75,32 @@ static const char* parallelLinearBvhCL= \
"{\n"
" return interleaveBits(x) << 0 | interleaveBits(y) << 1 | interleaveBits(z) << 2;\n"
"}\n"
"__kernel void findAllNodesMergedAabb(__global b3AabbCL* out_mergedAabb, int numAabbs)\n"
"//Should replace with an optimized parallel reduction\n"
"__kernel void findAllNodesMergedAabb(__global b3AabbCL* out_mergedAabb, int numAabbsNeedingMerge)\n"
"{\n"
" int aabbIndex = get_global_id(0);\n"
" if(aabbIndex >= numAabbs) return;\n"
" \n"
" //Find the most significant bit(msb)\n"
" int mostSignificantBit = 0;\n"
" {\n"
" int temp = numAabbs;\n"
" while(temp >>= 1) mostSignificantBit++; //Start counting from 0 (0 and 1 have msb 0, 2 has msb 1)\n"
" }\n"
" \n"
" int numberOfAabbsAboveMsbSplit = numAabbs & ~( ~(0) << mostSignificantBit );\n"
" int numRemainingAabbs = (1 << mostSignificantBit);\n"
" \n"
" //Merge AABBs above most significant bit so that the number of remaining AABBs is a power of 2\n"
" //For example, if there are 159 AABBs = 128 + 31, then merge indices [0, 30] and 128 + [0, 30]\n"
" if(aabbIndex < numberOfAabbsAboveMsbSplit)\n"
" {\n"
" int otherAabbIndex = numRemainingAabbs + aabbIndex;\n"
" \n"
" b3AabbCL aabb = out_mergedAabb[aabbIndex];\n"
" b3AabbCL otherAabb = out_mergedAabb[otherAabbIndex];\n"
" \n"
" b3AabbCL mergedAabb;\n"
" mergedAabb.m_min = b3Min(aabb.m_min, otherAabb.m_min);\n"
" mergedAabb.m_max = b3Max(aabb.m_max, otherAabb.m_max);\n"
" out_mergedAabb[aabbIndex] = mergedAabb;\n"
" }\n"
" \n"
" barrier(CLK_GLOBAL_MEM_FENCE);\n"
" \n"
" //Each time this kernel is added to the command queue, \n"
" //the number of AABBs needing to be merged is halved\n"
" //\n"
" int offset = numRemainingAabbs / 2;\n"
" while(offset >= 1)\n"
" {\n"
" if(aabbIndex < offset)\n"
" {\n"
" int otherAabbIndex = aabbIndex + offset;\n"
" //Example with 159 AABBs:\n"
" // numRemainingAabbs == 159 / 2 + 159 % 2 == 80\n"
" // numMergedAabbs == 159 - 80 == 79\n"
" //So, indices [0, 78] are merged with [0 + 80, 78 + 80]\n"
" \n"
" int numRemainingAabbs = numAabbsNeedingMerge / 2 + numAabbsNeedingMerge % 2;\n"
" int numMergedAabbs = numAabbsNeedingMerge - numRemainingAabbs;\n"
" \n"
" int aabbIndex = get_global_id(0);\n"
" if(aabbIndex >= numMergedAabbs) return;\n"
" \n"
" int otherAabbIndex = aabbIndex + numRemainingAabbs;\n"
" \n"
" b3AabbCL aabb = out_mergedAabb[aabbIndex];\n"
" b3AabbCL otherAabb = out_mergedAabb[otherAabbIndex];\n"
" \n"
" b3AabbCL aabb = out_mergedAabb[aabbIndex];\n"
" b3AabbCL otherAabb = out_mergedAabb[otherAabbIndex];\n"
" \n"
" b3AabbCL mergedAabb;\n"
" mergedAabb.m_min = b3Min(aabb.m_min, otherAabb.m_min);\n"
" mergedAabb.m_max = b3Max(aabb.m_max, otherAabb.m_max);\n"
" out_mergedAabb[aabbIndex] = mergedAabb;\n"
" }\n"
" \n"
" offset /= 2;\n"
" \n"
" barrier(CLK_GLOBAL_MEM_FENCE);\n"
" }\n"
" b3AabbCL mergedAabb;\n"
" mergedAabb.m_min = b3Min(aabb.m_min, otherAabb.m_min);\n"
" mergedAabb.m_max = b3Max(aabb.m_max, otherAabb.m_max);\n"
" out_mergedAabb[aabbIndex] = mergedAabb;\n"
"}\n"
"__kernel void assignMortonCodesAndAabbIndicies(__global b3AabbCL* worldSpaceAabbs, __global b3AabbCL* mergedAabbOfAllNodes, \n"
" __global SortDataCL* out_mortonCodesAndAabbIndices, int numAabbs)\n"
@@ -244,7 +217,7 @@ static const char* parallelLinearBvhCL= \
" {\n"
" int leafNodeLevel = numLevels - 1;\n"
" leftChildIndex = (isLeftChildLeaf) ? leftChildIndex - firstIndexOffsetPerLevel[leafNodeLevel] : leftChildIndex;\n"
" rightChildIndex = (isLeftChildLeaf) ? rightChildIndex - firstIndexOffsetPerLevel[leafNodeLevel] : rightChildIndex;\n"
" rightChildIndex = (isRightChildLeaf) ? rightChildIndex - firstIndexOffsetPerLevel[leafNodeLevel] : rightChildIndex;\n"
" }\n"
" \n"
" //Set the negative sign bit if the node is internal\n"
@@ -265,20 +238,19 @@ static const char* parallelLinearBvhCL= \
" __global int2* internalNodeChildIndices,\n"
" __global SortDataCL* mortonCodesAndAabbIndices,\n"
" __global b3AabbCL* leafNodeAabbs, \n"
" __global b3AabbCL* out_internalNodeAabbs, int numLevels, int numInternalNodes)\n"
" __global int2* out_internalNodeLeafIndexRanges,\n"
" __global b3AabbCL* out_internalNodeAabbs, \n"
" int numLevels, int numInternalNodes, int level)\n"
"{\n"
" int i = get_global_id(0);\n"
" if(i >= numInternalNodes) return;\n"
" \n"
" int numInternalLevels = numLevels - 1;\n"
" \n"
" //Starting from the level next to the leaf nodes, move towards the root(level 0)\n"
" for(int level = numInternalLevels - 1; level >= 0; --level)\n"
" //For each node in a level, check its child nodes to determine its AABB\n"
" {\n"
" int indexInLevel = i; //Index relative to firstIndexOffsetPerLevel[level]\n"
" \n"
" int numNodesInLevel = numNodesPerLevel[level];\n"
" if(i < numNodesInLevel)\n"
" if(indexInLevel < numNodesInLevel)\n"
" {\n"
" int internalNodeIndexGlobal = indexInLevel + firstIndexOffsetPerLevel[level];\n"
" int2 childIndicies = internalNodeChildIndices[internalNodeIndexGlobal];\n"
@@ -289,19 +261,26 @@ static const char* parallelLinearBvhCL= \
" int isLeftChildLeaf = isLeafNode(childIndicies.x);\n"
" int isRightChildLeaf = isLeafNode(childIndicies.y);\n"
" \n"
" //left/RightChildLeafIndex == Rigid body indicies\n"
" int leftChildLeafIndex = (isLeftChildLeaf) ? mortonCodesAndAabbIndices[leftChildIndex].m_value : -1;\n"
" int rightChildLeafIndex = (isRightChildLeaf) ? mortonCodesAndAabbIndices[rightChildIndex].m_value : -1;\n"
" \n"
" b3AabbCL leftChildAabb = (isLeftChildLeaf) ? leafNodeAabbs[leftChildLeafIndex] : out_internalNodeAabbs[leftChildIndex];\n"
" b3AabbCL rightChildAabb = (isRightChildLeaf) ? leafNodeAabbs[rightChildLeafIndex] : out_internalNodeAabbs[rightChildIndex];\n"
" \n"
" //\n"
" b3AabbCL internalNodeAabb;\n"
" internalNodeAabb.m_min = b3Min(leftChildAabb.m_min, rightChildAabb.m_min);\n"
" internalNodeAabb.m_max = b3Max(leftChildAabb.m_max, rightChildAabb.m_max);\n"
" out_internalNodeAabbs[internalNodeIndexGlobal] = internalNodeAabb;\n"
" \n"
" //For index range, x == min and y == max; left child always has lower index\n"
" int2 leafIndexRange;\n"
" leafIndexRange.x = (isLeftChildLeaf) ? leftChildIndex : out_internalNodeLeafIndexRanges[leftChildIndex].x;\n"
" leafIndexRange.y = (isRightChildLeaf) ? rightChildIndex : out_internalNodeLeafIndexRanges[rightChildIndex].y;\n"
" \n"
" out_internalNodeLeafIndexRanges[internalNodeIndexGlobal] = leafIndexRange;\n"
" }\n"
" \n"
" barrier(CLK_GLOBAL_MEM_FENCE);\n"
" }\n"
"}\n"
"//From sap.cl\n"
@@ -316,7 +295,9 @@ static const char* parallelLinearBvhCL= \
"}\n"
"//From sap.cl\n"
"__kernel void plbvhCalculateOverlappingPairs(__global b3AabbCL* rigidAabbs, \n"
" __global int2* internalNodeChildIndices, __global b3AabbCL* internalNodeAabbs,\n"
" __global int2* internalNodeChildIndices, \n"
" __global b3AabbCL* internalNodeAabbs,\n"
" __global int2* internalNodeLeafIndexRanges,\n"
" __global SortDataCL* mortonCodesAndAabbIndices,\n"
" __global int* out_numPairs, __global int4* out_overlappingPairs, \n"
" int maxPairs, int numQueryAabbs)\n"
@@ -326,7 +307,8 @@ static const char* parallelLinearBvhCL= \
" int queryRigidIndex = get_group_id(0) * get_local_size(0) + get_local_id(0);\n"
" if(queryRigidIndex >= numQueryAabbs) return;\n"
" \n"
" queryRigidIndex = mortonCodesAndAabbIndices[queryRigidIndex].m_value;\n"
" int queryBvhNodeIndex = queryRigidIndex;\n"
" queryRigidIndex = mortonCodesAndAabbIndices[queryRigidIndex].m_value; // fix queryRigidIndex naming for this branch\n"
"#else\n"
" int queryRigidIndex = get_global_id(0);\n"
" if(queryRigidIndex >= numQueryAabbs) return;\n"
@@ -347,7 +329,15 @@ static const char* parallelLinearBvhCL= \
" \n"
" int isLeaf = isLeafNode(internalOrLeafNodeIndex); //Internal node if false\n"
" int bvhNodeIndex = getIndexWithInternalNodeMarkerRemoved(internalOrLeafNodeIndex);\n"
" \n"
" \n"
" //Optimization - if the node is not a leaf, check whether the highest leaf index of that node\n"
" //is less than the queried node's index to avoid testing each pair twice.\n"
" {\n"
" // fix: produces duplicate pairs\n"
" // int highestLeafIndex = (isLeaf) ? numQueryAabbs : internalNodeLeafIndexRanges[bvhNodeIndex].y;\n"
" // if(highestLeafIndex < queryBvhNodeIndex) continue;\n"
" }\n"
" \n"
" //bvhRigidIndex is not used if internal node\n"
" int bvhRigidIndex = (isLeaf) ? mortonCodesAndAabbIndices[bvhNodeIndex].m_value : -1;\n"
" \n"