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

Experimenting with GUI in PairBench, still preliminary. Goal is to make it easier to add variables that can be tuned using GUI/gwen

Browse Source 
Pre-compile PairBench OpenCL kernel
This commit is contained in:
erwincoumans
2014-01-26 11:27:20 -08:00
parent ec4292ad75
commit 7cdda65822
15 changed files with 567 additions and 534 deletions
Download Patch File Download Diff File Expand all files Collapse all files

190
src/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphase.cl
View File

@@ -103,11 +103,10 @@ void findPairsInCell( int numObjects,
__global int2* pHash,
__global int* pCellStart,
__global float4* pAABB,
__global int* pPairBuff,
__global int2* pPairBuffStartCurr,
__global float4* pParams,
volatile __global int* pairCount,
__global int4* pPairBuff2
__global int4* pPairBuff2,
int maxPairs
)
{
int4 pGridDim = *((__global int4*)(pParams + 1));
@@ -125,11 +124,7 @@ void findPairsInCell( int numObjects,
float4 min0 = pAABB[unsorted_indx*2 + 0];
float4 max0 = pAABB[unsorted_indx*2 + 1];
int handleIndex = as_int(min0.w);
int2 start_curr = pPairBuffStartCurr[handleIndex];
int start = start_curr.x;
int curr = start_curr.y;
int2 start_curr_next = pPairBuffStartCurr[handleIndex+1];
int curr_max = start_curr_next.x - start - 1;
int bucketEnd = bucketStart + maxBodiesPerCell;
bucketEnd = (bucketEnd > numObjects) ? numObjects : bucketEnd;
for(int index2 = bucketStart; index2 < bucketEnd; index2++)
@@ -153,59 +148,31 @@ void findPairsInCell( int numObjects,
if (handleIndex<handleIndex2)
{
int curPair = atomic_add(pairCount,1);
int4 newpair;
newpair.x = handleIndex;
newpair.y = handleIndex2;
newpair.z = -1;
newpair.w = -1;
pPairBuff2[curPair] = newpair;
if (curPair<maxPairs)
{
int4 newpair;
newpair.x = handleIndex;
newpair.y = handleIndex2;
newpair.z = -1;
newpair.w = -1;
pPairBuff2[curPair] = newpair;
}
}
} else
{
int handleIndex2 = as_int(min1.w);
int k;
for(k = 0; k < curr; k++)
{
int old_pair = pPairBuff[start+k] & (~0x60000000);
if(old_pair == handleIndex2)
{
pPairBuff[start+k] |= 0x40000000;
break;
}
}
if(k == curr)
{
if(curr >= curr_max)
{ // not a good solution, but let's avoid crash
break;
}
pPairBuff[start+curr] = handleIndex2 | 0x20000000;
curr++;
}
}
}
}
}
if (!pairCount)
{
int2 newStartCurr;
newStartCurr.x = start;
newStartCurr.y = curr;
pPairBuffStartCurr[handleIndex] = newStartCurr;
}
}
__kernel void kFindOverlappingPairs( int numObjects,
__global float4* pAABB,
__global int2* pHash,
__global int* pCellStart,
__global int* pPairBuff,
__global int2* pPairBuffStartCurr,
__global float4* pParams ,
volatile __global int* pairCount,
__global int4* pPairBuff2
__global int4* pPairBuff2,
int maxPairs
)
{
@@ -235,140 +202,13 @@ __kernel void kFindOverlappingPairs( int numObjects,
for(int x=-1; x<=1; x++)
{
gridPosB.x = gridPosA.x + x;
findPairsInCell(numObjects, gridPosB, index, pHash, pCellStart, pAABB, pPairBuff, pPairBuffStartCurr, pParams, pairCount,pPairBuff2);
findPairsInCell(numObjects, gridPosB, index, pHash, pCellStart, pAABB, pParams, pairCount,pPairBuff2, maxPairs);
}
}
}
}
__kernel void kFindPairsLarge( int numObjects,
__global float4* pAABB,
__global int2* pHash,
__global int* pCellStart,
__global int* pPairBuff,
__global int2* pPairBuffStartCurr,
uint numLarge )
{
int index = get_global_id(0);
if(index >= numObjects)
{
return;
}
int2 sortedData = pHash[index];
int unsorted_indx = sortedData.y;
float4 min0 = pAABB[unsorted_indx*2 + 0];
float4 max0 = pAABB[unsorted_indx*2 + 1];
int handleIndex = as_int(min0.w);
int2 start_curr = pPairBuffStartCurr[handleIndex];
int start = start_curr.x;
int curr = start_curr.y;
int2 start_curr_next = pPairBuffStartCurr[handleIndex+1];
int curr_max = start_curr_next.x - start - 1;
for(uint i = 0; i < numLarge; i++)
{
int indx2 = numObjects + i;
float4 min1 = pAABB[indx2*2 + 0];
float4 max1 = pAABB[indx2*2 + 1];
if(testAABBOverlap(min0, max0, min1, max1))
{
int k;
int handleIndex2 = as_int(min1.w);
for(k = 0; k < curr; k++)
{
int old_pair = pPairBuff[start+k] & (~0x60000000);
if(old_pair == handleIndex2)
{
pPairBuff[start+k] |= 0x40000000;
break;
}
}
if(k == curr)
{
pPairBuff[start+curr] = handleIndex2 | 0x20000000;
if(curr >= curr_max)
{ // not a good solution, but let's avoid crash
break;
}
curr++;
}
}
}
int2 newStartCurr;
newStartCurr.x = start;
newStartCurr.y = curr;
pPairBuffStartCurr[handleIndex] = newStartCurr;
return;
}
__kernel void kComputePairCacheChanges( int numObjects,
__global int* pPairBuff,
__global int2* pPairBuffStartCurr,
__global int* pPairScan,
__global float4* pAABB )
{
int index = get_global_id(0);
if(index >= numObjects)
{
return;
}
float4 bbMin = pAABB[index * 2];
int handleIndex = as_int(bbMin.w);
int2 start_curr = pPairBuffStartCurr[handleIndex];
int start = start_curr.x;
int curr = start_curr.y;
__global int *pInp = pPairBuff + start;
int num_changes = 0;
for(int k = 0; k < curr; k++, pInp++)
{
if(!((*pInp) & 0x40000000))
{
num_changes++;
}
}
pPairScan[index+1] = num_changes;
}
__kernel void kSqueezeOverlappingPairBuff( int numObjects,
__global int* pPairBuff,
__global int2* pPairBuffStartCurr,
__global int* pPairScan,
__global int* pPairOut,
__global float4* pAABB )
{
int index = get_global_id(0);
if(index >= numObjects)
{
return;
}
float4 bbMin = pAABB[index * 2];
int handleIndex = as_int(bbMin.w);
int2 start_curr = pPairBuffStartCurr[handleIndex];
int start = start_curr.x;
int curr = start_curr.y;
__global int* pInp = pPairBuff + start;
__global int* pOut = pPairOut + pPairScan[index+1];
__global int* pOut2 = pInp;
int num = 0;
for(int k = 0; k < curr; k++, pInp++)
{
if(!((*pInp) & 0x40000000))
{
*pOut = *pInp;
pOut++;
}
if((*pInp) & 0x60000000)
{
*pOut2 = (*pInp) & (~0x60000000);
pOut2++;
num++;
}
}
int2 newStartCurr;
newStartCurr.x = start;
newStartCurr.y = num;
pPairBuffStartCurr[handleIndex] = newStartCurr;
}

188
src/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphaseKernels.h
View File

@@ -92,11 +92,10 @@ static const char* gridBroadphaseCL= \
" __global int2* pHash,\n"
" __global int* pCellStart,\n"
" __global float4* pAABB, \n"
" __global int* pPairBuff,\n"
" __global int2* pPairBuffStartCurr,\n"
" __global float4* pParams,\n"
" volatile __global int* pairCount,\n"
" __global int4* pPairBuff2\n"
" __global int4* pPairBuff2,\n"
" int maxPairs\n"
" )\n"
"{\n"
" int4 pGridDim = *((__global int4*)(pParams + 1));\n"
@@ -114,11 +113,7 @@ static const char* gridBroadphaseCL= \
" float4 min0 = pAABB[unsorted_indx*2 + 0]; \n"
" float4 max0 = pAABB[unsorted_indx*2 + 1];\n"
" int handleIndex = as_int(min0.w);\n"
" int2 start_curr = pPairBuffStartCurr[handleIndex];\n"
" int start = start_curr.x;\n"
" int curr = start_curr.y;\n"
" int2 start_curr_next = pPairBuffStartCurr[handleIndex+1];\n"
" int curr_max = start_curr_next.x - start - 1;\n"
" \n"
" int bucketEnd = bucketStart + maxBodiesPerCell;\n"
" bucketEnd = (bucketEnd > numObjects) ? numObjects : bucketEnd;\n"
" for(int index2 = bucketStart; index2 < bucketEnd; index2++) \n"
@@ -142,58 +137,30 @@ static const char* gridBroadphaseCL= \
" if (handleIndex<handleIndex2)\n"
" {\n"
" int curPair = atomic_add(pairCount,1);\n"
" int4 newpair;\n"
" newpair.x = handleIndex;\n"
" newpair.y = handleIndex2;\n"
" newpair.z = -1;\n"
" newpair.w = -1;\n"
" pPairBuff2[curPair] = newpair;\n"
" if (curPair<maxPairs)\n"
" {\n"
" int4 newpair;\n"
" newpair.x = handleIndex;\n"
" newpair.y = handleIndex2;\n"
" newpair.z = -1;\n"
" newpair.w = -1;\n"
" pPairBuff2[curPair] = newpair;\n"
" }\n"
" }\n"
" \n"
" } else\n"
" {\n"
" int handleIndex2 = as_int(min1.w);\n"
" int k;\n"
" for(k = 0; k < curr; k++)\n"
" {\n"
" int old_pair = pPairBuff[start+k] & (~0x60000000);\n"
" if(old_pair == handleIndex2)\n"
" {\n"
" pPairBuff[start+k] |= 0x40000000;\n"
" break;\n"
" }\n"
" }\n"
" if(k == curr)\n"
" {\n"
" if(curr >= curr_max) \n"
" { // not a good solution, but let's avoid crash\n"
" break;\n"
" }\n"
" pPairBuff[start+curr] = handleIndex2 | 0x20000000;\n"
" curr++;\n"
" }\n"
" }\n"
" }\n"
" }\n"
" }\n"
" if (!pairCount)\n"
" {\n"
" int2 newStartCurr;\n"
" newStartCurr.x = start;\n"
" newStartCurr.y = curr;\n"
" pPairBuffStartCurr[handleIndex] = newStartCurr;\n"
" }\n"
" \n"
"}\n"
"__kernel void kFindOverlappingPairs( int numObjects,\n"
" __global float4* pAABB, \n"
" __global int2* pHash, \n"
" __global int* pCellStart, \n"
" __global int* pPairBuff, \n"
" __global int2* pPairBuffStartCurr, \n"
" __global float4* pParams ,\n"
" volatile __global int* pairCount,\n"
" __global int4* pPairBuff2\n"
" __global int4* pPairBuff2,\n"
" int maxPairs\n"
" )\n"
"{\n"
" int index = get_global_id(0);\n"
@@ -222,134 +189,9 @@ static const char* gridBroadphaseCL= \
" for(int x=-1; x<=1; x++) \n"
" {\n"
" gridPosB.x = gridPosA.x + x;\n"
" findPairsInCell(numObjects, gridPosB, index, pHash, pCellStart, pAABB, pPairBuff, pPairBuffStartCurr, pParams, pairCount,pPairBuff2);\n"
" findPairsInCell(numObjects, gridPosB, index, pHash, pCellStart, pAABB, pParams, pairCount,pPairBuff2, maxPairs);\n"
" }\n"
" }\n"
" }\n"
"}\n"
"__kernel void kFindPairsLarge( int numObjects, \n"
" __global float4* pAABB, \n"
" __global int2* pHash, \n"
" __global int* pCellStart, \n"
" __global int* pPairBuff, \n"
" __global int2* pPairBuffStartCurr, \n"
" uint numLarge )\n"
"{\n"
" int index = get_global_id(0);\n"
" if(index >= numObjects)\n"
" {\n"
" return;\n"
" }\n"
" int2 sortedData = pHash[index];\n"
" int unsorted_indx = sortedData.y;\n"
" float4 min0 = pAABB[unsorted_indx*2 + 0];\n"
" float4 max0 = pAABB[unsorted_indx*2 + 1];\n"
" int handleIndex = as_int(min0.w);\n"
" int2 start_curr = pPairBuffStartCurr[handleIndex];\n"
" int start = start_curr.x;\n"
" int curr = start_curr.y;\n"
" int2 start_curr_next = pPairBuffStartCurr[handleIndex+1];\n"
" int curr_max = start_curr_next.x - start - 1;\n"
" for(uint i = 0; i < numLarge; i++)\n"
" {\n"
" int indx2 = numObjects + i;\n"
" float4 min1 = pAABB[indx2*2 + 0];\n"
" float4 max1 = pAABB[indx2*2 + 1];\n"
" if(testAABBOverlap(min0, max0, min1, max1))\n"
" {\n"
" int k;\n"
" int handleIndex2 = as_int(min1.w);\n"
" for(k = 0; k < curr; k++)\n"
" {\n"
" int old_pair = pPairBuff[start+k] & (~0x60000000);\n"
" if(old_pair == handleIndex2)\n"
" {\n"
" pPairBuff[start+k] |= 0x40000000;\n"
" break;\n"
" }\n"
" }\n"
" if(k == curr)\n"
" {\n"
" pPairBuff[start+curr] = handleIndex2 | 0x20000000;\n"
" if(curr >= curr_max) \n"
" { // not a good solution, but let's avoid crash\n"
" break;\n"
" }\n"
" curr++;\n"
" }\n"
" }\n"
" }\n"
" int2 newStartCurr;\n"
" newStartCurr.x = start;\n"
" newStartCurr.y = curr;\n"
" pPairBuffStartCurr[handleIndex] = newStartCurr;\n"
" return;\n"
"}\n"
"__kernel void kComputePairCacheChanges( int numObjects,\n"
" __global int* pPairBuff, \n"
" __global int2* pPairBuffStartCurr, \n"
" __global int* pPairScan, \n"
" __global float4* pAABB )\n"
"{\n"
" int index = get_global_id(0);\n"
" if(index >= numObjects)\n"
" {\n"
" return;\n"
" }\n"
" float4 bbMin = pAABB[index * 2];\n"
" int handleIndex = as_int(bbMin.w);\n"
" int2 start_curr = pPairBuffStartCurr[handleIndex];\n"
" int start = start_curr.x;\n"
" int curr = start_curr.y;\n"
" __global int *pInp = pPairBuff + start;\n"
" int num_changes = 0;\n"
" for(int k = 0; k < curr; k++, pInp++)\n"
" {\n"
" if(!((*pInp) & 0x40000000))\n"
" {\n"
" num_changes++;\n"
" }\n"
" }\n"
" pPairScan[index+1] = num_changes;\n"
"} \n"
"__kernel void kSqueezeOverlappingPairBuff( int numObjects,\n"
" __global int* pPairBuff, \n"
" __global int2* pPairBuffStartCurr, \n"
" __global int* pPairScan,\n"
" __global int* pPairOut, \n"
" __global float4* pAABB )\n"
"{\n"
" int index = get_global_id(0);\n"
" if(index >= numObjects)\n"
" {\n"
" return;\n"
" }\n"
" float4 bbMin = pAABB[index * 2];\n"
" int handleIndex = as_int(bbMin.w);\n"
" int2 start_curr = pPairBuffStartCurr[handleIndex];\n"
" int start = start_curr.x;\n"
" int curr = start_curr.y;\n"
" __global int* pInp = pPairBuff + start;\n"
" __global int* pOut = pPairOut + pPairScan[index+1];\n"
" __global int* pOut2 = pInp;\n"
" int num = 0; \n"
" for(int k = 0; k < curr; k++, pInp++)\n"
" {\n"
" if(!((*pInp) & 0x40000000))\n"
" {\n"
" *pOut = *pInp;\n"
" pOut++;\n"
" }\n"
" if((*pInp) & 0x60000000)\n"
" {\n"
" *pOut2 = (*pInp) & (~0x60000000);\n"
" pOut2++;\n"
" num++;\n"
" }\n"
" }\n"
" int2 newStartCurr;\n"
" newStartCurr.x = start;\n"
" newStartCurr.y = num;\n"
" pPairBuffStartCurr[handleIndex] = newStartCurr;\n"
"}\n"
;