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

made the simulation deterministic

Browse Source 
disable 'simdwidth' optimization for determinism (need to double-check)
made the spatial batching 3D
This commit is contained in:
erwin coumans
2013-07-14 19:16:33 -07:00
parent 4b293130bd
commit 1e31073f4b
14 changed files with 805 additions and 169 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
src/Bullet3OpenCL/RigidBody/kernels/solveContact.cl
View File

@@ -411,8 +411,8 @@ void BatchSolveKernelContact(__global Body* gBodies,
__global int* gN,
__global int* gOffsets,
int maxBatch,
int bIdx,
int nSplit
int cellBatch,
int4 nSplit
)
{
//__local int ldsBatchIdx[WG_SIZE+1];
@@ -428,17 +428,29 @@ void BatchSolveKernelContact(__global Body* gBodies,
//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;
int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);
int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);
int cellIdx = xIdx+yIdx*nSplit;
int zIdx = (wgIdx/((nSplit.x*nSplit.y)/4))*2+((cellBatch&4)>>2);
int remain= (wgIdx%((nSplit.x*nSplit.y)/4));
int yIdx = (remain/(nSplit.x/2))*2 + ((cellBatch&2)>>1);
int xIdx = (remain%(nSplit.x/2))*2 + (cellBatch&1);
int cellIdx = xIdx+yIdx*nSplit.x+zIdx*(nSplit.x*nSplit.y);
//int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);
//int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);
//int cellIdx = xIdx+yIdx*nSplit;
if( gN[cellIdx] == 0 )
return;
const int start = gOffsets[cellIdx];
const int end = start + gN[cellIdx];
//if (lIdx==0)
//printf("wgIdx = %d, start = %d, end=%d\n",wgIdx,start,end);
if( lIdx == 0 )
{
ldsCurBatch = 0;
@@ -456,6 +468,9 @@ void BatchSolveKernelContact(__global Body* gBodies,
{
if (gConstraints[idx].m_batchIdx == ldsCurBatch)
{
//if (wgIdx==0 && lIdx==0)
//printf("solved wgIdx=%d, ldsCurBatch=%d idx=%d \n", wgIdx, ldsCurBatch,idx);
solveContactConstraint( gBodies, gShapes, &gConstraints[idx] );
idx+=64;
@@ -465,6 +480,8 @@ void BatchSolveKernelContact(__global Body* gBodies,
}
}
GROUP_LDS_BARRIER;
// if (wgIdx==0 && lIdx==0)
// printf("-----------------------\n");
if( lIdx == 0 )
{
ldsCurBatch++;

27
src/Bullet3OpenCL/RigidBody/kernels/solveContact.h
View File

@@ -413,8 +413,8 @@ static const char* solveContactCL= \
" __global int* gN,\n"
" __global int* gOffsets,\n"
" int maxBatch,\n"
" int bIdx,\n"
" int nSplit\n"
" int cellBatch,\n"
" int4 nSplit\n"
" )\n"
"{\n"
" //__local int ldsBatchIdx[WG_SIZE+1];\n"
@@ -430,17 +430,29 @@ static const char* solveContactCL= \
" //debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;\n"
"\n"
"\n"
" int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);\n"
" int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);\n"
" int cellIdx = xIdx+yIdx*nSplit;\n"
" int zIdx = (wgIdx/((nSplit.x*nSplit.y)/4))*2+((cellBatch&4)>>2);\n"
" int remain= (wgIdx%((nSplit.x*nSplit.y)/4));\n"
" int yIdx = (remain/(nSplit.x/2))*2 + ((cellBatch&2)>>1);\n"
" int xIdx = (remain%(nSplit.x/2))*2 + (cellBatch&1);\n"
" int cellIdx = xIdx+yIdx*nSplit.x+zIdx*(nSplit.x*nSplit.y);\n"
"\n"
" //int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);\n"
" //int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);\n"
" //int cellIdx = xIdx+yIdx*nSplit;\n"
" \n"
" if( gN[cellIdx] == 0 ) \n"
" return;\n"
"\n"
" \n"
" \n"
" const int start = gOffsets[cellIdx];\n"
" const int end = start + gN[cellIdx];\n"
"\n"
" \n"
" //if (lIdx==0)\n"
" //printf(\"wgIdx = %d, start = %d, end=%d\n\",wgIdx,start,end);\n"
"\n"
" \n"
" if( lIdx == 0 )\n"
" {\n"
" ldsCurBatch = 0;\n"
@@ -458,6 +470,9 @@ static const char* solveContactCL= \
" {\n"
" if (gConstraints[idx].m_batchIdx == ldsCurBatch)\n"
" {\n"
" //if (wgIdx==0 && lIdx==0)\n"
" //printf(\"solved wgIdx=%d, ldsCurBatch=%d idx=%d \n\", wgIdx, ldsCurBatch,idx);\n"
" \n"
" solveContactConstraint( gBodies, gShapes, &gConstraints[idx] );\n"
"\n"
" idx+=64;\n"
@@ -467,6 +482,8 @@ static const char* solveContactCL= \
" }\n"
" }\n"
" GROUP_LDS_BARRIER;\n"
" // if (wgIdx==0 && lIdx==0)\n"
" // printf(\"-----------------------\n\");\n"
" if( lIdx == 0 )\n"
" {\n"
" ldsCurBatch++;\n"

13
src/Bullet3OpenCL/RigidBody/kernels/solveFriction.cl
View File

@@ -440,8 +440,8 @@ void BatchSolveKernelFriction(__global Body* gBodies,
__global int* gN,
__global int* gOffsets,
int maxBatch,
int bIdx,
int nSplit
int cellBatch,
int4 nSplit
)
{
//__local int ldsBatchIdx[WG_SIZE+1];
@@ -457,9 +457,12 @@ void BatchSolveKernelFriction(__global Body* gBodies,
//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;
int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);
int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);
int cellIdx = xIdx+yIdx*nSplit;
int zIdx = (wgIdx/((nSplit.x*nSplit.y)/4))*2+((cellBatch&4)>>2);
int remain= (wgIdx%((nSplit.x*nSplit.y)/4));
int yIdx = (remain/(nSplit.x/2))*2 + ((cellBatch&2)>>1);
int xIdx = (remain%(nSplit.x/2))*2 + (cellBatch&1);
int cellIdx = xIdx+yIdx*nSplit.x+zIdx*(nSplit.x*nSplit.y);
if( gN[cellIdx] == 0 )
return;

13
src/Bullet3OpenCL/RigidBody/kernels/solveFriction.h
View File

@@ -442,8 +442,8 @@ static const char* solveFrictionCL= \
" __global int* gN,\n"
" __global int* gOffsets,\n"
" int maxBatch,\n"
" int bIdx,\n"
" int nSplit\n"
" int cellBatch,\n"
" int4 nSplit\n"
" )\n"
"{\n"
" //__local int ldsBatchIdx[WG_SIZE+1];\n"
@@ -459,9 +459,12 @@ static const char* solveFrictionCL= \
" //debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;\n"
"\n"
"\n"
" int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);\n"
" int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);\n"
" int cellIdx = xIdx+yIdx*nSplit;\n"
" int zIdx = (wgIdx/((nSplit.x*nSplit.y)/4))*2+((cellBatch&4)>>2);\n"
" int remain= (wgIdx%((nSplit.x*nSplit.y)/4));\n"
" int yIdx = (remain/(nSplit.x/2))*2 + ((cellBatch&2)>>1);\n"
" int xIdx = (remain%(nSplit.x/2))*2 + (cellBatch&1);\n"
" int cellIdx = xIdx+yIdx*nSplit.x+zIdx*(nSplit.x*nSplit.y);\n"
"\n"
" \n"
" if( gN[cellIdx] == 0 ) \n"
" return;\n"

45
src/Bullet3OpenCL/RigidBody/kernels/solverSetup2.cl
View File

@@ -441,6 +441,42 @@ void ReorderContactKernel(__global Contact4* in, __global Contact4* out, __globa
}
}
__kernel
__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
void SetDeterminismSortDataBodyA(__global Contact4* contactsIn, __global int2* sortDataOut, int nContacts)
{
int gIdx = GET_GLOBAL_IDX;
if( gIdx < nContacts )
{
int2 sd;
sd.x = contactsIn[gIdx].m_bodyAPtrAndSignBit;
sd.y = gIdx;
sortDataOut[gIdx] = sd;
}
}
__kernel
__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
void SetDeterminismSortDataBodyB(__global Contact4* contactsIn, __global int2* sortDataInOut, int nContacts)
{
int gIdx = GET_GLOBAL_IDX;
if( gIdx < nContacts )
{
int2 sdIn;
sdIn = sortDataInOut[gIdx];
int2 sdOut;
sdOut.x = contactsIn[sdIn.y].m_bodyBPtrAndSignBit;
sdOut.y = sdIn.y;
sortDataInOut[gIdx] = sdOut;
}
}
typedef struct
{
int m_nContacts;
@@ -480,7 +516,7 @@ static __constant const int gridTable8x8[] =
__kernel
__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
void SetSortDataKernel(__global Contact4* gContact, __global Body* gBodies, __global int2* gSortDataOut,
int nContacts,float scale,int N_SPLIT, int staticIdx)
int nContacts,float scale,int4 nSplit,int staticIdx)
{
int gIdx = GET_GLOBAL_IDX;
@@ -499,9 +535,10 @@ int nContacts,float scale,int N_SPLIT, int staticIdx)
#if USE_SPATIAL_BATCHING
int idx = (aStatic)? bIdx: aIdx;
float4 p = gBodies[idx].m_pos;
int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (N_SPLIT-1);
int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (N_SPLIT-1);
int newIndex = (xIdx+zIdx*N_SPLIT);
int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (nSplit.x-1);
int yIdx = (int)((p.y-((p.y<0.f)?1.f:0.f))*scale) & (nSplit.y-1);
int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (nSplit.z-1);
int newIndex = (xIdx+yIdx*nSplit.x+zIdx*nSplit.x*nSplit.y);
#else//USE_SPATIAL_BATCHING
#if USE_4x4_GRID

45
src/Bullet3OpenCL/RigidBody/kernels/solverSetup2.h
View File

@@ -443,6 +443,42 @@ static const char* solverSetup2CL= \
" }\n"
"}\n"
"\n"
"\n"
"__kernel\n"
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
"void SetDeterminismSortDataBodyA(__global Contact4* contactsIn, __global int2* sortDataOut, int nContacts)\n"
"{\n"
" int gIdx = GET_GLOBAL_IDX;\n"
"\n"
" if( gIdx < nContacts )\n"
" {\n"
" int2 sd;\n"
" sd.x = contactsIn[gIdx].m_bodyAPtrAndSignBit;\n"
" sd.y = gIdx;\n"
" sortDataOut[gIdx] = sd;\n"
" }\n"
"}\n"
"\n"
"__kernel\n"
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
"void SetDeterminismSortDataBodyB(__global Contact4* contactsIn, __global int2* sortDataInOut, int nContacts)\n"
"{\n"
" int gIdx = GET_GLOBAL_IDX;\n"
"\n"
" if( gIdx < nContacts )\n"
" {\n"
" int2 sdIn;\n"
" sdIn = sortDataInOut[gIdx];\n"
" int2 sdOut;\n"
" sdOut.x = contactsIn[sdIn.y].m_bodyBPtrAndSignBit;\n"
" sdOut.y = sdIn.y;\n"
" sortDataInOut[gIdx] = sdOut;\n"
" }\n"
"}\n"
"\n"
"\n"
"\n"
"\n"
"typedef struct\n"
"{\n"
" int m_nContacts;\n"
@@ -482,7 +518,7 @@ static const char* solverSetup2CL= \
"__kernel\n"
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
"void SetSortDataKernel(__global Contact4* gContact, __global Body* gBodies, __global int2* gSortDataOut, \n"
"int nContacts,float scale,int N_SPLIT, int staticIdx)\n"
"int nContacts,float scale,int4 nSplit,int staticIdx)\n"
"\n"
"{\n"
" int gIdx = GET_GLOBAL_IDX;\n"
@@ -501,9 +537,10 @@ static const char* solverSetup2CL= \
"#if USE_SPATIAL_BATCHING \n"
" int idx = (aStatic)? bIdx: aIdx;\n"
" float4 p = gBodies[idx].m_pos;\n"
" int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (N_SPLIT-1);\n"
" int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (N_SPLIT-1);\n"
" int newIndex = (xIdx+zIdx*N_SPLIT);\n"
" int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (nSplit.x-1);\n"
" int yIdx = (int)((p.y-((p.y<0.f)?1.f:0.f))*scale) & (nSplit.y-1);\n"
" int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (nSplit.z-1);\n"
" int newIndex = (xIdx+yIdx*nSplit.x+zIdx*nSplit.x*nSplit.y);\n"
" \n"
"#else//USE_SPATIAL_BATCHING\n"
" #if USE_4x4_GRID\n"