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

Stop using wavefront/warp synchronization inside OpenCL/DirectCompute, it is not future proof.

Browse Source 
Similar to this commit for DX11: http://code.google.com/p/bullet/source/detail?spec=svn2343&r=2330
See also http://forum.beyond3d.com/archive/index.php/t-55751.html
This commit is contained in:
erwin.coumans
2011-03-15 22:41:55 +00:00
parent 21162f77e3
commit 1a222f642a
2 changed files with 2 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/BulletMultiThreaded/GpuSoftBodySolvers/OpenCL/OpenCLC/SolvePositionsSIMDBatched.cl
View File

@@ -46,18 +46,8 @@ SolvePositionsFromLinksKernel(
if( wavefront < (startWaveInBatch + numWaves) ) if( wavefront < (startWaveInBatch + numWaves) )
{ {
// Load the batch counts for the wavefronts // Load the batch counts for the wavefronts
// Mask out in case there's a stray "wavefront" at the end that's been forced in through the multiplier
if( laneInWavefront == 0 )
{
int2 batchesAndVertexCountsWithinWavefront = g_wavefrontBatchCountsVertexCounts[wavefront];
wavefrontBatchCountsVertexCounts[localWavefront] = batchesAndVertexCountsWithinWavefront;
}
mem_fence(CLK_LOCAL_MEM_FENCE); int2 batchesAndVerticesWithinWavefront = g_wavefrontBatchCountsVertexCounts[wavefront];
int2 batchesAndVerticesWithinWavefront = wavefrontBatchCountsVertexCounts[localWavefront];
int batchesWithinWavefront = batchesAndVerticesWithinWavefront.x; int batchesWithinWavefront = batchesAndVerticesWithinWavefront.x;
int verticesUsedByWave = batchesAndVerticesWithinWavefront.y; int verticesUsedByWave = batchesAndVerticesWithinWavefront.y;

12
src/BulletMultiThreaded/GpuSoftBodySolvers/OpenCL/OpenCLC10/SolvePositionsSIMDBatched.cl
View File

@@ -47,18 +47,8 @@ SolvePositionsFromLinksKernel(
if( wavefront < (startWaveInBatch + numWaves) ) if( wavefront < (startWaveInBatch + numWaves) )
{ {
// Load the batch counts for the wavefronts // Load the batch counts for the wavefronts
// Mask out in case there's a stray "wavefront" at the end that's been forced in through the multiplier
if( laneInWavefront == 0 )
{
int2 batchesAndVertexCountsWithinWavefront = g_wavefrontBatchCountsVertexCounts[wavefront];
wavefrontBatchCountsVertexCounts[localWavefront] = batchesAndVertexCountsWithinWavefront;
}
barrier(CLK_LOCAL_MEM_FENCE); int2 batchesAndVerticesWithinWavefront = g_wavefrontBatchCountsVertexCounts[wavefront];
int2 batchesAndVerticesWithinWavefront = wavefrontBatchCountsVertexCounts[localWavefront];
int batchesWithinWavefront = batchesAndVerticesWithinWavefront.x; int batchesWithinWavefront = batchesAndVerticesWithinWavefront.x;
int verticesUsedByWave = batchesAndVerticesWithinWavefront.y; int verticesUsedByWave = batchesAndVerticesWithinWavefront.y;