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

added a cellDmaLargeGetReadOnly/cellDmaGetReadOnly, which avoid memcpy on non-SPU platforms.

Browse Source 
enabled it for triangle mesh aabb DMA transfers. Todo: check if it is worth also modifying other parts of the code to avoid the memcpy into the 256kb Task memory.
This commit is contained in:
ejcoumans
2007-10-16 23:58:26 +00:00
parent 239d88985f
commit 830c096e10
3 changed files with 32 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
Extras/BulletMultiThreaded/SpuFakeDma.cpp
View File

@@ -5,6 +5,27 @@
#define USE_MEMCPY 1 #define USE_MEMCPY 1
void* cellDmaLargeGetReadOnly(void *ls, uint64_t ea, uint32_t size, uint32_t tag, uint32_t tid, uint32_t rid)
{
#if defined (WIN32) || defined (__PPU__)
return (void*)(uint32_t)ea;
#else
cellDmaLargeGet(ls,ea,size,tag,tid,rid);
return ls;
#endif //WIN32
}
void* cellDmaGetReadOnly(void *ls, uint64_t ea, uint32_t size, uint32_t tag, uint32_t tid, uint32_t rid)
{
#if defined (WIN32) || defined (__PPU__)
return (void*)(uint32_t)ea;
#else
cellDmaGet(ls,ea,size,tag,tid,rid);
return ls;
#endif //WIN32
}
///this unalignedDma should not be frequently used, only for small data. It handles alignment and performs check on size (<16 bytes) ///this unalignedDma should not be frequently used, only for small data. It handles alignment and performs check on size (<16 bytes)
int stallingUnalignedDmaSmallGet(void *ls, uint64_t ea, uint32_t size) int stallingUnalignedDmaSmallGet(void *ls, uint64_t ea, uint32_t size)

7
Extras/BulletMultiThreaded/SpuFakeDma.h
View File

@@ -4,7 +4,7 @@
#include "PlatformDefinitions.h" #include "PlatformDefinitions.h"
#include "LinearMath/btScalar.h"
#ifdef __CELLOS_LV2__ #ifdef __CELLOS_LV2__
@@ -118,4 +118,9 @@
///stallingUnalignedDmaSmallGet internally uses DMA_TAG(1) ///stallingUnalignedDmaSmallGet internally uses DMA_TAG(1)
int stallingUnalignedDmaSmallGet(void *ls, uint64_t ea, uint32_t size); int stallingUnalignedDmaSmallGet(void *ls, uint64_t ea, uint32_t size);
void* cellDmaLargeGetReadOnly(void *ls, uint64_t ea, uint32_t size, uint32_t tag, uint32_t tid, uint32_t rid);
void* cellDmaGetReadOnly(void *ls, uint64_t ea, uint32_t size, uint32_t tag, uint32_t tid, uint32_t rid);
#endif //FAKE_DMA_H #endif //FAKE_DMA_H

9
Extras/BulletMultiThreaded/SpuNarrowPhaseCollisionTask/SpuGatheringCollisionTask.cpp
View File

@@ -149,7 +149,8 @@ struct CollisionTask_LocalStoreMemory
return (btOptimizedBvh*) gOptimizedBvh; return (btOptimizedBvh*) gOptimizedBvh;
} }
ATTRIBUTE_ALIGNED16(btTriangleIndexVertexArray gTriangleMeshInterface); ATTRIBUTE_ALIGNED16(btTriangleIndexVertexArray gTriangleMeshInterfaceStorage);
btTriangleIndexVertexArray* gTriangleMeshInterfacePtr;
///only a single mesh part for now, we can add support for multiple parts, but quantized trees don't support this at the moment ///only a single mesh part for now, we can add support for multiple parts, but quantized trees don't support this at the moment
ATTRIBUTE_ALIGNED16(btIndexedMesh gIndexMesh); ATTRIBUTE_ALIGNED16(btIndexedMesh gIndexMesh);
@@ -388,7 +389,7 @@ public:
// spu_printf("SPU index2=%d ,",spuIndices[2]); // spu_printf("SPU index2=%d ,",spuIndices[2]);
// spu_printf("SPU: indexBasePtr=%llx\n",indexBasePtr); // spu_printf("SPU: indexBasePtr=%llx\n",indexBasePtr);
const btVector3& meshScaling = m_lsMemPtr->gTriangleMeshInterface.getScaling(); const btVector3& meshScaling = m_lsMemPtr->gTriangleMeshInterfacePtr->getScaling();
for (int j=2;btLikely( j>=0 );j--) for (int j=2;btLikely( j>=0 );j--)
{ {
int graphicsindex = m_lsMemPtr->spuIndices[j]; int graphicsindex = m_lsMemPtr->spuIndices[j];
@@ -482,7 +483,7 @@ void ProcessConvexConcaveSpuCollision(SpuCollisionPairInput* wuInput, CollisionT
dmaPpuAddress2 = reinterpret_cast<uint32_t>(trimeshShape->getMeshInterface()); dmaPpuAddress2 = reinterpret_cast<uint32_t>(trimeshShape->getMeshInterface());
#endif #endif
// spu_printf("trimeshShape->getMeshInterface() == %llx\n",dmaPpuAddress2); // spu_printf("trimeshShape->getMeshInterface() == %llx\n",dmaPpuAddress2);
cellDmaGet(&lsMemPtr->gTriangleMeshInterface, dmaPpuAddress2 , dmaSize, DMA_TAG(1), 0, 0); lsMemPtr->gTriangleMeshInterfacePtr = (btTriangleIndexVertexArray*)cellDmaGetReadOnly(&lsMemPtr->gTriangleMeshInterfaceStorage, dmaPpuAddress2 , dmaSize, DMA_TAG(1), 0, 0);
//cellDmaWaitTagStatusAll(DMA_MASK(1)); //cellDmaWaitTagStatusAll(DMA_MASK(1));
@@ -611,7 +612,7 @@ void ProcessConvexConcaveSpuCollision(SpuCollisionPairInput* wuInput, CollisionT
BvhSubtreeInfoArray& subTrees = lsMemPtr->getOptimizedBvh()->getSubtreeInfoArray(); BvhSubtreeInfoArray& subTrees = lsMemPtr->getOptimizedBvh()->getSubtreeInfoArray();
spuNodeCallback nodeCallback(wuInput,lsMemPtr,spuContacts); spuNodeCallback nodeCallback(wuInput,lsMemPtr,spuContacts);
IndexedMeshArray& indexArray = lsMemPtr->gTriangleMeshInterface.getIndexedMeshArray(); IndexedMeshArray& indexArray = lsMemPtr->gTriangleMeshInterfacePtr->getIndexedMeshArray();
//spu_printf("SPU:indexArray.size() = %d\n",indexArray.size()); //spu_printf("SPU:indexArray.size() = %d\n",indexArray.size());