SIMD_FORCE_INLINE void small_cache_read_triple( void* ls0, ppu_address_t ea0,
void* ls1, ppu_address_t ea1,
void* ls2, ppu_address_t ea2,
size_t size)
{
btAssert(size<16);
ATTRIBUTE_ALIGNED16(char tmpBuffer0[32]);
ATTRIBUTE_ALIGNED16(char tmpBuffer1[32]);
ATTRIBUTE_ALIGNED16(char tmpBuffer2[32]);
uint32_t i;
///make sure last 4 bits are the same, for cellDmaSmallGet
char* localStore0 = (char*)ls0;
uint32_t last4BitsOffset = ea0 & 0x0f;
char* tmpTarget0 = tmpBuffer0 + last4BitsOffset;
#ifdef __SPU__
cellDmaSmallGet(tmpTarget0,ea0,size,DMA_TAG(1),0,0);
#else
tmpTarget0 = (char*)cellDmaSmallGetReadOnly(tmpTarget0,ea0,size,DMA_TAG(1),0,0);
#endif
char* localStore1 = (char*)ls1;
last4BitsOffset = ea1 & 0x0f;
char* tmpTarget1 = tmpBuffer1 + last4BitsOffset;
#ifdef __SPU__
cellDmaSmallGet(tmpTarget1,ea1,size,DMA_TAG(1),0,0);
#else
tmpTarget1 = (char*)cellDmaSmallGetReadOnly(tmpTarget1,ea1,size,DMA_TAG(1),0,0);
#endif
char* localStore2 = (char*)ls2;
last4BitsOffset = ea2 & 0x0f;
char* tmpTarget2 = tmpBuffer2 + last4BitsOffset;
#ifdef __SPU__
cellDmaSmallGet(tmpTarget2,ea2,size,DMA_TAG(1),0,0);
#else
tmpTarget2 = (char*)cellDmaSmallGetReadOnly(tmpTarget2,ea2,size,DMA_TAG(1),0,0);
#endif
cellDmaWaitTagStatusAll( DMA_MASK(1) );
//this is slowish, perhaps memcpy on SPU is smarter?
for (i=0; btLikely( i<size );i++)
{
localStore0[i] = tmpTarget0[i];
localStore1[i] = tmpTarget1[i];
localStore2[i] = tmpTarget2[i];
}
}
///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)
{
btAssert(size<32);
ATTRIBUTE_ALIGNED16(char tmpBuffer[32]);
char* mainMem = (char*)ea;
char* localStore = (char*)ls;
uint32_t i;
///make sure last 4 bits are the same, for cellDmaSmallGet
uint32_t last4BitsOffset = ea & 0x0f;
char* tmpTarget = tmpBuffer + last4BitsOffset;
#if defined (__SPU__) || defined (USE_LIBSPE2)
int remainingSize = size;
//#define FORCE_cellDmaUnalignedGet 1
#ifdef FORCE_cellDmaUnalignedGet
cellDmaUnalignedGet(tmpTarget,ea,size,DMA_TAG(1),0,0);
#else
char* remainingTmpTarget = tmpTarget;
uint64_t remainingEa = ea;
while (remainingSize)
{
switch (remainingSize)
{
case 1:
case 2:
case 4:
case 8:
case 16:
{
mfc_get(remainingTmpTarget,remainingEa,remainingSize,DMA_TAG(1),0,0);
remainingSize=0;
break;
}
default:
{
//spu_printf("unaligned DMA with non-natural size:%d\n",remainingSize);
int actualSize = 0;
if (remainingSize > 16)
actualSize = 16;
else
if (remainingSize >8)
actualSize=8;
else
if (remainingSize >4)
actualSize=4;
else
if (remainingSize >2)
actualSize=2;
mfc_get(remainingTmpTarget,remainingEa,actualSize,DMA_TAG(1),0,0);
remainingSize-=actualSize;
remainingTmpTarget+=actualSize;
remainingEa += actualSize;
}
}
}
#endif//FORCE_cellDmaUnalignedGet
#else
//copy into final destination
#ifdef USE_MEMCPY
memcpy(tmpTarget,mainMem,size);
#else
for ( i=0;i<size;i++)
{
tmpTarget[i] = mainMem[i];
}
#endif //USE_MEMCPY
#endif
cellDmaWaitTagStatusAll(DMA_MASK(1));
//this is slowish, perhaps memcpy on SPU is smarter?
for (i=0; btLikely( i<size );i++)
{
localStore[i] = tmpTarget[i];
}
return 0;
}
virtual void processNode(int subPart, int triangleIndex)
{
///Create a triangle on the stack, call process collision, with GJK
///DMA the vertices, can benefit from software caching
// spu_printf("processNode with triangleIndex %d\n",triangleIndex);
// ugly solution to support both 16bit and 32bit indices
if (m_lsMemPtr->bvhShapeData.gIndexMesh.m_indexType == PHY_SHORT)
{
short int* indexBasePtr = (short int*)(m_lsMemPtr->bvhShapeData.gIndexMesh.m_triangleIndexBase+triangleIndex*m_lsMemPtr->bvhShapeData.gIndexMesh.m_triangleIndexStride);
ATTRIBUTE_ALIGNED16(short int tmpIndices[3]);
small_cache_read_triple(&tmpIndices[0],(ppu_address_t)&indexBasePtr[0],
&tmpIndices[1],(ppu_address_t)&indexBasePtr[1],
&tmpIndices[2],(ppu_address_t)&indexBasePtr[2],
sizeof(short int));
m_lsMemPtr->spuIndices[0] = int(tmpIndices[0]);
m_lsMemPtr->spuIndices[1] = int(tmpIndices[1]);
m_lsMemPtr->spuIndices[2] = int(tmpIndices[2]);
} else
{
int* indexBasePtr = (int*)(m_lsMemPtr->bvhShapeData.gIndexMesh.m_triangleIndexBase+triangleIndex*m_lsMemPtr->bvhShapeData.gIndexMesh.m_triangleIndexStride);
small_cache_read_triple(&m_lsMemPtr->spuIndices[0],(ppu_address_t)&indexBasePtr[0],
&m_lsMemPtr->spuIndices[1],(ppu_address_t)&indexBasePtr[1],
&m_lsMemPtr->spuIndices[2],(ppu_address_t)&indexBasePtr[2],
sizeof(int));
}
//printf("%d %d %d\n", m_lsMemPtr->spuIndices[0], m_lsMemPtr->spuIndices[1], m_lsMemPtr->spuIndices[2]);
// spu_printf("SPU index0=%d ,",spuIndices[0]);
// spu_printf("SPU index1=%d ,",spuIndices[1]);
// spu_printf("SPU index2=%d ,",spuIndices[2]);
// spu_printf("SPU: indexBasePtr=%llx\n",indexBasePtr);
const btVector3& meshScaling = m_lsMemPtr->bvhShapeData.gTriangleMeshInterfacePtr->getScaling();
for (int j=2;btLikely( j>=0 );j--)
{
int graphicsindex = m_lsMemPtr->spuIndices[j];
//spu_printf("SPU index=%d ,",graphicsindex);
btScalar* graphicsbasePtr = (btScalar*)(m_lsMemPtr->bvhShapeData.gIndexMesh.m_vertexBase+graphicsindex*m_lsMemPtr->bvhShapeData.gIndexMesh.m_vertexStride);
// spu_printf("SPU graphicsbasePtr=%llx\n",graphicsbasePtr);
///handle un-aligned vertices...
//another DMA for each vertex
small_cache_read_triple(&spuUnscaledVertex[0],(ppu_address_t)&graphicsbasePtr[0],
&spuUnscaledVertex[1],(ppu_address_t)&graphicsbasePtr[1],
&spuUnscaledVertex[2],(ppu_address_t)&graphicsbasePtr[2],
sizeof(btScalar));
//printf("%f %f %f\n", spuUnscaledVertex[0],spuUnscaledVertex[1],spuUnscaledVertex[2]);
spuTriangleVertices[j] = btVector3(
spuUnscaledVertex[0]*meshScaling.getX(),
spuUnscaledVertex[1]*meshScaling.getY(),
spuUnscaledVertex[2]*meshScaling.getZ());
//spu_printf("SPU:triangle vertices:%f,%f,%f\n",spuTriangleVertices[j].x(),spuTriangleVertices[j].y(),spuTriangleVertices[j].z());
}
RaycastGatheredObjectData triangleGatheredObjectData (*m_gatheredObjectData);
triangleGatheredObjectData.m_shapeType = TRIANGLE_SHAPE_PROXYTYPE;
triangleGatheredObjectData.m_spuCollisionShape = &spuTriangleVertices[0];
//printf("%f %f %f\n", spuTriangleVertices[0][0],spuTriangleVertices[0][1],spuTriangleVertices[0][2]);
//printf("%f %f %f\n", spuTriangleVertices[1][0],spuTriangleVertices[1][1],spuTriangleVertices[1][2]);
//printf("%f %f %f\n", spuTriangleVertices[2][0],spuTriangleVertices[2][1],spuTriangleVertices[2][2]);
for (int i = 0; i < m_numWorkUnits; i++)
{
SpuRaycastTaskWorkUnitOut out;
out.hitFraction = 1.0;
performRaycastAgainstConvex (&triangleGatheredObjectData, m_workUnits[i], &out, m_lsMemPtr);
/* XXX: For now only take the closest hit */
if (out.hitFraction < m_workUnitsOut[i].hitFraction)
{
m_workUnitsOut[i].hitFraction = out.hitFraction;
m_workUnitsOut[i].hitNormal = out.hitNormal;
}
}
}
virtual void processNode(int subPart, int triangleIndex)
{
///Create a triangle on the stack, call process collision, with GJK
///DMA the vertices, can benefit from software caching
// spu_printf("processNode with triangleIndex %d\n",triangleIndex);
if (m_lsMemPtr->bvhShapeData.gIndexMesh.m_indexType == PHY_SHORT)
{
unsigned short int* indexBasePtr = (unsigned short int*)(m_lsMemPtr->bvhShapeData.gIndexMesh.m_triangleIndexBase+triangleIndex*m_lsMemPtr->bvhShapeData.gIndexMesh.m_triangleIndexStride);
ATTRIBUTE_ALIGNED16(unsigned short int tmpIndices[3]);
small_cache_read_triple(&tmpIndices[0],(ppu_address_t)&indexBasePtr[0],
&tmpIndices[1],(ppu_address_t)&indexBasePtr[1],
&tmpIndices[2],(ppu_address_t)&indexBasePtr[2],
sizeof(unsigned short int));
m_lsMemPtr->spuIndices[0] = int(tmpIndices[0]);
m_lsMemPtr->spuIndices[1] = int(tmpIndices[1]);
m_lsMemPtr->spuIndices[2] = int(tmpIndices[2]);
} else
{
unsigned int* indexBasePtr = (unsigned int*)(m_lsMemPtr->bvhShapeData.gIndexMesh.m_triangleIndexBase+triangleIndex*m_lsMemPtr->bvhShapeData.gIndexMesh.m_triangleIndexStride);
small_cache_read_triple(&m_lsMemPtr->spuIndices[0],(ppu_address_t)&indexBasePtr[0],
&m_lsMemPtr->spuIndices[1],(ppu_address_t)&indexBasePtr[1],
&m_lsMemPtr->spuIndices[2],(ppu_address_t)&indexBasePtr[2],
sizeof(int));
}
// spu_printf("SPU index0=%d ,",spuIndices[0]);
// spu_printf("SPU index1=%d ,",spuIndices[1]);
// spu_printf("SPU index2=%d ,",spuIndices[2]);
// spu_printf("SPU: indexBasePtr=%llx\n",indexBasePtr);
const btVector3& meshScaling = m_lsMemPtr->bvhShapeData.gTriangleMeshInterfacePtr->getScaling();
for (int j=2;btLikely( j>=0 );j--)
{
int graphicsindex = m_lsMemPtr->spuIndices[j];
// spu_printf("SPU index=%d ,",graphicsindex);
btScalar* graphicsbasePtr = (btScalar*)(m_lsMemPtr->bvhShapeData.gIndexMesh.m_vertexBase+graphicsindex*m_lsMemPtr->bvhShapeData.gIndexMesh.m_vertexStride);
// spu_printf("SPU graphicsbasePtr=%llx\n",graphicsbasePtr);
///handle un-aligned vertices...
//another DMA for each vertex
small_cache_read_triple(&spuUnscaledVertex[0],(ppu_address_t)&graphicsbasePtr[0],
&spuUnscaledVertex[1],(ppu_address_t)&graphicsbasePtr[1],
&spuUnscaledVertex[2],(ppu_address_t)&graphicsbasePtr[2],
sizeof(btScalar));
m_tmpTriangleShape.getVertexPtr(j).setValue(spuUnscaledVertex[0]*meshScaling.getX(),
spuUnscaledVertex[1]*meshScaling.getY(),
spuUnscaledVertex[2]*meshScaling.getZ());
// spu_printf("SPU:triangle vertices:%f,%f,%f\n",spuTriangleVertices[j].x(),spuTriangleVertices[j].y(),spuTriangleVertices[j].z());
}
SpuCollisionPairInput triangleConcaveInput(*m_wuInput);
// triangleConcaveInput.m_spuCollisionShapes[1] = &spuTriangleVertices[0];
triangleConcaveInput.m_spuCollisionShapes[1] = &m_tmpTriangleShape;
triangleConcaveInput.m_shapeType1 = TRIANGLE_SHAPE_PROXYTYPE;
m_spuContacts.setShapeIdentifiersB(subPart,triangleIndex);
// m_spuContacts.flush();
ProcessSpuConvexConvexCollision(&triangleConcaveInput, m_lsMemPtr,m_spuContacts);
///this flush should be automatic
// m_spuContacts.flush();
}