change lcpp Lua preprocessor, to keep #defines and comments, remove empty lines
remove duplicate data in b3Contact4 (now in btContact4Data shared between CPU/C++ and OpenCL) OpenCL kernels use #include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h" Increase number of batches back to 250 (from 50), need to fix this hard coded number (see https://github.com/erwincoumans/bullet3/issues/12) Work towards GJK/EPA, in addition to SAT/clipping (early on)
This commit is contained in:
erwincoumans
@@ -2,59 +2,45 @@
|
||||
static const char* boundSearchKernelsCL= \
|
||||
"/*\n"
|
||||
"Copyright (c) 2012 Advanced Micro Devices, Inc. \n"
|
||||
"\n"
|
||||
"This software is provided 'as-is', without any express or implied warranty.\n"
|
||||
"In no event will the authors be held liable for any damages arising from the use of this software.\n"
|
||||
"Permission is granted to anyone to use this software for any purpose, \n"
|
||||
"including commercial applications, and to alter it and redistribute it freely, \n"
|
||||
"subject to the following restrictions:\n"
|
||||
"\n"
|
||||
"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
|
||||
"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
|
||||
"3. This notice may not be removed or altered from any source distribution.\n"
|
||||
"*/\n"
|
||||
"//Originally written by Takahiro Harada\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"typedef unsigned int u32;\n"
|
||||
"#define GET_GROUP_IDX get_group_id(0)\n"
|
||||
"#define GET_LOCAL_IDX get_local_id(0)\n"
|
||||
"#define GET_GLOBAL_IDX get_global_id(0)\n"
|
||||
"#define GET_GROUP_SIZE get_local_size(0)\n"
|
||||
"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
|
||||
"\n"
|
||||
"typedef struct\n"
|
||||
"{\n"
|
||||
" u32 m_key; \n"
|
||||
" u32 m_value;\n"
|
||||
"}SortData;\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"typedef struct\n"
|
||||
"{\n"
|
||||
" u32 m_nSrc;\n"
|
||||
" u32 m_nDst;\n"
|
||||
" u32 m_padding[2];\n"
|
||||
"} ConstBuffer;\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void SearchSortDataLowerKernel(__global SortData* src, __global u32 *dst, \n"
|
||||
" unsigned int nSrc, unsigned int nDst)\n"
|
||||
"{\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
"\n"
|
||||
" if( gIdx < nSrc )\n"
|
||||
" {\n"
|
||||
" SortData first; first.m_key = (u32)(-1); first.m_value = (u32)(-1);\n"
|
||||
" SortData end; end.m_key = nDst; end.m_value = nDst;\n"
|
||||
"\n"
|
||||
" SortData iData = (gIdx==0)? first: src[gIdx-1];\n"
|
||||
" SortData jData = (gIdx==nSrc)? end: src[gIdx];\n"
|
||||
"\n"
|
||||
" if( iData.m_key != jData.m_key )\n"
|
||||
" {\n"
|
||||
"// for(u32 k=iData.m_key+1; k<=min(jData.m_key, nDst-1); k++)\n"
|
||||
@@ -65,23 +51,18 @@ static const char* boundSearchKernelsCL= \
|
||||
" }\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void SearchSortDataUpperKernel(__global SortData* src, __global u32 *dst, \n"
|
||||
" unsigned int nSrc, unsigned int nDst)\n"
|
||||
"{\n"
|
||||
" int gIdx = GET_GLOBAL_IDX+1;\n"
|
||||
"\n"
|
||||
" if( gIdx < nSrc+1 )\n"
|
||||
" {\n"
|
||||
" SortData first; first.m_key = 0; first.m_value = 0;\n"
|
||||
" SortData end; end.m_key = nDst; end.m_value = nDst;\n"
|
||||
"\n"
|
||||
" SortData iData = src[gIdx-1];\n"
|
||||
" SortData jData = (gIdx==nSrc)? end: src[gIdx];\n"
|
||||
"\n"
|
||||
" if( iData.m_key != jData.m_key )\n"
|
||||
" {\n"
|
||||
" u32 k = iData.m_key;\n"
|
||||
@@ -91,7 +72,6 @@ static const char* boundSearchKernelsCL= \
|
||||
" }\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void SubtractKernel(__global u32* A, __global u32 *B, __global u32 *C, \n"
|
||||
@@ -99,11 +79,9 @@ static const char* boundSearchKernelsCL= \
|
||||
"{\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
" \n"
|
||||
"\n"
|
||||
" if( gIdx < nDst )\n"
|
||||
" {\n"
|
||||
" C[gIdx] = A[gIdx] - B[gIdx];\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
;
|
||||
|
||||
@@ -2,23 +2,18 @@
|
||||
static const char* fillKernelsCL= \
|
||||
"/*\n"
|
||||
"Copyright (c) 2012 Advanced Micro Devices, Inc. \n"
|
||||
"\n"
|
||||
"This software is provided 'as-is', without any express or implied warranty.\n"
|
||||
"In no event will the authors be held liable for any damages arising from the use of this software.\n"
|
||||
"Permission is granted to anyone to use this software for any purpose, \n"
|
||||
"including commercial applications, and to alter it and redistribute it freely, \n"
|
||||
"subject to the following restrictions:\n"
|
||||
"\n"
|
||||
"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
|
||||
"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
|
||||
"3. This notice may not be removed or altered from any source distribution.\n"
|
||||
"*/\n"
|
||||
"//Originally written by Takahiro Harada\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
|
||||
"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
|
||||
"\n"
|
||||
"typedef unsigned int u32;\n"
|
||||
"#define GET_GROUP_IDX get_group_id(0)\n"
|
||||
"#define GET_LOCAL_IDX get_local_id(0)\n"
|
||||
@@ -28,11 +23,9 @@ static const char* fillKernelsCL= \
|
||||
"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
|
||||
"#define AtomInc(x) atom_inc(&(x))\n"
|
||||
"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
|
||||
"\n"
|
||||
"#define make_uint4 (uint4)\n"
|
||||
"#define make_uint2 (uint2)\n"
|
||||
"#define make_int2 (int2)\n"
|
||||
"\n"
|
||||
"typedef struct\n"
|
||||
"{\n"
|
||||
" union\n"
|
||||
@@ -45,66 +38,54 @@ static const char* fillKernelsCL= \
|
||||
" int m_n;\n"
|
||||
" int m_padding[2];\n"
|
||||
"} ConstBuffer;\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"void FillIntKernel(__global int* dstInt, int num_elements, int value, const int offset)\n"
|
||||
"{\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
"\n"
|
||||
" if( gIdx < num_elements )\n"
|
||||
" {\n"
|
||||
" dstInt[ offset+gIdx ] = value;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"void FillFloatKernel(__global float* dstFloat, int num_elements, float value, const int offset)\n"
|
||||
"{\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
"\n"
|
||||
" if( gIdx < num_elements )\n"
|
||||
" {\n"
|
||||
" dstFloat[ offset+gIdx ] = value;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"void FillUnsignedIntKernel(__global unsigned int* dstInt, const int num, const unsigned int value, const int offset)\n"
|
||||
"{\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
"\n"
|
||||
" if( gIdx < num )\n"
|
||||
" {\n"
|
||||
" dstInt[ offset+gIdx ] = value;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"void FillInt2Kernel(__global int2* dstInt2, const int num, const int2 value, const int offset)\n"
|
||||
"{\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
"\n"
|
||||
" if( gIdx < num )\n"
|
||||
" {\n"
|
||||
" dstInt2[ gIdx + offset] = make_int2( value.x, value.y );\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"void FillInt4Kernel(__global int4* dstInt4, const int num, const int4 value, const int offset)\n"
|
||||
"{\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
"\n"
|
||||
" if( gIdx < num )\n"
|
||||
" {\n"
|
||||
" dstInt4[ offset+gIdx ] = value;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
;
|
||||
|
||||
@@ -2,33 +2,27 @@
|
||||
static const char* prefixScanKernelsCL= \
|
||||
"/*\n"
|
||||
"Copyright (c) 2012 Advanced Micro Devices, Inc. \n"
|
||||
"\n"
|
||||
"This software is provided 'as-is', without any express or implied warranty.\n"
|
||||
"In no event will the authors be held liable for any damages arising from the use of this software.\n"
|
||||
"Permission is granted to anyone to use this software for any purpose, \n"
|
||||
"including commercial applications, and to alter it and redistribute it freely, \n"
|
||||
"subject to the following restrictions:\n"
|
||||
"\n"
|
||||
"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
|
||||
"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
|
||||
"3. This notice may not be removed or altered from any source distribution.\n"
|
||||
"*/\n"
|
||||
"//Originally written by Takahiro Harada\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"typedef unsigned int u32;\n"
|
||||
"#define GET_GROUP_IDX get_group_id(0)\n"
|
||||
"#define GET_LOCAL_IDX get_local_id(0)\n"
|
||||
"#define GET_GLOBAL_IDX get_global_id(0)\n"
|
||||
"#define GET_GROUP_SIZE get_local_size(0)\n"
|
||||
"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
|
||||
"\n"
|
||||
"// takahiro end\n"
|
||||
"#define WG_SIZE 128 \n"
|
||||
"#define m_numElems x\n"
|
||||
"#define m_numBlocks y\n"
|
||||
"#define m_numScanBlocks z\n"
|
||||
"\n"
|
||||
"/*typedef struct\n"
|
||||
"{\n"
|
||||
" uint m_numElems;\n"
|
||||
@@ -37,7 +31,6 @@ static const char* prefixScanKernelsCL= \
|
||||
" uint m_padding[1];\n"
|
||||
"} ConstBuffer;\n"
|
||||
"*/\n"
|
||||
"\n"
|
||||
"u32 ScanExclusive(__local u32* data, u32 n, int lIdx, int lSize)\n"
|
||||
"{\n"
|
||||
" u32 blocksum;\n"
|
||||
@@ -52,17 +45,13 @@ static const char* prefixScanKernelsCL= \
|
||||
" data[bi] += data[ai];\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" if( lIdx == 0 )\n"
|
||||
" {\n"
|
||||
" blocksum = data[ n-1 ];\n"
|
||||
" data[ n-1 ] = 0;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" offset >>= 1;\n"
|
||||
" for(int nActive=1; nActive<n; nActive<<=1, offset>>=1 )\n"
|
||||
" {\n"
|
||||
@@ -77,27 +66,20 @@ static const char* prefixScanKernelsCL= \
|
||||
" }\n"
|
||||
" }\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" return blocksum;\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void LocalScanKernel(__global u32* dst, __global u32 *src, __global u32 *sumBuffer,\n"
|
||||
" uint4 cb)\n"
|
||||
"{\n"
|
||||
" __local u32 ldsData[WG_SIZE*2];\n"
|
||||
"\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
" int lIdx = GET_LOCAL_IDX;\n"
|
||||
"\n"
|
||||
" ldsData[2*lIdx] = ( 2*gIdx < cb.m_numElems )? src[2*gIdx]: 0;\n"
|
||||
" ldsData[2*lIdx + 1] = ( 2*gIdx+1 < cb.m_numElems )? src[2*gIdx + 1]: 0;\n"
|
||||
"\n"
|
||||
" u32 sum = ScanExclusive(ldsData, WG_SIZE*2, GET_LOCAL_IDX, GET_GROUP_SIZE);\n"
|
||||
"\n"
|
||||
" if( lIdx == 0 ) sumBuffer[GET_GROUP_IDX] = sum;\n"
|
||||
"\n"
|
||||
" if( (2*gIdx) < cb.m_numElems )\n"
|
||||
" {\n"
|
||||
" dst[2*gIdx] = ldsData[2*lIdx];\n"
|
||||
@@ -107,25 +89,20 @@ static const char* prefixScanKernelsCL= \
|
||||
" dst[2*gIdx + 1] = ldsData[2*lIdx + 1];\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void AddOffsetKernel(__global u32 *dst, __global u32 *blockSum, uint4 cb)\n"
|
||||
"{\n"
|
||||
" const u32 blockSize = WG_SIZE*2;\n"
|
||||
"\n"
|
||||
" int myIdx = GET_GROUP_IDX+1;\n"
|
||||
" int lIdx = GET_LOCAL_IDX;\n"
|
||||
"\n"
|
||||
" u32 iBlockSum = blockSum[myIdx];\n"
|
||||
"\n"
|
||||
" int endValue = min((myIdx+1)*(blockSize), cb.m_numElems);\n"
|
||||
" for(int i=myIdx*blockSize+lIdx; i<endValue; i+=GET_GROUP_SIZE)\n"
|
||||
" {\n"
|
||||
" dst[i] += iBlockSum;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void TopLevelScanKernel(__global u32* dst, uint4 cb)\n"
|
||||
@@ -134,21 +111,16 @@ static const char* prefixScanKernelsCL= \
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
" int lIdx = GET_LOCAL_IDX;\n"
|
||||
" int lSize = GET_GROUP_SIZE;\n"
|
||||
"\n"
|
||||
" for(int i=lIdx; i<cb.m_numScanBlocks; i+=lSize )\n"
|
||||
" {\n"
|
||||
" ldsData[i] = (i<cb.m_numBlocks)? dst[i]:0;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" u32 sum = ScanExclusive(ldsData, cb.m_numScanBlocks, GET_LOCAL_IDX, GET_GROUP_SIZE);\n"
|
||||
"\n"
|
||||
" for(int i=lIdx; i<cb.m_numBlocks; i+=lSize )\n"
|
||||
" {\n"
|
||||
" dst[i] = ldsData[i];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" if( gIdx == 0 )\n"
|
||||
" {\n"
|
||||
" dst[cb.m_numBlocks] = sum;\n"
|
||||
|
||||
@@ -2,33 +2,27 @@
|
||||
static const char* prefixScanKernelsFloat4CL= \
|
||||
"/*\n"
|
||||
"Copyright (c) 2012 Advanced Micro Devices, Inc. \n"
|
||||
"\n"
|
||||
"This software is provided 'as-is', without any express or implied warranty.\n"
|
||||
"In no event will the authors be held liable for any damages arising from the use of this software.\n"
|
||||
"Permission is granted to anyone to use this software for any purpose, \n"
|
||||
"including commercial applications, and to alter it and redistribute it freely, \n"
|
||||
"subject to the following restrictions:\n"
|
||||
"\n"
|
||||
"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
|
||||
"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
|
||||
"3. This notice may not be removed or altered from any source distribution.\n"
|
||||
"*/\n"
|
||||
"//Originally written by Takahiro Harada\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"typedef unsigned int u32;\n"
|
||||
"#define GET_GROUP_IDX get_group_id(0)\n"
|
||||
"#define GET_LOCAL_IDX get_local_id(0)\n"
|
||||
"#define GET_GLOBAL_IDX get_global_id(0)\n"
|
||||
"#define GET_GROUP_SIZE get_local_size(0)\n"
|
||||
"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
|
||||
"\n"
|
||||
"// takahiro end\n"
|
||||
"#define WG_SIZE 128 \n"
|
||||
"#define m_numElems x\n"
|
||||
"#define m_numBlocks y\n"
|
||||
"#define m_numScanBlocks z\n"
|
||||
"\n"
|
||||
"/*typedef struct\n"
|
||||
"{\n"
|
||||
" uint m_numElems;\n"
|
||||
@@ -37,7 +31,6 @@ static const char* prefixScanKernelsFloat4CL= \
|
||||
" uint m_padding[1];\n"
|
||||
"} ConstBuffer;\n"
|
||||
"*/\n"
|
||||
"\n"
|
||||
"float4 ScanExclusiveFloat4(__local float4* data, u32 n, int lIdx, int lSize)\n"
|
||||
"{\n"
|
||||
" float4 blocksum;\n"
|
||||
@@ -52,17 +45,13 @@ static const char* prefixScanKernelsFloat4CL= \
|
||||
" data[bi] += data[ai];\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" if( lIdx == 0 )\n"
|
||||
" {\n"
|
||||
" blocksum = data[ n-1 ];\n"
|
||||
" data[ n-1 ] = 0;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" offset >>= 1;\n"
|
||||
" for(int nActive=1; nActive<n; nActive<<=1, offset>>=1 )\n"
|
||||
" {\n"
|
||||
@@ -77,27 +66,20 @@ static const char* prefixScanKernelsFloat4CL= \
|
||||
" }\n"
|
||||
" }\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" return blocksum;\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void LocalScanKernel(__global float4* dst, __global float4* src, __global float4* sumBuffer, uint4 cb)\n"
|
||||
"{\n"
|
||||
" __local float4 ldsData[WG_SIZE*2];\n"
|
||||
"\n"
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
" int lIdx = GET_LOCAL_IDX;\n"
|
||||
"\n"
|
||||
" ldsData[2*lIdx] = ( 2*gIdx < cb.m_numElems )? src[2*gIdx]: 0;\n"
|
||||
" ldsData[2*lIdx + 1] = ( 2*gIdx+1 < cb.m_numElems )? src[2*gIdx + 1]: 0;\n"
|
||||
"\n"
|
||||
" float4 sum = ScanExclusiveFloat4(ldsData, WG_SIZE*2, GET_LOCAL_IDX, GET_GROUP_SIZE);\n"
|
||||
"\n"
|
||||
" if( lIdx == 0 ) \n"
|
||||
" sumBuffer[GET_GROUP_IDX] = sum;\n"
|
||||
"\n"
|
||||
" if( (2*gIdx) < cb.m_numElems )\n"
|
||||
" {\n"
|
||||
" dst[2*gIdx] = ldsData[2*lIdx];\n"
|
||||
@@ -107,25 +89,20 @@ static const char* prefixScanKernelsFloat4CL= \
|
||||
" dst[2*gIdx + 1] = ldsData[2*lIdx + 1];\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void AddOffsetKernel(__global float4* dst, __global float4* blockSum, uint4 cb)\n"
|
||||
"{\n"
|
||||
" const u32 blockSize = WG_SIZE*2;\n"
|
||||
"\n"
|
||||
" int myIdx = GET_GROUP_IDX+1;\n"
|
||||
" int lIdx = GET_LOCAL_IDX;\n"
|
||||
"\n"
|
||||
" float4 iBlockSum = blockSum[myIdx];\n"
|
||||
"\n"
|
||||
" int endValue = min((myIdx+1)*(blockSize), cb.m_numElems);\n"
|
||||
" for(int i=myIdx*blockSize+lIdx; i<endValue; i+=GET_GROUP_SIZE)\n"
|
||||
" {\n"
|
||||
" dst[i] += iBlockSum;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"__kernel\n"
|
||||
"void TopLevelScanKernel(__global float4* dst, uint4 cb)\n"
|
||||
@@ -134,21 +111,16 @@ static const char* prefixScanKernelsFloat4CL= \
|
||||
" int gIdx = GET_GLOBAL_IDX;\n"
|
||||
" int lIdx = GET_LOCAL_IDX;\n"
|
||||
" int lSize = GET_GROUP_SIZE;\n"
|
||||
"\n"
|
||||
" for(int i=lIdx; i<cb.m_numScanBlocks; i+=lSize )\n"
|
||||
" {\n"
|
||||
" ldsData[i] = (i<cb.m_numBlocks)? dst[i]:0;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" float4 sum = ScanExclusiveFloat4(ldsData, cb.m_numScanBlocks, GET_LOCAL_IDX, GET_GROUP_SIZE);\n"
|
||||
"\n"
|
||||
" for(int i=lIdx; i<cb.m_numBlocks; i+=lSize )\n"
|
||||
" {\n"
|
||||
" dst[i] = ldsData[i];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" if( gIdx == 0 )\n"
|
||||
" {\n"
|
||||
" dst[cb.m_numBlocks] = sum;\n"
|
||||
|
||||
@@ -3,24 +3,19 @@ static const char* radixSort32KernelsCL= \
|
||||
"/*\n"
|
||||
"Bullet Continuous Collision Detection and Physics Library\n"
|
||||
"Copyright (c) 2011 Advanced Micro Devices, Inc. http://bulletphysics.org\n"
|
||||
"\n"
|
||||
"This software is provided 'as-is', without any express or implied warranty.\n"
|
||||
"In no event will the authors be held liable for any damages arising from the use of this software.\n"
|
||||
"Permission is granted to anyone to use this software for any purpose, \n"
|
||||
"including commercial applications, and to alter it and redistribute it freely, \n"
|
||||
"subject to the following restrictions:\n"
|
||||
"\n"
|
||||
"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
|
||||
"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
|
||||
"3. This notice may not be removed or altered from any source distribution.\n"
|
||||
"*/\n"
|
||||
"//Author Takahiro Harada\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"//#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
|
||||
"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
|
||||
"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
|
||||
"\n"
|
||||
"typedef unsigned int u32;\n"
|
||||
"#define GET_GROUP_IDX get_group_id(0)\n"
|
||||
"#define GET_LOCAL_IDX get_local_id(0)\n"
|
||||
@@ -31,38 +26,27 @@ static const char* radixSort32KernelsCL= \
|
||||
"#define AtomInc(x) atom_inc(&(x))\n"
|
||||
"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
|
||||
"#define AtomAdd(x, value) atom_add(&(x), value)\n"
|
||||
"\n"
|
||||
"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"#define make_uint4 (uint4)\n"
|
||||
"#define make_uint2 (uint2)\n"
|
||||
"#define make_int2 (int2)\n"
|
||||
"\n"
|
||||
"#define WG_SIZE 64\n"
|
||||
"#define ELEMENTS_PER_WORK_ITEM (256/WG_SIZE)\n"
|
||||
"#define BITS_PER_PASS 4\n"
|
||||
"#define NUM_BUCKET (1<<BITS_PER_PASS)\n"
|
||||
"typedef uchar u8;\n"
|
||||
"\n"
|
||||
"// this isn't optimization for VLIW. But just reducing writes. \n"
|
||||
"#define USE_2LEVEL_REDUCE 1\n"
|
||||
"\n"
|
||||
"//#define CHECK_BOUNDARY 1\n"
|
||||
"\n"
|
||||
"//#define NV_GPU 1\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"// Cypress\n"
|
||||
"#define nPerWI 16\n"
|
||||
"// Cayman\n"
|
||||
"//#define nPerWI 20\n"
|
||||
"\n"
|
||||
"#define m_n x\n"
|
||||
"#define m_nWGs y\n"
|
||||
"#define m_startBit z\n"
|
||||
"#define m_nBlocksPerWG w\n"
|
||||
"\n"
|
||||
"/*\n"
|
||||
"typedef struct\n"
|
||||
"{\n"
|
||||
@@ -72,14 +56,11 @@ static const char* radixSort32KernelsCL= \
|
||||
" int m_nBlocksPerWG;\n"
|
||||
"} ConstBuffer;\n"
|
||||
"*/\n"
|
||||
"\n"
|
||||
"typedef struct\n"
|
||||
"{\n"
|
||||
" unsigned int m_key;\n"
|
||||
" unsigned int m_value;\n"
|
||||
"} SortDataCL;\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"uint prefixScanVectorEx( uint4* data )\n"
|
||||
"{\n"
|
||||
" u32 sum = 0;\n"
|
||||
@@ -97,16 +78,13 @@ static const char* radixSort32KernelsCL= \
|
||||
" sum += tmp;\n"
|
||||
" return sum;\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"u32 localPrefixSum( u32 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory, int wgSize /*64 or 128*/ )\n"
|
||||
"{\n"
|
||||
" { // Set data\n"
|
||||
" sorterSharedMemory[lIdx] = 0;\n"
|
||||
" sorterSharedMemory[lIdx+wgSize] = pData;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" { // Prefix sum\n"
|
||||
" int idx = 2*lIdx + (wgSize+1);\n"
|
||||
"#if defined(USE_2LEVEL_REDUCE)\n"
|
||||
@@ -118,13 +96,11 @@ static const char* radixSort32KernelsCL= \
|
||||
" u2 = sorterSharedMemory[idx-1];\n"
|
||||
" AtomAdd( sorterSharedMemory[idx], u0+u1+u2 ); \n"
|
||||
" GROUP_MEM_FENCE;\n"
|
||||
"\n"
|
||||
" u0 = sorterSharedMemory[idx-12];\n"
|
||||
" u1 = sorterSharedMemory[idx-8];\n"
|
||||
" u2 = sorterSharedMemory[idx-4];\n"
|
||||
" AtomAdd( sorterSharedMemory[idx], u0+u1+u2 ); \n"
|
||||
" GROUP_MEM_FENCE;\n"
|
||||
"\n"
|
||||
" u0 = sorterSharedMemory[idx-48];\n"
|
||||
" u1 = sorterSharedMemory[idx-32];\n"
|
||||
" u2 = sorterSharedMemory[idx-16];\n"
|
||||
@@ -135,7 +111,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" sorterSharedMemory[idx] += sorterSharedMemory[idx-64];\n"
|
||||
" GROUP_MEM_FENCE;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" sorterSharedMemory[idx-1] += sorterSharedMemory[idx-2];\n"
|
||||
" GROUP_MEM_FENCE;\n"
|
||||
" }\n"
|
||||
@@ -159,20 +134,16 @@ static const char* radixSort32KernelsCL= \
|
||||
" sorterSharedMemory[idx] += sorterSharedMemory[idx-64];\n"
|
||||
" GROUP_MEM_FENCE;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" sorterSharedMemory[idx-1] += sorterSharedMemory[idx-2];\n"
|
||||
" GROUP_MEM_FENCE;\n"
|
||||
" }\n"
|
||||
"#endif\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" *totalSum = sorterSharedMemory[wgSize*2-1];\n"
|
||||
" u32 addValue = sorterSharedMemory[lIdx+wgSize-1];\n"
|
||||
" return addValue;\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"//__attribute__((reqd_work_group_size(128,1,1)))\n"
|
||||
"uint4 localPrefixSum128V( uint4 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory )\n"
|
||||
"{\n"
|
||||
@@ -180,8 +151,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" u32 rank = localPrefixSum( s4, lIdx, totalSum, sorterSharedMemory, 128 );\n"
|
||||
" return pData + make_uint4( rank, rank, rank, rank );\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"//__attribute__((reqd_work_group_size(64,1,1)))\n"
|
||||
"uint4 localPrefixSum64V( uint4 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory )\n"
|
||||
"{\n"
|
||||
@@ -189,28 +158,18 @@ static const char* radixSort32KernelsCL= \
|
||||
" u32 rank = localPrefixSum( s4, lIdx, totalSum, sorterSharedMemory, 64 );\n"
|
||||
" return pData + make_uint4( rank, rank, rank, rank );\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"u32 unpack4Key( u32 key, int keyIdx ){ return (key>>(keyIdx*8)) & 0xff;}\n"
|
||||
"\n"
|
||||
"u32 bit8Scan(u32 v)\n"
|
||||
"{\n"
|
||||
" return (v<<8) + (v<<16) + (v<<24);\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"//===\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"#define MY_HISTOGRAM(idx) localHistogramMat[(idx)*WG_SIZE+lIdx]\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"void StreamCountKernel( __global u32* gSrc, __global u32* histogramOut, int4 cb )\n"
|
||||
"{\n"
|
||||
" __local u32 localHistogramMat[NUM_BUCKET*WG_SIZE];\n"
|
||||
"\n"
|
||||
" u32 gIdx = GET_GLOBAL_IDX;\n"
|
||||
" u32 lIdx = GET_LOCAL_IDX;\n"
|
||||
" u32 wgIdx = GET_GROUP_IDX;\n"
|
||||
@@ -219,21 +178,15 @@ static const char* radixSort32KernelsCL= \
|
||||
" const int n = cb.m_n;\n"
|
||||
" const int nWGs = cb.m_nWGs;\n"
|
||||
" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
|
||||
"\n"
|
||||
" for(int i=0; i<NUM_BUCKET; i++)\n"
|
||||
" {\n"
|
||||
" MY_HISTOGRAM(i) = 0;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
|
||||
" u32 localKey;\n"
|
||||
"\n"
|
||||
" int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n"
|
||||
"\n"
|
||||
" int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
|
||||
"\n"
|
||||
" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n"
|
||||
" {\n"
|
||||
" // MY_HISTOGRAM( localKeys.x ) ++ is much expensive than atomic add as it requires read and write while atomics can just add on AMD\n"
|
||||
@@ -254,7 +207,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" }\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" \n"
|
||||
" if( lIdx < NUM_BUCKET )\n"
|
||||
@@ -267,13 +219,11 @@ static const char* radixSort32KernelsCL= \
|
||||
" histogramOut[lIdx*nWGs+wgIdx] = sum;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"void StreamCountSortDataKernel( __global SortDataCL* gSrc, __global u32* histogramOut, int4 cb )\n"
|
||||
"{\n"
|
||||
" __local u32 localHistogramMat[NUM_BUCKET*WG_SIZE];\n"
|
||||
"\n"
|
||||
" u32 gIdx = GET_GLOBAL_IDX;\n"
|
||||
" u32 lIdx = GET_LOCAL_IDX;\n"
|
||||
" u32 wgIdx = GET_GROUP_IDX;\n"
|
||||
@@ -282,21 +232,15 @@ static const char* radixSort32KernelsCL= \
|
||||
" const int n = cb.m_n;\n"
|
||||
" const int nWGs = cb.m_nWGs;\n"
|
||||
" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
|
||||
"\n"
|
||||
" for(int i=0; i<NUM_BUCKET; i++)\n"
|
||||
" {\n"
|
||||
" MY_HISTOGRAM(i) = 0;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
|
||||
" u32 localKey;\n"
|
||||
"\n"
|
||||
" int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n"
|
||||
"\n"
|
||||
" int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
|
||||
"\n"
|
||||
" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n"
|
||||
" {\n"
|
||||
" // MY_HISTOGRAM( localKeys.x ) ++ is much expensive than atomic add as it requires read and write while atomics can just add on AMD\n"
|
||||
@@ -317,7 +261,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" }\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" \n"
|
||||
" if( lIdx < NUM_BUCKET )\n"
|
||||
@@ -330,20 +273,16 @@ static const char* radixSort32KernelsCL= \
|
||||
" histogramOut[lIdx*nWGs+wgIdx] = sum;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"#define nPerLane (nPerWI/4)\n"
|
||||
"\n"
|
||||
"// NUM_BUCKET*nWGs < 128*nPerWI\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(128,1,1)))\n"
|
||||
"void PrefixScanKernel( __global u32* wHistogram1, int4 cb )\n"
|
||||
"{\n"
|
||||
" __local u32 ldsTopScanData[128*2];\n"
|
||||
"\n"
|
||||
" u32 lIdx = GET_LOCAL_IDX;\n"
|
||||
" u32 wgIdx = GET_GROUP_IDX;\n"
|
||||
" const int nWGs = cb.m_nWGs;\n"
|
||||
"\n"
|
||||
" u32 data[nPerWI];\n"
|
||||
" for(int i=0; i<nPerWI; i++)\n"
|
||||
" {\n"
|
||||
@@ -351,9 +290,7 @@ static const char* radixSort32KernelsCL= \
|
||||
" if( (nPerWI*lIdx+i) < NUM_BUCKET*nWGs )\n"
|
||||
" data[i] = wHistogram1[nPerWI*lIdx+i];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" uint4 myData = make_uint4(0,0,0,0);\n"
|
||||
"\n"
|
||||
" for(int i=0; i<nPerLane; i++)\n"
|
||||
" {\n"
|
||||
" myData.x += data[nPerLane*0+i];\n"
|
||||
@@ -361,10 +298,8 @@ static const char* radixSort32KernelsCL= \
|
||||
" myData.z += data[nPerLane*2+i];\n"
|
||||
" myData.w += data[nPerLane*3+i];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" uint totalSum;\n"
|
||||
" uint4 scanned = localPrefixSum128V( myData, lIdx, &totalSum, ldsTopScanData );\n"
|
||||
"\n"
|
||||
"// for(int j=0; j<4; j++) // somehow it introduces a lot of branches\n"
|
||||
" { int j = 0;\n"
|
||||
" u32 sum = 0;\n"
|
||||
@@ -402,7 +337,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" sum += tmp;\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" for(int i=0; i<nPerLane; i++)\n"
|
||||
" {\n"
|
||||
" data[nPerLane*0+i] += scanned.x;\n"
|
||||
@@ -410,7 +344,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" data[nPerLane*2+i] += scanned.z;\n"
|
||||
" data[nPerLane*3+i] += scanned.w;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" for(int i=0; i<nPerWI; i++)\n"
|
||||
" {\n"
|
||||
" int index = nPerWI*lIdx+i;\n"
|
||||
@@ -418,7 +351,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" wHistogram1[nPerWI*lIdx+i] = data[i];\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"// 4 scan, 4 exchange\n"
|
||||
"void sort4Bits(u32 sortData[4], int startBit, int lIdx, __local u32* ldsSortData)\n"
|
||||
"{\n"
|
||||
@@ -433,26 +365,20 @@ static const char* radixSort32KernelsCL= \
|
||||
" uint4 localAddr = make_uint4(lIdx*4+0,lIdx*4+1,lIdx*4+2,lIdx*4+3);\n"
|
||||
" uint4 dstAddr = localAddr - prefixSum + make_uint4( total, total, total, total );\n"
|
||||
" dstAddr = SELECT_UINT4( prefixSum, dstAddr, cmpResult != make_uint4(0, 0, 0, 0) );\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" ldsSortData[dstAddr.x] = sortData[0];\n"
|
||||
" ldsSortData[dstAddr.y] = sortData[1];\n"
|
||||
" ldsSortData[dstAddr.z] = sortData[2];\n"
|
||||
" ldsSortData[dstAddr.w] = sortData[3];\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" sortData[0] = ldsSortData[localAddr.x];\n"
|
||||
" sortData[1] = ldsSortData[localAddr.y];\n"
|
||||
" sortData[2] = ldsSortData[localAddr.z];\n"
|
||||
" sortData[3] = ldsSortData[localAddr.w];\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"// 2 scan, 2 exchange\n"
|
||||
"void sort4Bits1(u32 sortData[4], int startBit, int lIdx, __local u32* ldsSortData)\n"
|
||||
"{\n"
|
||||
@@ -462,7 +388,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" (sortData[1]>>(startBit+ibit)) & 0x3, \n"
|
||||
" (sortData[2]>>(startBit+ibit)) & 0x3, \n"
|
||||
" (sortData[3]>>(startBit+ibit)) & 0x3);\n"
|
||||
"\n"
|
||||
" u32 key4;\n"
|
||||
" u32 sKeyPacked[4] = { 0, 0, 0, 0 };\n"
|
||||
" {\n"
|
||||
@@ -470,22 +395,17 @@ static const char* radixSort32KernelsCL= \
|
||||
" sKeyPacked[1] |= 1<<(8*b.y);\n"
|
||||
" sKeyPacked[2] |= 1<<(8*b.z);\n"
|
||||
" sKeyPacked[3] |= 1<<(8*b.w);\n"
|
||||
"\n"
|
||||
" key4 = sKeyPacked[0] + sKeyPacked[1] + sKeyPacked[2] + sKeyPacked[3];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" u32 rankPacked;\n"
|
||||
" u32 sumPacked;\n"
|
||||
" {\n"
|
||||
" rankPacked = localPrefixSum( key4, lIdx, &sumPacked, ldsSortData, WG_SIZE );\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" u32 newOffset[4] = { 0,0,0,0 };\n"
|
||||
" {\n"
|
||||
" u32 sumScanned = bit8Scan( sumPacked );\n"
|
||||
"\n"
|
||||
" u32 scannedKeys[4];\n"
|
||||
" scannedKeys[0] = 1<<(8*b.x);\n"
|
||||
" scannedKeys[1] = 1<<(8*b.y);\n"
|
||||
@@ -500,7 +420,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" sum4 += tmp;\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" {\n"
|
||||
" u32 sumPlusRank = sumScanned + rankPacked;\n"
|
||||
" { u32 ie = b.x;\n"
|
||||
@@ -521,31 +440,23 @@ static const char* radixSort32KernelsCL= \
|
||||
" }\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" {\n"
|
||||
" ldsSortData[newOffset[0]] = sortData[0];\n"
|
||||
" ldsSortData[newOffset[1]] = sortData[1];\n"
|
||||
" ldsSortData[newOffset[2]] = sortData[2];\n"
|
||||
" ldsSortData[newOffset[3]] = sortData[3];\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" u32 dstAddr = 4*lIdx;\n"
|
||||
" sortData[0] = ldsSortData[dstAddr+0];\n"
|
||||
" sortData[1] = ldsSortData[dstAddr+1];\n"
|
||||
" sortData[2] = ldsSortData[dstAddr+2];\n"
|
||||
" sortData[3] = ldsSortData[dstAddr+3];\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"#define SET_HISTOGRAM(setIdx, key) ldsSortData[(setIdx)*NUM_BUCKET+key]\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"void SortAndScatterKernel( __global const u32* restrict gSrc, __global const u32* rHistogram, __global u32* restrict gDst, int4 cb )\n"
|
||||
@@ -553,34 +464,25 @@ static const char* radixSort32KernelsCL= \
|
||||
" __local u32 ldsSortData[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];\n"
|
||||
" __local u32 localHistogramToCarry[NUM_BUCKET];\n"
|
||||
" __local u32 localHistogram[NUM_BUCKET*2];\n"
|
||||
"\n"
|
||||
" u32 gIdx = GET_GLOBAL_IDX;\n"
|
||||
" u32 lIdx = GET_LOCAL_IDX;\n"
|
||||
" u32 wgIdx = GET_GROUP_IDX;\n"
|
||||
" u32 wgSize = GET_GROUP_SIZE;\n"
|
||||
"\n"
|
||||
" const int n = cb.m_n;\n"
|
||||
" const int nWGs = cb.m_nWGs;\n"
|
||||
" const int startBit = cb.m_startBit;\n"
|
||||
" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
|
||||
"\n"
|
||||
" if( lIdx < (NUM_BUCKET) )\n"
|
||||
" {\n"
|
||||
" localHistogramToCarry[lIdx] = rHistogram[lIdx*nWGs + wgIdx];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
|
||||
"\n"
|
||||
" int nBlocks = n/blockSize - nBlocksPerWG*wgIdx;\n"
|
||||
"\n"
|
||||
" int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
|
||||
"\n"
|
||||
" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n"
|
||||
" {\n"
|
||||
" u32 myHistogram = 0;\n"
|
||||
"\n"
|
||||
" u32 sortData[ELEMENTS_PER_WORK_ITEM];\n"
|
||||
" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
|
||||
"#if defined(CHECK_BOUNDARY)\n"
|
||||
@@ -588,13 +490,10 @@ static const char* radixSort32KernelsCL= \
|
||||
"#else\n"
|
||||
" sortData[i] = gSrc[ addr+i ];\n"
|
||||
"#endif\n"
|
||||
"\n"
|
||||
" sort4Bits(sortData, startBit, lIdx, ldsSortData);\n"
|
||||
"\n"
|
||||
" u32 keys[ELEMENTS_PER_WORK_ITEM];\n"
|
||||
" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
|
||||
" keys[i] = (sortData[i]>>startBit) & 0xf;\n"
|
||||
"\n"
|
||||
" { // create histogram\n"
|
||||
" u32 setIdx = lIdx/16;\n"
|
||||
" if( lIdx < NUM_BUCKET )\n"
|
||||
@@ -603,12 +502,10 @@ static const char* radixSort32KernelsCL= \
|
||||
" }\n"
|
||||
" ldsSortData[lIdx] = 0;\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
|
||||
"#if defined(CHECK_BOUNDARY)\n"
|
||||
" if( addr+i < n )\n"
|
||||
"#endif\n"
|
||||
"\n"
|
||||
"#if defined(NV_GPU)\n"
|
||||
" SET_HISTOGRAM( setIdx, keys[i] )++;\n"
|
||||
"#else\n"
|
||||
@@ -629,13 +526,11 @@ static const char* radixSort32KernelsCL= \
|
||||
" localHistogram[hIdx] = sum;\n"
|
||||
" }\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
"#if defined(USE_2LEVEL_REDUCE)\n"
|
||||
" if( lIdx < NUM_BUCKET )\n"
|
||||
" {\n"
|
||||
" localHistogram[hIdx] = localHistogram[hIdx-1];\n"
|
||||
" GROUP_MEM_FENCE;\n"
|
||||
"\n"
|
||||
" u32 u0, u1, u2;\n"
|
||||
" u0 = localHistogram[hIdx-3];\n"
|
||||
" u1 = localHistogram[hIdx-2];\n"
|
||||
@@ -665,7 +560,6 @@ static const char* radixSort32KernelsCL= \
|
||||
"#endif\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" {\n"
|
||||
" for(int ie=0; ie<ELEMENTS_PER_WORK_ITEM; ie++)\n"
|
||||
" {\n"
|
||||
@@ -679,9 +573,7 @@ static const char* radixSort32KernelsCL= \
|
||||
" gDst[ groupOffset + myIdx ] = sortData[ie];\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" if( lIdx < NUM_BUCKET )\n"
|
||||
" {\n"
|
||||
" localHistogramToCarry[lIdx] += myHistogram;\n"
|
||||
@@ -689,7 +581,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"// 2 scan, 2 exchange\n"
|
||||
"void sort4Bits1KeyValue(u32 sortData[4], int sortVal[4], int startBit, int lIdx, __local u32* ldsSortData, __local int *ldsSortVal)\n"
|
||||
"{\n"
|
||||
@@ -699,7 +590,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" (sortData[1]>>(startBit+ibit)) & 0x3, \n"
|
||||
" (sortData[2]>>(startBit+ibit)) & 0x3, \n"
|
||||
" (sortData[3]>>(startBit+ibit)) & 0x3);\n"
|
||||
"\n"
|
||||
" u32 key4;\n"
|
||||
" u32 sKeyPacked[4] = { 0, 0, 0, 0 };\n"
|
||||
" {\n"
|
||||
@@ -707,22 +597,17 @@ static const char* radixSort32KernelsCL= \
|
||||
" sKeyPacked[1] |= 1<<(8*b.y);\n"
|
||||
" sKeyPacked[2] |= 1<<(8*b.z);\n"
|
||||
" sKeyPacked[3] |= 1<<(8*b.w);\n"
|
||||
"\n"
|
||||
" key4 = sKeyPacked[0] + sKeyPacked[1] + sKeyPacked[2] + sKeyPacked[3];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" u32 rankPacked;\n"
|
||||
" u32 sumPacked;\n"
|
||||
" {\n"
|
||||
" rankPacked = localPrefixSum( key4, lIdx, &sumPacked, ldsSortData, WG_SIZE );\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" u32 newOffset[4] = { 0,0,0,0 };\n"
|
||||
" {\n"
|
||||
" u32 sumScanned = bit8Scan( sumPacked );\n"
|
||||
"\n"
|
||||
" u32 scannedKeys[4];\n"
|
||||
" scannedKeys[0] = 1<<(8*b.x);\n"
|
||||
" scannedKeys[1] = 1<<(8*b.y);\n"
|
||||
@@ -737,7 +622,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" sum4 += tmp;\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" {\n"
|
||||
" u32 sumPlusRank = sumScanned + rankPacked;\n"
|
||||
" { u32 ie = b.x;\n"
|
||||
@@ -758,42 +642,30 @@ static const char* radixSort32KernelsCL= \
|
||||
" }\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" {\n"
|
||||
" ldsSortData[newOffset[0]] = sortData[0];\n"
|
||||
" ldsSortData[newOffset[1]] = sortData[1];\n"
|
||||
" ldsSortData[newOffset[2]] = sortData[2];\n"
|
||||
" ldsSortData[newOffset[3]] = sortData[3];\n"
|
||||
"\n"
|
||||
" ldsSortVal[newOffset[0]] = sortVal[0];\n"
|
||||
" ldsSortVal[newOffset[1]] = sortVal[1];\n"
|
||||
" ldsSortVal[newOffset[2]] = sortVal[2];\n"
|
||||
" ldsSortVal[newOffset[3]] = sortVal[3];\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" u32 dstAddr = 4*lIdx;\n"
|
||||
" sortData[0] = ldsSortData[dstAddr+0];\n"
|
||||
" sortData[1] = ldsSortData[dstAddr+1];\n"
|
||||
" sortData[2] = ldsSortData[dstAddr+2];\n"
|
||||
" sortData[3] = ldsSortData[dstAddr+3];\n"
|
||||
"\n"
|
||||
" sortVal[0] = ldsSortVal[dstAddr+0];\n"
|
||||
" sortVal[1] = ldsSortVal[dstAddr+1];\n"
|
||||
" sortVal[2] = ldsSortVal[dstAddr+2];\n"
|
||||
" sortVal[3] = ldsSortVal[dstAddr+3];\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"void SortAndScatterSortDataKernel( __global const SortDataCL* restrict gSrc, __global const u32* rHistogram, __global SortDataCL* restrict gDst, int4 cb)\n"
|
||||
@@ -802,39 +674,28 @@ static const char* radixSort32KernelsCL= \
|
||||
" __local int ldsSortVal[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];\n"
|
||||
" __local u32 localHistogramToCarry[NUM_BUCKET];\n"
|
||||
" __local u32 localHistogram[NUM_BUCKET*2];\n"
|
||||
"\n"
|
||||
" u32 gIdx = GET_GLOBAL_IDX;\n"
|
||||
" u32 lIdx = GET_LOCAL_IDX;\n"
|
||||
" u32 wgIdx = GET_GROUP_IDX;\n"
|
||||
" u32 wgSize = GET_GROUP_SIZE;\n"
|
||||
"\n"
|
||||
" const int n = cb.m_n;\n"
|
||||
" const int nWGs = cb.m_nWGs;\n"
|
||||
" const int startBit = cb.m_startBit;\n"
|
||||
" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
|
||||
"\n"
|
||||
" if( lIdx < (NUM_BUCKET) )\n"
|
||||
" {\n"
|
||||
" localHistogramToCarry[lIdx] = rHistogram[lIdx*nWGs + wgIdx];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" \n"
|
||||
"\n"
|
||||
" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
|
||||
"\n"
|
||||
" int nBlocks = n/blockSize - nBlocksPerWG*wgIdx;\n"
|
||||
"\n"
|
||||
" int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
|
||||
"\n"
|
||||
" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n"
|
||||
" {\n"
|
||||
"\n"
|
||||
" u32 myHistogram = 0;\n"
|
||||
"\n"
|
||||
" int sortData[ELEMENTS_PER_WORK_ITEM];\n"
|
||||
" int sortVal[ELEMENTS_PER_WORK_ITEM];\n"
|
||||
"\n"
|
||||
" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
|
||||
"#if defined(CHECK_BOUNDARY)\n"
|
||||
" {\n"
|
||||
@@ -847,13 +708,10 @@ static const char* radixSort32KernelsCL= \
|
||||
" sortVal[i] = gSrc[ addr+i ].m_value;\n"
|
||||
" }\n"
|
||||
"#endif\n"
|
||||
"\n"
|
||||
" sort4Bits1KeyValue(sortData, sortVal, startBit, lIdx, ldsSortData, ldsSortVal);\n"
|
||||
"\n"
|
||||
" u32 keys[ELEMENTS_PER_WORK_ITEM];\n"
|
||||
" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
|
||||
" keys[i] = (sortData[i]>>startBit) & 0xf;\n"
|
||||
"\n"
|
||||
" { // create histogram\n"
|
||||
" u32 setIdx = lIdx/16;\n"
|
||||
" if( lIdx < NUM_BUCKET )\n"
|
||||
@@ -862,12 +720,10 @@ static const char* radixSort32KernelsCL= \
|
||||
" }\n"
|
||||
" ldsSortData[lIdx] = 0;\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
|
||||
"#if defined(CHECK_BOUNDARY)\n"
|
||||
" if( addr+i < n )\n"
|
||||
"#endif\n"
|
||||
"\n"
|
||||
"#if defined(NV_GPU)\n"
|
||||
" SET_HISTOGRAM( setIdx, keys[i] )++;\n"
|
||||
"#else\n"
|
||||
@@ -888,13 +744,11 @@ static const char* radixSort32KernelsCL= \
|
||||
" localHistogram[hIdx] = sum;\n"
|
||||
" }\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
"#if defined(USE_2LEVEL_REDUCE)\n"
|
||||
" if( lIdx < NUM_BUCKET )\n"
|
||||
" {\n"
|
||||
" localHistogram[hIdx] = localHistogram[hIdx-1];\n"
|
||||
" GROUP_MEM_FENCE;\n"
|
||||
"\n"
|
||||
" u32 u0, u1, u2;\n"
|
||||
" u0 = localHistogram[hIdx-3];\n"
|
||||
" u1 = localHistogram[hIdx-2];\n"
|
||||
@@ -924,7 +778,6 @@ static const char* radixSort32KernelsCL= \
|
||||
"#endif\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" {\n"
|
||||
" for(int ie=0; ie<ELEMENTS_PER_WORK_ITEM; ie++)\n"
|
||||
" {\n"
|
||||
@@ -958,9 +811,7 @@ static const char* radixSort32KernelsCL= \
|
||||
"#endif\n"
|
||||
" }\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
"\n"
|
||||
" if( lIdx < NUM_BUCKET )\n"
|
||||
" {\n"
|
||||
" localHistogramToCarry[lIdx] += myHistogram;\n"
|
||||
@@ -968,13 +819,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" GROUP_LDS_BARRIER;\n"
|
||||
" }\n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"void SortAndScatterSortDataKernelSerial( __global const SortDataCL* restrict gSrc, __global const u32* rHistogram, __global SortDataCL* restrict gDst, int4 cb)\n"
|
||||
@@ -988,7 +832,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" const int n = cb.m_n;\n"
|
||||
" const int nWGs = cb.m_nWGs;\n"
|
||||
" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
|
||||
"\n"
|
||||
" int counter[NUM_BUCKET];\n"
|
||||
" \n"
|
||||
" if (realLocalIdx>0)\n"
|
||||
@@ -996,11 +839,9 @@ static const char* radixSort32KernelsCL= \
|
||||
" \n"
|
||||
" for (int c=0;c<NUM_BUCKET;c++)\n"
|
||||
" counter[c]=0;\n"
|
||||
"\n"
|
||||
" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
|
||||
" \n"
|
||||
" int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n"
|
||||
"\n"
|
||||
" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++)\n"
|
||||
" {\n"
|
||||
" for (int lIdx=0;lIdx<WG_SIZE;lIdx++)\n"
|
||||
@@ -1022,8 +863,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" }\n"
|
||||
" \n"
|
||||
"}\n"
|
||||
"\n"
|
||||
"\n"
|
||||
"__kernel\n"
|
||||
"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
|
||||
"void SortAndScatterKernelSerial( __global const u32* restrict gSrc, __global const u32* rHistogram, __global u32* restrict gDst, int4 cb )\n"
|
||||
@@ -1037,7 +876,6 @@ static const char* radixSort32KernelsCL= \
|
||||
" const int n = cb.m_n;\n"
|
||||
" const int nWGs = cb.m_nWGs;\n"
|
||||
" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
|
||||
"\n"
|
||||
" int counter[NUM_BUCKET];\n"
|
||||
" \n"
|
||||
" if (realLocalIdx>0)\n"
|
||||
@@ -1045,11 +883,9 @@ static const char* radixSort32KernelsCL= \
|
||||
" \n"
|
||||
" for (int c=0;c<NUM_BUCKET;c++)\n"
|
||||
" counter[c]=0;\n"
|
||||
"\n"
|
||||
" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
|
||||
" \n"
|
||||
" int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n"
|
||||
"\n"
|
||||
" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++)\n"
|
||||
" {\n"
|
||||
" for (int lIdx=0;lIdx<WG_SIZE;lIdx++)\n"
|
||||
|
||||
Reference in New Issue
Block a user