diff options
Diffstat (limited to 'thirdparty/bullet/src/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32KernelsCL.h')
| -rw-r--r-- | thirdparty/bullet/src/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32KernelsCL.h | 910 |
1 files changed, 910 insertions, 0 deletions
diff --git a/thirdparty/bullet/src/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32KernelsCL.h b/thirdparty/bullet/src/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32KernelsCL.h new file mode 100644 index 0000000000..8876c16aa6 --- /dev/null +++ b/thirdparty/bullet/src/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32KernelsCL.h @@ -0,0 +1,910 @@ +//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project +static const char* radixSort32KernelsCL= \ +"/*\n" +"Bullet Continuous Collision Detection and Physics Library\n" +"Copyright (c) 2011 Advanced Micro Devices, Inc. http://bulletphysics.org\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" +"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" +"//#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" +"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" +"#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" +"#define AtomAdd(x, value) atom_add(&(x), value)\n" +"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n" +"#define make_uint4 (uint4)\n" +"#define make_uint2 (uint2)\n" +"#define make_int2 (int2)\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" +"// this isn't optimization for VLIW. But just reducing writes. \n" +"#define USE_2LEVEL_REDUCE 1\n" +"//#define CHECK_BOUNDARY 1\n" +"//#define NV_GPU 1\n" +"// Cypress\n" +"#define nPerWI 16\n" +"// Cayman\n" +"//#define nPerWI 20\n" +"#define m_n x\n" +"#define m_nWGs y\n" +"#define m_startBit z\n" +"#define m_nBlocksPerWG w\n" +"/*\n" +"typedef struct\n" +"{\n" +" int m_n;\n" +" int m_nWGs;\n" +" int m_startBit;\n" +" int m_nBlocksPerWG;\n" +"} ConstBuffer;\n" +"*/\n" +"typedef struct\n" +"{\n" +" unsigned int m_key;\n" +" unsigned int m_value;\n" +"} SortDataCL;\n" +"uint prefixScanVectorEx( uint4* data )\n" +"{\n" +" u32 sum = 0;\n" +" u32 tmp = data[0].x;\n" +" data[0].x = sum;\n" +" sum += tmp;\n" +" tmp = data[0].y;\n" +" data[0].y = sum;\n" +" sum += tmp;\n" +" tmp = data[0].z;\n" +" data[0].z = sum;\n" +" sum += tmp;\n" +" tmp = data[0].w;\n" +" data[0].w = sum;\n" +" sum += tmp;\n" +" return sum;\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" +" GROUP_LDS_BARRIER;\n" +" { // Prefix sum\n" +" int idx = 2*lIdx + (wgSize+1);\n" +"#if defined(USE_2LEVEL_REDUCE)\n" +" if( lIdx < 64 )\n" +" {\n" +" u32 u0, u1, u2;\n" +" u0 = sorterSharedMemory[idx-3];\n" +" u1 = sorterSharedMemory[idx-2];\n" +" u2 = sorterSharedMemory[idx-1];\n" +" AtomAdd( sorterSharedMemory[idx], u0+u1+u2 ); \n" +" GROUP_MEM_FENCE;\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" +" u0 = sorterSharedMemory[idx-48];\n" +" u1 = sorterSharedMemory[idx-32];\n" +" u2 = sorterSharedMemory[idx-16];\n" +" AtomAdd( sorterSharedMemory[idx], u0+u1+u2 ); \n" +" GROUP_MEM_FENCE;\n" +" if( wgSize > 64 )\n" +" {\n" +" sorterSharedMemory[idx] += sorterSharedMemory[idx-64];\n" +" GROUP_MEM_FENCE;\n" +" }\n" +" sorterSharedMemory[idx-1] += sorterSharedMemory[idx-2];\n" +" GROUP_MEM_FENCE;\n" +" }\n" +"#else\n" +" if( lIdx < 64 )\n" +" {\n" +" sorterSharedMemory[idx] += sorterSharedMemory[idx-1];\n" +" GROUP_MEM_FENCE;\n" +" sorterSharedMemory[idx] += sorterSharedMemory[idx-2]; \n" +" GROUP_MEM_FENCE;\n" +" sorterSharedMemory[idx] += sorterSharedMemory[idx-4];\n" +" GROUP_MEM_FENCE;\n" +" sorterSharedMemory[idx] += sorterSharedMemory[idx-8];\n" +" GROUP_MEM_FENCE;\n" +" sorterSharedMemory[idx] += sorterSharedMemory[idx-16];\n" +" GROUP_MEM_FENCE;\n" +" sorterSharedMemory[idx] += sorterSharedMemory[idx-32];\n" +" GROUP_MEM_FENCE;\n" +" if( wgSize > 64 )\n" +" {\n" +" sorterSharedMemory[idx] += sorterSharedMemory[idx-64];\n" +" GROUP_MEM_FENCE;\n" +" }\n" +" sorterSharedMemory[idx-1] += sorterSharedMemory[idx-2];\n" +" GROUP_MEM_FENCE;\n" +" }\n" +"#endif\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" *totalSum = sorterSharedMemory[wgSize*2-1];\n" +" u32 addValue = sorterSharedMemory[lIdx+wgSize-1];\n" +" return addValue;\n" +"}\n" +"//__attribute__((reqd_work_group_size(128,1,1)))\n" +"uint4 localPrefixSum128V( uint4 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory )\n" +"{\n" +" u32 s4 = prefixScanVectorEx( &pData );\n" +" u32 rank = localPrefixSum( s4, lIdx, totalSum, sorterSharedMemory, 128 );\n" +" return pData + make_uint4( rank, rank, rank, rank );\n" +"}\n" +"//__attribute__((reqd_work_group_size(64,1,1)))\n" +"uint4 localPrefixSum64V( uint4 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory )\n" +"{\n" +" u32 s4 = prefixScanVectorEx( &pData );\n" +" u32 rank = localPrefixSum( s4, lIdx, totalSum, sorterSharedMemory, 64 );\n" +" return pData + make_uint4( rank, rank, rank, rank );\n" +"}\n" +"u32 unpack4Key( u32 key, int keyIdx ){ return (key>>(keyIdx*8)) & 0xff;}\n" +"u32 bit8Scan(u32 v)\n" +"{\n" +" return (v<<8) + (v<<16) + (v<<24);\n" +"}\n" +"//===\n" +"#define MY_HISTOGRAM(idx) localHistogramMat[(idx)*WG_SIZE+lIdx]\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" +" u32 gIdx = GET_GLOBAL_IDX;\n" +" u32 lIdx = GET_LOCAL_IDX;\n" +" u32 wgIdx = GET_GROUP_IDX;\n" +" u32 wgSize = GET_GROUP_SIZE;\n" +" const int startBit = cb.m_startBit;\n" +" const int n = cb.m_n;\n" +" const int nWGs = cb.m_nWGs;\n" +" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n" +" for(int i=0; i<NUM_BUCKET; i++)\n" +" {\n" +" MY_HISTOGRAM(i) = 0;\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n" +" u32 localKey;\n" +" int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n" +" int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\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" +" // Using registers didn't perform well. It seems like use localKeys to address requires a lot of alu ops\n" +" // AMD: AtomInc performs better while NV prefers ++\n" +" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n" +" {\n" +"#if defined(CHECK_BOUNDARY)\n" +" if( addr+i < n )\n" +"#endif\n" +" {\n" +" localKey = (gSrc[addr+i]>>startBit) & 0xf;\n" +"#if defined(NV_GPU)\n" +" MY_HISTOGRAM( localKey )++;\n" +"#else\n" +" AtomInc( MY_HISTOGRAM( localKey ) );\n" +"#endif\n" +" }\n" +" }\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" \n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" u32 sum = 0;\n" +" for(int i=0; i<GET_GROUP_SIZE; i++)\n" +" {\n" +" sum += localHistogramMat[lIdx*WG_SIZE+(i+lIdx)%GET_GROUP_SIZE];\n" +" }\n" +" histogramOut[lIdx*nWGs+wgIdx] = sum;\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" +" u32 gIdx = GET_GLOBAL_IDX;\n" +" u32 lIdx = GET_LOCAL_IDX;\n" +" u32 wgIdx = GET_GROUP_IDX;\n" +" u32 wgSize = GET_GROUP_SIZE;\n" +" const int startBit = cb.m_startBit;\n" +" const int n = cb.m_n;\n" +" const int nWGs = cb.m_nWGs;\n" +" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n" +" for(int i=0; i<NUM_BUCKET; i++)\n" +" {\n" +" MY_HISTOGRAM(i) = 0;\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n" +" u32 localKey;\n" +" int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n" +" int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\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" +" // Using registers didn't perform well. It seems like use localKeys to address requires a lot of alu ops\n" +" // AMD: AtomInc performs better while NV prefers ++\n" +" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n" +" {\n" +"#if defined(CHECK_BOUNDARY)\n" +" if( addr+i < n )\n" +"#endif\n" +" {\n" +" localKey = (gSrc[addr+i].m_key>>startBit) & 0xf;\n" +"#if defined(NV_GPU)\n" +" MY_HISTOGRAM( localKey )++;\n" +"#else\n" +" AtomInc( MY_HISTOGRAM( localKey ) );\n" +"#endif\n" +" }\n" +" }\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" \n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" u32 sum = 0;\n" +" for(int i=0; i<GET_GROUP_SIZE; i++)\n" +" {\n" +" sum += localHistogramMat[lIdx*WG_SIZE+(i+lIdx)%GET_GROUP_SIZE];\n" +" }\n" +" histogramOut[lIdx*nWGs+wgIdx] = sum;\n" +" }\n" +"}\n" +"#define nPerLane (nPerWI/4)\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" +" u32 lIdx = GET_LOCAL_IDX;\n" +" u32 wgIdx = GET_GROUP_IDX;\n" +" const int nWGs = cb.m_nWGs;\n" +" u32 data[nPerWI];\n" +" for(int i=0; i<nPerWI; i++)\n" +" {\n" +" data[i] = 0;\n" +" if( (nPerWI*lIdx+i) < NUM_BUCKET*nWGs )\n" +" data[i] = wHistogram1[nPerWI*lIdx+i];\n" +" }\n" +" uint4 myData = make_uint4(0,0,0,0);\n" +" for(int i=0; i<nPerLane; i++)\n" +" {\n" +" myData.x += data[nPerLane*0+i];\n" +" myData.y += data[nPerLane*1+i];\n" +" myData.z += data[nPerLane*2+i];\n" +" myData.w += data[nPerLane*3+i];\n" +" }\n" +" uint totalSum;\n" +" uint4 scanned = localPrefixSum128V( myData, lIdx, &totalSum, ldsTopScanData );\n" +"// for(int j=0; j<4; j++) // somehow it introduces a lot of branches\n" +" { int j = 0;\n" +" u32 sum = 0;\n" +" for(int i=0; i<nPerLane; i++)\n" +" {\n" +" u32 tmp = data[nPerLane*j+i];\n" +" data[nPerLane*j+i] = sum;\n" +" sum += tmp;\n" +" }\n" +" }\n" +" { int j = 1;\n" +" u32 sum = 0;\n" +" for(int i=0; i<nPerLane; i++)\n" +" {\n" +" u32 tmp = data[nPerLane*j+i];\n" +" data[nPerLane*j+i] = sum;\n" +" sum += tmp;\n" +" }\n" +" }\n" +" { int j = 2;\n" +" u32 sum = 0;\n" +" for(int i=0; i<nPerLane; i++)\n" +" {\n" +" u32 tmp = data[nPerLane*j+i];\n" +" data[nPerLane*j+i] = sum;\n" +" sum += tmp;\n" +" }\n" +" }\n" +" { int j = 3;\n" +" u32 sum = 0;\n" +" for(int i=0; i<nPerLane; i++)\n" +" {\n" +" u32 tmp = data[nPerLane*j+i];\n" +" data[nPerLane*j+i] = sum;\n" +" sum += tmp;\n" +" }\n" +" }\n" +" for(int i=0; i<nPerLane; i++)\n" +" {\n" +" data[nPerLane*0+i] += scanned.x;\n" +" data[nPerLane*1+i] += scanned.y;\n" +" data[nPerLane*2+i] += scanned.z;\n" +" data[nPerLane*3+i] += scanned.w;\n" +" }\n" +" for(int i=0; i<nPerWI; i++)\n" +" {\n" +" int index = nPerWI*lIdx+i;\n" +" if (index < NUM_BUCKET*nWGs)\n" +" wHistogram1[nPerWI*lIdx+i] = data[i];\n" +" }\n" +"}\n" +"// 4 scan, 4 exchange\n" +"void sort4Bits(u32 sortData[4], int startBit, int lIdx, __local u32* ldsSortData)\n" +"{\n" +" for(int bitIdx=0; bitIdx<BITS_PER_PASS; bitIdx++)\n" +" {\n" +" u32 mask = (1<<bitIdx);\n" +" uint4 cmpResult = make_uint4( (sortData[0]>>startBit) & mask, (sortData[1]>>startBit) & mask, (sortData[2]>>startBit) & mask, (sortData[3]>>startBit) & mask );\n" +" uint4 prefixSum = SELECT_UINT4( make_uint4(1,1,1,1), make_uint4(0,0,0,0), cmpResult != make_uint4(0,0,0,0) );\n" +" u32 total;\n" +" prefixSum = localPrefixSum64V( prefixSum, lIdx, &total, ldsSortData );\n" +" {\n" +" 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" +" GROUP_LDS_BARRIER;\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" +" GROUP_LDS_BARRIER;\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" +" GROUP_LDS_BARRIER;\n" +" }\n" +" }\n" +"}\n" +"// 2 scan, 2 exchange\n" +"void sort4Bits1(u32 sortData[4], int startBit, int lIdx, __local u32* ldsSortData)\n" +"{\n" +" for(uint ibit=0; ibit<BITS_PER_PASS; ibit+=2)\n" +" {\n" +" uint4 b = make_uint4((sortData[0]>>(startBit+ibit)) & 0x3, \n" +" (sortData[1]>>(startBit+ibit)) & 0x3, \n" +" (sortData[2]>>(startBit+ibit)) & 0x3, \n" +" (sortData[3]>>(startBit+ibit)) & 0x3);\n" +" u32 key4;\n" +" u32 sKeyPacked[4] = { 0, 0, 0, 0 };\n" +" {\n" +" sKeyPacked[0] |= 1<<(8*b.x);\n" +" sKeyPacked[1] |= 1<<(8*b.y);\n" +" sKeyPacked[2] |= 1<<(8*b.z);\n" +" sKeyPacked[3] |= 1<<(8*b.w);\n" +" key4 = sKeyPacked[0] + sKeyPacked[1] + sKeyPacked[2] + sKeyPacked[3];\n" +" }\n" +" u32 rankPacked;\n" +" u32 sumPacked;\n" +" {\n" +" rankPacked = localPrefixSum( key4, lIdx, &sumPacked, ldsSortData, WG_SIZE );\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" u32 newOffset[4] = { 0,0,0,0 };\n" +" {\n" +" u32 sumScanned = bit8Scan( sumPacked );\n" +" u32 scannedKeys[4];\n" +" scannedKeys[0] = 1<<(8*b.x);\n" +" scannedKeys[1] = 1<<(8*b.y);\n" +" scannedKeys[2] = 1<<(8*b.z);\n" +" scannedKeys[3] = 1<<(8*b.w);\n" +" { // 4 scans at once\n" +" u32 sum4 = 0;\n" +" for(int ie=0; ie<4; ie++)\n" +" {\n" +" u32 tmp = scannedKeys[ie];\n" +" scannedKeys[ie] = sum4;\n" +" sum4 += tmp;\n" +" }\n" +" }\n" +" {\n" +" u32 sumPlusRank = sumScanned + rankPacked;\n" +" { u32 ie = b.x;\n" +" scannedKeys[0] += sumPlusRank;\n" +" newOffset[0] = unpack4Key( scannedKeys[0], ie );\n" +" }\n" +" { u32 ie = b.y;\n" +" scannedKeys[1] += sumPlusRank;\n" +" newOffset[1] = unpack4Key( scannedKeys[1], ie );\n" +" }\n" +" { u32 ie = b.z;\n" +" scannedKeys[2] += sumPlusRank;\n" +" newOffset[2] = unpack4Key( scannedKeys[2], ie );\n" +" }\n" +" { u32 ie = b.w;\n" +" scannedKeys[3] += sumPlusRank;\n" +" newOffset[3] = unpack4Key( scannedKeys[3], ie );\n" +" }\n" +" }\n" +" }\n" +" GROUP_LDS_BARRIER;\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" +" GROUP_LDS_BARRIER;\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" +" GROUP_LDS_BARRIER;\n" +" }\n" +" }\n" +"}\n" +"#define SET_HISTOGRAM(setIdx, key) ldsSortData[(setIdx)*NUM_BUCKET+key]\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" +"{\n" +" __local u32 ldsSortData[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];\n" +" __local u32 localHistogramToCarry[NUM_BUCKET];\n" +" __local u32 localHistogram[NUM_BUCKET*2];\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" +" 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" +" if( lIdx < (NUM_BUCKET) )\n" +" {\n" +" localHistogramToCarry[lIdx] = rHistogram[lIdx*nWGs + wgIdx];\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n" +" int nBlocks = n/blockSize - nBlocksPerWG*wgIdx;\n" +" int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n" +" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n" +" {\n" +" u32 myHistogram = 0;\n" +" u32 sortData[ELEMENTS_PER_WORK_ITEM];\n" +" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n" +"#if defined(CHECK_BOUNDARY)\n" +" sortData[i] = ( addr+i < n )? gSrc[ addr+i ] : 0xffffffff;\n" +"#else\n" +" sortData[i] = gSrc[ addr+i ];\n" +"#endif\n" +" sort4Bits(sortData, startBit, lIdx, ldsSortData);\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" +" { // create histogram\n" +" u32 setIdx = lIdx/16;\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" localHistogram[lIdx] = 0;\n" +" }\n" +" ldsSortData[lIdx] = 0;\n" +" GROUP_LDS_BARRIER;\n" +" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n" +"#if defined(CHECK_BOUNDARY)\n" +" if( addr+i < n )\n" +"#endif\n" +"#if defined(NV_GPU)\n" +" SET_HISTOGRAM( setIdx, keys[i] )++;\n" +"#else\n" +" AtomInc( SET_HISTOGRAM( setIdx, keys[i] ) );\n" +"#endif\n" +" \n" +" GROUP_LDS_BARRIER;\n" +" \n" +" uint hIdx = NUM_BUCKET+lIdx;\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" u32 sum = 0;\n" +" for(int i=0; i<WG_SIZE/16; i++)\n" +" {\n" +" sum += SET_HISTOGRAM( i, lIdx );\n" +" }\n" +" myHistogram = sum;\n" +" localHistogram[hIdx] = sum;\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +"#if defined(USE_2LEVEL_REDUCE)\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" localHistogram[hIdx] = localHistogram[hIdx-1];\n" +" GROUP_MEM_FENCE;\n" +" u32 u0, u1, u2;\n" +" u0 = localHistogram[hIdx-3];\n" +" u1 = localHistogram[hIdx-2];\n" +" u2 = localHistogram[hIdx-1];\n" +" AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );\n" +" GROUP_MEM_FENCE;\n" +" u0 = localHistogram[hIdx-12];\n" +" u1 = localHistogram[hIdx-8];\n" +" u2 = localHistogram[hIdx-4];\n" +" AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );\n" +" GROUP_MEM_FENCE;\n" +" }\n" +"#else\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" localHistogram[hIdx] = localHistogram[hIdx-1];\n" +" GROUP_MEM_FENCE;\n" +" localHistogram[hIdx] += localHistogram[hIdx-1];\n" +" GROUP_MEM_FENCE;\n" +" localHistogram[hIdx] += localHistogram[hIdx-2];\n" +" GROUP_MEM_FENCE;\n" +" localHistogram[hIdx] += localHistogram[hIdx-4];\n" +" GROUP_MEM_FENCE;\n" +" localHistogram[hIdx] += localHistogram[hIdx-8];\n" +" GROUP_MEM_FENCE;\n" +" }\n" +"#endif\n" +" GROUP_LDS_BARRIER;\n" +" }\n" +" {\n" +" for(int ie=0; ie<ELEMENTS_PER_WORK_ITEM; ie++)\n" +" {\n" +" int dataIdx = ELEMENTS_PER_WORK_ITEM*lIdx+ie;\n" +" int binIdx = keys[ie];\n" +" int groupOffset = localHistogramToCarry[binIdx];\n" +" int myIdx = dataIdx - localHistogram[NUM_BUCKET+binIdx];\n" +"#if defined(CHECK_BOUNDARY)\n" +" if( addr+ie < n )\n" +"#endif\n" +" gDst[ groupOffset + myIdx ] = sortData[ie];\n" +" }\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" localHistogramToCarry[lIdx] += myHistogram;\n" +" }\n" +" GROUP_LDS_BARRIER;\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" +" for(uint ibit=0; ibit<BITS_PER_PASS; ibit+=2)\n" +" {\n" +" uint4 b = make_uint4((sortData[0]>>(startBit+ibit)) & 0x3, \n" +" (sortData[1]>>(startBit+ibit)) & 0x3, \n" +" (sortData[2]>>(startBit+ibit)) & 0x3, \n" +" (sortData[3]>>(startBit+ibit)) & 0x3);\n" +" u32 key4;\n" +" u32 sKeyPacked[4] = { 0, 0, 0, 0 };\n" +" {\n" +" sKeyPacked[0] |= 1<<(8*b.x);\n" +" sKeyPacked[1] |= 1<<(8*b.y);\n" +" sKeyPacked[2] |= 1<<(8*b.z);\n" +" sKeyPacked[3] |= 1<<(8*b.w);\n" +" key4 = sKeyPacked[0] + sKeyPacked[1] + sKeyPacked[2] + sKeyPacked[3];\n" +" }\n" +" u32 rankPacked;\n" +" u32 sumPacked;\n" +" {\n" +" rankPacked = localPrefixSum( key4, lIdx, &sumPacked, ldsSortData, WG_SIZE );\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" u32 newOffset[4] = { 0,0,0,0 };\n" +" {\n" +" u32 sumScanned = bit8Scan( sumPacked );\n" +" u32 scannedKeys[4];\n" +" scannedKeys[0] = 1<<(8*b.x);\n" +" scannedKeys[1] = 1<<(8*b.y);\n" +" scannedKeys[2] = 1<<(8*b.z);\n" +" scannedKeys[3] = 1<<(8*b.w);\n" +" { // 4 scans at once\n" +" u32 sum4 = 0;\n" +" for(int ie=0; ie<4; ie++)\n" +" {\n" +" u32 tmp = scannedKeys[ie];\n" +" scannedKeys[ie] = sum4;\n" +" sum4 += tmp;\n" +" }\n" +" }\n" +" {\n" +" u32 sumPlusRank = sumScanned + rankPacked;\n" +" { u32 ie = b.x;\n" +" scannedKeys[0] += sumPlusRank;\n" +" newOffset[0] = unpack4Key( scannedKeys[0], ie );\n" +" }\n" +" { u32 ie = b.y;\n" +" scannedKeys[1] += sumPlusRank;\n" +" newOffset[1] = unpack4Key( scannedKeys[1], ie );\n" +" }\n" +" { u32 ie = b.z;\n" +" scannedKeys[2] += sumPlusRank;\n" +" newOffset[2] = unpack4Key( scannedKeys[2], ie );\n" +" }\n" +" { u32 ie = b.w;\n" +" scannedKeys[3] += sumPlusRank;\n" +" newOffset[3] = unpack4Key( scannedKeys[3], ie );\n" +" }\n" +" }\n" +" }\n" +" GROUP_LDS_BARRIER;\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" +" ldsSortVal[newOffset[0]] = sortVal[0];\n" +" ldsSortVal[newOffset[1]] = sortVal[1];\n" +" ldsSortVal[newOffset[2]] = sortVal[2];\n" +" ldsSortVal[newOffset[3]] = sortVal[3];\n" +" GROUP_LDS_BARRIER;\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" +" sortVal[0] = ldsSortVal[dstAddr+0];\n" +" sortVal[1] = ldsSortVal[dstAddr+1];\n" +" sortVal[2] = ldsSortVal[dstAddr+2];\n" +" sortVal[3] = ldsSortVal[dstAddr+3];\n" +" GROUP_LDS_BARRIER;\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" +"{\n" +" __local int ldsSortData[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];\n" +" __local int ldsSortVal[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];\n" +" __local u32 localHistogramToCarry[NUM_BUCKET];\n" +" __local u32 localHistogram[NUM_BUCKET*2];\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" +" 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" +" if( lIdx < (NUM_BUCKET) )\n" +" {\n" +" localHistogramToCarry[lIdx] = rHistogram[lIdx*nWGs + wgIdx];\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" \n" +" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n" +" int nBlocks = n/blockSize - nBlocksPerWG*wgIdx;\n" +" int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n" +" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n" +" {\n" +" u32 myHistogram = 0;\n" +" int sortData[ELEMENTS_PER_WORK_ITEM];\n" +" int sortVal[ELEMENTS_PER_WORK_ITEM];\n" +" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n" +"#if defined(CHECK_BOUNDARY)\n" +" {\n" +" sortData[i] = ( addr+i < n )? gSrc[ addr+i ].m_key : 0xffffffff;\n" +" sortVal[i] = ( addr+i < n )? gSrc[ addr+i ].m_value : 0xffffffff;\n" +" }\n" +"#else\n" +" {\n" +" sortData[i] = gSrc[ addr+i ].m_key;\n" +" sortVal[i] = gSrc[ addr+i ].m_value;\n" +" }\n" +"#endif\n" +" sort4Bits1KeyValue(sortData, sortVal, startBit, lIdx, ldsSortData, ldsSortVal);\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" +" { // create histogram\n" +" u32 setIdx = lIdx/16;\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" localHistogram[lIdx] = 0;\n" +" }\n" +" ldsSortData[lIdx] = 0;\n" +" GROUP_LDS_BARRIER;\n" +" for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n" +"#if defined(CHECK_BOUNDARY)\n" +" if( addr+i < n )\n" +"#endif\n" +"#if defined(NV_GPU)\n" +" SET_HISTOGRAM( setIdx, keys[i] )++;\n" +"#else\n" +" AtomInc( SET_HISTOGRAM( setIdx, keys[i] ) );\n" +"#endif\n" +" \n" +" GROUP_LDS_BARRIER;\n" +" \n" +" uint hIdx = NUM_BUCKET+lIdx;\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" u32 sum = 0;\n" +" for(int i=0; i<WG_SIZE/16; i++)\n" +" {\n" +" sum += SET_HISTOGRAM( i, lIdx );\n" +" }\n" +" myHistogram = sum;\n" +" localHistogram[hIdx] = sum;\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +"#if defined(USE_2LEVEL_REDUCE)\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" localHistogram[hIdx] = localHistogram[hIdx-1];\n" +" GROUP_MEM_FENCE;\n" +" u32 u0, u1, u2;\n" +" u0 = localHistogram[hIdx-3];\n" +" u1 = localHistogram[hIdx-2];\n" +" u2 = localHistogram[hIdx-1];\n" +" AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );\n" +" GROUP_MEM_FENCE;\n" +" u0 = localHistogram[hIdx-12];\n" +" u1 = localHistogram[hIdx-8];\n" +" u2 = localHistogram[hIdx-4];\n" +" AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );\n" +" GROUP_MEM_FENCE;\n" +" }\n" +"#else\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" localHistogram[hIdx] = localHistogram[hIdx-1];\n" +" GROUP_MEM_FENCE;\n" +" localHistogram[hIdx] += localHistogram[hIdx-1];\n" +" GROUP_MEM_FENCE;\n" +" localHistogram[hIdx] += localHistogram[hIdx-2];\n" +" GROUP_MEM_FENCE;\n" +" localHistogram[hIdx] += localHistogram[hIdx-4];\n" +" GROUP_MEM_FENCE;\n" +" localHistogram[hIdx] += localHistogram[hIdx-8];\n" +" GROUP_MEM_FENCE;\n" +" }\n" +"#endif\n" +" GROUP_LDS_BARRIER;\n" +" }\n" +" {\n" +" for(int ie=0; ie<ELEMENTS_PER_WORK_ITEM; ie++)\n" +" {\n" +" int dataIdx = ELEMENTS_PER_WORK_ITEM*lIdx+ie;\n" +" int binIdx = keys[ie];\n" +" int groupOffset = localHistogramToCarry[binIdx];\n" +" int myIdx = dataIdx - localHistogram[NUM_BUCKET+binIdx];\n" +"#if defined(CHECK_BOUNDARY)\n" +" if( addr+ie < n )\n" +" {\n" +" if ((groupOffset + myIdx)<n)\n" +" {\n" +" if (sortData[ie]==sortVal[ie])\n" +" {\n" +" \n" +" SortDataCL tmp;\n" +" tmp.m_key = sortData[ie];\n" +" tmp.m_value = sortVal[ie];\n" +" if (tmp.m_key == tmp.m_value)\n" +" gDst[groupOffset + myIdx ] = tmp;\n" +" }\n" +" \n" +" }\n" +" }\n" +"#else\n" +" if ((groupOffset + myIdx)<n)\n" +" {\n" +" gDst[ groupOffset + myIdx ].m_key = sortData[ie];\n" +" gDst[ groupOffset + myIdx ].m_value = sortVal[ie];\n" +" }\n" +"#endif\n" +" }\n" +" }\n" +" GROUP_LDS_BARRIER;\n" +" if( lIdx < NUM_BUCKET )\n" +" {\n" +" localHistogramToCarry[lIdx] += myHistogram;\n" +" }\n" +" GROUP_LDS_BARRIER;\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" +"{\n" +" \n" +" u32 gIdx = GET_GLOBAL_IDX;\n" +" u32 realLocalIdx = GET_LOCAL_IDX;\n" +" u32 wgIdx = GET_GROUP_IDX;\n" +" u32 wgSize = GET_GROUP_SIZE;\n" +" const int startBit = cb.m_startBit;\n" +" const int n = cb.m_n;\n" +" const int nWGs = cb.m_nWGs;\n" +" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n" +" int counter[NUM_BUCKET];\n" +" \n" +" if (realLocalIdx>0)\n" +" return;\n" +" \n" +" for (int c=0;c<NUM_BUCKET;c++)\n" +" counter[c]=0;\n" +" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n" +" \n" +" int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n" +" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++)\n" +" {\n" +" for (int lIdx=0;lIdx<WG_SIZE;lIdx++)\n" +" {\n" +" int addr2 = iblock*blockSize + blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n" +" \n" +" for(int j=0; j<ELEMENTS_PER_WORK_ITEM; j++)\n" +" {\n" +" int i = addr2+j;\n" +" if( i < n )\n" +" {\n" +" int tableIdx;\n" +" tableIdx = (gSrc[i].m_key>>startBit) & 0xf;//0xf = NUM_TABLES-1\n" +" gDst[rHistogram[tableIdx*nWGs+wgIdx] + counter[tableIdx]] = gSrc[i];\n" +" counter[tableIdx] ++;\n" +" }\n" +" }\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" +"{\n" +" \n" +" u32 gIdx = GET_GLOBAL_IDX;\n" +" u32 realLocalIdx = GET_LOCAL_IDX;\n" +" u32 wgIdx = GET_GROUP_IDX;\n" +" u32 wgSize = GET_GROUP_SIZE;\n" +" const int startBit = cb.m_startBit;\n" +" const int n = cb.m_n;\n" +" const int nWGs = cb.m_nWGs;\n" +" const int nBlocksPerWG = cb.m_nBlocksPerWG;\n" +" int counter[NUM_BUCKET];\n" +" \n" +" if (realLocalIdx>0)\n" +" return;\n" +" \n" +" for (int c=0;c<NUM_BUCKET;c++)\n" +" counter[c]=0;\n" +" const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n" +" \n" +" int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n" +" for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++)\n" +" {\n" +" for (int lIdx=0;lIdx<WG_SIZE;lIdx++)\n" +" {\n" +" int addr2 = iblock*blockSize + blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n" +" \n" +" for(int j=0; j<ELEMENTS_PER_WORK_ITEM; j++)\n" +" {\n" +" int i = addr2+j;\n" +" if( i < n )\n" +" {\n" +" int tableIdx;\n" +" tableIdx = (gSrc[i]>>startBit) & 0xf;//0xf = NUM_TABLES-1\n" +" gDst[rHistogram[tableIdx*nWGs+wgIdx] + counter[tableIdx]] = gSrc[i];\n" +" counter[tableIdx] ++;\n" +" }\n" +" }\n" +" }\n" +" }\n" +" \n" +"}\n" +; |