diff options
Diffstat (limited to 'external/clBLAS/src/library/blas/gens/gemm_cached.cpp')
| -rw-r--r-- | external/clBLAS/src/library/blas/gens/gemm_cached.cpp | 503 |
1 files changed, 0 insertions, 503 deletions
diff --git a/external/clBLAS/src/library/blas/gens/gemm_cached.cpp b/external/clBLAS/src/library/blas/gens/gemm_cached.cpp deleted file mode 100644 index 5c7c3526..00000000 --- a/external/clBLAS/src/library/blas/gens/gemm_cached.cpp +++ /dev/null @@ -1,503 +0,0 @@ -/* ************************************************************************ - * Copyright 2013 Advanced Micro Devices, Inc. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * ************************************************************************/ - - -/* - * Cached global buffers based gemm generator - */ - -#include <string.h> -#include <stdio.h> -#include <assert.h> -#include <clblas_stddef.h> -#include <clBLAS.h> -#include <blas_mempat.h> -#include <clkern.h> -#include <clblas-internal.h> -#include <kprintf.hpp> -#include <gemm.clT> -#include <gemm_helper.clT> -#include <symm_helper.clT> -#include <solution_seq.h> -#include "tuned_numbers.h" - -//#define DEBUG_GEMM_2 -static CLBLASMpatExtra mpatExtra; - -static char Prefix[4]; - -/* Function, finding default decomposition */ -static int -getDefaultDecomposition( - PGranularity *pgran, - SubproblemDim *subdims, - unsigned int subdimsNum, - void *pArgs); - -static ssize_t -generator( - char *buf, - size_t buflen, - const struct SubproblemDim *subdims, - const struct PGranularity *pgran, - void *extra); - -static void -assignKargs(KernelArg *args, const void *params, const void *extra); - -static SolverFlags -solverFlags(void); - -static void -setBuildOpts( - char * buildOptStr, - const void *args); - -static void -calcNrThreads( - size_t threads[2], - const SubproblemDim *subdims, - const PGranularity *pgran, - const void *args, - const void *extra); - -static SolverOps gemmSops = { - generator, - assignKargs, - NULL, - NULL, - NULL, - calcNrThreads, - NULL, - solverFlags, - NULL, - getDefaultDecomposition, - NULL, - setBuildOpts, - NULL -}; - -static void -calcNrThreads( - size_t threads[2], - const SubproblemDim *subdims, - const PGranularity *pgran, - const void *args, - const void *extra) -{ - const CLBlasKargs *kargs = (const CLBlasKargs *)args; - //const CLBLASKernExtra *kextra = ( CLBLASKernExtra *)extra; - //KernelExtraFlags kflags = kextra->flags; - size_t M, N; - - M = kargs->M; - N = kargs->N; - - threads[1] = 1; - - if ((subdims->x != SUBDIM_UNUSED) && - (subdims->y != SUBDIM_UNUSED)) { - - size_t groupWorkX, groupWorkY; - size_t nrGroupsX, nrGroupsY; - int nrDims; - - groupWorkX = subdims->x; - groupWorkY = subdims->y; - - nrGroupsX = N / groupWorkX; - if (N % groupWorkX) { - nrGroupsX++; - } - - nrGroupsY = M / groupWorkY; - if (M % groupWorkY) { - nrGroupsY++; - } - nrDims = (pgran == NULL) ? 1 : pgran->wgDim; - threads[0] = nrGroupsX * nrGroupsY; - - if(kargs->pigFuncID == CLBLAS_HERK) - { - threads[0] = (nrGroupsY * (nrGroupsY + 1)) / 2; - } - - } - - if (pgran != NULL) { - threads[0] *= pgran->wgSize[0]; - threads[1] *= pgran->wgSize[1]; - } -} - -static void -setBuildOpts( - char * buildOptStr, - const void *args) -{ - SolutionStep *step = (SolutionStep *)args; - const CLBlasKargs *kargs = (const CLBlasKargs *)(&step->args); - const SubproblemDim *dims = step->subdims; - //size_t vecLen = sizeof(cl_float4)/dtypeSize(kargs->dtype); - KernelExtraFlags kflags = step->extraFlags; - - blockSizes bestSize = bestBlockSizeForDevice( step ); - - if ( kargs->dtype == TYPE_DOUBLE || kargs->dtype == TYPE_COMPLEX_DOUBLE) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DDOUBLE_PRECISION"); - } - - if (isComplexType(kargs->dtype)) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DCOMPLEX"); - } - - if ((bestSize.useBarrier) == 1) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DGEMM_NEEDS_BARRIER"); - } - - if (kargs->M % dims->y) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DM_TAIL_PRESENT"); - } - - if (kargs->N % dims->x) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DN_TAIL_PRESENT"); - } - - if (kflags & KEXTRA_CONJUGATE_A) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DCONJUGATE_A"); - } - if (kflags & KEXTRA_CONJUGATE_B) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DCONJUGATE_B"); - } - - switch(kargs->pigFuncID) - { - case CLBLAS_HEMM: - case CLBLAS_SYMM: - case CLBLAS_SYMM_DIAGONAL: - case CLBLAS_HEMM_DIAGONAL: - #ifdef DEBUG_GEMM_2 - printf("GEMM2: setBuildOpts: Setting options for SYMM\n"); - #endif - if (kargs->side == clblasLeft) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-D__SYMM_LEFT__"); - } - if (kargs->side == clblasRight) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-D__SYMM_RIGHT__"); - } - if (kargs->uplo == clblasLower) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-D__SYMM_LOWER__"); - } - if (kargs->uplo == clblasUpper) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-D__SYMM_UPPER__"); - } - // Define the order for Legacy sake. - if (kargs->order == clblasColumnMajor) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-D__SYMM_COLMAJOR__"); - } else { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-D__SYMM_ROWMAJOR__"); - } - if ((kargs->pigFuncID == CLBLAS_SYMM_DIAGONAL) || (kargs->pigFuncID == CLBLAS_HEMM_DIAGONAL)) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-D__SYMM_DIAGONAL__"); - } - if (kargs->pigFuncID == CLBLAS_HEMM_DIAGONAL) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-D__HEMM__"); - } - break; - - case CLBLAS_HERK: - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DHERK"); - if(kargs->uplo == clblasLower) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DHERK_LOWER_TRIANGLE"); - } - else if(kargs->uplo == clblasUpper) - { - addBuildOpt( buildOptStr, BUILD_OPTS_MAXLEN, "-DHERK_UPPER_TRIANGLE"); - } - break; - - default: - break; - } - - #ifdef DEBUG_GEMM_2 - printf("buildStr: %s\n", buildOptStr); - #endif - return; -} - -static ssize_t -generator( - char *buf, - size_t buflen, - const struct SubproblemDim *subdims, - const struct PGranularity *pgran, - void *extra) -{ - CLBLASKernExtra *kextra = (CLBLASKernExtra*)extra; - KernelExtraFlags kflags = kextra->flags; - DataType dtype = kextra->dtype; - char tempTemplate[64*1024]; //PENDING: Is it safe to have 64K in stack for threadSafety? - char itemx[10], itemy[10], width[10], itemy_by_width[10], itemx_by_width[10]; - char bwidth[10], panel_by_v[10]; - size_t Y, X, BLOCKSIZE, ITEMX, ITEMY; - bool doVLOAD = false; - unsigned int veclen; - - if (buf == NULL) - { - buflen = 64*1024*sizeof(char); - return (ssize_t)buflen; - } - - // - // PENDING: Add Support for Row Major - // - if ((kflags & KEXTRA_COLUMN_MAJOR) == 0) - { - return 0; - } - - if ((kflags & KEXTRA_NO_COPY_VEC_A) || (kflags & KEXTRA_NO_COPY_VEC_B) || (kflags & KEXTRA_NO_COPY_VEC_C)) - { - #ifdef DEBUG_GEMM_2 - printf("GEMM2: Doing un-aligned access\n"); - #endif - doVLOAD= true; - } else { - #ifdef DEBUG_GEMM_2 - printf("GEMM2: Doing Aligned access\n"); - #endif - } - - - BLOCKSIZE = pgran->wgSize[0]; - #ifdef DEBUG_GEMM_2 - printf("GEMM2- generator(): Blocksize passed = %lu, subdimy = %lu, subdimx = %lu, veclen = %d \n", - BLOCKSIZE, subdims->y, subdims->x, kextra->vecLen); - #endif - - veclen = kextra->vecLen; - - ITEMY = subdims->itemY; - ITEMX = subdims->itemX; - Y = subdims->y / ITEMY; - X = subdims->x / ITEMX; - - // - // Handle in-compatible subdims and workgroup sizes - // We will use "veclen" of 1 as our shield against these in-compatible - // geometries. - // - if ( (ITEMY % kextra->vecLen) || ((ITEMX % kextra->vecLen) && (kflags & KEXTRA_TRANS_B)) ) - { - // - // FIXME: - // This kernel must be stored against vecLen of 1 in Kernel Cache. - // This needs change in EXTRA structure. However, this is against the API. - // We are going against the API by changing fields in EXTRA structure. - // One alternate FIX is to return an error. - // - kextra->vecLen = kextra->vecLenA = kextra->vecLenB = kextra->vecLenC = 1; - - doVLOAD = true; - veclen = 1; - } - - // - // PENDING: Selective Vectorization for A, B and C access has to be added - // in KPRINTF module (VLOADA, VLOADB, VLOADC, VSTOREC) - // - kprintf kobj(Prefix[dtype], veclen, doVLOAD, doVLOAD); // Only Vectored Access - sprintf(width, "%lu", Y); - sprintf(itemy, "%lu", ITEMY); - sprintf(itemx, "%lu", ITEMX); - sprintf(itemy_by_width, "%lu", (size_t) ITEMY/veclen); - sprintf(itemx_by_width, "%lu", (size_t) ITEMX/veclen); - //sprintf(bwidth, "%lu", subdims->bwidth); - //sprintf(panel_by_v, "%lu", (subdims->bwidth / veclen)); - sprintf(bwidth, "%lu", (size_t) veclen); - sprintf(panel_by_v, "%lu", (size_t) 1); - - kobj.put("%WIDTH", width); - kobj.put("%ITEMX", itemx); - kobj.put("%ITEMY", itemy); - kobj.put("%ITEMY_BY_V", itemy_by_width); - kobj.put("%ITEMX_BY_V", itemx_by_width); - kobj.put("%PANEL", bwidth); - kobj.put("%PANEL_BY_V", panel_by_v); - #ifdef DEBUG_GEMM_2 - printf("ColMajor GEMM - WIDTH = %s, PANEL = %lu, ITEMX = %s, ITEMY = %s, Veclen = %lu\n", width, subdims->bwidth, itemx, itemy, veclen); - #endif - - strcpy(tempTemplate, SYMM_HEMM_HELPER); - if ((kflags & KEXTRA_TRANS_A) == 0) - { - if (kflags & KEXTRA_TRANS_B) - { - #ifdef DEBUG_GEMM_2 - printf("Using GEMM_NT_KERNEL\n"); - #endif - strcat(tempTemplate, GEMM_HELPER); - strcat(tempTemplate, GEMM_NT_KERNEL); - } else { - #ifdef DEBUG_GEMM_2 - printf("Using GEMM_NN_KERNEL\n"); - #endif - strcat(tempTemplate, GEMM_HELPER); - strcat(tempTemplate, GEMM_NN_KERNEL); - } - } else { - // PENDING: - if (kflags & KEXTRA_TRANS_B) - { - tempTemplate[0] = 0; - } else { - #ifdef DEBUG_GEMM_2 - printf("Using GEMM_TN_KERNEL\n"); - #endif - strcat(tempTemplate, GEMM_HELPER); - strcat(tempTemplate, GEMM_TN_KERNEL); - } - } - kobj.spit(buf, tempTemplate); - #ifdef DEBUG_GEMM_KPRINTF - printf("Kernel = \n%s\n", buf); - #endif - size_t tail = strlen(buf) + 1; - while(tail < 64*1024) - { - buf[tail++] = 0; - } - return 64*1024*sizeof(char); -} - -static void -assignKargs(KernelArg *args, const void *params, const void*) -{ - CLBlasKargs *blasArgs = (CLBlasKargs*)params; - - #ifdef DEBUG_GEMM_2 - printf("SAlpha=%f, DAlpha=%f, CAlpha =<%f, %f>, DAlpha=<%f, %f>\n", - blasArgs->alpha.argFloat, blasArgs->alpha.argDouble, CREAL(blasArgs->alpha.argFloatComplex), CIMAG(blasArgs->alpha.argFloatComplex), - CREAL(blasArgs->alpha.argDoubleComplex) , CIMAG(blasArgs->alpha.argDoubleComplex)); - printf("SBeta=%f, DBeta=%f, CBeta=<%f, %f>, DBeta=<%f, %f>\n", - blasArgs->beta.argFloat, blasArgs->beta.argDouble, CREAL(blasArgs->beta.argFloatComplex), CIMAG(blasArgs->beta.argFloatComplex), - CREAL(blasArgs->beta.argDoubleComplex) , CIMAG(blasArgs->beta.argDoubleComplex)); - #endif - - INIT_KARG(&args[0], blasArgs->A); //A - input matrix - argument - INIT_KARG(&args[1], blasArgs->B); //x - result buffer = _xnew argument - INIT_KARG(&args[2], blasArgs->C); //y - scratch == _x_vector argument - initSizeKarg(&args[3], blasArgs->M); - initSizeKarg(&args[4], blasArgs->N); - initSizeKarg(&args[5], blasArgs->K); - initSizeKarg(&args[6], blasArgs->lda.matrix); - initSizeKarg(&args[7], blasArgs->ldb.matrix); - initSizeKarg(&args[8], blasArgs->ldc.matrix); - initSizeKarg(&args[9], blasArgs->offA); - initSizeKarg(&args[10], blasArgs->offBX); - initSizeKarg(&args[11], blasArgs->offCY); - assignScalarKarg(&args[12], &(blasArgs->alpha), blasArgs->dtype); - assignScalarKarg(&args[13], &(blasArgs->beta), blasArgs->dtype); - return; -} - -static SolverFlags -solverFlags(void) -{ - return (SF_WSPACE_1D); -} - -extern "C" -void -initGemmV2CachedPattern(MemoryPattern *mempat) -{ - mempat->name = "Cached global memory based block gemm"; - mempat->nrLevels = 2; - mempat->cuLevel = 0; - mempat->thLevel = 1; - mempat->sops = &gemmSops; - - mpatExtra.aMset = CLMEM_LEVEL_L1; - mpatExtra.bMset = CLMEM_LEVEL_L1; - mpatExtra.mobjA = CLMEM_BUFFER; - mpatExtra.mobjB = CLMEM_BUFFER; - mempat->extra = &mpatExtra; - - - Prefix[TYPE_FLOAT] = 'S'; - Prefix[TYPE_DOUBLE] = 'D'; - Prefix[TYPE_COMPLEX_FLOAT] = 'C'; - Prefix[TYPE_COMPLEX_DOUBLE] = 'Z'; -} - -static int -getDefaultDecomposition( - PGranularity *pgran, - SubproblemDim *subdims, - unsigned int subdimsNum, - void *pArgs) -{ - - DUMMY_ARG_USAGE(pArgs); - // - // FIXME: container_of() - Counts on the fact that "getDefaultDecomposition" is called - // with step->pgran, step->subdims - // - SolutionStep *step = container_of( pgran , pgran, SolutionStep); - - blockSizes bestSize = bestBlockSizeForDevice( step ); - - pgran->wgSize[0] = bestSize.TY * bestSize.TX; - pgran->wgSize[1] = 1; - pgran->wgDim = 1; - - if (subdimsNum >= 1) - { - subdims[0].y = bestSize.TY * bestSize.ITEMY; - subdims[0].x = bestSize.TX * bestSize.ITEMX; - subdims[0].itemY = bestSize.ITEMY; - subdims[0].itemX = bestSize.ITEMX; - subdims[0].bwidth = 4; - } - if (subdimsNum >= 2) - { - subdims[1].y = bestSize.TY * bestSize.ITEMY; - subdims[1].x = bestSize.TX * bestSize.ITEMX; - subdims[1].itemY = bestSize.ITEMY; - subdims[1].itemX = bestSize.ITEMX; - subdims[1].bwidth = 4; - } - - return 0; -} - |