floraachy/OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
3d7debbb280d0e671691469da650a124bedda219
OpenBLAS/interface/gemm.c
Arjan van de Ven cdc668d82b Add a "sgemm direct" mode for small matrixes 
OpenBLAS has a fancy algorithm for copying the input data while laying
it out in a more CPU friendly memory layout.

This is great for large matrixes; the cost of the copy is easily
ammortized by the gains from the better memory layout.

But for small matrixes (on CPUs that can do efficient unaligned loads) this
copy can be a net loss.

This patch adds (for SKYLAKEX initially) a "sgemm direct" mode, that bypasses
the whole copy machinary for ALPHA=1/BETA=0/... standard arguments,
for small matrixes only.

What is small? For the non-threaded case this has been measured to be
in the M*N*K = 28 * 512 * 512 range, while in the threaded case it's
less, around M*N*K = 1 * 512 * 512
2018-12-13 13:47:31 +00:00

478 lines
12 KiB
C
Raw Blame History

/*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */
/* All rights reserved. */
/* */
/* Redistribution and use in source and binary forms, with or */
/* without modification, are permitted provided that the following */
/* conditions are met: */
/* */
/* 1. Redistributions of source code must retain the above */
/* copyright notice, this list of conditions and the following */
/* disclaimer. */
/* */
/* 2. Redistributions in binary form must reproduce the above */
/* copyright notice, this list of conditions and the following */
/* disclaimer in the documentation and/or other materials */
/* provided with the distribution. */
/* */
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
/* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
/* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
/* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
/* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
/* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
/* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
/* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
/* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
/* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
/* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
/* POSSIBILITY OF SUCH DAMAGE. */
/* */
/* The views and conclusions contained in the software and */
/* documentation are those of the authors and should not be */
/* interpreted as representing official policies, either expressed */
/* or implied, of The University of Texas at Austin. */
/*********************************************************************/
#include <stdio.h>
#include <stdlib.h>
#include "common.h"
#ifdef FUNCTION_PROFILE
#include "functable.h"
#endif
#ifndef COMPLEX
#define SMP_THRESHOLD_MIN 65536.0
#ifdef XDOUBLE
#define ERROR_NAME "QGEMM "
#elif defined(DOUBLE)
#define ERROR_NAME "DGEMM "
#else
#define ERROR_NAME "SGEMM "
#endif
#else
#define SMP_THRESHOLD_MIN 8192.0
#ifndef GEMM3M
#ifdef XDOUBLE
#define ERROR_NAME "XGEMM "
#elif defined(DOUBLE)
#define ERROR_NAME "ZGEMM "
#else
#define ERROR_NAME "CGEMM "
#endif
#else
#ifdef XDOUBLE
#define ERROR_NAME "XGEMM3M "
#elif defined(DOUBLE)
#define ERROR_NAME "ZGEMM3M "
#else
#define ERROR_NAME "CGEMM3M "
#endif
#endif
#endif
#ifndef GEMM_MULTITHREAD_THRESHOLD
#define GEMM_MULTITHREAD_THRESHOLD 4
#endif
static int (*gemm[])(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLASLONG) = {
#ifndef GEMM3M
GEMM_NN, GEMM_TN, GEMM_RN, GEMM_CN,
GEMM_NT, GEMM_TT, GEMM_RT, GEMM_CT,
GEMM_NR, GEMM_TR, GEMM_RR, GEMM_CR,
GEMM_NC, GEMM_TC, GEMM_RC, GEMM_CC,
#if defined(SMP) && !defined(USE_SIMPLE_THREADED_LEVEL3)
GEMM_THREAD_NN, GEMM_THREAD_TN, GEMM_THREAD_RN, GEMM_THREAD_CN,
GEMM_THREAD_NT, GEMM_THREAD_TT, GEMM_THREAD_RT, GEMM_THREAD_CT,
GEMM_THREAD_NR, GEMM_THREAD_TR, GEMM_THREAD_RR, GEMM_THREAD_CR,
GEMM_THREAD_NC, GEMM_THREAD_TC, GEMM_THREAD_RC, GEMM_THREAD_CC,
#endif
#else
GEMM3M_NN, GEMM3M_TN, GEMM3M_RN, GEMM3M_CN,
GEMM3M_NT, GEMM3M_TT, GEMM3M_RT, GEMM3M_CT,
GEMM3M_NR, GEMM3M_TR, GEMM3M_RR, GEMM3M_CR,
GEMM3M_NC, GEMM3M_TC, GEMM3M_RC, GEMM3M_CC,
#if defined(SMP) && !defined(USE_SIMPLE_THREADED_LEVEL3)
GEMM3M_THREAD_NN, GEMM3M_THREAD_TN, GEMM3M_THREAD_RN, GEMM3M_THREAD_CN,
GEMM3M_THREAD_NT, GEMM3M_THREAD_TT, GEMM3M_THREAD_RT, GEMM3M_THREAD_CT,
GEMM3M_THREAD_NR, GEMM3M_THREAD_TR, GEMM3M_THREAD_RR, GEMM3M_THREAD_CR,
GEMM3M_THREAD_NC, GEMM3M_THREAD_TC, GEMM3M_THREAD_RC, GEMM3M_THREAD_CC,
#endif
#endif
};
#ifndef CBLAS
void NAME(char *TRANSA, char *TRANSB,
blasint *M, blasint *N, blasint *K,
FLOAT *alpha,
FLOAT *a, blasint *ldA,
FLOAT *b, blasint *ldB,
FLOAT *beta,
FLOAT *c, blasint *ldC){
blas_arg_t args;
int transa, transb, nrowa, nrowb;
blasint info;
char transA, transB;
FLOAT *buffer;
FLOAT *sa, *sb;
#ifdef SMP
double MNK;
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
#if defined(SMP) && !defined(NO_AFFINITY) && !defined(USE_SIMPLE_THREADED_LEVEL3)
int nodes;
#endif
PRINT_DEBUG_NAME;
args.m = *M;
args.n = *N;
args.k = *K;
args.a = (void *)a;
args.b = (void *)b;
args.c = (void *)c;
args.lda = *ldA;
args.ldb = *ldB;
args.ldc = *ldC;
args.alpha = (void *)alpha;
args.beta = (void *)beta;
transA = *TRANSA;
transB = *TRANSB;
TOUPPER(transA);
TOUPPER(transB);
transa = -1;
transb = -1;
if (transA == 'N') transa = 0;
if (transA == 'T') transa = 1;
#ifndef COMPLEX
if (transA == 'R') transa = 0;
if (transA == 'C') transa = 1;
#else
if (transA == 'R') transa = 2;
if (transA == 'C') transa = 3;
#endif
if (transB == 'N') transb = 0;
if (transB == 'T') transb = 1;
#ifndef COMPLEX
if (transB == 'R') transb = 0;
if (transB == 'C') transb = 1;
#else
if (transB == 'R') transb = 2;
if (transB == 'C') transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = 0;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
if (info){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB,
blasint m, blasint n, blasint k,
#ifndef COMPLEX
FLOAT alpha,
FLOAT *a, blasint lda,
FLOAT *b, blasint ldb,
FLOAT beta,
FLOAT *c, blasint ldc) {
#else
void *valpha,
void *va, blasint lda,
void *vb, blasint ldb,
void *vbeta,
void *vc, blasint ldc) {
FLOAT *alpha = (FLOAT*) valpha;
FLOAT *beta = (FLOAT*) vbeta;
FLOAT *a = (FLOAT*) va;
FLOAT *b = (FLOAT*) vb;
FLOAT *c = (FLOAT*) vc;
#endif
blas_arg_t args;
int transa, transb;
blasint nrowa, nrowb, info;
XFLOAT *buffer;
XFLOAT *sa, *sb;
#ifdef SMP
double MNK;
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
#if defined(SMP) && !defined(NO_AFFINITY) && !defined(USE_SIMPLE_THREADED_LEVEL3)
int nodes;
#endif
PRINT_DEBUG_CNAME;
#if !defined(COMPLEX) && !defined(DOUBLE) && defined(USE_SGEMM_KERNEL_DIRECT)
if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans && sgemm_kernel_direct_performant(m,n,k)) {
sgemm_kernel_direct(m, n, k, a, lda, b, ldb, c, ldc);
return;
}
#endif
#ifndef COMPLEX
args.alpha = (void *)&alpha;
args.beta = (void *)&beta;
#else
args.alpha = (void *)alpha;
args.beta = (void *)beta;
#endif
transa = -1;
transb = -1;
info = 0;
if (order == CblasColMajor) {
args.m = m;
args.n = n;
args.k = k;
args.a = (void *)a;
args.b = (void *)b;
args.c = (void *)c;
args.lda = lda;
args.ldb = ldb;
args.ldc = ldc;
if (TransA == CblasNoTrans) transa = 0;
if (TransA == CblasTrans) transa = 1;
#ifndef COMPLEX
if (TransA == CblasConjNoTrans) transa = 0;
if (TransA == CblasConjTrans) transa = 1;
#else
if (TransA == CblasConjNoTrans) transa = 2;
if (TransA == CblasConjTrans) transa = 3;
#endif
if (TransB == CblasNoTrans) transb = 0;
if (TransB == CblasTrans) transb = 1;
#ifndef COMPLEX
if (TransB == CblasConjNoTrans) transb = 0;
if (TransB == CblasConjTrans) transb = 1;
#else
if (TransB == CblasConjNoTrans) transb = 2;
if (TransB == CblasConjTrans) transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = -1;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
}
if (order == CblasRowMajor) {
args.m = n;
args.n = m;
args.k = k;
args.a = (void *)b;
args.b = (void *)a;
args.c = (void *)c;
args.lda = ldb;
args.ldb = lda;
args.ldc = ldc;
if (TransB == CblasNoTrans) transa = 0;
if (TransB == CblasTrans) transa = 1;
#ifndef COMPLEX
if (TransB == CblasConjNoTrans) transa = 0;
if (TransB == CblasConjTrans) transa = 1;
#else
if (TransB == CblasConjNoTrans) transa = 2;
if (TransB == CblasConjTrans) transa = 3;
#endif
if (TransA == CblasNoTrans) transb = 0;
if (TransA == CblasTrans) transb = 1;
#ifndef COMPLEX
if (TransA == CblasConjNoTrans) transb = 0;
if (TransA == CblasConjTrans) transb = 1;
#else
if (TransA == CblasConjNoTrans) transb = 2;
if (TransA == CblasConjTrans) transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = -1;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if ((args.m == 0) || (args.n == 0)) return;
#if 0
fprintf(stderr, "m = %4d n = %d k = %d lda = %4d ldb = %4d ldc = %4d\n",
args.m, args.n, args.k, args.lda, args.ldb, args.ldc);
#endif
IDEBUG_START;
FUNCTION_PROFILE_START();
buffer = (XFLOAT *)blas_memory_alloc(0);
sa = (XFLOAT *)((BLASLONG)buffer +GEMM_OFFSET_A);
sb = (XFLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#ifdef SMP
mode |= (transa << BLAS_TRANSA_SHIFT);
mode |= (transb << BLAS_TRANSB_SHIFT);
MNK = (double) args.m * (double) args.n * (double) args.k;
if ( MNK <= (SMP_THRESHOLD_MIN * (double) GEMM_MULTITHREAD_THRESHOLD) )
args.nthreads = 1;
else
args.nthreads = num_cpu_avail(3);
args.common = NULL;
if (args.nthreads == 1) {
#endif
(gemm[(transb << 2) | transa])(&args, NULL, NULL, sa, sb, 0);
#ifdef SMP
} else {
#ifndef USE_SIMPLE_THREADED_LEVEL3
#ifndef NO_AFFINITY
nodes = get_num_nodes();
if ((nodes > 1) && get_node_equal()) {
args.nthreads /= nodes;
gemm_thread_mn(mode, &args, NULL, NULL, gemm[16 | (transb << 2) | transa], sa, sb, nodes);
} else {
#endif
(gemm[16 | (transb << 2) | transa])(&args, NULL, NULL, sa, sb, 0);
#else
GEMM_THREAD(mode, &args, NULL, NULL, gemm[(transb << 2) | transa], sa, sb, args.nthreads);
#endif
#ifndef USE_SIMPLE_THREADED_LEVEL3
#ifndef NO_AFFINITY
}
#endif
#endif
#endif
#ifdef SMP
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(COMPSIZE * COMPSIZE, args.m * args.k + args.k * args.n + args.m * args.n, 2 * args.m * args.n * args.k);
IDEBUG_END;
return;
}
Reference in New Issue View Git Blame Copy Permalink