|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
/* Copyright 2009, 2010 The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/* All rights reserved. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* Redistribution and use in source and binary forms, with or */
|
|
kusano |
2b45e8 |
/* without modification, are permitted provided that the following */
|
|
kusano |
2b45e8 |
/* conditions are met: */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 1. Redistributions of source code must retain the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 2. Redistributions in binary form must reproduce the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer in the documentation and/or other materials */
|
|
kusano |
2b45e8 |
/* provided with the distribution. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
|
|
kusano |
2b45e8 |
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
|
|
kusano |
2b45e8 |
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
|
|
kusano |
2b45e8 |
/* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
|
|
kusano |
2b45e8 |
/* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
|
|
kusano |
2b45e8 |
/* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
|
|
kusano |
2b45e8 |
/* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
|
|
kusano |
2b45e8 |
/* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
|
|
kusano |
2b45e8 |
/* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
|
|
kusano |
2b45e8 |
/* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
|
|
kusano |
2b45e8 |
/* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
|
|
kusano |
2b45e8 |
/* POSSIBILITY OF SUCH DAMAGE. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* The views and conclusions contained in the software and */
|
|
kusano |
2b45e8 |
/* documentation are those of the authors and should not be */
|
|
kusano |
2b45e8 |
/* interpreted as representing official policies, either expressed */
|
|
kusano |
2b45e8 |
/* or implied, of The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#include <stdio.h></stdio.h>
|
|
kusano |
2b45e8 |
#include <ctype.h></ctype.h>
|
|
kusano |
2b45e8 |
#include "common.h"
|
|
kusano |
2b45e8 |
#ifdef FUNCTION_PROFILE
|
|
kusano |
2b45e8 |
#include "functable.h"
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef XDOUBLE
|
|
kusano |
2b45e8 |
#define ERROR_NAME "XTPMV "
|
|
kusano |
2b45e8 |
#elif defined(DOUBLE)
|
|
kusano |
2b45e8 |
#define ERROR_NAME "ZTPMV "
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define ERROR_NAME "CTPMV "
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
static int (*tpmv[])(BLASLONG, FLOAT *, FLOAT *, BLASLONG, void *) = {
|
|
kusano |
2b45e8 |
#ifdef XDOUBLE
|
|
kusano |
2b45e8 |
xtpmv_NUU, xtpmv_NUN, xtpmv_NLU, xtpmv_NLN,
|
|
kusano |
2b45e8 |
xtpmv_TUU, xtpmv_TUN, xtpmv_TLU, xtpmv_TLN,
|
|
kusano |
2b45e8 |
xtpmv_RUU, xtpmv_RUN, xtpmv_RLU, xtpmv_RLN,
|
|
kusano |
2b45e8 |
xtpmv_CUU, xtpmv_CUN, xtpmv_CLU, xtpmv_CLN,
|
|
kusano |
2b45e8 |
#elif defined(DOUBLE)
|
|
kusano |
2b45e8 |
ztpmv_NUU, ztpmv_NUN, ztpmv_NLU, ztpmv_NLN,
|
|
kusano |
2b45e8 |
ztpmv_TUU, ztpmv_TUN, ztpmv_TLU, ztpmv_TLN,
|
|
kusano |
2b45e8 |
ztpmv_RUU, ztpmv_RUN, ztpmv_RLU, ztpmv_RLN,
|
|
kusano |
2b45e8 |
ztpmv_CUU, ztpmv_CUN, ztpmv_CLU, ztpmv_CLN,
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
ctpmv_NUU, ctpmv_NUN, ctpmv_NLU, ctpmv_NLN,
|
|
kusano |
2b45e8 |
ctpmv_TUU, ctpmv_TUN, ctpmv_TLU, ctpmv_TLN,
|
|
kusano |
2b45e8 |
ctpmv_RUU, ctpmv_RUN, ctpmv_RLU, ctpmv_RLN,
|
|
kusano |
2b45e8 |
ctpmv_CUU, ctpmv_CUN, ctpmv_CLU, ctpmv_CLN,
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
};
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef SMP
|
|
kusano |
2b45e8 |
static int (*tpmv_thread[])(BLASLONG, FLOAT *, FLOAT *, BLASLONG, FLOAT *, int) = {
|
|
kusano |
2b45e8 |
#ifdef XDOUBLE
|
|
kusano |
2b45e8 |
xtpmv_thread_NUU, xtpmv_thread_NUN, xtpmv_thread_NLU, xtpmv_thread_NLN,
|
|
kusano |
2b45e8 |
xtpmv_thread_TUU, xtpmv_thread_TUN, xtpmv_thread_TLU, xtpmv_thread_TLN,
|
|
kusano |
2b45e8 |
xtpmv_thread_RUU, xtpmv_thread_RUN, xtpmv_thread_RLU, xtpmv_thread_RLN,
|
|
kusano |
2b45e8 |
xtpmv_thread_CUU, xtpmv_thread_CUN, xtpmv_thread_CLU, xtpmv_thread_CLN,
|
|
kusano |
2b45e8 |
#elif defined(DOUBLE)
|
|
kusano |
2b45e8 |
ztpmv_thread_NUU, ztpmv_thread_NUN, ztpmv_thread_NLU, ztpmv_thread_NLN,
|
|
kusano |
2b45e8 |
ztpmv_thread_TUU, ztpmv_thread_TUN, ztpmv_thread_TLU, ztpmv_thread_TLN,
|
|
kusano |
2b45e8 |
ztpmv_thread_RUU, ztpmv_thread_RUN, ztpmv_thread_RLU, ztpmv_thread_RLN,
|
|
kusano |
2b45e8 |
ztpmv_thread_CUU, ztpmv_thread_CUN, ztpmv_thread_CLU, ztpmv_thread_CLN,
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
ctpmv_thread_NUU, ctpmv_thread_NUN, ctpmv_thread_NLU, ctpmv_thread_NLN,
|
|
kusano |
2b45e8 |
ctpmv_thread_TUU, ctpmv_thread_TUN, ctpmv_thread_TLU, ctpmv_thread_TLN,
|
|
kusano |
2b45e8 |
ctpmv_thread_RUU, ctpmv_thread_RUN, ctpmv_thread_RLU, ctpmv_thread_RLN,
|
|
kusano |
2b45e8 |
ctpmv_thread_CUU, ctpmv_thread_CUN, ctpmv_thread_CLU, ctpmv_thread_CLN,
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
};
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CBLAS
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
void NAME(char *UPLO, char *TRANS, char *DIAG,
|
|
kusano |
2b45e8 |
blasint *N, FLOAT *a, FLOAT *x, blasint *INCX){
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
char uplo_arg = *UPLO;
|
|
kusano |
2b45e8 |
char trans_arg = *TRANS;
|
|
kusano |
2b45e8 |
char diag_arg = *DIAG;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blasint n = *N;
|
|
kusano |
2b45e8 |
blasint incx = *INCX;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blasint info;
|
|
kusano |
2b45e8 |
int uplo;
|
|
kusano |
2b45e8 |
int unit;
|
|
kusano |
2b45e8 |
int trans;
|
|
kusano |
2b45e8 |
FLOAT *buffer;
|
|
kusano |
2b45e8 |
#ifdef SMP
|
|
kusano |
2b45e8 |
int nthreads;
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PRINT_DEBUG_NAME;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
TOUPPER(uplo_arg);
|
|
kusano |
2b45e8 |
TOUPPER(trans_arg);
|
|
kusano |
2b45e8 |
TOUPPER(diag_arg);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
trans = -1;
|
|
kusano |
2b45e8 |
unit = -1;
|
|
kusano |
2b45e8 |
uplo = -1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (trans_arg == 'N') trans = 0;
|
|
kusano |
2b45e8 |
if (trans_arg == 'T') trans = 1;
|
|
kusano |
2b45e8 |
if (trans_arg == 'R') trans = 2;
|
|
kusano |
2b45e8 |
if (trans_arg == 'C') trans = 3;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (diag_arg == 'U') unit = 0;
|
|
kusano |
2b45e8 |
if (diag_arg == 'N') unit = 1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (uplo_arg == 'U') uplo = 0;
|
|
kusano |
2b45e8 |
if (uplo_arg == 'L') uplo = 1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
info = 0;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (incx == 0) info = 7;
|
|
kusano |
2b45e8 |
if (n < 0) info = 4;
|
|
kusano |
2b45e8 |
if (unit < 0) info = 3;
|
|
kusano |
2b45e8 |
if (trans < 0) info = 2;
|
|
kusano |
2b45e8 |
if (uplo < 0) info = 1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (info != 0) {
|
|
kusano |
2b45e8 |
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
|
|
kusano |
2b45e8 |
return;
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
|
|
kusano |
2b45e8 |
enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
|
|
kusano |
2b45e8 |
blasint n, FLOAT *a, FLOAT *x, blasint incx) {
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
int trans, uplo, unit;
|
|
kusano |
2b45e8 |
blasint info;
|
|
kusano |
2b45e8 |
FLOAT *buffer;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PRINT_DEBUG_CNAME;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
unit = -1;
|
|
kusano |
2b45e8 |
uplo = -1;
|
|
kusano |
2b45e8 |
trans = -1;
|
|
kusano |
2b45e8 |
info = 0;
|
|
kusano |
2b45e8 |
#ifdef SMP
|
|
kusano |
2b45e8 |
int nthreads;
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (order == CblasColMajor) {
|
|
kusano |
2b45e8 |
if (Uplo == CblasUpper) uplo = 0;
|
|
kusano |
2b45e8 |
if (Uplo == CblasLower) uplo = 1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (TransA == CblasNoTrans) trans = 0;
|
|
kusano |
2b45e8 |
if (TransA == CblasTrans) trans = 1;
|
|
kusano |
2b45e8 |
if (TransA == CblasConjNoTrans) trans = 2;
|
|
kusano |
2b45e8 |
if (TransA == CblasConjTrans) trans = 3;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (Diag == CblasUnit) unit = 0;
|
|
kusano |
2b45e8 |
if (Diag == CblasNonUnit) unit = 1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
info = -1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (incx == 0) info = 7;
|
|
kusano |
2b45e8 |
if (n < 0) info = 4;
|
|
kusano |
2b45e8 |
if (unit < 0) info = 3;
|
|
kusano |
2b45e8 |
if (trans < 0) info = 2;
|
|
kusano |
2b45e8 |
if (uplo < 0) info = 1;
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (order == CblasRowMajor) {
|
|
kusano |
2b45e8 |
if (Uplo == CblasUpper) uplo = 1;
|
|
kusano |
2b45e8 |
if (Uplo == CblasLower) uplo = 0;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (TransA == CblasNoTrans) trans = 1;
|
|
kusano |
2b45e8 |
if (TransA == CblasTrans) trans = 0;
|
|
kusano |
2b45e8 |
if (TransA == CblasConjNoTrans) trans = 3;
|
|
kusano |
2b45e8 |
if (TransA == CblasConjTrans) trans = 2;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (Diag == CblasUnit) unit = 0;
|
|
kusano |
2b45e8 |
if (Diag == CblasNonUnit) unit = 1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
info = -1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (incx == 0) info = 7;
|
|
kusano |
2b45e8 |
if (n < 0) info = 4;
|
|
kusano |
2b45e8 |
if (unit < 0) info = 3;
|
|
kusano |
2b45e8 |
if (trans < 0) info = 2;
|
|
kusano |
2b45e8 |
if (uplo < 0) info = 1;
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (info >= 0) {
|
|
kusano |
2b45e8 |
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
|
|
kusano |
2b45e8 |
return;
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (n == 0) return;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
IDEBUG_START;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FUNCTION_PROFILE_START();
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (incx < 0 ) x -= (n - 1) * incx * 2;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
buffer = (FLOAT *)blas_memory_alloc(1);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef SMP
|
|
kusano |
2b45e8 |
nthreads = num_cpu_avail(2);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (nthreads == 1) {
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
(tpmv[(trans<<2) | (uplo<<1) | unit])(n, a, x, incx, buffer);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef SMP
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
} else {
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
(tpmv_thread[(trans<<2) | (uplo<<1) | unit])(n, a, x, incx, buffer, nthreads);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blas_memory_free(buffer);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FUNCTION_PROFILE_END(4, n * n / 2 + n, n * n);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
IDEBUG_END;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
return;
|
|
kusano |
2b45e8 |
}
|