|
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 "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 "QTRTRI"
|
|
kusano |
2b45e8 |
#elif defined(DOUBLE)
|
|
kusano |
2b45e8 |
#define ERROR_NAME "DTRTRI"
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define ERROR_NAME "STRTRI"
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
static blasint (*trtri_single[])(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLASLONG) ={
|
|
kusano |
2b45e8 |
TRTRI_UU_SINGLE, TRTRI_UN_SINGLE, TRTRI_LU_SINGLE, TRTRI_LN_SINGLE,
|
|
kusano |
2b45e8 |
};
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef SMP
|
|
kusano |
2b45e8 |
static blasint (*trtri_parallel[])(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLASLONG) ={
|
|
kusano |
2b45e8 |
TRTRI_UU_PARALLEL, TRTRI_UN_PARALLEL, TRTRI_LU_PARALLEL, TRTRI_LN_PARALLEL,
|
|
kusano |
2b45e8 |
};
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
int NAME(char *UPLO, char *DIAG, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blas_arg_t args;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blasint uplo_arg = *UPLO;
|
|
kusano |
2b45e8 |
blasint diag_arg = *DIAG;
|
|
kusano |
2b45e8 |
blasint uplo, diag;
|
|
kusano |
2b45e8 |
blasint info;
|
|
kusano |
2b45e8 |
FLOAT *buffer;
|
|
kusano |
2b45e8 |
#ifdef PPC440
|
|
kusano |
2b45e8 |
extern
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
FLOAT *sa, *sb;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PRINT_DEBUG_NAME;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
args.n = *N;
|
|
kusano |
2b45e8 |
args.a = (void *)a;
|
|
kusano |
2b45e8 |
args.lda = *ldA;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
TOUPPER(uplo_arg);
|
|
kusano |
2b45e8 |
TOUPPER(diag_arg);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
uplo = -1;
|
|
kusano |
2b45e8 |
if (uplo_arg == 'U') uplo = 0;
|
|
kusano |
2b45e8 |
if (uplo_arg == 'L') uplo = 1;
|
|
kusano |
2b45e8 |
diag = -1;
|
|
kusano |
2b45e8 |
if (diag_arg == 'U') diag = 0;
|
|
kusano |
2b45e8 |
if (diag_arg == 'N') diag = 1;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
info = 0;
|
|
kusano |
2b45e8 |
if (args.lda < MAX(1,args.n)) info = 5;
|
|
kusano |
2b45e8 |
if (args.n < 0) info = 3;
|
|
kusano |
2b45e8 |
if (diag < 0) info = 2;
|
|
kusano |
2b45e8 |
if (uplo < 0) info = 1;
|
|
kusano |
2b45e8 |
if (info) {
|
|
kusano |
2b45e8 |
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
|
|
kusano |
2b45e8 |
*Info = - info;
|
|
kusano |
2b45e8 |
return 0;
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
*Info = 0;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (args.n == 0) return 0;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (diag) {
|
|
kusano |
2b45e8 |
if (AMIN_K(args.n, args.a, args.lda + 1) == ZERO) {
|
|
kusano |
2b45e8 |
*Info = IAMIN_K(args.n, args.a, args.lda + 1);
|
|
kusano |
2b45e8 |
return 0;
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
IDEBUG_START;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FUNCTION_PROFILE_START();
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef PPC440
|
|
kusano |
2b45e8 |
buffer = (FLOAT *)blas_memory_alloc(1);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
sa = (FLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
|
|
kusano |
2b45e8 |
sb = (FLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef SMP
|
|
kusano |
2b45e8 |
args.nthreads = num_cpu_avail(4);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (args.nthreads == 1) {
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
*Info = (trtri_single[(uplo << 1) | diag])(&args, NULL, NULL, sa, sb, 0);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef SMP
|
|
kusano |
2b45e8 |
} else {
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
*Info = (trtri_parallel[(uplo << 1) | diag])(&args, NULL, NULL, sa, sb, 0);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef PPC440
|
|
kusano |
2b45e8 |
blas_memory_free(buffer);
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FUNCTION_PROFILE_END(COMPSIZE * COMPSIZE, .5 * args.n * args.n,
|
|
kusano |
2b45e8 |
args.n * (1./3. + args.n * ( 1./2. + args.n * 1./6.))
|
|
kusano |
2b45e8 |
+ args.n * (1./3. + args.n * (-1./2. + args.n * 1./6.)));
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
IDEBUG_END;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
return 0;
|
|
kusano |
2b45e8 |
}
|