Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/lapack/lauum/lauum_L_parallel.c

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
kusano 2b45e8
blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG myid) {
kusano 2b45e8
kusano 2b45e8
  BLASLONG n, bk, i, blocking, lda;
kusano 2b45e8
  int mode;
kusano 2b45e8
  blas_arg_t newarg;
kusano 2b45e8
  FLOAT *a;
kusano 2b45e8
  FLOAT alpha[2] = { ONE, ZERO};
kusano 2b45e8
kusano 2b45e8
#ifndef COMPLEX
kusano 2b45e8
#ifdef XDOUBLE
kusano 2b45e8
  mode  =  BLAS_XDOUBLE | BLAS_REAL;
kusano 2b45e8
#elif defined(DOUBLE)
kusano 2b45e8
  mode  =  BLAS_DOUBLE  | BLAS_REAL;
kusano 2b45e8
#else
kusano 2b45e8
  mode  =  BLAS_SINGLE  | BLAS_REAL;
kusano 2b45e8
#endif  
kusano 2b45e8
#else
kusano 2b45e8
#ifdef XDOUBLE
kusano 2b45e8
  mode  =  BLAS_XDOUBLE | BLAS_COMPLEX;
kusano 2b45e8
#elif defined(DOUBLE)
kusano 2b45e8
  mode  =  BLAS_DOUBLE  | BLAS_COMPLEX;
kusano 2b45e8
#else
kusano 2b45e8
  mode  =  BLAS_SINGLE  | BLAS_COMPLEX;
kusano 2b45e8
#endif  
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
  if (args -> nthreads  == 1) {
kusano 2b45e8
    LAUUM_L_SINGLE(args, NULL, NULL, sa, sb, 0); 
kusano 2b45e8
    return 0;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  n  = args -> n;
kusano 2b45e8
  a  = (FLOAT *)args -> a;
kusano 2b45e8
  lda = args -> lda;
kusano 2b45e8
kusano 2b45e8
  if (range_n) n  = range_n[1] - range_n[0];
kusano 2b45e8
kusano 2b45e8
  if (n <= GEMM_UNROLL_N * 2) {
kusano 2b45e8
    LAUUM_L_SINGLE(args, NULL, range_n, sa, sb, 0);
kusano 2b45e8
    return 0;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  newarg.lda = lda;
kusano 2b45e8
  newarg.ldb = lda;
kusano 2b45e8
  newarg.ldc = lda;
kusano 2b45e8
  newarg.alpha = alpha;
kusano 2b45e8
  newarg.beta = NULL;
kusano 2b45e8
  newarg.nthreads = args -> nthreads;
kusano 2b45e8
  
kusano 2b45e8
  blocking = (n / 2 + GEMM_UNROLL_N - 1) & ~(GEMM_UNROLL_N - 1);
kusano 2b45e8
  if (blocking > GEMM_Q) blocking = GEMM_Q;
kusano 2b45e8
kusano 2b45e8
  for (i = 0; i < n; i += blocking) {
kusano 2b45e8
kusano 2b45e8
    bk = n - i;
kusano 2b45e8
    if (bk > blocking) bk = blocking;
kusano 2b45e8
    
kusano 2b45e8
    newarg.n = i;
kusano 2b45e8
    newarg.k = bk;
kusano 2b45e8
    newarg.a = a + i * COMPSIZE;
kusano 2b45e8
    newarg.c = a;
kusano 2b45e8
kusano 2b45e8
    syrk_thread(mode | BLAS_TRANSA_T | BLAS_TRANSB_N | BLAS_UPLO,
kusano 2b45e8
		&newarg, NULL, NULL, (void *)HERK_LC, sa, sb, args -> nthreads);
kusano 2b45e8
kusano 2b45e8
    newarg.m = bk;
kusano 2b45e8
    newarg.n = i;
kusano 2b45e8
    newarg.a = a + (i + i * lda) * COMPSIZE;
kusano 2b45e8
    newarg.b = a + (i          ) * COMPSIZE;
kusano 2b45e8
kusano 2b45e8
    gemm_thread_n(mode | BLAS_TRANSA_T,
kusano 2b45e8
		  &newarg, NULL, NULL, (void *)TRMM_LCLN, sa, sb, args -> nthreads);
kusano 2b45e8
kusano 2b45e8
    newarg.m = bk;
kusano 2b45e8
    newarg.n = bk;
kusano 2b45e8
    newarg.a = a + (i + i * lda) * COMPSIZE;
kusano 2b45e8
kusano 2b45e8
    CNAME(&newarg, NULL, NULL, sa, sb, 0);
kusano 2b45e8
  }
kusano 2b45e8
  
kusano 2b45e8
  return 0;
kusano 2b45e8
}