|
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 |
|
|
kusano |
2b45e8 |
#ifndef XCONJ
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
#define ZAXPY AXPYU_K
|
|
kusano |
2b45e8 |
#define ZDOT DOTU_K
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define ZAXPY AXPYC_K
|
|
kusano |
2b45e8 |
#define ZDOT DOTC_K
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
#define ZAXPY AXPYU_K
|
|
kusano |
2b45e8 |
#define ZDOT DOTC_K
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define ZAXPY AXPYC_K
|
|
kusano |
2b45e8 |
#define ZDOT DOTU_K
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRANS
|
|
kusano |
2b45e8 |
#define M m
|
|
kusano |
2b45e8 |
#define N n
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define N m
|
|
kusano |
2b45e8 |
#define M n
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
void CNAME(BLASLONG m, BLASLONG n, BLASLONG ku, BLASLONG kl, FLOAT alpha_r, FLOAT alpha_i,
|
|
kusano |
2b45e8 |
FLOAT *a, BLASLONG lda,
|
|
kusano |
2b45e8 |
FLOAT *x, BLASLONG incx, FLOAT *y, BLASLONG incy, void *buffer){
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
BLASLONG i, offset_u, offset_l, start, end, length;
|
|
kusano |
2b45e8 |
FLOAT *X = x;
|
|
kusano |
2b45e8 |
FLOAT *Y = y;
|
|
kusano |
2b45e8 |
FLOAT *gemvbuffer = (FLOAT *)buffer;
|
|
kusano |
2b45e8 |
FLOAT *bufferY = gemvbuffer;
|
|
kusano |
2b45e8 |
FLOAT *bufferX = gemvbuffer;
|
|
kusano |
2b45e8 |
#ifdef TRANS
|
|
kusano |
2b45e8 |
FLOAT _Complex temp;
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (incy != 1) {
|
|
kusano |
2b45e8 |
Y = bufferY;
|
|
kusano |
2b45e8 |
bufferX = (FLOAT *)(((BLASLONG)bufferY + M * sizeof(FLOAT) * 2 + 4095) & ~4095);
|
|
kusano |
2b45e8 |
gemvbuffer = bufferX;
|
|
kusano |
2b45e8 |
COPY_K(M, y, incy, Y, 1);
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (incx != 1) {
|
|
kusano |
2b45e8 |
X = bufferX;
|
|
kusano |
2b45e8 |
gemvbuffer = (FLOAT *)(((BLASLONG)bufferX + N * sizeof(FLOAT) * 2 + 4095) & ~4095);
|
|
kusano |
2b45e8 |
COPY_K(N, x, incx, X, 1);
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
offset_u = ku;
|
|
kusano |
2b45e8 |
offset_l = ku + m;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
for (i = 0; i < MIN(n, m + ku); i++) {
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
start = MAX(offset_u, 0);
|
|
kusano |
2b45e8 |
end = MIN(offset_l, ku + kl + 1);
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
length = end - start;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRANS
|
|
kusano |
2b45e8 |
ZAXPY(length, 0, 0,
|
|
kusano |
2b45e8 |
#ifndef XCONJ
|
|
kusano |
2b45e8 |
alpha_r * X[i * 2 + 0] - alpha_i * X[i * 2 + 1],
|
|
kusano |
2b45e8 |
alpha_i * X[i * 2 + 0] + alpha_r * X[i * 2 + 1],
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
alpha_r * X[i * 2 + 0] + alpha_i * X[i * 2 + 1],
|
|
kusano |
2b45e8 |
alpha_i * X[i * 2 + 0] - alpha_r * X[i * 2 + 1],
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
a + start * 2, 1, Y + (start - offset_u) * 2, 1, NULL, 0);
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef XCONJ
|
|
kusano |
2b45e8 |
temp = ZDOT(length, a + start * 2, 1, X + (start - offset_u) * 2, 1);
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
temp = ZDOT(length, X + (start - offset_u) * 2, 1, a + start * 2, 1);
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if !defined(XCONJ) || !defined(CONJ)
|
|
kusano |
2b45e8 |
Y[i * 2 + 0] += alpha_r * CREAL(temp) - alpha_i * CIMAG(temp);
|
|
kusano |
2b45e8 |
Y[i * 2 + 1] += alpha_i * CREAL(temp) + alpha_r * CIMAG(temp);
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
Y[i * 2 + 0] += alpha_r * CREAL(temp) + alpha_i * CIMAG(temp);
|
|
kusano |
2b45e8 |
Y[i * 2 + 1] += alpha_i * CREAL(temp) - alpha_r * CIMAG(temp);
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
offset_u --;
|
|
kusano |
2b45e8 |
offset_l --;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
a += lda * 2;
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
if (incy != 1) {
|
|
kusano |
2b45e8 |
COPY_K(M, Y, 1, y, incy);
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
return;
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
|