OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Optimize Zgemv

tags/v0.3.6^2
maamountki GitHub 7 years ago
parent
453bfa7e71
commit
e6c0e39492
No known key found for this signature in database GPG Key ID: 4AEE18F83AFDEB23
26 changed files with 2866 additions and 1713 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +4
    -4
      cpuid_zarch.c
  2. +2
    -2
      kernel/zarch/KERNEL.Z13
  3. +4
    -4
      kernel/zarch/KERNEL.Z14
  4. +24
    -22
      kernel/zarch/camax.c
  5. +24
    -22
      kernel/zarch/camin.c
  6. +2
    -2
      kernel/zarch/caxpy.c
  7. +743
    -0
      kernel/zarch/cgemv_n_4.c
  8. +671
    -0
      kernel/zarch/cgemv_t_4.c
  9. +7
    -2
      kernel/zarch/icamax.c
  10. +7
    -2
      kernel/zarch/icamin.c
  11. +10
    -1
      kernel/zarch/idamax.c
  12. +10
    -1
      kernel/zarch/idamin.c
  13. +10
    -1
      kernel/zarch/idmax.c
  14. +10
    -1
      kernel/zarch/idmin.c
  15. +10
    -1
      kernel/zarch/isamax.c
  16. +10
    -1
      kernel/zarch/isamin.c
  17. +10
    -1
      kernel/zarch/ismax.c
  18. +10
    -1
      kernel/zarch/ismin.c
  19. +7
    -2
      kernel/zarch/izamax.c
  20. +7
    -2
      kernel/zarch/izamin.c
  21. +25
    -23
      kernel/zarch/zamax.c
  22. +24
    -22
      kernel/zarch/zamin.c
  23. +2
    -2
      kernel/zarch/zaxpy.c
  24. +598
    -803
      kernel/zarch/zgemv_n_4.c
  25. +548
    -719
      kernel/zarch/zgemv_t_4.c
  26. +87
    -72
      ztest/gemv.c

+ 4
- 4
cpuid_zarch.c View File

@@ -27,9 +27,9 @@


#include <string.h> #include <string.h>


#define CPU_GENERIC 0
#define CPU_Z13 1
#define CPU_Z14 2
#define CPU_GENERIC 0
#define CPU_Z13 1
#define CPU_Z14 2


static char *cpuname[] = { static char *cpuname[] = {
"ZARCH_GENERIC", "ZARCH_GENERIC",
@@ -112,7 +112,7 @@ void get_cpuconfig(void)
printf("#define Z13\n"); printf("#define Z13\n");
printf("#define DTB_DEFAULT_ENTRIES 64\n"); printf("#define DTB_DEFAULT_ENTRIES 64\n");
break; break;
case CPU_Z14:
case CPU_Z14:
printf("#define Z14\n"); printf("#define Z14\n");
printf("#define DTB_DEFAULT_ENTRIES 64\n"); printf("#define DTB_DEFAULT_ENTRIES 64\n");
break; break;


+ 2
- 2
kernel/zarch/KERNEL.Z13 View File

@@ -74,12 +74,12 @@ ZSWAPKERNEL = zswap.c
SGEMVNKERNEL = ../arm/gemv_n.c SGEMVNKERNEL = ../arm/gemv_n.c
DGEMVNKERNEL = dgemv_n_4.c DGEMVNKERNEL = dgemv_n_4.c
CGEMVNKERNEL = ../arm/zgemv_n.c CGEMVNKERNEL = ../arm/zgemv_n.c
ZGEMVNKERNEL = ../arm/zgemv_n.c
ZGEMVNKERNEL = zgemv_n_4.c


SGEMVTKERNEL = ../arm/gemv_t.c SGEMVTKERNEL = ../arm/gemv_t.c
DGEMVTKERNEL = dgemv_t_4.c DGEMVTKERNEL = dgemv_t_4.c
CGEMVTKERNEL = ../arm/zgemv_t.c CGEMVTKERNEL = ../arm/zgemv_t.c
ZGEMVTKERNEL = ../arm/zgemv_t.c
ZGEMVTKERNEL = zgemv_t_4.c


STRMMKERNEL = strmm8x4V.S STRMMKERNEL = strmm8x4V.S
DTRMMKERNEL = trmm8x4V.S DTRMMKERNEL = trmm8x4V.S


+ 4
- 4
kernel/zarch/KERNEL.Z14 View File

@@ -73,13 +73,13 @@ ZSWAPKERNEL = zswap.c


SGEMVNKERNEL = sgemv_n_4.c SGEMVNKERNEL = sgemv_n_4.c
DGEMVNKERNEL = dgemv_n_4.c DGEMVNKERNEL = dgemv_n_4.c
CGEMVNKERNEL = ../arm/zgemv_n.c
ZGEMVNKERNEL = ../arm/zgemv_n.c
CGEMVNKERNEL = cgemv_n_4.c
ZGEMVNKERNEL = zgemv_n_4.c


SGEMVTKERNEL = sgemv_t_4.c SGEMVTKERNEL = sgemv_t_4.c
DGEMVTKERNEL = dgemv_t_4.c DGEMVTKERNEL = dgemv_t_4.c
CGEMVTKERNEL = ../arm/zgemv_t.c
ZGEMVTKERNEL = ../arm/zgemv_t.c
CGEMVTKERNEL = cgemv_t_4.c
ZGEMVTKERNEL = zgemv_t_4.c


STRMMKERNEL = strmm8x4V.S STRMMKERNEL = strmm8x4V.S
DTRMMKERNEL = trmm8x4V.S DTRMMKERNEL = trmm8x4V.S


+ 24
- 22
kernel/zarch/camax.c View File

@@ -198,7 +198,7 @@ static FLOAT camax_kernel_32(BLASLONG n, FLOAT *x)
FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) { FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
BLASLONG i = 0; BLASLONG i = 0;
BLASLONG j = 0;
BLASLONG ix = 0;
FLOAT maxf = 0.0; FLOAT maxf = 0.0;
BLASLONG inc_x2; BLASLONG inc_x2;


@@ -216,53 +216,55 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
else else
{ {
maxf=CABS1(x,0); maxf=CABS1(x,0);
ix += 2;
i++; i++;
} }


while (i < n) { while (i < n) {
if (ABS(x[i*2]) > maxf) {
maxf = ABS(x[i*2]);
if (CABS1(x,ix) > maxf) {
maxf = CABS1(x,ix);
} }
ix += 2;
i++; i++;
} }
return (maxf); return (maxf);


} else { } else {


inc_x2 = 2 * inc_x;
maxf=CABS1(x,0); maxf=CABS1(x,0);
i += inc_x2;
j++;
inc_x2 = 2 * inc_x;
ix += inc_x2;
i++;


BLASLONG n1 = (n - 1) & -4; BLASLONG n1 = (n - 1) & -4;
while (j < n1) {
while (i < n1) {


if (CABS1(x,i) > maxf) {
maxf = CABS1(x,i);
if (CABS1(x,ix) > maxf) {
maxf = CABS1(x,ix);
} }
if (CABS1(x,i+inc_x2) > maxf) {
maxf = CABS1(x,i+inc_x2);
if (CABS1(x,ix+inc_x2) > maxf) {
maxf = CABS1(x,ix+inc_x2);
} }
if (CABS1(x,i+inc_x2*2) > maxf) {
maxf = CABS1(x,i+inc_x2*2);
if (CABS1(x,ix+inc_x2*2) > maxf) {
maxf = CABS1(x,ix+inc_x2*2);
} }
if (CABS1(x,i+inc_x2*3) > maxf) {
maxf = CABS1(x,i+inc_x2*3);
if (CABS1(x,ix+inc_x2*3) > maxf) {
maxf = CABS1(x,ix+inc_x2*3);
} }


i += inc_x2 * 4;
ix += inc_x2 * 4;


j += 4;
i += 4;


} }




while (j < n) {
if (CABS1(x,i) > maxf) {
maxf = CABS1(x,i);
while (i < n) {
if (CABS1(x,ix) > maxf) {
maxf = CABS1(x,ix);
} }
i += inc_x2;
j++;
ix += inc_x2;
i++;
} }
return (maxf); return (maxf);
} }


+ 24
- 22
kernel/zarch/camin.c View File

@@ -198,7 +198,7 @@ static FLOAT camin_kernel_32(BLASLONG n, FLOAT *x)
FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) { FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
BLASLONG i = 0; BLASLONG i = 0;
BLASLONG j = 0;
BLASLONG ix = 0;
FLOAT minf = 0.0; FLOAT minf = 0.0;
BLASLONG inc_x2; BLASLONG inc_x2;


@@ -216,53 +216,55 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
else else
{ {
minf=CABS1(x,0); minf=CABS1(x,0);
ix += 2;
i++; i++;
} }


while (i < n) { while (i < n) {
if (ABS(x[i*2]) < minf) {
minf = ABS(x[i*2]);
if (CABS1(x,ix) < minf) {
minf = CABS1(x,ix);
} }
ix += 2;
i++; i++;
} }
return (minf); return (minf);


} else { } else {


inc_x2 = 2 * inc_x;
minf=CABS1(x,0); minf=CABS1(x,0);
i += inc_x2;
j++;
inc_x2 = 2 * inc_x;
ix += inc_x2;
i++;


BLASLONG n1 = (n - 1) & -4; BLASLONG n1 = (n - 1) & -4;
while (j < n1) {
while (i < n1) {


if (CABS1(x,i) < minf) {
minf = CABS1(x,i);
if (CABS1(x,ix) < minf) {
minf = CABS1(x,ix);
} }
if (CABS1(x,i+inc_x2) < minf) {
minf = CABS1(x,i+inc_x2);
if (CABS1(x,ix+inc_x2) < minf) {
minf = CABS1(x,ix+inc_x2);
} }
if (CABS1(x,i+inc_x2*2) < minf) {
minf = CABS1(x,i+inc_x2*2);
if (CABS1(x,ix+inc_x2*2) < minf) {
minf = CABS1(x,ix+inc_x2*2);
} }
if (CABS1(x,i+inc_x2*3) < minf) {
minf = CABS1(x,i+inc_x2*3);
if (CABS1(x,ix+inc_x2*3) < minf) {
minf = CABS1(x,ix+inc_x2*3);
} }


i += inc_x2 * 4;
ix += inc_x2 * 4;


j += 4;
i += 4;


} }




while (j < n) {
if (CABS1(x,i) < minf) {
minf = CABS1(x,i);
while (i < n) {
if (CABS1(x,ix) < minf) {
minf = CABS1(x,ix);
} }
i += inc_x2;
j++;
ix += inc_x2;
i++;
} }
return (minf); return (minf);
} }


+ 2
- 2
kernel/zarch/caxpy.c View File

@@ -110,7 +110,7 @@ static void caxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"agfi %%r1,128 \n\t" "agfi %%r1,128 \n\t"
"brctg %%r0,0b " "brctg %%r0,0b "
: :
:"r"(n),"ZR"((const FLOAT (*)[n * 2])x),"ZR"((FLOAT (*)[n * 2])y),"a"(alpha)
:"r"(n),"ZR"((const FLOAT (*)[n * 2])x),"ZR"((FLOAT (*)[n * 2])y),"ZQ"((const FLOAT (*)[2])alpha)
:"memory","cc","r0","r1","v0","v1","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31" :"memory","cc","r0","r1","v0","v1","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
); );
} }
@@ -118,7 +118,7 @@ static void caxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2) { int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2) {
BLASLONG i = 0; BLASLONG i = 0;
BLASLONG ix = 0, iy = 0; BLASLONG ix = 0, iy = 0;
FLOAT da[2];
FLOAT da[2] __attribute__ ((aligned(16)));


if (n <= 0) return (0); if (n <= 0) return (0);




+ 743
- 0
kernel/zarch/cgemv_n_4.c View File

@@ -0,0 +1,743 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
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.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "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 OPENBLAS PROJECT 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.
*****************************************************************************/

#include <stdlib.h>
#include <stdio.h>
#include "common.h"

#define NBMAX 1024

static void cgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
__asm__ volatile (
"vlrepg %%v16,0(%5) \n\t"
"vlrepg %%v17,8(%5) \n\t"
"vlrepg %%v18,16(%5) \n\t"
"vlrepg %%v19,24(%5) \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vlef %%v20,4(%5),0 \n\t"
"vlef %%v20,4(%5),2 \n\t"
"vflcsb %%v20,%%v20 \n\t"
"vlef %%v20,0(%5),1 \n\t"
"vlef %%v20,0(%5),3 \n\t"

"vlef %%v21,12(%5),0 \n\t"
"vlef %%v21,12(%5),2 \n\t"
"vflcsb %%v21,%%v21 \n\t"
"vlef %%v21,8(%5),1 \n\t"
"vlef %%v21,8(%5),3 \n\t"

"vlef %%v22,20(%5),0 \n\t"
"vlef %%v22,20(%5),2 \n\t"
"vflcsb %%v22,%%v22 \n\t"
"vlef %%v22,16(%5),1 \n\t"
"vlef %%v22,16(%5),3 \n\t"

"vlef %%v23,28(%5),0 \n\t"
"vlef %%v23,28(%5),2 \n\t"
"vflcsb %%v23,%%v23 \n\t"
"vlef %%v23,24(%5),1 \n\t"
"vlef %%v23,24(%5),3 \n\t"
#else
"vlef %%v20,0(%5),1 \n\t"
"vlef %%v20,0(%5),3 \n\t"
"vflcsb %%v20,%%v20 \n\t"
"vlef %%v20,4(%5),0 \n\t"
"vlef %%v20,4(%5),2 \n\t"

"vlef %%v21,8(%5),1 \n\t"
"vlef %%v21,8(%5),3 \n\t"
"vflcsb %%v21,%%v21 \n\t"
"vlef %%v21,12(%5),0 \n\t"
"vlef %%v21,12(%5),2 \n\t"

"vlef %%v22,16(%5),1 \n\t"
"vlef %%v22,16(%5),3 \n\t"
"vflcsb %%v22,%%v22 \n\t"
"vlef %%v22,20(%5),0 \n\t"
"vlef %%v22,20(%5),2 \n\t"

"vlef %%v23,24(%5),1 \n\t"
"vlef %%v23,24(%5),3 \n\t"
"vflcsb %%v23,%%v23 \n\t"
"vlef %%v23,28(%5),0 \n\t"
"vlef %%v23,28(%5),2 \n\t"
#endif
"xgr %%r1,%%r1 \n\t"
"srlg %%r0,%%r0,1 \n\t"
"0: \n\t"
"pfd 1,1024(%%r1,%1) \n\t"
"pfd 1,1024(%%r1,%2) \n\t"
"pfd 1,1024(%%r1,%3) \n\t"
"pfd 1,1024(%%r1,%4) \n\t"
"pfd 2,1024(%%r1,%6) \n\t"

"vlef %%v24,0(%%r1,%1),0 \n\t"
"vlef %%v24,0(%%r1,%1),1 \n\t"
"vlef %%v24,8(%%r1,%1),2 \n\t"
"vlef %%v24,8(%%r1,%1),3 \n\t"
"vlef %%v25,4(%%r1,%1),0 \n\t"
"vlef %%v25,4(%%r1,%1),1 \n\t"
"vlef %%v25,12(%%r1,%1),2 \n\t"
"vlef %%v25,12(%%r1,%1),3 \n\t"
"vlef %%v26,0(%%r1,%2),0 \n\t"
"vlef %%v26,0(%%r1,%2),1 \n\t"
"vlef %%v26,8(%%r1,%2),2 \n\t"
"vlef %%v26,8(%%r1,%2),3 \n\t"
"vlef %%v27,4(%%r1,%2),0 \n\t"
"vlef %%v27,4(%%r1,%2),1 \n\t"
"vlef %%v27,12(%%r1,%2),2 \n\t"
"vlef %%v27,12(%%r1,%2),3 \n\t"

"vl %%v0,0(%%r1,%6) \n\t"
"vfmasb %%v0,%%v24,%%v16,%%v0 \n\t"
"vfmasb %%v0,%%v25,%%v20,%%v0 \n\t"
"vfmasb %%v0,%%v26,%%v17,%%v0 \n\t"
"vfmasb %%v0,%%v27,%%v21,%%v0 \n\t"

"vlef %%v28,0(%%r1,%1),0 \n\t"
"vlef %%v28,0(%%r1,%1),1 \n\t"
"vlef %%v28,8(%%r1,%1),2 \n\t"
"vlef %%v28,8(%%r1,%1),3 \n\t"
"vlef %%v29,4(%%r1,%1),0 \n\t"
"vlef %%v29,4(%%r1,%1),1 \n\t"
"vlef %%v29,12(%%r1,%1),2 \n\t"
"vlef %%v29,12(%%r1,%1),3 \n\t"
"vlef %%v30,0(%%r1,%2),0 \n\t"
"vlef %%v30,0(%%r1,%2),1 \n\t"
"vlef %%v30,8(%%r1,%2),2 \n\t"
"vlef %%v30,8(%%r1,%2),3 \n\t"
"vlef %%v31,4(%%r1,%2),0 \n\t"
"vlef %%v31,4(%%r1,%2),1 \n\t"
"vlef %%v31,12(%%r1,%2),2 \n\t"
"vlef %%v31,12(%%r1,%2),3 \n\t"

"vfmasb %%v0,%%v28,%%v18,%%v0 \n\t"
"vfmasb %%v0,%%v29,%%v22,%%v0 \n\t"
"vfmasb %%v0,%%v30,%%v19,%%v0 \n\t"
"vfmasb %%v0,%%v31,%%v23,%%v0 \n\t"
"vst %%v0,0(%%r1,%6) \n\t"
"agfi %%r1,16 \n\t"
"brctg %%r0,0b \n\t"
:
:"r"(n),"ZR"((const FLOAT (*)[n * 2])ap[0]),"ZR"((const FLOAT (*)[n * 2])ap[1]),"ZR"((const FLOAT (*)[n * 2])ap[2]),"ZR"((const FLOAT (*)[n * 2])ap[3]),"ZQ"((const FLOAT (*)[8])x),"ZR"((FLOAT (*)[n * 2])y)
:"memory","cc","r0","r1","v0","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
);
}

static void cgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{
__asm__ volatile (
"vlrepg %%v16,0(%3) \n\t"
"vlrepg %%v17,8(%3) \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vlef %%v18,4(%3),0 \n\t"
"vlef %%v18,4(%3),2 \n\t"
"vflcsb %%v18,%%v18 \n\t"
"vlef %%v18,0(%3),1 \n\t"
"vlef %%v18,0(%3),3 \n\t"

"vlef %%v19,12(%3),0 \n\t"
"vlef %%v19,12(%3),2 \n\t"
"vflcsb %%v19,%%v19 \n\t"
"vlef %%v19,8(%3),1 \n\t"
"vlef %%v19,8(%3),3 \n\t"
#else
"vlef %%v18,0(%3),1 \n\t"
"vlef %%v18,0(%3),3 \n\t"
"vflcsb %%v18,%%v18 \n\t"
"vlef %%v18,4(%3),0 \n\t"
"vlef %%v18,4(%3),2 \n\t"

"vlef %%v19,8(%3),1 \n\t"
"vlef %%v19,8(%3),3 \n\t"
"vflcsb %%v19,%%v19 \n\t"
"vlef %%v19,12(%3),0 \n\t"
"vlef %%v19,12(%3),2 \n\t"
#endif
"xgr %%r1,%%r1 \n\t"
"srlg %%r0,%%r0,1 \n\t"
"0: \n\t"
"pfd 1,1024(%%r1,%1) \n\t"
"pfd 1,1024(%%r1,%2) \n\t"
"pfd 2,1024(%%r1,%4) \n\t"

"vlef %%v20,0(%%r1,%1),0 \n\t"
"vlef %%v20,0(%%r1,%1),1 \n\t"
"vlef %%v20,8(%%r1,%1),2 \n\t"
"vlef %%v20,8(%%r1,%1),3 \n\t"
"vlef %%v21,4(%%r1,%1),0 \n\t"
"vlef %%v21,4(%%r1,%1),1 \n\t"
"vlef %%v21,12(%%r1,%1),2 \n\t"
"vlef %%v21,12(%%r1,%1),3 \n\t"
"vlef %%v22,0(%%r1,%2),0 \n\t"
"vlef %%v22,0(%%r1,%2),1 \n\t"
"vlef %%v22,8(%%r1,%2),2 \n\t"
"vlef %%v22,8(%%r1,%2),3 \n\t"
"vlef %%v23,4(%%r1,%2),0 \n\t"
"vlef %%v23,4(%%r1,%2),1 \n\t"
"vlef %%v23,12(%%r1,%2),2 \n\t"
"vlef %%v23,12(%%r1,%2),3 \n\t"

"vl %%v0,0(%%r1,%4) \n\t"
"vfmasb %%v0,%%v20,%%v16,%%v0 \n\t"
"vfmasb %%v0,%%v21,%%v18,%%v0 \n\t"
"vfmasb %%v0,%%v22,%%v17,%%v0 \n\t"
"vfmasb %%v0,%%v23,%%v19,%%v0 \n\t"
"vst %%v0,0(%%r1,%4) \n\t"
"agfi %%r1,16 \n\t"
"brctg %%r0,0b \n\t"
:
:"r"(n),"ZR"((const FLOAT (*)[n * 2])ap[0]),"ZR"((const FLOAT (*)[n * 2])ap[1]),"ZQ"((const FLOAT (*)[4])x),"ZR"((FLOAT (*)[n * 2])y)
:"memory","cc","r0","r1","v0","v16","v17","v18","v19","v20","v21","v22","v23"
);
}

static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
__asm__ volatile (
"vlrepg %%v16,0(%2) \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vlef %%v17,4(%2),0 \n\t"
"vlef %%v17,4(%2),2 \n\t"
"vflcsb %%v17,%%v17 \n\t"
"vlef %%v17,0(%2),1 \n\t"
"vlef %%v17,0(%2),3 \n\t"
#else
"vlef %%v17,0(%2),1 \n\t"
"vlef %%v17,0(%2),3 \n\t"
"vflcsb %%v17,%%v17 \n\t"
"vlef %%v17,4(%2),0 \n\t"
"vlef %%v17,4(%2),2 \n\t"
#endif
"xgr %%r1,%%r1 \n\t"
"srlg %%r0,%%r0,1 \n\t"
"0: \n\t"
"pfd 1,1024(%%r1,%1) \n\t"
"pfd 2,1024(%%r1,%3) \n\t"

"vlef %%v18,0(%%r1,%1),0 \n\t"
"vlef %%v18,0(%%r1,%1),1 \n\t"
"vlef %%v18,8(%%r1,%1),2 \n\t"
"vlef %%v18,8(%%r1,%1),3 \n\t"
"vlef %%v19,4(%%r1,%1),0 \n\t"
"vlef %%v19,4(%%r1,%1),1 \n\t"
"vlef %%v19,12(%%r1,%1),2 \n\t"
"vlef %%v19,12(%%r1,%1),3 \n\t"

"vl %%v0,0(%%r1,%3) \n\t"
"vfmasb %%v0,%%v18,%%v16,%%v0 \n\t"
"vfmasb %%v0,%%v19,%%v17,%%v0 \n\t"
"vst %%v0,0(%%r1,%3) \n\t"
"agfi %%r1,16 \n\t"
"brctg %%r0,0b \n\t"
:
:"r"(n),"ZR"((const FLOAT (*)[n * 2])ap),"ZQ"((const FLOAT (*)[2])x),"ZR"((FLOAT (*)[n * 2])y)
:"memory","cc","r0","r1","v0","v16","v17","v18","v19"
);
}

static void add_y_4(BLASLONG n, FLOAT *src, FLOAT *dest, FLOAT alpha_r, FLOAT alpha_i)
{
__asm__ volatile (
#if !defined(XCONJ)
"vlrepf %%v0,%3 \n\t"
"vlef %%v1,%4,0 \n\t"
"vlef %%v1,%4,2 \n\t"
"vflcsb %%v1,%%v1 \n\t"
"vlef %%v1,%4,1 \n\t"
"vlef %%v1,%4,3 \n\t"
#else
"vlef %%v0,%3,1 \n\t"
"vlef %%v0,%3,3 \n\t"
"vflcsb %%v0,%%v0 \n\t"
"vlef %%v0,%3,0 \n\t"
"vlef %%v0,%3,2 \n\t"
"vlrepf %%v1,%4 \n\t"
#endif
"xgr %%r1,%%r1 \n\t"
"srlg %%r0,%0,2 \n\t"
"0: \n\t"
"pfd 1,1024(%%r1,%1) \n\t"
"pfd 2,1024(%%r1,%2) \n\t"

"vl %%v16,0(%%r1,%1) \n\t"
"vl %%v17,16(%%r1,%1) \n\t"
"vl %%v18,0(%%r1,%2) \n\t"
"vl %%v19,16(%%r1,%2) \n\t"
"verllg %%v20,%%v16,32 \n\t"
"verllg %%v21,%%v17,32 \n\t"

"vfmasb %%v22,%%v16,%%v0,%%v18 \n\t"
"vfmasb %%v23,%%v17,%%v0,%%v19 \n\t"

"vfmasb %%v22,%%v20,%%v1,%%v22 \n\t"
"vfmasb %%v23,%%v21,%%v1,%%v23 \n\t"

"vst %%v22,0(%%r1,%2) \n\t"
"vst %%v23,16(%%r1,%2) \n\t"
"agfi %%r1,32 \n\t"
"brctg %%r0,0b "
:
:"r"(n),"ZR"((const FLOAT (*)[n * 2])src),"ZR"((FLOAT (*)[n * 2])dest),"m"(alpha_r),"m"(alpha_i)
:"memory","cc","r0","r1","v0","v1","v16","v17","v18","v19","v20","v21","v22","v23"
);
}

static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest, FLOAT alpha_r, FLOAT alpha_i)
{
BLASLONG i;

if ( inc_dest != 2 )
{

FLOAT temp_r;
FLOAT temp_i;
for ( i=0; i<n; i++ )
{
#if !defined(XCONJ)
temp_r = alpha_r * src[0] - alpha_i * src[1];
temp_i = alpha_r * src[1] + alpha_i * src[0];
#else
temp_r = alpha_r * src[0] + alpha_i * src[1];
temp_i = -alpha_r * src[1] + alpha_i * src[0];
#endif

*dest += temp_r;
*(dest+1) += temp_i;

src+=2;
dest += inc_dest;
}
return;
}

add_y_4(n, src, dest, alpha_r, alpha_i);
}

int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
{
BLASLONG i;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
FLOAT *ap[4];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG m3;
BLASLONG n2;
BLASLONG lda4;
FLOAT xbuffer[8],*ybuffer;

if ( m < 1 ) return(0);
if ( n < 1 ) return(0);

ybuffer = buffer;
inc_x *= 2;
inc_y *= 2;
lda *= 2;
lda4 = 4 * lda;

n1 = n / 4 ;
n2 = n % 4 ;
m3 = m % 4;
m1 = m - ( m % 4 );
m2 = (m % NBMAX) - (m % 4) ;
y_ptr = y;

BLASLONG NB = NBMAX;

while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
x_ptr = x;
//zero_y(NB,ybuffer);
memset(ybuffer,0,NB*16);

if ( inc_x == 2 )
{

for( i = 0; i < n1 ; i++)
{
cgemv_kernel_4x4(NB,ap,x_ptr,ybuffer);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
x_ptr += 8;
}

if ( n2 & 2 )
{
cgemv_kernel_4x2(NB,ap,x_ptr,ybuffer);
x_ptr += 4;
a_ptr += 2 * lda;

}

if ( n2 & 1 )
{
cgemv_kernel_4x1(NB,a_ptr,x_ptr,ybuffer);
/* x_ptr += 2;
a_ptr += lda; */

}
}
else
{

for( i = 0; i < n1 ; i++)
{

xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
xbuffer[2] = x_ptr[0];
xbuffer[3] = x_ptr[1];
x_ptr += inc_x;
xbuffer[4] = x_ptr[0];
xbuffer[5] = x_ptr[1];
x_ptr += inc_x;
xbuffer[6] = x_ptr[0];
xbuffer[7] = x_ptr[1];
x_ptr += inc_x;

cgemv_kernel_4x4(NB,ap,xbuffer,ybuffer);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
}

for( i = 0; i < n2 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
cgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;

}

}

add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
a += 2 * NB;
y_ptr += NB * inc_y;
}

if ( m3 == 0 ) return(0);

if ( m3 == 1 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;

if ( lda == 2 && inc_x == 2 )
{


for( i=0 ; i < (n & -2); i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r += a_ptr[2] * x_ptr[2] - a_ptr[3] * x_ptr[3];
temp_i += a_ptr[2] * x_ptr[3] + a_ptr[3] * x_ptr[2];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r += a_ptr[2] * x_ptr[2] + a_ptr[3] * x_ptr[3];
temp_i += a_ptr[2] * x_ptr[3] - a_ptr[3] * x_ptr[2];
#endif

a_ptr += 4;
x_ptr += 4;
}



for( ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif

a_ptr += 2;
x_ptr += 2;
}


}
else
{

for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif

a_ptr += lda;
x_ptr += inc_x;
}

}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
return(0);
}

if ( m3 == 2 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i1 = 0.0;

if ( lda == 4 && inc_x == 2 )
{

for( i = 0; i < (n & -2); i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )

temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];

temp_r0 += a_ptr[4] * x_ptr[2] - a_ptr[5] * x_ptr[3];
temp_i0 += a_ptr[4] * x_ptr[3] + a_ptr[5] * x_ptr[2];
temp_r1 += a_ptr[6] * x_ptr[2] - a_ptr[7] * x_ptr[3];
temp_i1 += a_ptr[6] * x_ptr[3] + a_ptr[7] * x_ptr[2];

#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];

temp_r0 += a_ptr[4] * x_ptr[2] + a_ptr[5] * x_ptr[3];
temp_i0 += a_ptr[4] * x_ptr[3] - a_ptr[5] * x_ptr[2];
temp_r1 += a_ptr[6] * x_ptr[2] + a_ptr[7] * x_ptr[3];
temp_i1 += a_ptr[6] * x_ptr[3] - a_ptr[7] * x_ptr[2];

#endif

a_ptr += 8;
x_ptr += 4;
}


for( ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
#endif

a_ptr += 4;
x_ptr += 2;
}


}
else
{

for( i=0 ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
#endif

a_ptr += lda;
x_ptr += inc_x;
}


}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
#else
y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
#endif
return(0);
}


if ( m3 == 3 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i1 = 0.0;
FLOAT temp_r2 = 0.0;
FLOAT temp_i2 = 0.0;

if ( lda == 6 && inc_x == 2 )
{

for( i=0 ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
#endif

a_ptr += 6;
x_ptr += 2;
}


}
else
{

for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
#endif

a_ptr += lda;
x_ptr += inc_x;
}

}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r2 - alpha_i * temp_i2;
y_ptr[1] += alpha_r * temp_i2 + alpha_i * temp_r2;
#else
y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r2 + alpha_i * temp_i2;
y_ptr[1] -= alpha_r * temp_i2 - alpha_i * temp_r2;
#endif
return(0);
}

return(0);
}

+ 671
- 0
kernel/zarch/cgemv_t_4.c View File

@@ -0,0 +1,671 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
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.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "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 OPENBLAS PROJECT 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.
*****************************************************************************/

#include "common.h"

#define NBMAX 1024

static void cgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
__asm__ volatile (
"vzero %%v16 \n\t"
"vzero %%v17 \n\t"
"vzero %%v18 \n\t"
"vzero %%v19 \n\t"
"xgr %%r1,%%r1 \n\t"
"srlg %%r0,%0,1 \n\t"
"0: \n\t"
"pfd 1,1024(%%r1,%1) \n\t"
"pfd 1,1024(%%r1,%2) \n\t"
"pfd 1,1024(%%r1,%3) \n\t"
"pfd 1,1024(%%r1,%4) \n\t"
"pfd 1,1024(%%r1,%5) \n\t"

"vl %%v20,0(%%r1,%5) \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vlef %%v21,4(%%r1,%5),0 \n\t"
"vlef %%v21,12(%%r1,%5),2 \n\t"
"vflcsb %%v21,%%v21 \n\t"
"vlef %%v21,0(%%r1,%5),1 \n\t"
"vlef %%v21,8(%%r1,%5),3 \n\t"
#else
"vlef %%v21,0(%%r1,%5),1 \n\t"
"vlef %%v21,8(%%r1,%5),3 \n\t"
"vflcsb %%v21,%%v21 \n\t"
"vlef %%v21,4(%%r1,%5),0 \n\t"
"vlef %%v21,12(%%r1,%5),2 \n\t"
#endif

"vlef %%v22,0(%%r1,%1),0 \n\t"
"vlef %%v22,0(%%r1,%1),1 \n\t"
"vlef %%v22,8(%%r1,%1),2 \n\t"
"vlef %%v22,8(%%r1,%1),3 \n\t"
"vlef %%v23,4(%%r1,%1),0 \n\t"
"vlef %%v23,4(%%r1,%1),1 \n\t"
"vlef %%v23,12(%%r1,%1),2 \n\t"
"vlef %%v23,12(%%r1,%1),3 \n\t"
"vlef %%v24,0(%%r1,%2),0 \n\t"
"vlef %%v24,0(%%r1,%2),1 \n\t"
"vlef %%v24,8(%%r1,%2),2 \n\t"
"vlef %%v24,8(%%r1,%2),3 \n\t"
"vlef %%v25,4(%%r1,%2),0 \n\t"
"vlef %%v25,4(%%r1,%2),1 \n\t"
"vlef %%v25,12(%%r1,%2),2 \n\t"
"vlef %%v25,12(%%r1,%2),3 \n\t"

"vfmasb %%v16,%%v22,%%v20,%%v16 \n\t"
"vfmasb %%v16,%%v23,%%v21,%%v16 \n\t"
"vfmasb %%v17,%%v24,%%v20,%%v17 \n\t"
"vfmasb %%v17,%%v25,%%v21,%%v17 \n\t"

"vlef %%v26,0(%%r1,%3),0 \n\t"
"vlef %%v26,0(%%r1,%3),1 \n\t"
"vlef %%v26,8(%%r1,%3),2 \n\t"
"vlef %%v26,8(%%r1,%3),3 \n\t"
"vlef %%v27,4(%%r1,%3),0 \n\t"
"vlef %%v27,4(%%r1,%3),1 \n\t"
"vlef %%v27,12(%%r1,%3),2 \n\t"
"vlef %%v27,12(%%r1,%3),3 \n\t"
"vlef %%v28,0(%%r1,%4),0 \n\t"
"vlef %%v28,0(%%r1,%4),1 \n\t"
"vlef %%v28,8(%%r1,%4),2 \n\t"
"vlef %%v28,8(%%r1,%4),3 \n\t"
"vlef %%v29,4(%%r1,%4),0 \n\t"
"vlef %%v29,4(%%r1,%4),1 \n\t"
"vlef %%v29,12(%%r1,%4),2 \n\t"
"vlef %%v29,12(%%r1,%4),3 \n\t"
"vfmasb %%v18,%%v26,%%v20,%%v18 \n\t"
"vfmasb %%v18,%%v27,%%v21,%%v18 \n\t"
"vfmasb %%v19,%%v28,%%v20,%%v19 \n\t"
"vfmasb %%v19,%%v29,%%v21,%%v19 \n\t"

"agfi %%r1,16 \n\t"
"brctg %%r0,0b \n\t"

"vrepg %%v20,%%v16,1 \n\t"
"vrepg %%v21,%%v17,1 \n\t"
"vrepg %%v22,%%v18,1 \n\t"
"vrepg %%v23,%%v19,1 \n\t"
"vfasb %%v16,%%v16,%%v20 \n\t"
"vfasb %%v17,%%v17,%%v21 \n\t"
"vfasb %%v18,%%v18,%%v22 \n\t"
"vfasb %%v19,%%v19,%%v23 \n\t"
"vmrhg %%v16,%%v16,%%v17 \n\t"
"vmrhg %%v17,%%v18,%%v19 \n\t"
"verllg %%v18,%%v16,32 \n\t"
"verllg %%v19,%%v17,32 \n\t"
#if !defined(XCONJ)
"vlrepf %%v20,0(%7) \n\t"
"vlef %%v21,4(%7),0 \n\t"
"vlef %%v21,4(%7),2 \n\t"
"vflcsb %%v21,%%v21 \n\t"
"vlef %%v21,4(%7),1 \n\t"
"vlef %%v21,4(%7),3 \n\t"
#else
"vlef %%v20,0(%7),1 \n\t"
"vlef %%v20,0(%7),3 \n\t"
"vflcsb %%v20,%%v20 \n\t"
"vlef %%v20,0(%7),0 \n\t"
"vlef %%v20,0(%7),2 \n\t"
"vlrepf %%v21,4(%7) \n\t"
#endif
"vl %%v22,0(%6) \n\t"
"vl %%v23,16(%6) \n\t"
"vfmasb %%v22,%%v16,%%v20,%%v22 \n\t"
"vfmasb %%v22,%%v18,%%v21,%%v22 \n\t"
"vfmasb %%v23,%%v17,%%v20,%%v23 \n\t"
"vfmasb %%v23,%%v19,%%v21,%%v23 \n\t"
"vst %%v22,0(%6) \n\t"
"vst %%v23,16(%6) "
:
:"r"(n),"ZR"((const FLOAT (*)[n * 2])ap[0]),"ZR"((const FLOAT (*)[n * 2])ap[1]),"ZR"((const FLOAT (*)[n * 2])ap[2]),"ZR"((const FLOAT (*)[n * 2])ap[3]),"ZR"((const FLOAT (*)[n * 2])x),"ZQ"((FLOAT (*)[8])y),"ZQ"((const FLOAT (*)[2])alpha)
:"memory","cc","r0","r1","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29"
);
}

static void cgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
__asm__ volatile (
"vzero %%v16 \n\t"
"vzero %%v17 \n\t"
"xgr %%r1,%%r1 \n\t"
"srlg %%r0,%0,1 \n\t"
"0: \n\t"
"pfd 1,1024(%%r1,%1) \n\t"
"pfd 1,1024(%%r1,%2) \n\t"
"pfd 1,1024(%%r1,%3) \n\t"

"vl %%v18,0(%%r1,%3) \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vlef %%v19,4(%%r1,%3),0 \n\t"
"vlef %%v19,12(%%r1,%3),2 \n\t"
"vflcsb %%v19,%%v19 \n\t"
"vlef %%v19,0(%%r1,%3),1 \n\t"
"vlef %%v19,8(%%r1,%3),3 \n\t"
#else
"vlef %%v19,0(%%r1,%3),1 \n\t"
"vlef %%v19,8(%%r1,%3),3 \n\t"
"vflcsb %%v19,%%v19 \n\t"
"vlef %%v19,4(%%r1,%3),0 \n\t"
"vlef %%v19,12(%%r1,%3),2 \n\t"
#endif

"vlef %%v20,0(%%r1,%1),0 \n\t"
"vlef %%v20,0(%%r1,%1),1 \n\t"
"vlef %%v20,8(%%r1,%1),2 \n\t"
"vlef %%v20,8(%%r1,%1),3 \n\t"
"vlef %%v21,4(%%r1,%1),0 \n\t"
"vlef %%v21,4(%%r1,%1),1 \n\t"
"vlef %%v21,12(%%r1,%1),2 \n\t"
"vlef %%v21,12(%%r1,%1),3 \n\t"
"vlef %%v22,0(%%r1,%2),0 \n\t"
"vlef %%v22,0(%%r1,%2),1 \n\t"
"vlef %%v22,8(%%r1,%2),2 \n\t"
"vlef %%v22,8(%%r1,%2),3 \n\t"
"vlef %%v23,4(%%r1,%2),0 \n\t"
"vlef %%v23,4(%%r1,%2),1 \n\t"
"vlef %%v23,12(%%r1,%2),2 \n\t"
"vlef %%v23,12(%%r1,%2),3 \n\t"

"vfmasb %%v16,%%v20,%%v18,%%v16 \n\t"
"vfmasb %%v16,%%v21,%%v19,%%v16 \n\t"
"vfmasb %%v17,%%v22,%%v18,%%v17 \n\t"
"vfmasb %%v17,%%v23,%%v19,%%v17 \n\t"

"agfi %%r1,16 \n\t"
"brctg %%r0,0b \n\t"

"vrepg %%v18,%%v16,1 \n\t"
"vrepg %%v19,%%v17,1 \n\t"
"vfasb %%v16,%%v16,%%v18 \n\t"
"vfasb %%v17,%%v17,%%v19 \n\t"
"vmrhg %%v16,%%v16,%%v17 \n\t"
"verllg %%v17,%%v16,32 \n\t"
#if !defined(XCONJ)
"vlrepf %%v18,0(%5) \n\t"
"vlef %%v19,4(%5),0 \n\t"
"vlef %%v19,4(%5),2 \n\t"
"vflcsb %%v19,%%v19 \n\t"
"vlef %%v19,4(%5),1 \n\t"
"vlef %%v19,4(%5),3 \n\t"
#else
"vlef %%v18,0(%5),1 \n\t"
"vlef %%v18,0(%5),3 \n\t"
"vflcsb %%v18,%%v18 \n\t"
"vlef %%v18,0(%5),0 \n\t"
"vlef %%v18,0(%5),2 \n\t"
"vlrepf %%v19,4(%5) \n\t"
#endif
"vl %%v20,0(%4) \n\t"
"vfmasb %%v20,%%v16,%%v18,%%v20 \n\t"
"vfmasb %%v20,%%v17,%%v19,%%v20 \n\t"
"vst %%v20,0(%4) "
:
:"r"(n),"ZR"((const FLOAT (*)[n * 2])ap[0]),"ZR"((const FLOAT (*)[n * 2])ap[1]),"ZR"((const FLOAT (*)[n * 2])x),"ZQ"((FLOAT (*)[4])y),"ZQ"((const FLOAT (*)[2])alpha)
:"memory","cc","r0","r1","v16","v17","v18","v19","v20","v21","v22","v23"
);
}

static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
__asm__ volatile (
"vzero %%v16 \n\t"
"xgr %%r1,%%r1 \n\t"
"srlg %%r0,%0,1 \n\t"
"0: \n\t"
"pfd 1,1024(%%r1,%1) \n\t"
"pfd 1,1024(%%r1,%2) \n\t"

"vl %%v17,0(%%r1,%2) \n\t"
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vlef %%v18,4(%%r1,%2),0 \n\t"
"vlef %%v18,12(%%r1,%2),2 \n\t"
"vflcsb %%v18,%%v18 \n\t"
"vlef %%v18,0(%%r1,%2),1 \n\t"
"vlef %%v18,8(%%r1,%2),3 \n\t"
#else
"vlef %%v18,0(%%r1,%2),1 \n\t"
"vlef %%v18,8(%%r1,%2),3 \n\t"
"vflcsb %%v18,%%v18 \n\t"
"vlef %%v18,4(%%r1,%2),0 \n\t"
"vlef %%v18,12(%%r1,%2),2 \n\t"
#endif

"vlef %%v19,0(%%r1,%1),0 \n\t"
"vlef %%v19,0(%%r1,%1),1 \n\t"
"vlef %%v19,8(%%r1,%1),2 \n\t"
"vlef %%v19,8(%%r1,%1),3 \n\t"
"vlef %%v20,4(%%r1,%1),0 \n\t"
"vlef %%v20,4(%%r1,%1),1 \n\t"
"vlef %%v20,12(%%r1,%1),2 \n\t"
"vlef %%v20,12(%%r1,%1),3 \n\t"

"vfmasb %%v16,%%v19,%%v17,%%v16 \n\t"
"vfmasb %%v16,%%v20,%%v18,%%v16 \n\t"

"agfi %%r1,16 \n\t"
"brctg %%r0,0b \n\t"

"vrepg %%v17,%%v16,1 \n\t"
"vfasb %%v16,%%v16,%%v17 \n\t"
"verllg %%v17,%%v16,32 \n\t"
#if !defined(XCONJ)
"vlrepf %%v18,0(%4) \n\t"
"vlef %%v19,4(%4),0 \n\t"
"vflcsb %%v19,%%v19 \n\t"
"vlef %%v19,4(%4),1 \n\t"
#else
"vlef %%v18,0(%4),1 \n\t"
"vflcsb %%v18,%%v18 \n\t"
"vlef %%v18,0(%4),0 \n\t"
"vlrepf %%v19,4(%4) \n\t"
#endif
"vleg %%v20,0(%3),0 \n\t"
"vfmasb %%v20,%%v16,%%v18,%%v20 \n\t"
"vfmasb %%v20,%%v17,%%v19,%%v20 \n\t"
"vsteg %%v20,0(%3),0 "
:
:"r"(n),"ZR"((const FLOAT (*)[n * 2])ap),"ZR"((const FLOAT (*)[n * 2])x),"ZQ"((FLOAT (*)[2])y),"ZQ"((const FLOAT (*)[2])alpha)
:"memory","cc","r0","r1","v16","v17","v18","v19","v20","v21","v22","v23"
);
}

static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest = *src;
*(dest+1) = *(src+1);
dest+=2;
src += inc_src;
}
}

int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
{
BLASLONG i;
BLASLONG j;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
FLOAT *ap[8];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG m3;
BLASLONG n2;
BLASLONG lda4;
FLOAT ybuffer[8],*xbuffer;
FLOAT alpha[2];

if ( m < 1 ) return(0);
if ( n < 1 ) return(0);

inc_x <<= 1;
inc_y <<= 1;
lda <<= 1;
lda4 = lda << 2;

xbuffer = buffer;
n1 = n >> 2 ;
n2 = n & 3 ;
m3 = m & 3 ;
m1 = m - m3;
m2 = (m & (NBMAX-1)) - m3 ;
alpha[0] = alpha_r;
alpha[1] = alpha_i;

BLASLONG NB = NBMAX;

while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
y_ptr = y;
a_ptr = a;
x_ptr = x;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
if ( inc_x != 2 )
copy_x(NB,x_ptr,xbuffer,inc_x);
else
xbuffer = x_ptr;
if ( inc_y == 2 )
{

for( i = 0; i < n1 ; i++)
{
cgemv_kernel_4x4(NB,ap,xbuffer,y_ptr,alpha);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
y_ptr += 8;
}

if ( n2 & 2 )
{
cgemv_kernel_4x2(NB,ap,xbuffer,y_ptr,alpha);
a_ptr += lda * 2;
y_ptr += 4;

}

if ( n2 & 1 )
{
cgemv_kernel_4x1(NB,a_ptr,xbuffer,y_ptr,alpha);
/* a_ptr += lda;
y_ptr += 2; */

}

}
else
{

for( i = 0; i < n1 ; i++)
{
memset(ybuffer,0,sizeof(ybuffer));
cgemv_kernel_4x4(NB,ap,xbuffer,ybuffer,alpha);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;

y_ptr[0] += ybuffer[0];
y_ptr[1] += ybuffer[1];
y_ptr += inc_y;
y_ptr[0] += ybuffer[2];
y_ptr[1] += ybuffer[3];
y_ptr += inc_y;
y_ptr[0] += ybuffer[4];
y_ptr[1] += ybuffer[5];
y_ptr += inc_y;
y_ptr[0] += ybuffer[6];
y_ptr[1] += ybuffer[7];
y_ptr += inc_y;

}

for( i = 0; i < n2 ; i++)
{
memset(ybuffer,0,sizeof(ybuffer));
cgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer,alpha);
a_ptr += lda;
y_ptr[0] += ybuffer[0];
y_ptr[1] += ybuffer[1];
y_ptr += inc_y;

}

}
a += 2 * NB;
x += NB * inc_x;
}



if ( m3 == 0 ) return(0);

x_ptr = x;
j=0;
a_ptr = a;
y_ptr = y;

if ( m3 == 3 )
{

FLOAT temp_r ;
FLOAT temp_i ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
x_ptr += inc_x;
FLOAT x2 = x_ptr[0];
FLOAT x3 = x_ptr[1];
x_ptr += inc_x;
FLOAT x4 = x_ptr[0];
FLOAT x5 = x_ptr[1];
while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
temp_r += a_ptr[4] * x4 - a_ptr[5] * x5;
temp_i += a_ptr[4] * x5 + a_ptr[5] * x4;
#else

temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
temp_r += a_ptr[4] * x4 + a_ptr[5] * x5;
temp_i += a_ptr[4] * x5 - a_ptr[5] * x4;
#endif

#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif

a_ptr += lda;
y_ptr += inc_y;
j++;
}
return(0);
}


if ( m3 == 2 )
{

FLOAT temp_r ;
FLOAT temp_i ;
FLOAT temp_r1 ;
FLOAT temp_i1 ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
x_ptr += inc_x;
FLOAT x2 = x_ptr[0];
FLOAT x3 = x_ptr[1];
FLOAT ar = alpha[0];
FLOAT ai = alpha[1];

while ( j < ( n & -2 ))
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r1 += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i1 += a_ptr[2] * x3 + a_ptr[3] * x2;
#else

temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r1 += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i1 += a_ptr[2] * x3 - a_ptr[3] * x2;
#endif

#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 - ai * temp_i1;
y_ptr[1] += ar * temp_i1 + ai * temp_r1;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 + ai * temp_i1;
y_ptr[1] -= ar * temp_i1 - ai * temp_r1;
#endif

a_ptr += lda;
y_ptr += inc_y;
j+=2;
}


while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
#else

temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
#endif

#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
#endif

a_ptr += lda;
y_ptr += inc_y;
j++;
}

return(0);
}


if ( m3 == 1 )
{

FLOAT temp_r ;
FLOAT temp_i ;
FLOAT temp_r1 ;
FLOAT temp_i1 ;
FLOAT x0 = x_ptr[0];
FLOAT x1 = x_ptr[1];
FLOAT ar = alpha[0];
FLOAT ai = alpha[1];

while ( j < ( n & -2 ))
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
#else

temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
a_ptr += lda;
temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
#endif

#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 - ai * temp_i1;
y_ptr[1] += ar * temp_i1 + ai * temp_r1;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
y_ptr += inc_y;
y_ptr[0] += ar * temp_r1 + ai * temp_i1;
y_ptr[1] -= ar * temp_i1 - ai * temp_r1;
#endif

a_ptr += lda;
y_ptr += inc_y;
j+=2;
}

while ( j < n)
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
#else

temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
#endif

#if !defined(XCONJ)
y_ptr[0] += ar * temp_r - ai * temp_i;
y_ptr[1] += ar * temp_i + ai * temp_r;
#else
y_ptr[0] += ar * temp_r + ai * temp_i;
y_ptr[1] -= ar * temp_i - ai * temp_r;
#endif

a_ptr += lda;
y_ptr += inc_y;
j++;
}
return(0);
}

return(0);
}

+ 7
- 2
kernel/zarch/icamax.c View File

@@ -281,6 +281,12 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)


i = n1; i = n1;
} }
else
{
maxf = CABS1(x,0);
ix += 2;
i++;
}


while(i < n) while(i < n)
{ {
@@ -296,9 +302,8 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)


} else { } else {
inc_x2 = 2 * inc_x;

maxf = CABS1(x,0); maxf = CABS1(x,0);
inc_x2 = 2 * inc_x;
ix += inc_x2; ix += inc_x2;
i++; i++;




+ 7
- 2
kernel/zarch/icamin.c View File

@@ -281,6 +281,12 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)


i = n1; i = n1;
} }
else
{
minf = CABS1(x,0);
ix += 2;
i++;
}


while(i < n) while(i < n)
{ {
@@ -296,9 +302,8 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)


} else { } else {
inc_x2 = 2 * inc_x;

minf = CABS1(x,0); minf = CABS1(x,0);
inc_x2 = 2 * inc_x;
ix += inc_x2; ix += inc_x2;
i++; i++;




+ 10
- 1
kernel/zarch/idamax.c View File

@@ -204,6 +204,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


i = n1; i = n1;
} }
else
{
maxf = ABS(x[0]);
i++;
}


while (i < n) { while (i < n) {
if (ABS(x[i]) > maxf) { if (ABS(x[i]) > maxf) {
@@ -216,7 +221,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


} else { } else {


BLASLONG n1 = n & -4;
maxf = ABS(x[0]);
i += inc_x;
j++;

BLASLONG n1 = (n - 1) & -4;
while (j < n1) { while (j < n1) {


if (ABS(x[i]) > maxf) { if (ABS(x[i]) > maxf) {


+ 10
- 1
kernel/zarch/idamin.c View File

@@ -204,6 +204,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


i = n1; i = n1;
} }
else
{
minf = ABS(x[0]);
i++;
}


while (i < n) { while (i < n) {
if (ABS(x[i]) < minf) { if (ABS(x[i]) < minf) {
@@ -216,7 +221,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


} else { } else {


BLASLONG n1 = n & -4;
minf = ABS(x[0]);
i += inc_x;
j++;

BLASLONG n1 = (n - 1) & -4;
while (j < n1) { while (j < n1) {


if (ABS(x[i]) < minf) { if (ABS(x[i]) < minf) {


+ 10
- 1
kernel/zarch/idmax.c View File

@@ -180,6 +180,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


i = n1; i = n1;
} }
else
{
maxf = x[0];
i++;
}


while (i < n) { while (i < n) {
if (x[i] > maxf) { if (x[i] > maxf) {
@@ -192,7 +197,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


} else { } else {


BLASLONG n1 = n & -4;
maxf = x[0];
i += inc_x;
j++;

BLASLONG n1 = (n - 1) & -4;
while (j < n1) { while (j < n1) {


if (x[i] > maxf) { if (x[i] > maxf) {


+ 10
- 1
kernel/zarch/idmin.c View File

@@ -180,6 +180,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


i = n1; i = n1;
} }
else
{
minf = x[0];
i++;
}


while (i < n) { while (i < n) {
if (x[i] < minf) { if (x[i] < minf) {
@@ -192,7 +197,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


} else { } else {


BLASLONG n1 = n & -4;
minf = x[0];
i += inc_x;
j++;

BLASLONG n1 = (n - 1) & -4;
while (j < n1) { while (j < n1) {


if (x[i] < minf) { if (x[i] < minf) {


+ 10
- 1
kernel/zarch/isamax.c View File

@@ -247,6 +247,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


i = n1; i = n1;
} }
else
{
maxf = ABS(x[0]);
i++;
}


while (i < n) { while (i < n) {
if (ABS(x[i]) > maxf) { if (ABS(x[i]) > maxf) {
@@ -259,7 +264,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


} else { } else {


BLASLONG n1 = n & -4;
maxf = ABS(x[0]);
i += inc_x;
j++;

BLASLONG n1 = (n - 1) & -4;
while (j < n1) { while (j < n1) {


if (ABS(x[i]) > maxf) { if (ABS(x[i]) > maxf) {


+ 10
- 1
kernel/zarch/isamin.c View File

@@ -247,6 +247,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


i = n1; i = n1;
} }
else
{
minf = ABS(x[0]);
i++;
}


while (i < n) { while (i < n) {
if (ABS(x[i]) < minf) { if (ABS(x[i]) < minf) {
@@ -259,7 +264,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


} else { } else {


BLASLONG n1 = n & -4;
minf = ABS(x[0]);
i += inc_x;
j++;

BLASLONG n1 = (n - 1) & -4;
while (j < n1) { while (j < n1) {


if (ABS(x[i]) < minf) { if (ABS(x[i]) < minf) {


+ 10
- 1
kernel/zarch/ismax.c View File

@@ -223,6 +223,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


i = n1; i = n1;
} }
else
{
maxf = x[0];
i++;
}


while (i < n) { while (i < n) {
if (x[i] > maxf) { if (x[i] > maxf) {
@@ -235,7 +240,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


} else { } else {


BLASLONG n1 = n & -4;
maxf = x[0];
i += inc_x;
j++;

BLASLONG n1 = (n - 1) & -4;
while (j < n1) { while (j < n1) {


if (x[i] > maxf) { if (x[i] > maxf) {


+ 10
- 1
kernel/zarch/ismin.c View File

@@ -223,6 +223,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


i = n1; i = n1;
} }
else
{
minf = x[0];
i++;
}


while (i < n) { while (i < n) {
if (x[i] < minf) { if (x[i] < minf) {
@@ -235,7 +240,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {


} else { } else {


BLASLONG n1 = n & -4;
minf = x[0];
i += inc_x;
j++;

BLASLONG n1 = (n - 1) & -4;
while (j < n1) { while (j < n1) {


if (x[i] < minf) { if (x[i] < minf) {


+ 7
- 2
kernel/zarch/izamax.c View File

@@ -202,6 +202,12 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)


i = n1; i = n1;
} }
else
{
maxf = CABS1(x,0);
ix += 2;
i++;
}


while(i < n) while(i < n)
{ {
@@ -217,9 +223,8 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)


} else { } else {
inc_x2 = 2 * inc_x;

maxf = CABS1(x,0); maxf = CABS1(x,0);
inc_x2 = 2 * inc_x;
ix += inc_x2; ix += inc_x2;
i++; i++;




+ 7
- 2
kernel/zarch/izamin.c View File

@@ -202,6 +202,12 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)


i = n1; i = n1;
} }
else
{
minf = CABS1(x,0);
ix += 2;
i++;
}


while(i < n) while(i < n)
{ {
@@ -217,9 +223,8 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)


} else { } else {
inc_x2 = 2 * inc_x;

minf = CABS1(x,0); minf = CABS1(x,0);
inc_x2 = 2 * inc_x;
ix += inc_x2; ix += inc_x2;
i++; i++;




+ 25
- 23
kernel/zarch/zamax.c View File

@@ -150,7 +150,7 @@ static FLOAT zamax_kernel_16(BLASLONG n, FLOAT *x)
FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) { FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
BLASLONG i = 0; BLASLONG i = 0;
BLASLONG j = 0;
BLASLONG ix = 0;
FLOAT maxf = 0.0; FLOAT maxf = 0.0;
BLASLONG inc_x2; BLASLONG inc_x2;


@@ -168,53 +168,55 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
else else
{ {
maxf=CABS1(x,0); maxf=CABS1(x,0);
ix += 2;
i++; i++;
} }


while (i < n) { while (i < n) {
if (ABS(x[i*2]) > maxf) {
maxf = ABS(x[i*2]);
if (CABS1(x,ix) > maxf) {
maxf = CABS1(x,ix);
} }
ix += 2;
i++; i++;
} }
return (maxf); return (maxf);


} else { } else {

inc_x2 = 2 * inc_x;
maxf=CABS1(x,0); maxf=CABS1(x,0);
i += inc_x2;
j++;
inc_x2 = 2 * inc_x;
ix += inc_x2;
i++;


BLASLONG n1 = (n - 1) & -4; BLASLONG n1 = (n - 1) & -4;
while (j < n1) {
while (i < n1) {


if (CABS1(x,i) > maxf) {
maxf = CABS1(x,i);
if (CABS1(x,ix) > maxf) {
maxf = CABS1(x,ix);
} }
if (CABS1(x,i+inc_x2) > maxf) {
maxf = CABS1(x,i+inc_x2);
if (CABS1(x,ix+inc_x2) > maxf) {
maxf = CABS1(x,ix+inc_x2);
} }
if (CABS1(x,i+inc_x2*2) > maxf) {
maxf = CABS1(x,i+inc_x2*2);
if (CABS1(x,ix+inc_x2*2) > maxf) {
maxf = CABS1(x,ix+inc_x2*2);
} }
if (CABS1(x,i+inc_x2*3) > maxf) {
maxf = CABS1(x,i+inc_x2*3);
if (CABS1(x,ix+inc_x2*3) > maxf) {
maxf = CABS1(x,ix+inc_x2*3);
} }


i += inc_x2 * 4;
ix += inc_x2 * 4;


j += 4;
i += 4;


} }




while (j < n) {
if (CABS1(x,i) > maxf) {
maxf = CABS1(x,i);
while (i < n) {
if (CABS1(x,ix) > maxf) {
maxf = CABS1(x,ix);
} }
i += inc_x2;
j++;
ix += inc_x2;
i++;
} }
return (maxf); return (maxf);
} }


+ 24
- 22
kernel/zarch/zamin.c View File

@@ -150,7 +150,7 @@ static FLOAT zamin_kernel_16(BLASLONG n, FLOAT *x)
FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) { FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
BLASLONG i = 0; BLASLONG i = 0;
BLASLONG j = 0;
BLASLONG ix = 0;
FLOAT minf = 0.0; FLOAT minf = 0.0;
BLASLONG inc_x2; BLASLONG inc_x2;


@@ -168,53 +168,55 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
else else
{ {
minf=CABS1(x,0); minf=CABS1(x,0);
ix += 2;
i++; i++;
} }


while (i < n) { while (i < n) {
if (ABS(x[i*2]) < minf) {
minf = ABS(x[i*2]);
if (CABS1(x,ix) < minf) {
minf = CABS1(x,ix);
} }
ix += 2;
i++; i++;
} }
return (minf); return (minf);


} else { } else {


inc_x2 = 2 * inc_x;
minf=CABS1(x,0); minf=CABS1(x,0);
i += inc_x2;
j++;
inc_x2 = 2 * inc_x;
ix += inc_x2;
i++;


BLASLONG n1 = (n - 1) & -4; BLASLONG n1 = (n - 1) & -4;
while (j < n1) {
while (i < n1) {


if (CABS1(x,i) < minf) {
minf = CABS1(x,i);
if (CABS1(x,ix) < minf) {
minf = CABS1(x,ix);
} }
if (CABS1(x,i+inc_x2) < minf) {
minf = CABS1(x,i+inc_x2);
if (CABS1(x,ix+inc_x2) < minf) {
minf = CABS1(x,ix+inc_x2);
} }
if (CABS1(x,i+inc_x2*2) < minf) {
minf = CABS1(x,i+inc_x2*2);
if (CABS1(x,ix+inc_x2*2) < minf) {
minf = CABS1(x,ix+inc_x2*2);
} }
if (CABS1(x,i+inc_x2*3) < minf) {
minf = CABS1(x,i+inc_x2*3);
if (CABS1(x,ix+inc_x2*3) < minf) {
minf = CABS1(x,ix+inc_x2*3);
} }


i += inc_x2 * 4;
ix += inc_x2 * 4;


j += 4;
i += 4;


} }




while (j < n) {
if (CABS1(x,i) < minf) {
minf = CABS1(x,i);
while (i < n) {
if (CABS1(x,ix) < minf) {
minf = CABS1(x,ix);
} }
i += inc_x2;
j++;
ix += inc_x2;
i++;
} }
return (minf); return (minf);
} }


+ 2
- 2
kernel/zarch/zaxpy.c View File

@@ -106,7 +106,7 @@ static void zaxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"agfi %%r1,128 \n\t" "agfi %%r1,128 \n\t"
"brctg %%r0,0b " "brctg %%r0,0b "
: :
:"r"(n),"ZR"((const FLOAT (*)[n * 2])x),"ZR"((FLOAT (*)[n * 2])y),"a"(alpha)
:"r"(n),"ZR"((const FLOAT (*)[n * 2])x),"ZR"((FLOAT (*)[n * 2])y),"ZQ"((const FLOAT (*)[2])alpha)
:"memory","cc","r0","r1","v0","v1","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31" :"memory","cc","r0","r1","v0","v1","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
); );
} }
@@ -114,7 +114,7 @@ static void zaxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2) { int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2) {
BLASLONG i = 0; BLASLONG i = 0;
BLASLONG ix = 0, iy = 0; BLASLONG ix = 0, iy = 0;
FLOAT da[2];
FLOAT da[2] __attribute__ ((aligned(16)));


if (n <= 0) return (0); if (n <= 0) return (0);




+ 598
- 803
kernel/zarch/zgemv_n_4.c
File diff suppressed because it is too large
View File


+ 548
- 719
kernel/zarch/zgemv_t_4.c
File diff suppressed because it is too large
View File


+ 87
- 72
ztest/gemv.c View File

@@ -52,67 +52,66 @@ int assert_dbl_near(double exp, double real, double tol) {
int zgemv_n_c(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) int zgemv_n_c(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{ {
BLASLONG i; BLASLONG i;
BLASLONG ix,iy;
BLASLONG ix, iy;
BLASLONG j; BLASLONG j;
FLOAT *a_ptr; FLOAT *a_ptr;
FLOAT temp_r,temp_i;
BLASLONG inc_x2,inc_y2;
FLOAT temp_r, temp_i;
BLASLONG inc_x2, inc_y2;
BLASLONG lda2; BLASLONG lda2;
BLASLONG i2; BLASLONG i2;


lda2 = 2*lda;
lda2 = 2 * lda;


ix = 0; ix = 0;
a_ptr = a; a_ptr = a;


if ( inc_x == 1 && inc_y == 1 )
if (inc_x == 1 && inc_y == 1)
{ {


for (j=0; j<n; j++)
{
for (j = 0; j<n; j++)
{


#if !defined(XCONJ) #if !defined(XCONJ)
temp_r = alpha_r * x[ix] - alpha_i * x[ix+1];
temp_i = alpha_r * x[ix+1] + alpha_i * x[ix];
temp_r = alpha_r * x[ix] - alpha_i * x[ix + 1];
temp_i = alpha_r * x[ix + 1] + alpha_i * x[ix];
#else #else
temp_r = alpha_r * x[ix] + alpha_i * x[ix+1];
temp_i = alpha_r * x[ix+1] - alpha_i * x[ix];
temp_r = alpha_r * x[ix] + alpha_i * x[ix + 1];
temp_i = alpha_r * x[ix + 1] - alpha_i * x[ix];
#endif #endif
iy = 0;
i2=0;
iy = 0;
i2 = 0;


for (i=0; i<m; i++)
{
for (i = 0; i<m; i++)
{
#if !defined(CONJ) #if !defined(CONJ)


#if !defined(XCONJ) #if !defined(XCONJ)
printf("\nParO: %f %f %f %f\n", a_ptr[i2], a_ptr[i2+1], temp_r, temp_i);
y[iy] += temp_r * a_ptr[i2] - temp_i * a_ptr[i2+1];
y[iy+1] += temp_r * a_ptr[i2+1] + temp_i * a_ptr[i2];
y[iy] += temp_r * a_ptr[i2] - temp_i * a_ptr[i2 + 1];
y[iy + 1] += temp_r * a_ptr[i2 + 1] + temp_i * a_ptr[i2];
#else #else
y[iy] += temp_r * a_ptr[i2] + temp_i * a_ptr[i2+1];
y[iy+1] += temp_r * a_ptr[i2+1] - temp_i * a_ptr[i2];
y[iy] += temp_r * a_ptr[i2] + temp_i * a_ptr[i2 + 1];
y[iy + 1] += temp_r * a_ptr[i2 + 1] - temp_i * a_ptr[i2];
#endif #endif


#else #else


#if !defined(XCONJ) #if !defined(XCONJ)
y[iy] += temp_r * a_ptr[i2] + temp_i * a_ptr[i2+1];
y[iy+1] -= temp_r * a_ptr[i2+1] - temp_i * a_ptr[i2];
y[iy] += temp_r * a_ptr[i2] + temp_i * a_ptr[i2 + 1];
y[iy + 1] -= temp_r * a_ptr[i2 + 1] - temp_i * a_ptr[i2];
#else #else
y[iy] += temp_r * a_ptr[i2] - temp_i * a_ptr[i2+1];
y[iy+1] -= temp_r * a_ptr[i2+1] + temp_i * a_ptr[i2];
y[iy] += temp_r * a_ptr[i2] - temp_i * a_ptr[i2 + 1];
y[iy + 1] -= temp_r * a_ptr[i2 + 1] + temp_i * a_ptr[i2];
#endif #endif


#endif #endif
i2 += 2;
iy += 2;
i2 += 2;
iy += 2;
}
a_ptr += lda2;
ix += 2;
} }
a_ptr += lda2;
ix += 2;
}


return(0);
return(0);


} }


@@ -120,39 +119,39 @@ int zgemv_n_c(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alph
inc_x2 = 2 * inc_x; inc_x2 = 2 * inc_x;
inc_y2 = 2 * inc_y; inc_y2 = 2 * inc_y;


for (j=0; j<n; j++)
for (j = 0; j<n; j++)
{ {


#if !defined(XCONJ) #if !defined(XCONJ)
temp_r = alpha_r * x[ix] - alpha_i * x[ix+1];
temp_i = alpha_r * x[ix+1] + alpha_i * x[ix];
temp_r = alpha_r * x[ix] - alpha_i * x[ix + 1];
temp_i = alpha_r * x[ix + 1] + alpha_i * x[ix];
#else #else
temp_r = alpha_r * x[ix] + alpha_i * x[ix+1];
temp_i = alpha_r * x[ix+1] - alpha_i * x[ix];
temp_r = alpha_r * x[ix] + alpha_i * x[ix + 1];
temp_i = alpha_r * x[ix + 1] - alpha_i * x[ix];
#endif #endif
iy = 0; iy = 0;
i2=0;
i2 = 0;


for (i=0; i<m; i++)
for (i = 0; i<m; i++)
{ {
#if !defined(CONJ) #if !defined(CONJ)


#if !defined(XCONJ) #if !defined(XCONJ)
y[iy] += temp_r * a_ptr[i2] - temp_i * a_ptr[i2+1];
y[iy+1] += temp_r * a_ptr[i2+1] + temp_i * a_ptr[i2];
y[iy] += temp_r * a_ptr[i2] - temp_i * a_ptr[i2 + 1];
y[iy + 1] += temp_r * a_ptr[i2 + 1] + temp_i * a_ptr[i2];
#else #else
y[iy] += temp_r * a_ptr[i2] + temp_i * a_ptr[i2+1];
y[iy+1] += temp_r * a_ptr[i2+1] - temp_i * a_ptr[i2];
y[iy] += temp_r * a_ptr[i2] + temp_i * a_ptr[i2 + 1];
y[iy + 1] += temp_r * a_ptr[i2 + 1] - temp_i * a_ptr[i2];
#endif #endif


#else #else


#if !defined(XCONJ) #if !defined(XCONJ)
y[iy] += temp_r * a_ptr[i2] + temp_i * a_ptr[i2+1];
y[iy+1] -= temp_r * a_ptr[i2+1] - temp_i * a_ptr[i2];
y[iy] += temp_r * a_ptr[i2] + temp_i * a_ptr[i2 + 1];
y[iy + 1] -= temp_r * a_ptr[i2 + 1] - temp_i * a_ptr[i2];
#else #else
y[iy] += temp_r * a_ptr[i2] - temp_i * a_ptr[i2+1];
y[iy+1] -= temp_r * a_ptr[i2+1] + temp_i * a_ptr[i2];
y[iy] += temp_r * a_ptr[i2] - temp_i * a_ptr[i2 + 1];
y[iy + 1] -= temp_r * a_ptr[i2 + 1] + temp_i * a_ptr[i2];
#endif #endif


#endif #endif
@@ -160,7 +159,7 @@ int zgemv_n_c(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alph
iy += inc_y2; iy += inc_y2;
} }
a_ptr += lda2; a_ptr += lda2;
ix += inc_x2;
ix += inc_x2;
} }




@@ -404,7 +403,7 @@ int main(int argc, char *argv[]){


FLOAT *a, *x, *y, *y_c; FLOAT *a, *x, *y, *y_c;
FLOAT alpha[] = {1.0, 1.0}; FLOAT alpha[] = {1.0, 1.0};
FLOAT beta [] = {1.0, 1.0};
FLOAT beta [] = {1.0, 0.0};
char trans='N'; char trans='N';
blasint m, i, j; blasint m, i, j;
blasint inc_x=1,inc_y=1; blasint inc_x=1,inc_y=1;
@@ -422,6 +421,7 @@ int main(int argc, char *argv[]){
struct timeval start, stop; struct timeval start, stop;
double time1,timeg,timeg_c; double time1,timeg,timeg_c;


blasint y_size;
blasint iy; blasint iy;
int test = 1; int test = 1;


@@ -500,13 +500,15 @@ int main(int argc, char *argv[]){


for(i = 0; i < n * COMPSIZE * abs(inc_y); i++){ for(i = 0; i < n * COMPSIZE * abs(inc_y); i++){
y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
y_c[i]= y[i];
y_c[i]= y[i];
} }
gettimeofday( &start, (struct timezone *)0);
GEMV (&trans, &m, &n, alpha, a, &m, x, &inc_x, beta, y, &inc_y );
gettimeofday( &stop, (struct timezone *)0);
time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
timeg += time1;

gettimeofday( &start, (struct timezone *)0);
GEMV (&trans, &m, &n, alpha, a, &m, x, &inc_x, beta, y, &inc_y );
gettimeofday( &stop, (struct timezone *)0);
time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
timeg += time1;


gettimeofday( &start, (struct timezone *)0); gettimeofday( &start, (struct timezone *)0);
#ifdef COMPLEX #ifdef COMPLEX
@@ -524,21 +526,27 @@ int main(int argc, char *argv[]){
time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
timeg_c += time1; timeg_c += time1;


iy = 0;
if (trans == 'N')
y_size = m;
else
y_size = n;
iy = 0;
for (i = 0; i < y_size; i++)
{
#ifdef COMPLEX #ifdef COMPLEX
for (i = 0; i < m * 2; i++)
#else
for (i = 0; i < m; i++)
#endif
{
test &= assert_dbl_near(y[iy], y_c[iy], SINGLE_EPS); test &= assert_dbl_near(y[iy], y_c[iy], SINGLE_EPS);
test &= assert_dbl_near(y[iy + 1], y_c[iy + 1], SINGLE_EPS);
iy += (inc_y * 2);
#else
test &= assert_dbl_near(y[iy], y_c[iy], SINGLE_EPS);
iy += inc_y; iy += inc_y;
#endif
} }


} }


timeg /= loops; timeg /= loops;
timeg_c /= loops;
timeg_c /= loops;


fprintf(stderr, "%10.2f MFlops %10.6f sec %10.6f sec %s\n", 2. * (double)m / timeg * 1.e-6, timeg, timeg_c, test ? "PASS" : "FAILD"); fprintf(stderr, "%10.2f MFlops %10.6f sec %10.6f sec %s\n", 2. * (double)m / timeg * 1.e-6, timeg, timeg_c, test ? "PASS" : "FAILD");


@@ -567,13 +575,14 @@ int main(int argc, char *argv[]){


for(i = 0; i < n * COMPSIZE * abs(inc_y); i++){ for(i = 0; i < n * COMPSIZE * abs(inc_y); i++){
y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
y_c[i]= y[i];
y_c[i]= y[i];
} }
gettimeofday( &start, (struct timezone *)0);
GEMV (&trans, &m, &n, alpha, a, &m, x, &inc_x, beta, y, &inc_y );
gettimeofday( &stop, (struct timezone *)0);
time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
timeg += time1;
gettimeofday( &start, (struct timezone *)0);
GEMV (&trans, &m, &n, alpha, a, &m, x, &inc_x, beta, y, &inc_y );
gettimeofday( &stop, (struct timezone *)0);
time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
timeg += time1;


gettimeofday( &start, (struct timezone *)0); gettimeofday( &start, (struct timezone *)0);
#ifdef COMPLEX #ifdef COMPLEX
@@ -591,21 +600,27 @@ int main(int argc, char *argv[]){
time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6; time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
timeg_c += time1; timeg_c += time1;


if (trans == 'N')
y_size = m;
else
y_size = n;
iy = 0; iy = 0;
for (i = 0; i < y_size; i++)
{
#ifdef COMPLEX #ifdef COMPLEX
for (i = 0; i < m * 2; i++)
#else
for (i = 0; i < m; i++)
#endif
{
test &= assert_dbl_near(y[iy], y_c[iy], SINGLE_EPS); test &= assert_dbl_near(y[iy], y_c[iy], SINGLE_EPS);
test &= assert_dbl_near(y[iy + 1], y_c[iy + 1], SINGLE_EPS);
iy += (inc_y * 2);
#else
test &= assert_dbl_near(y[iy], y_c[iy], SINGLE_EPS);
iy += inc_y; iy += inc_y;
#endif
} }


} }


timeg /= loops; timeg /= loops;
timeg_c /= loops;
timeg_c /= loops;


fprintf(stderr, "%10.2f MFlops %10.6f sec %10.6f sec %s\n", 2. * (double)m / timeg * 1.e-6, timeg, timeg_c, test ? "PASS" : "FAILD"); fprintf(stderr, "%10.2f MFlops %10.6f sec %10.6f sec %s\n", 2. * (double)m / timeg * 1.e-6, timeg, timeg_c, test ? "PASS" : "FAILD");




Write Preview
Loading…
Cancel
Save