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

powerpc: Optimized SGEMM/DGEMM/CGEMM for POWER10 

This patch introduces new optimized version of SGEMM, CGEMM and DGEMM
using power10 Matrix-Multiply Assist (MMA) feature introduced in
POWER ISA v3.1. This patch makes use of new POWER10 compute instructions
for matrix multiplication operation.

Tested on simulator and there are no new test failures.
Cycles count reduced by 30-50%  compared to POWER9 version depending on
M/N/K sizes.
MMA GCC patch for reference:
https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=8ee2640bfdc62f835ec9740278f948034bc7d9f1
tags/v0.3.11^2
Rajalakshmi Srinivasaraghavan 5 years ago
parent
93592d1260
commit
571eadb880
6 changed files with 7435 additions and 6 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +6
    -6
      kernel/power/KERNEL.POWER10
  2. +286
    -0
      kernel/power/cgemm_kernel_power10.S
  3. +2814
    -0
      kernel/power/cgemm_logic_power10.S
  4. +2131
    -0
      kernel/power/cgemm_macros_power10.S
  5. +864
    -0
      kernel/power/dgemm_kernel_power10.c
  6. +1334
    -0
      kernel/power/sgemm_kernel_power10.c

+ 6
- 6
kernel/power/KERNEL.POWER10 View File

@@ -7,12 +7,12 @@ else
#CGEMM_BETA = ../generic/zgemm_beta.c
#ZGEMM_BETA = ../generic/zgemm_beta.c

STRMMKERNEL = sgemm_kernel_power9.S
DTRMMKERNEL = dgemm_kernel_power9.S
CTRMMKERNEL = cgemm_kernel_power9.S
STRMMKERNEL = sgemm_kernel_power10.c
DTRMMKERNEL = dgemm_kernel_power10.c
CTRMMKERNEL = cgemm_kernel_power10.S
ZTRMMKERNEL = zgemm_kernel_power9.S

SGEMMKERNEL = sgemm_kernel_power9.S
SGEMMKERNEL = sgemm_kernel_power10.c
SGEMMINCOPY = ../generic/gemm_ncopy_16.c
SGEMMITCOPY = sgemm_tcopy_16_power8.S
SGEMMONCOPY = ../generic/gemm_ncopy_8.c
@@ -22,7 +22,7 @@ SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)

DGEMMKERNEL = dgemm_kernel_power9.S
DGEMMKERNEL = dgemm_kernel_power10.c
DGEMMINCOPY = ../generic/gemm_ncopy_16.c
DGEMMITCOPY = dgemm_tcopy_16_power8.S
DGEMMONCOPY = dgemm_ncopy_4_power8.S
@@ -32,7 +32,7 @@ DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)

CGEMMKERNEL = cgemm_kernel_power9.S
CGEMMKERNEL = cgemm_kernel_power10.S
CGEMMINCOPY = ../generic/zgemm_ncopy_8.c
CGEMMITCOPY = ../generic/zgemm_tcopy_8.c
CGEMMONCOPY = ../generic/zgemm_ncopy_4.c


+ 286
- 0
kernel/power/cgemm_kernel_power10.S View File

@@ -0,0 +1,286 @@
/***************************************************************************
Copyright (c) 2013-2020, 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.
*****************************************************************************/

#define ASSEMBLER
#include "common.h"
#include "def_vsx.h"

#define LOAD ld
#define STACKSIZE (512 )
#define FLINK_SAVE (STACKSIZE+16) /* 16($r12) */
#define M r3
#define N r4
#define K r5


#define A r8
#define B r9
#define C r10
#define LDC r6
#define OFFSET r7


#define alpha_r vs51
#define alpha_i vs55
#define save_permute_1 vs59
#define permute_mask vs63
#define o0 0

#define T1 r11
#define T2 r12
#define T3 r14
#define T4 r15
#define T5 r16
#define T6 r17
#define L r18
#define T7 r19
#define T8 r20
#define TEMP_REG r21
#define I r22
#define J r23
#define AO r24
#define BO r25
#define CO r26
#define T9 r27
#define T10 r28
#define PRE r29

#define T12 r30
#define T13 r31

#include "cgemm_macros_power10.S"

.equ perm_const1, 0x0405060700010203
.equ perm_const2, 0x0c0d0e0f08090a0b
.equ save_permute_12, 0x0c0d0e0f1c1d1e1f
.equ save_permute_11, 0x0405060714151617



#ifndef NEEDPARAM

PROLOGUE
PROFCODE


addi SP, SP, -STACKSIZE
mflr r0


stfd f14, 0(SP)
stfd f15, 8(SP)
stfd f16, 16(SP)
stfd f17, 24(SP)

stfd f18, 32(SP)
stfd f19, 40(SP)
stfd f20, 48(SP)
stfd f21, 56(SP)

stfd f22, 64(SP)
stfd f23, 72(SP)
stfd f24, 80(SP)
stfd f25, 88(SP)

stfd f26, 96(SP)
stfd f27, 104(SP)
stfd f28, 112(SP)
stfd f29, 120(SP)

stfd f30, 128(SP)
stfd f31, 136(SP)


std r31, 144(SP)
std r30, 152(SP)
std r29, 160(SP)
std r28, 168(SP)
std r27, 176(SP)
std r26, 184(SP)
std r25, 192(SP)
std r24, 200(SP)
std r23, 208(SP)
std r22, 216(SP)
std r21, 224(SP)
std r20, 232(SP)
std r19, 240(SP)
std r18, 248(SP)
std r17, 256(SP)
std r16, 264(SP)
std r15, 272(SP)
std r14, 280(SP)
stxv vs52, 288(SP)
stxv vs53, 304(SP)
stxv vs54, 320(SP)
stxv vs55, 336(SP)
stxv vs56, 352(SP)
stxv vs57, 368(SP)
stxv vs58, 384(SP)
stxv vs59, 400(SP)
stxv vs60, 416(SP)
stxv vs61, 432(SP)
stxv vs62, 448(SP)
stxv vs63, 464(SP)
std r0, FLINK_SAVE(SP)


ld LDC, FRAMESLOT(0) + STACKSIZE(SP)



#ifdef TRMMKERNEL
ld OFFSET, FRAMESLOT(1) + STACKSIZE(SP)
#endif
slwi LDC, LDC, ZBASE_SHIFT

/*alpha is stored in f1. convert to single and splat*/
xscvdpspn alpha_r,vs1
xscvdpspn alpha_i,vs2
xxspltw alpha_r,alpha_r,0
xxspltw alpha_i,alpha_i,0
/*load reverse permute mask for big endian
uint128 = 0xc0d0e0f08090a0b0405060700010203
*/
lis T2, perm_const2@highest
lis T1, perm_const1@highest
lis T3, save_permute_12@highest
lis T4, save_permute_11@highest

ori T2, T2, perm_const2@higher
ori T1, T1, perm_const1@higher
ori T3, T3, save_permute_12@higher
ori T4, T4, save_permute_11@higher

rldicr T2, T2, 32, 31
rldicr T1, T1, 32, 31
rldicr T3, T3, 32, 31
rldicr T4, T4, 32, 31

oris T2, T2, perm_const2@h
oris T1, T1, perm_const1@h
oris T3, T3, save_permute_12@h
oris T4, T4, save_permute_11@h

ori T2, T2, perm_const2@l
ori T1, T1, perm_const1@l
ori T3, T3, save_permute_12@l
ori T4, T4, save_permute_11@l

li r0,0
li PRE,512

#if defined(CC) || defined(CR) || defined(RC) || defined(RR)
/*negate for this case as we will use addition -1*(a+b) */
xvnegsp alpha_r,alpha_r
xvnegsp alpha_i,alpha_i
#endif

mtvsrdd permute_mask,T2,T1
mtvsrdd save_permute_1,T3,T4

/*mask is reverse permute so we have to make it inner permute */
xxpermdi permute_mask, permute_mask, permute_mask,2

#include "cgemm_logic_power10.S"

.L999:
lfd f14, 0(SP)
lfd f15, 8(SP)
lfd f16, 16(SP)
lfd f17, 24(SP)

lfd f18, 32(SP)
lfd f19, 40(SP)
lfd f20, 48(SP)
lfd f21, 56(SP)

lfd f22, 64(SP)
lfd f23, 72(SP)
lfd f24, 80(SP)
lfd f25, 88(SP)

lfd f26, 96(SP)
lfd f27, 104(SP)
lfd f28, 112(SP)
lfd f29, 120(SP)

lfd f30, 128(SP)
lfd f31, 136(SP)

ld r31, 144(SP)
ld r30, 152(SP)
ld r29, 160(SP)
ld r28, 168(SP)
ld r27, 176(SP)
ld r26, 184(SP)
ld r25, 192(SP)
ld r24, 200(SP)
ld r23, 208(SP)
ld r22, 216(SP)
ld r21, 224(SP)
ld r20, 232(SP)
ld r19, 240(SP)
ld r18, 248(SP)
ld r17, 256(SP)
ld r16, 264(SP)
ld r15, 272(SP)
ld r14, 280(SP)

ld r0, FLINK_SAVE(SP)
lxv vs52, 288(SP)
lxv vs53, 304(SP)
lxv vs54, 320(SP)
lxv vs55, 336(SP)
lxv vs56, 352(SP)
lxv vs57, 368(SP)
lxv vs58, 384(SP)
lxv vs59, 400(SP)
mtlr r0
lxv vs60, 416(SP)
lxv vs61, 432(SP)
lxv vs62, 448(SP)
lxv vs63, 464(SP)

addi SP, SP, STACKSIZE
blr


EPILOGUE
#endif

+ 2814
- 0
kernel/power/cgemm_logic_power10.S
File diff suppressed because it is too large
View File


+ 2131
- 0
kernel/power/cgemm_macros_power10.S
File diff suppressed because it is too large
View File


+ 864
- 0
kernel/power/dgemm_kernel_power10.c View File

@@ -0,0 +1,864 @@
/*********************************************************************************
Copyright (c) 2020, 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"
#include <altivec.h>

typedef unsigned char vec_t __attribute__ ((vector_size (16)));
typedef FLOAT v4sf_t __attribute__ ((vector_size (16)));
typedef FLOAT v2sf_t __attribute__ ((vector_size (8)));

#ifdef TRMMKERNEL
#define SAVE_ACC(ACC, J) \
__builtin_mma_disassemble_acc (result, ACC); \
rowC = (v4sf_t *) &CO[0* ldc+J]; \
rowC[0] = result[3] * alpha; \
rowC = (v4sf_t *) &CO[1*ldc+J]; \
rowC[0] = result[2] * alpha; \
rowC = (v4sf_t *) &CO[2*ldc+J]; \
rowC[0] = result[1] * alpha; \
rowC = (v4sf_t *) &CO[3*ldc+J]; \
rowC[0] = result[0] * alpha;
#define SAVE_ACC1(ACC, J) \
__builtin_mma_disassemble_acc (result, ACC); \
rowC = (v4sf_t *) &CO[4* ldc+J]; \
rowC[0] = result[3] * alpha; \
rowC = (v4sf_t *) &CO[5*ldc+J]; \
rowC[0] = result[2] * alpha; \
rowC = (v4sf_t *) &CO[6*ldc+J]; \
rowC[0] = result[1] * alpha; \
rowC = (v4sf_t *) &CO[7*ldc+J]; \
rowC[0] = result[0] * alpha;
#define SAVE2x4_ACC(ACC, J) \
__builtin_mma_disassemble_acc (result, ACC); \
rowC = (v4sf_t *) &CO[0* ldc+J]; \
rowC[0] = result[3] * alpha; \
rowC = (v4sf_t *) &CO[1* ldc+J]; \
rowC[0] = result[2] * alpha;
#else
#define SAVE_ACC(ACC, J) \
__builtin_mma_disassemble_acc (result, ACC); \
rowC = (v4sf_t *) &CO[0* ldc+J]; \
rowC[0] += result[3] * alpha; \
rowC = (v4sf_t *) &CO[1*ldc+J]; \
rowC[0] += result[2] * alpha; \
rowC = (v4sf_t *) &CO[2*ldc+J]; \
rowC[0] += result[1] * alpha; \
rowC = (v4sf_t *) &CO[3*ldc+J]; \
rowC[0] += result[0] * alpha;
#define SAVE_ACC1(ACC, J) \
__builtin_mma_disassemble_acc (result, ACC); \
rowC = (v4sf_t *) &CO[4* ldc+J]; \
rowC[0] += result[3] * alpha; \
rowC = (v4sf_t *) &CO[5*ldc+J]; \
rowC[0] += result[2] * alpha; \
rowC = (v4sf_t *) &CO[6*ldc+J]; \
rowC[0] += result[1] * alpha; \
rowC = (v4sf_t *) &CO[7*ldc+J]; \
rowC[0] += result[0] * alpha;
#define SAVE2x4_ACC(ACC, J) \
__builtin_mma_disassemble_acc (result, ACC); \
rowC = (v4sf_t *) &CO[0* ldc+J]; \
rowC[0] += result[3] * alpha; \
rowC = (v4sf_t *) &CO[1* ldc+J]; \
rowC[0] += result[2] * alpha;
#endif

#define SET_ACC_ZERO4() \
__builtin_mma_xxsetaccz (&acc0); \
__builtin_mma_xxsetaccz (&acc1); \
__builtin_mma_xxsetaccz (&acc2); \
__builtin_mma_xxsetaccz (&acc3);

#define SET_ACC_ZERO8() \
__builtin_mma_xxsetaccz (&acc0); \
__builtin_mma_xxsetaccz (&acc1); \
__builtin_mma_xxsetaccz (&acc2); \
__builtin_mma_xxsetaccz (&acc3); \
__builtin_mma_xxsetaccz (&acc4); \
__builtin_mma_xxsetaccz (&acc5); \
__builtin_mma_xxsetaccz (&acc6); \
__builtin_mma_xxsetaccz (&acc7);

#define PREFETCH1(x, y) asm volatile ("dcbt %0, %1" : : "r" (x), "b" (y) : "memory");

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
#define REFRESH_TEMP_BK(x, y) \
temp = k - off;
#elif defined(LEFT)
#define REFRESH_TEMP_BK(x, y) \
temp = off + x;
#else
#define REFRESH_TEMP_BK(x, y) \
temp = off + y;
#endif
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
#define REFRESH_POINTERS(x, y) \
BO = B; \
REFRESH_TEMP_BK(x, y)
#else
#define REFRESH_POINTERS(x, y) \
AO += off * x; \
BO = B + off * y; \
REFRESH_TEMP_BK(x, y)
#endif

#ifdef LEFT
#define REFRESH_OFF(x) \
off += x;
#else
#define REFRESH_OFF(x)
#endif

#ifdef LEFT
#define UPDATE_TEMP(x, y) \
temp -= x;
#else
#define UPDATE_TEMP(x, y) \
temp -= y;
#endif

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
#define REFRESH_TMP_AFTER_SAVE(x, y) \
temp = k - off; \
UPDATE_TEMP(x, y) \
AO += temp * x; \
BO += temp * y;
#else
#define REFRESH_TMP_AFTER_SAVE(x, y)
#endif

#define REFRESH_AFTER_SAVE(x,y) \
REFRESH_TMP_AFTER_SAVE(x, y) \
REFRESH_OFF(x)
/*************************************************************************************
* GEMM Kernel
*************************************************************************************/
int
CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
FLOAT * C, BLASLONG ldc
#ifdef TRMMKERNEL
, BLASLONG offset
#endif
)
{
BLASLONG N = n;
BLASLONG i1;
#if defined(TRMMKERNEL)
BLASLONG off;
#endif
#if defined(TRMMKERNEL) && !defined(LEFT)
off = -offset;
#endif
v4sf_t valpha = { alpha, alpha };
N = n >> 2;
for (i1 = 0; i1 < N; i1++)
{
BLASLONG i, j, temp;
FLOAT *CO;
FLOAT *AO;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
CO = C;
C += ldc << 2;
AO = A;
PREFETCH1 (A, 128);
PREFETCH1 (A, 256);
i = m >> 4;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (16, 4);
#else
BO = B;
temp = k;
#endif
v4sf_t *rowC;
v4sf_t result[4];
BLASLONG l = 0;
PREFETCH1 (CO, 0);
PREFETCH1 (CO + ldc, 0);
PREFETCH1 (CO + ldc + ldc, 0);
PREFETCH1 (CO + ldc + ldc + ldc, 0);
PREFETCH1 (CO, 128);
PREFETCH1 (CO + ldc, 128);
PREFETCH1 (CO + ldc + ldc, 128);
PREFETCH1 (CO + ldc + ldc + ldc, 128);
__vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;
SET_ACC_ZERO8 ();
for (l = 0; l < temp; l++)
{
vec_t *rowA = (vec_t *) & AO[l << 4];
__vector_pair rowB;
vec_t *rb = (vec_t *) & BO[l << 2];
__builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
__builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
__builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
__builtin_mma_xvf64gerpp (&acc4, rowB, rowA[4]);
__builtin_mma_xvf64gerpp (&acc5, rowB, rowA[5]);
__builtin_mma_xvf64gerpp (&acc6, rowB, rowA[6]);
__builtin_mma_xvf64gerpp (&acc7, rowB, rowA[7]);
}
SAVE_ACC (&acc0, 0);
SAVE_ACC (&acc2, 4);
SAVE_ACC (&acc1, 2);
SAVE_ACC (&acc3, 6);
SAVE_ACC (&acc4, 8);
SAVE_ACC (&acc6, 12);
SAVE_ACC (&acc5, 10);
SAVE_ACC (&acc7, 14);
AO += temp << 4;
BO += temp << 2;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (16, 4)
#endif
CO += 16;
}
i = (m & 15) >> 3;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (8, 4);
#else
BO = B;
temp = k;
#endif
v4sf_t *rowC;
v4sf_t result[4];
__vector_quad acc0, acc1, acc2, acc3;
SET_ACC_ZERO4 ();
BLASLONG l = 0;
for (l = 0; l < temp; l++)
{
vec_t *rowA = (vec_t *) & AO[l << 3];
__vector_pair rowB;
vec_t *rb = (vec_t *) & BO[l << 2];
__builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
__builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
__builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
}
SAVE_ACC (&acc0, 0);
SAVE_ACC (&acc2, 4);
SAVE_ACC (&acc1, 2);
SAVE_ACC (&acc3, 6);
CO += 8;
AO += temp << 3;
BO += temp << 2;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (8, 4)
#endif
}
i = (m & 7) >> 2;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (4, 4);
#else
BO = B;
temp = k;
#endif
v4sf_t *rowC;
v4sf_t result[4];
__vector_quad acc0, acc1;
__builtin_mma_xxsetaccz (&acc0);
__builtin_mma_xxsetaccz (&acc1);
BLASLONG l = 0;
for (l = 0; l < temp; l++)
{
vec_t *rowA = (vec_t *) & AO[l << 2];
__vector_pair rowB;
vec_t *rb = (vec_t *) & BO[l << 2];
__builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
}
SAVE_ACC (&acc0, 0);
SAVE_ACC (&acc1, 2);
CO += 4;
AO += temp << 2;
BO += temp << 2;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (4, 4)
#endif
}
i = (m & 3) >> 1;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (2, 4);
#else
BO = B;
temp = k;
#endif
v4sf_t *rowC;
v4sf_t result[4];
__vector_quad acc0;
__builtin_mma_xxsetaccz (&acc0);
BLASLONG l = 0;
for (l = 0; l < temp; l++)
{
vec_t *rowA = (vec_t *) & AO[l << 1];
__vector_pair rowB;
vec_t *rb = (vec_t *) & BO[l << 2];
__builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
}
SAVE_ACC (&acc0, 0);
CO += 2;
AO += temp << 1;
BO += temp << 2;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (2, 4)
#endif
}
i = (m & 1) >> 0;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (1, 4);
#else
BO = B;
temp = k;
#endif
BLASLONG l = 0;
v4sf_t t = { 0, 0 };
v4sf_t t1 = { 0, 0 };
for (l = 0; l < temp; l++)
{
v4sf_t rowA = { AO[l], AO[l] };
v4sf_t rowB = { BO[l << 2], BO[(l << 2) + 1] };
v4sf_t rowB1 = { BO[(l << 2) + 2], BO[(l << 2) + 3] };
t += rowA * rowB;
t1 += rowA * rowB1;
}
t = t * valpha;
t1 = t1 * valpha;
#if defined(TRMMKERNEL)
CO[0 * ldc] = t[0];
CO[1 * ldc] = t[1];
CO[2 * ldc] = t1[0];
CO[3 * ldc] = t1[1];
#else
CO[0 * ldc] += t[0];
CO[1 * ldc] += t[1];
CO[2 * ldc] += t1[0];
CO[3 * ldc] += t1[1];
#endif
CO += 1;
AO += temp;
BO += temp << 2;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (1, 4)
#endif
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 4; // number of values in A
#endif
B += k << 2;
}
N = (n & 3) >> 1;
for (i1 = 0; i1 < N; i1++)
{
BLASLONG i, j, temp;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
FLOAT *CO;
FLOAT *AO;
CO = C;
C += ldc << 1;
AO = A;
i = m >> 4;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (16, 2);
#else
BO = B;
temp = k;
#endif
v4sf_t *rowC;
v4sf_t result[4];
__vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;
SET_ACC_ZERO8 ();
BLASLONG l = 0;
for (l = 0; l < temp; l++)
{
FLOAT t[4] = { 0, 0, 0, 0 };
t[0] = BO[l << 1], t[1] = BO[(l << 1) + 1];
__vector_pair rowB;
vec_t *rb = (vec_t *) & t[0];
__builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
vec_t *rowA = (vec_t *) & AO[l << 4];
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
__builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
__builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
__builtin_mma_xvf64gerpp (&acc4, rowB, rowA[4]);
__builtin_mma_xvf64gerpp (&acc5, rowB, rowA[5]);
__builtin_mma_xvf64gerpp (&acc6, rowB, rowA[6]);
__builtin_mma_xvf64gerpp (&acc7, rowB, rowA[7]);
}
SAVE2x4_ACC (&acc0, 0);
SAVE2x4_ACC (&acc1, 2);
SAVE2x4_ACC (&acc2, 4);
SAVE2x4_ACC (&acc3, 6);
SAVE2x4_ACC (&acc4, 8);
SAVE2x4_ACC (&acc5, 10);
SAVE2x4_ACC (&acc6, 12);
SAVE2x4_ACC (&acc7, 14);
CO += 16;
AO += temp << 4;
BO += temp << 1;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (16, 2)
#endif
}
i = (m & 15) >> 3;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (8, 2);
#else
BO = B;
temp = k;
#endif
v4sf_t *rowC;
v4sf_t result[4];
__vector_quad acc0, acc1, acc2, acc3;
SET_ACC_ZERO4 ();
BLASLONG l = 0;
for (l = 0; l < temp; l++)
{
FLOAT t[4] = { 0, 0, 0, 0 };
t[0] = BO[l << 1], t[1] = BO[(l << 1) + 1];
__vector_pair rowB;
vec_t *rb = (vec_t *) & t[0];
__builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
vec_t *rowA = (vec_t *) & AO[l << 3];
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
__builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
__builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
}
SAVE2x4_ACC (&acc0, 0);
SAVE2x4_ACC (&acc1, 2);
SAVE2x4_ACC (&acc2, 4);
SAVE2x4_ACC (&acc3, 6);
CO += 8;
AO += temp << 3;
BO += temp << 1;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (8, 2)
#endif
}
i = (m & 7) >> 2;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (4, 2);
#else
BO = B;
temp = k;
#endif
v4sf_t *rowC;
v4sf_t result[4];
__vector_quad acc0, acc1;
__builtin_mma_xxsetaccz (&acc0);
__builtin_mma_xxsetaccz (&acc1);
BLASLONG l = 0;
for (l = 0; l < temp; l++)
{
FLOAT t[4] = { 0, 0, 0, 0 };
t[0] = BO[l << 1], t[1] = BO[(l << 1) + 1];
__vector_pair rowB;
vec_t *rb = (vec_t *) & t[0];
__builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
vec_t *rowA = (vec_t *) & AO[l << 2];
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
}
SAVE2x4_ACC (&acc0, 0);
SAVE2x4_ACC (&acc1, 2);
CO += 4;
AO += temp << 2;
BO += temp << 1;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (4, 2)
#endif
}
i = (m & 3) >> 1;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (2, 2);
#else
BO = B;
temp = k;
#endif
v4sf_t *rowC;
v4sf_t result[4];
__vector_quad acc0;
__builtin_mma_xxsetaccz (&acc0);
BLASLONG l = 0;
for (l = 0; l < temp; l++)
{
FLOAT t[4] = { 0, 0, 0, 0 };
t[0] = BO[l << 1], t[1] = BO[(l << 1) + 1];
__vector_pair rowB;
vec_t *rb = (vec_t *) & t[0];
__builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
vec_t *rowA = (vec_t *) & AO[l << 1];
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
}
SAVE2x4_ACC (&acc0, 0);
CO += 2;
AO += temp << 1;
BO += temp << 1;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (2, 2)
#endif
}
i = (m & 1) >> 0;
for (j = 0; j < i; j++)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (1, 2);
#else
BO = B;
temp = k;
#endif
BLASLONG l = 0;
v4sf_t t = { 0, 0 };
for (l = 0; l < temp; l++)
{
v4sf_t rowA = { AO[l], AO[l] };
v4sf_t rowB = { BO[l << 1], BO[(l << 1) + 1] };
t += rowA * rowB;
}
t = t * valpha;
#if defined(TRMMKERNEL)
CO[0 * ldc] = t[0];
CO[1 * ldc] = t[1];
#else
CO[0 * ldc] += t[0];
CO[1 * ldc] += t[1];
#endif
CO += 1;
AO += temp;
BO += temp << 1;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (1, 2)
#endif
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 2; // number of values in A
#endif
B += k << 1;
}
N = (n & 1) >> 0;
for (i1 = 0; i1 < N; i1++)
{
BLASLONG i, temp;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
FLOAT *CO;
FLOAT *AO;
CO = C;
C += ldc;
AO = A;
i = m;
while (i >= 16)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (16, 1)
#else
BO = B;
temp = k;
#endif
BLASLONG l = 0;
v4sf_t t = { 0, 0 };
v4sf_t t1 = { 0, 0 };
v4sf_t t2 = { 0, 0 };
v4sf_t t3 = { 0, 0 };
v4sf_t t4 = { 0, 0 };
v4sf_t t5 = { 0, 0 };
v4sf_t t6 = { 0, 0 };
v4sf_t t7 = { 0, 0 };
for (l = 0; l < temp; l++)
{
v4sf_t rowB = { BO[l], BO[l] };
v4sf_t rowA = { AO[l << 4], AO[(l << 4) + 1] };
v4sf_t rowA1 = { AO[(l << 4) + 2], AO[(l << 4) + 3] };
v4sf_t rowA2 = { AO[(l << 4) + 4], AO[(l << 4) + 5] };
v4sf_t rowA3 = { AO[(l << 4) + 6], AO[(l << 4) + 7] };
v4sf_t rowA4 = { AO[(l << 4) + 8], AO[(l << 4) + 9] };
v4sf_t rowA5 = { AO[(l << 4) + 10], AO[(l << 4) + 11] };
v4sf_t rowA6 = { AO[(l << 4) + 12], AO[(l << 4) + 13] };
v4sf_t rowA7 = { AO[(l << 4) + 14], AO[(l << 4) + 15] };
t += rowA * rowB;
t1 += rowA1 * rowB;
t2 += rowA2 * rowB;
t3 += rowA3 * rowB;
t4 += rowA4 * rowB;
t5 += rowA5 * rowB;
t6 += rowA6 * rowB;
t7 += rowA7 * rowB;
}
t = t * valpha;
t1 = t1 * valpha;
t2 = t2 * valpha;
t3 = t3 * valpha;
t4 = t4 * valpha;
t5 = t5 * valpha;
t6 = t6 * valpha;
t7 = t7 * valpha;
#if defined(TRMMKERNEL)
CO[0] = t[0];
CO[1] = t[1];
CO[2] = t1[0];
CO[3] = t1[1];
CO[4] = t2[0];
CO[5] = t2[1];
CO[6] = t3[0];
CO[7] = t3[1];
CO[8] = t4[0];
CO[9] = t4[1];
CO[10] = t5[0];
CO[11] = t5[1];
CO[12] = t6[0];
CO[13] = t6[1];
CO[14] = t7[0];
CO[15] = t7[1];
#else
CO[0] += t[0];
CO[1] += t[1];
CO[2] += t1[0];
CO[3] += t1[1];
CO[4] += t2[0];
CO[5] += t2[1];
CO[6] += t3[0];
CO[7] += t3[1];
CO[8] += t4[0];
CO[9] += t4[1];
CO[10] += t5[0];
CO[11] += t5[1];
CO[12] += t6[0];
CO[13] += t6[1];
CO[14] += t7[0];
CO[15] += t7[1];
#endif
AO += temp << 4;
BO += temp;
CO += 16;
i -= 16;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (16, 1)
#endif
}
while (i >= 8)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (8, 1)
#else
BO = B;
temp = k;
#endif
BLASLONG l = 0;
v4sf_t t = { 0, 0 };
v4sf_t t1 = { 0, 0 };
v4sf_t t2 = { 0, 0 };
v4sf_t t3 = { 0, 0 };
for (l = 0; l < temp; l++)
{
v4sf_t rowB = { BO[l], BO[l] };
v4sf_t rowA = { AO[l << 3], AO[(l << 3) + 1] };
v4sf_t rowA1 = { AO[(l << 3) + 2], AO[(l << 3) + 3] };
v4sf_t rowA2 = { AO[(l << 3) + 4], AO[(l << 3) + 5] };
v4sf_t rowA3 = { AO[(l << 3) + 6], AO[(l << 3) + 7] };
t += rowA * rowB;
t1 += rowA1 * rowB;
t2 += rowA2 * rowB;
t3 += rowA3 * rowB;
}
t = t * valpha;
t1 = t1 * valpha;
t2 = t2 * valpha;
t3 = t3 * valpha;
#if defined(TRMMKERNEL)
CO[0] = t[0];
CO[1] = t[1];
CO[2] = t1[0];
CO[3] = t1[1];
CO[4] = t2[0];
CO[5] = t2[1];
CO[6] = t3[0];
CO[7] = t3[1];
#else
CO[0] += t[0];
CO[1] += t[1];
CO[2] += t1[0];
CO[3] += t1[1];
CO[4] += t2[0];
CO[5] += t2[1];
CO[6] += t3[0];
CO[7] += t3[1];
#endif
AO += temp << 3;
BO += temp;
CO += 8;
i -= 8;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (8, 1)
#endif
}
while (i >= 4)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (4, 1)
#else
BO = B;
temp = k;
#endif
BLASLONG l = 0;
v4sf_t t = { 0, 0 };
v4sf_t t1 = { 0, 0 };
for (l = 0; l < temp; l++)
{
v4sf_t rowB = { BO[l], BO[l] };
v4sf_t rowA = { AO[l << 2], AO[(l << 2) + 1] };
v4sf_t rowA1 = { AO[(l << 2) + 2], AO[(l << 2) + 3] };
t += rowA * rowB;
t1 += rowA1 * rowB;
}
t = t * valpha;
t1 = t1 * valpha;
#if defined(TRMMKERNEL)
CO[0] = t[0];
CO[1] = t[1];
CO[2] = t1[0];
CO[3] = t1[1];
#else
CO[0] += t[0];
CO[1] += t[1];
CO[2] += t1[0];
CO[3] += t1[1];
#endif
AO += temp << 2;
BO += temp;
CO += 4;
i -= 4;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (4, 1)
#endif
}
while (i >= 2)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (2, 1)
#else
BO = B;
temp = k;
#endif
BLASLONG l = 0;
v4sf_t t = { 0, 0 };
for (l = 0; l < temp; l++)
{
v4sf_t rowB = { BO[l], BO[l] };
v4sf_t rowA = { AO[l << 1], AO[(l << 1) + 1] };
t += rowA * rowB;
}
t = t * valpha;
#if defined(TRMMKERNEL)
CO[0] = t[0];
CO[1] = t[1];
#else
CO[0] += t[0];
CO[1] += t[1];
#endif
AO += temp << 1;
BO += temp;
CO += 2;
i -= 2;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (2, 1)
#endif
}
while (i >= 1)
{
FLOAT *BO;
#if defined(TRMMKERNEL)
REFRESH_POINTERS (1, 1)
#else
BO = B;
temp = k;
#endif
BLASLONG l = 0;
FLOAT t = 0;
for (l = 0; l < temp; l++)
{
t += AO[l] * BO[l];
}
AO += temp;
BO += temp;
#if defined(TRMMKERNEL)
CO[0] = t * alpha;
#else
CO[0] += t * alpha;
#endif
CO += 1;
i -= 1;
#if defined(TRMMKERNEL)
REFRESH_AFTER_SAVE (1, 1)
#endif
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 1; // number of values in A
#endif
B += k;
}
return 0;
}

+ 1334
- 0
kernel/power/sgemm_kernel_power10.c
File diff suppressed because it is too large
View File


Write Preview
Loading…
Cancel
Save