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

Merge pull request #1061 from ashwinyes/develop_aarch64_vulcan_thunderx_patch 

Add new targets for ARM64
tags/v0.2.20^2
Zhang Xianyi GitHub 8 years ago
parent
2e5f906f41 4b55fae337
commit
0863a0d4b4
20 changed files with 4482 additions and 20 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +14
    -0
      Makefile.arm64
  2. +3
    -0
      TargetList.txt
  3. +1
    -1
      common_macro.h
  4. +91
    -17
      cpuid_arm64.c
  5. +1
    -1
      driver/others/memory.c
  6. +23
    -0
      driver/others/parameter.c
  7. +49
    -1
      getarch.c
  8. +15
    -0
      kernel/arm64/KERNEL.CORTEXA57
  9. +6
    -0
      kernel/arm64/KERNEL.THUNDERX
  10. +2
    -0
      kernel/arm64/KERNEL.THUNDERX2T99
  11. +4
    -0
      kernel/arm64/KERNEL.VULCAN
  12. +151
    -0
      kernel/arm64/daxpy-thunderx.c
  13. +119
    -0
      kernel/arm64/ddot-thunderx.c
  14. +1817
    -0
      kernel/arm64/dgemm_kernel_8x4_vulcan.S
  15. +340
    -0
      kernel/arm64/dgemm_ncopy_4.S
  16. +544
    -0
      kernel/arm64/dgemm_ncopy_8.S
  17. +402
    -0
      kernel/arm64/dgemm_tcopy_4.S
  18. +682
    -0
      kernel/arm64/dgemm_tcopy_8.S
  19. +104
    -0
      kernel/arm64/dot-thunderx.c
  20. +114
    -0
      param.h

+ 14
- 0
Makefile.arm64 View File

@@ -9,3 +9,17 @@ CCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57
FCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57 FCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57
endif endif


ifeq ($(CORE), VULCAN)
CCOMMON_OPT += -mtune=vulcan -mcpu=vulcan
FCOMMON_OPT += -mtune=vulcan -mcpu=vulcan
endif

ifeq ($(CORE), THUNDERX)
CCOMMON_OPT += -mtune=thunderx -mcpu=thunderx
FCOMMON_OPT += -mtune=thunderx -mcpu=thunderx
endif

ifeq ($(CORE), THUNDERX2T99)
CCOMMON_OPT += -mtune=vulcan -mcpu=vulcan
FCOMMON_OPT += -mtune=vulcan -mcpu=vulcan
endif

+ 3
- 0
TargetList.txt View File

@@ -80,4 +80,7 @@ ARMV5
8.ARM 64-bit CPU: 8.ARM 64-bit CPU:
ARMV8 ARMV8
CORTEXA57 CORTEXA57
VULCAN
THUNDERX
THUNDERX2T99



+ 1
- 1
common_macro.h View File

@@ -2193,7 +2193,7 @@
#endif #endif


#ifndef ASSEMBLER #ifndef ASSEMBLER
#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64)
#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64)
extern BLASLONG gemm_offset_a; extern BLASLONG gemm_offset_a;
extern BLASLONG gemm_offset_b; extern BLASLONG gemm_offset_b;
extern BLASLONG sgemm_p; extern BLASLONG sgemm_p;


+ 91
- 17
cpuid_arm64.c View File

@@ -30,17 +30,26 @@
#define CPU_UNKNOWN 0 #define CPU_UNKNOWN 0
#define CPU_ARMV8 1 #define CPU_ARMV8 1
#define CPU_CORTEXA57 2 #define CPU_CORTEXA57 2
#define CPU_VULCAN 3
#define CPU_THUNDERX 4
#define CPU_THUNDERX2T99 5


static char *cpuname[] = { static char *cpuname[] = {
"UNKNOWN", "UNKNOWN",
"ARMV8" , "ARMV8" ,
"CORTEXA57"
"CORTEXA57",
"VULCAN",
"THUNDERX",
"THUNDERX2T99"
}; };


static char *cpuname_lower[] = { static char *cpuname_lower[] = {
"unknown", "unknown",
"armv8" , "armv8" ,
"cortexa57"
"cortexa57",
"vulcan",
"thunderx",
"thunderx2t99"
}; };


int get_feature(char *search) int get_feature(char *search)
@@ -85,25 +94,34 @@ int detect(void)
#ifdef linux #ifdef linux


FILE *infile; FILE *infile;
char buffer[512], *p;
p = (char *) NULL ;

infile = fopen("/proc/cpuinfo", "r");
while (fgets(buffer, sizeof(buffer), infile))
{
char buffer[512], *p, *cpu_part = NULL, *cpu_implementer = NULL;
p = (char *) NULL ;


if (!strncmp("CPU part", buffer, 8))
{
p = strchr(buffer, ':') + 2;
infile = fopen("/proc/cpuinfo", "r");
while (fgets(buffer, sizeof(buffer), infile)) {
if ((cpu_part != NULL) && (cpu_implementer != NULL)) {
break; break;
} }

if ((cpu_part == NULL) && !strncmp("CPU part", buffer, 8)) {
cpu_part = strchr(buffer, ':') + 2;
cpu_part = strdup(cpu_part);
} else if ((cpu_implementer == NULL) && !strncmp("CPU implementer", buffer, 15)) {
cpu_implementer = strchr(buffer, ':') + 2;
cpu_implementer = strdup(cpu_implementer);
}
} }


fclose(infile); fclose(infile);
if(p != NULL) {
if (strstr(p, "0xd07")) {
return CPU_CORTEXA57;
}
if(cpu_part != NULL && cpu_implementer != NULL) {
if (strstr(cpu_part, "0xd07") && strstr(cpu_implementer, "0x41"))
return CPU_CORTEXA57;
else if (strstr(cpu_part, "0x516") && strstr(cpu_implementer, "0x42"))
return CPU_VULCAN;
else if (strstr(cpu_part, "0x0a1") && strstr(cpu_implementer, "0x43"))
return CPU_THUNDERX;
else if (strstr(cpu_part, "0xFFF") && strstr(cpu_implementer, "0x43")) /* TODO */
return CPU_THUNDERX2T99;
} }


p = (char *) NULL ; p = (char *) NULL ;
@@ -176,6 +194,28 @@ void get_cpuconfig(void)
printf("#define L2_ASSOCIATIVE 4\n"); printf("#define L2_ASSOCIATIVE 4\n");
break; break;


case CPU_VULCAN:
printf("#define VULCAN \n");
printf("#define HAVE_VFP \n");
printf("#define HAVE_VFPV3 \n");
printf("#define HAVE_NEON \n");
printf("#define HAVE_VFPV4 \n");
printf("#define L1_CODE_SIZE 32768 \n");
printf("#define L1_CODE_LINESIZE 64 \n");
printf("#define L1_CODE_ASSOCIATIVE 8 \n");
printf("#define L1_DATA_SIZE 32768 \n");
printf("#define L1_DATA_LINESIZE 64 \n");
printf("#define L1_DATA_ASSOCIATIVE 8 \n");
printf("#define L2_SIZE 262144 \n");
printf("#define L2_LINESIZE 64 \n");
printf("#define L2_ASSOCIATIVE 8 \n");
printf("#define L3_SIZE 33554432 \n");
printf("#define L3_LINESIZE 64 \n");
printf("#define L3_ASSOCIATIVE 32 \n");
printf("#define DTB_DEFAULT_ENTRIES 64 \n");
printf("#define DTB_SIZE 4096 \n");
break;

case CPU_CORTEXA57: case CPU_CORTEXA57:
printf("#define CORTEXA57\n"); printf("#define CORTEXA57\n");
printf("#define HAVE_VFP\n"); printf("#define HAVE_VFP\n");
@@ -191,8 +231,42 @@ void get_cpuconfig(void)
printf("#define L2_SIZE 2097152\n"); printf("#define L2_SIZE 2097152\n");
printf("#define L2_LINESIZE 64\n"); printf("#define L2_LINESIZE 64\n");
printf("#define L2_ASSOCIATIVE 16\n"); printf("#define L2_ASSOCIATIVE 16\n");
printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 4096\n");
printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 4096\n");
break;

case CPU_THUNDERX:
printf("#define ARMV8\n");
printf("#define THUNDERX\n");
printf("#define L1_DATA_SIZE 32768\n");
printf("#define L1_DATA_LINESIZE 128\n");
printf("#define L2_SIZE 16777216\n");
printf("#define L2_LINESIZE 128\n");
printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 4096\n");
printf("#define L2_ASSOCIATIVE 16\n");
break;

case CPU_THUNDERX2T99:
printf("#define VULCAN \n");
printf("#define HAVE_VFP \n");
printf("#define HAVE_VFPV3 \n");
printf("#define HAVE_NEON \n");
printf("#define HAVE_VFPV4 \n");
printf("#define L1_CODE_SIZE 32768 \n");
printf("#define L1_CODE_LINESIZE 64 \n");
printf("#define L1_CODE_ASSOCIATIVE 8 \n");
printf("#define L1_DATA_SIZE 32768 \n");
printf("#define L1_DATA_LINESIZE 64 \n");
printf("#define L1_DATA_ASSOCIATIVE 8 \n");
printf("#define L2_SIZE 262144 \n");
printf("#define L2_LINESIZE 64 \n");
printf("#define L2_ASSOCIATIVE 8 \n");
printf("#define L3_SIZE 33554432 \n");
printf("#define L3_LINESIZE 64 \n");
printf("#define L3_ASSOCIATIVE 32 \n");
printf("#define DTB_DEFAULT_ENTRIES 64 \n");
printf("#define DTB_SIZE 4096 \n");
break; break;
} }
} }


+ 1
- 1
driver/others/memory.c View File

@@ -995,7 +995,7 @@ void *blas_memory_alloc(int procpos){
if (!blas_num_threads) blas_cpu_number = blas_get_cpu_number(); if (!blas_num_threads) blas_cpu_number = blas_get_cpu_number();
#endif #endif


#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64)
#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64)
#ifndef DYNAMIC_ARCH #ifndef DYNAMIC_ARCH
blas_set_parameter(); blas_set_parameter();
#endif #endif


+ 23
- 0
driver/others/parameter.c View File

@@ -727,3 +727,26 @@ void blas_set_parameter(void){


} }
#endif #endif

#if defined(ARCH_ARM64)

#if defined(VULCAN) || defined(THUNDERX2T99)
unsigned long dgemm_prefetch_size_a;
unsigned long dgemm_prefetch_size_b;
unsigned long dgemm_prefetch_size_c;
#endif

void blas_set_parameter(void)
{
#if defined(VULCAN) || defined(THUNDERX2T99)
dgemm_p = 160;
dgemm_q = 128;
dgemm_r = 4096;

dgemm_prefetch_size_a = 3584;
dgemm_prefetch_size_b = 512;
dgemm_prefetch_size_c = 128;
#endif
}

#endif

+ 49
- 1
getarch.c View File

@@ -884,7 +884,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#ifdef FORCE_CORTEXA57 #ifdef FORCE_CORTEXA57
#define FORCE #define FORCE
#define ARCHITECTURE "ARM64" #define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "ARMV8"
#define SUBARCHITECTURE "CORTEXA57"
#define SUBDIRNAME "arm64" #define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DCORTEXA57 " \ #define ARCHCONFIG "-DCORTEXA57 " \
"-DL1_CODE_SIZE=49152 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=3 " \ "-DL1_CODE_SIZE=49152 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=3 " \
@@ -897,6 +897,54 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#else #else
#endif #endif


#ifdef FORCE_VULCAN
#define FORCE
#define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "VULCAN"
#define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DVULCAN " \
"-DL1_CODE_SIZE=32768 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=8 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=8 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=8 " \
"-DL3_SIZE=33554432 -DL3_LINESIZE=64 -DL3_ASSOCIATIVE=32 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON"
#define LIBNAME "vulcan"
#define CORENAME "VULCAN"
#else
#endif

#ifdef FORCE_THUNDERX
#define FORCE
#define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "THUNDERX"
#define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DTHUNDERX " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=16777216 -DL2_LINESIZE=128 -DL2_ASSOCIATIVE=16 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 "
#define LIBNAME "thunderx"
#define CORENAME "THUNDERX"
#else
#endif

#ifdef FORCE_THUNDERX2T99
#define FORCE
#define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "THUNDERX2T99"
#define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DTHUNDERX2T99 " \
"-DL1_CODE_SIZE=32768 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=8 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=8 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=8 " \
"-DL3_SIZE=33554432 -DL3_LINESIZE=64 -DL3_ASSOCIATIVE=32 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON"
#define LIBNAME "thunderx2t99"
#define CORENAME "THUNDERX2T99"
#else
#endif

#ifndef FORCE #ifndef FORCE


#if defined(__powerpc__) || defined(__powerpc) || defined(powerpc) || \ #if defined(__powerpc__) || defined(__powerpc) || defined(powerpc) || \


+ 15
- 0
kernel/arm64/KERNEL.CORTEXA57 View File

@@ -75,14 +75,29 @@ SGEMMOTCOPYOBJ = sgemm_otcopy.o


DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
DTRMMKERNEL = dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S DTRMMKERNEL = dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S

ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N)) ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))

ifeq ($(DGEMM_UNROLL_M), 8)
DGEMMINCOPY = dgemm_ncopy_$(DGEMM_UNROLL_M).S
DGEMMITCOPY = dgemm_tcopy_$(DGEMM_UNROLL_M).S
else
DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
endif

DGEMMINCOPYOBJ = dgemm_incopy.o DGEMMINCOPYOBJ = dgemm_incopy.o
DGEMMITCOPYOBJ = dgemm_itcopy.o DGEMMITCOPYOBJ = dgemm_itcopy.o
endif endif

ifeq ($(DGEMM_UNROLL_N), 4)
DGEMMONCOPY = dgemm_ncopy_$(DGEMM_UNROLL_N).S
DGEMMOTCOPY = dgemm_tcopy_$(DGEMM_UNROLL_N).S
else
DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
endif

DGEMMONCOPYOBJ = dgemm_oncopy.o DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o DGEMMOTCOPYOBJ = dgemm_otcopy.o




+ 6
- 0
kernel/arm64/KERNEL.THUNDERX View File

@@ -0,0 +1,6 @@
include $(KERNELDIR)/KERNEL.ARMV8

SDOTKERNEL=dot-thunderx.c
DDOTKERNEL=ddot-thunderx.c
DAXPYKERNEL=daxpy-thunderx.c


+ 2
- 0
kernel/arm64/KERNEL.THUNDERX2T99 View File

@@ -0,0 +1,2 @@
include $(KERNELDIR)/KERNEL.VULCAN


+ 4
- 0
kernel/arm64/KERNEL.VULCAN View File

@@ -0,0 +1,4 @@
include $(KERNELDIR)/KERNEL.CORTEXA57

DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N)_vulcan.S


+ 151
- 0
kernel/arm64/daxpy-thunderx.c View File

@@ -0,0 +1,151 @@
/***************************************************************************
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"
#include <arm_neon.h>

#define prefetch(a) __asm__("prfm PLDL1STRM, [%0]"::"r"(a):"memory");
//#define prefetch(a)

static void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
BLASLONG register i = 0;
double a = *alpha;

#if 0
prefetch(x + 128/sizeof(*x));
prefetch(y + 128/sizeof(*y));
#endif
prefetch(x + 2*128/sizeof(*x));
prefetch(y + 2*128/sizeof(*y));
prefetch(x + 3*128/sizeof(*x));
prefetch(y + 3*128/sizeof(*y));
prefetch(x + 4*128/sizeof(*x));
prefetch(y + 4*128/sizeof(*y));

while(i < n)
{
double y0, y1, y2, y3;
double y4, y5, y6, y7;
double *xx;
double *yy;
y0 = a * x[0] + y[0];
y1 = a * x[1] + y[1];
y2 = a * x[2] + y[2];
y3 = a * x[3] + y[3];
y4 = a * x[4] + y[4];
y5 = a * x[5] + y[5];
y6 = a * x[6] + y[6];
y7 = a * x[7] + y[7];
asm("":"+w"(y0),"+w"(y1),"+w"(y2),"+w"(y3),"+w"(y4),"+w"(y5),"+w"(y6),"+w"(y7));
y[0] = y0;
y[1] = y1;
y[2] = y2;
y[3] = y3;
y[4] = y4;
y[5] = y5;
y[6] = y6;
y[7] = y7;

xx = (x + 4*128/sizeof(*x));
yy = (y + 4*128/sizeof(*y));
asm("":"+r"(yy)::"memory");
prefetch(xx);
prefetch(yy);

y += 8;
x += 8;
i += 8 ;
}

}


int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
{
BLASLONG i=0;
BLASLONG ix=0,iy=0;

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

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

BLASLONG n1 = n & -32;

if ( n1 )
daxpy_kernel_8(n1, x, y , &da );

i = n1;
while(i < n)
{

y[i] += da * x[i] ;
i++ ;

}
return(0);


}

BLASLONG n1 = n & -4;

while(i < n1)
{

FLOAT m1 = da * x[ix] ;
FLOAT m2 = da * x[ix+inc_x] ;
FLOAT m3 = da * x[ix+2*inc_x] ;
FLOAT m4 = da * x[ix+3*inc_x] ;

y[iy] += m1 ;
y[iy+inc_y] += m2 ;
y[iy+2*inc_y] += m3 ;
y[iy+3*inc_y] += m4 ;

ix += inc_x*4 ;
iy += inc_y*4 ;
i+=4 ;

}

while(i < n)
{

y[iy] += da * x[ix] ;
ix += inc_x ;
iy += inc_y ;
i++ ;

}
return(0);

}



+ 119
- 0
kernel/arm64/ddot-thunderx.c View File

@@ -0,0 +1,119 @@
/***************************************************************************
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"

#include <arm_neon.h>

#define prefetch(a) __asm__("prfm PLDL1STRM, [%0]"::"r"(a):"memory");

FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{
BLASLONG i=0;
BLASLONG ix=0,iy=0;

FLOAT dot = 0.0 ;

if ( n < 0 ) return(dot);

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

float64x2_t vdot0 = {0.0, 0.0};
float64x2_t vdot1 = {0.0, 0.0};
float64x2_t vdot2 = {0.0, 0.0};
float64x2_t vdot3 = {0.0, 0.0};
float64x2_t *vx = (float64x2_t*)x;
float64x2_t *vy = (float64x2_t*)y;
#if 0
prefetch(x + 128/sizeof(*x));
prefetch(y + 128/sizeof(*y));
#endif
prefetch(x + 2*128/sizeof(*x));
prefetch(y + 2*128/sizeof(*y));
prefetch(x + 3*128/sizeof(*x));
prefetch(y + 3*128/sizeof(*y));

int n1 = n&-8;

while(i < n1)
{
#if 0
vdot0 = vfmaq_f64 (vdot0,
vy[0],
vx[0]);
vdot1 = vfmaq_f64 (vdot1,
vy[1],
vx[1]);
vdot2 = vfmaq_f64 (vdot2,
vy[2],
vx[2]);
vdot3 = vfmaq_f64 (vdot3,
vy[3],
vx[3]);
#else
vdot0 = vy[0] * vx[0] + vdot0;
vdot1 = vy[1] * vx[1] + vdot1;
vdot2 = vy[2] * vx[2] + vdot2;
vdot3 = vy[3] * vx[3] + vdot3;
#endif
vy += 4;
vx += 4;
i += 8;
prefetch(vx + 3*128/sizeof(*x));
prefetch(vy + 3*128/sizeof(*y));

}
dot = vaddvq_f64 (vdot0 + vdot1);
dot += vaddvq_f64 (vdot2 + vdot3);
i = n1;

while(i < n)
{
dot += y[i] * x[i] ;
i++ ;

}
return(dot);


}

while(i < n)
{
dot += y[iy] * x[ix] ;
ix += inc_x ;
iy += inc_y ;
i++ ;

}
return(dot);

}



+ 1817
- 0
kernel/arm64/dgemm_kernel_8x4_vulcan.S
File diff suppressed because it is too large
View File


+ 340
- 0
kernel/arm64/dgemm_ncopy_4.S View File

@@ -0,0 +1,340 @@
/***************************************************************************
Copyright (c) 2016, 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 A00 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"

#define M x0
#define N x1
#define A00 x2
#define LDA x3
#define B00 x4

#define A01 x5
#define A02 x6
#define A03 x7
#define A04 x8

#define I x9
#define J x10

#define TEMP1 x11
#define TEMP2 x12

#define A_PREFETCH 2560

/**************************************************************************************
* Macro definitions
**************************************************************************************/

.macro SAVE_REGS
add sp, sp, #-(11 * 16)
stp d8, d9, [sp, #(0 * 16)]
stp d10, d11, [sp, #(1 * 16)]
stp d12, d13, [sp, #(2 * 16)]
stp d14, d15, [sp, #(3 * 16)]
stp d16, d17, [sp, #(4 * 16)]
stp x18, x19, [sp, #(5 * 16)]
stp x20, x21, [sp, #(6 * 16)]
stp x22, x23, [sp, #(7 * 16)]
stp x24, x25, [sp, #(8 * 16)]
stp x26, x27, [sp, #(9 * 16)]
str x28, [sp, #(10 * 16)]
.endm

.macro RESTORE_REGS
ldp d8, d9, [sp, #(0 * 16)]
ldp d10, d11, [sp, #(1 * 16)]
ldp d12, d13, [sp, #(2 * 16)]
ldp d14, d15, [sp, #(3 * 16)]
ldp d16, d17, [sp, #(4 * 16)]
ldp x18, x19, [sp, #(5 * 16)]
ldp x20, x21, [sp, #(6 * 16)]
ldp x22, x23, [sp, #(7 * 16)]
ldp x24, x25, [sp, #(8 * 16)]
ldp x26, x27, [sp, #(9 * 16)]
ldr x28, [sp, #(10 * 16)]
add sp, sp, #(11*16)
.endm

.macro COPY4x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldp q0, q1, [A01], #32
ins v8.d[0], v0.d[0]
ins v10.d[0], v0.d[1]
ins v12.d[0], v1.d[0]
ins v14.d[0], v1.d[1]

ldp q2, q3, [A02], #32
ins v8.d[1], v2.d[0]
ins v10.d[1], v2.d[1]
ins v12.d[1], v3.d[0]
ins v14.d[1], v3.d[1]

ldp q4, q5, [A03], #32
ins v9.d[0], v4.d[0]
ins v11.d[0], v4.d[1]
ins v13.d[0], v5.d[0]
ins v15.d[0], v5.d[1]

ldp q6, q7, [A04], #32
ins v9.d[1], v6.d[0]
ins v11.d[1], v6.d[1]
ins v13.d[1], v7.d[0]
ins v15.d[1], v7.d[1]

st1 {v8.2d, v9.2d, v10.2d, v11.2d}, [B00]
add B00, B00, #64

st1 {v12.2d, v13.2d, v14.2d, v15.2d}, [B00]
add B00, B00, #64
.endm

.macro COPY1x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8
ldr d2, [A03], #8
ldr d3, [A04], #8

st1 {v0.1d, v1.1d, v2.1d, v3.1d}, [B00]
add B00, B00, #32
.endm

.macro COPY4x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldp q0, q1, [A01], #32
ins v8.d[0], v0.d[0]
ins v9.d[0], v0.d[1]
ins v10.d[0], v1.d[0]
ins v11.d[0], v1.d[1]

ldp q2, q3, [A02], #32
ins v8.d[1], v2.d[0]
ins v9.d[1], v2.d[1]
ins v10.d[1], v3.d[0]
ins v11.d[1], v3.d[1]

st1 {v8.2d, v9.2d, v10.2d, v11.2d}, [B00]
add B00, B00, #64
.endm


.macro COPY1x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8

stp d0, d1, [B00]
add B00, B00, #16
.endm

.macro COPY4x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldp q0, q1, [A01], #32
stp q0, q1, [B00], #32
.endm


.macro COPY1x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldr d0, [A01], #8
str d0, [B00], #8
.endm

/**************************************************************************************
* End of macro definitions
**************************************************************************************/

PROLOGUE

.align 5

SAVE_REGS

lsl LDA, LDA, #3 // LDA = LDA * SIZE

dgemm_ncopy_L4_BEGIN:

asr J, N, #2 // J = N / 4
cmp J, #0
ble dgemm_ncopy_L2_BEGIN

.align 5
dgemm_ncopy_L4_M4_BEGIN:

mov A01, A00
add A02, A01, LDA
add A03, A02, LDA
add A04, A03, LDA
add A00, A04, LDA

asr I, M, #2 // I = M / 4
cmp I, #0
ble dgemm_ncopy_L4_M4_40

.align 5
dgemm_ncopy_L4_M4_20:

COPY4x4

subs I , I , #1
bne dgemm_ncopy_L4_M4_20


dgemm_ncopy_L4_M4_40:

and I, M , #3
cmp I, #0
ble dgemm_ncopy_L4_M4_END

.align 5
dgemm_ncopy_L4_M4_60:

COPY1x4

subs I , I , #1
bne dgemm_ncopy_L4_M4_60


dgemm_ncopy_L4_M4_END:

subs J , J, #1 // j--
bne dgemm_ncopy_L4_M4_BEGIN



/*********************************************************************************************/

dgemm_ncopy_L2_BEGIN:

tst N, #3
ble dgemm_ncopy_L999

tst N, #2
ble dgemm_ncopy_L1_BEGIN

dgemm_ncopy_L2_M4_BEGIN:
mov A01, A00
add A02, A01, LDA
add A00, A02, LDA

asr I, M, #2 // I = M / 4
cmp I, #0
ble dgemm_ncopy_L2_M4_40

.align 5
dgemm_ncopy_L2_M4_20:

COPY4x2

subs I , I , #1
bne dgemm_ncopy_L2_M4_20


dgemm_ncopy_L2_M4_40:

and I, M , #3
cmp I, #0
ble dgemm_ncopy_L2_M4_END

.align 5
dgemm_ncopy_L2_M4_60:

COPY1x2

subs I , I , #1
bne dgemm_ncopy_L2_M4_60


dgemm_ncopy_L2_M4_END:


/*********************************************************************************************/

dgemm_ncopy_L1_BEGIN:

tst N, #1
ble dgemm_ncopy_L999


dgemm_ncopy_L1_M4_BEGIN:

mov A01, A00

asr I, M, #2 // I = M / 4
cmp I, #0
ble dgemm_ncopy_L1_M4_40

.align 5
dgemm_ncopy_L1_M4_20:

COPY4x1

subs I , I , #1
bne dgemm_ncopy_L1_M4_20


dgemm_ncopy_L1_M4_40:

and I, M , #3
cmp I, #0
ble dgemm_ncopy_L1_M4_END

.align 5
dgemm_ncopy_L1_M4_60:

COPY1x1

subs I , I , #1
bne dgemm_ncopy_L1_M4_60


dgemm_ncopy_L1_M4_END:

dgemm_ncopy_L999:

mov x0, #0
RESTORE_REGS
ret

EPILOGUE


+ 544
- 0
kernel/arm64/dgemm_ncopy_8.S View File

@@ -0,0 +1,544 @@
/***************************************************************************
Copyright (c) 2016, 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 A00 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"

#define M x0
#define N x1
#define A00 x2
#define LDA x3
#define B00 x4

#define A01 x5
#define A02 x6
#define A03 x7
#define A04 x8
#define A05 x9
#define A06 x10
#define A07 x11
#define A08 x12

#define I x13
#define J x14

#define TEMP1 x15
#define TEMP2 x16

#define A_PREFETCH 2560

/**************************************************************************************
* Macro definitions
**************************************************************************************/

.macro SAVE_REGS
add sp, sp, #-(11 * 16)
stp d8, d9, [sp, #(0 * 16)]
stp d10, d11, [sp, #(1 * 16)]
stp d12, d13, [sp, #(2 * 16)]
stp d14, d15, [sp, #(3 * 16)]
stp d16, d17, [sp, #(4 * 16)]
stp x18, x19, [sp, #(5 * 16)]
stp x20, x21, [sp, #(6 * 16)]
stp x22, x23, [sp, #(7 * 16)]
stp x24, x25, [sp, #(8 * 16)]
stp x26, x27, [sp, #(9 * 16)]
str x28, [sp, #(10 * 16)]
.endm

.macro RESTORE_REGS
ldp d8, d9, [sp, #(0 * 16)]
ldp d10, d11, [sp, #(1 * 16)]
ldp d12, d13, [sp, #(2 * 16)]
ldp d14, d15, [sp, #(3 * 16)]
ldp d16, d17, [sp, #(4 * 16)]
ldp x18, x19, [sp, #(5 * 16)]
ldp x20, x21, [sp, #(6 * 16)]
ldp x22, x23, [sp, #(7 * 16)]
ldp x24, x25, [sp, #(8 * 16)]
ldp x26, x27, [sp, #(9 * 16)]
ldr x28, [sp, #(10 * 16)]
add sp, sp, #(11*16)
.endm

/*************************************************************************************/

.macro COPY8x8
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]
//prfm PLDL1KEEP, [A05, #A_PREFETCH]
//prfm PLDL1KEEP, [A06, #A_PREFETCH]
//prfm PLDL1KEEP, [A07, #A_PREFETCH]
//prfm PLDL1KEEP, [A08, #A_PREFETCH]

COPY4x8
COPY4x8
.endm

.macro COPY4x8
ldp q0, q1, [A01], #32
ins v16.d[0], v0.d[0]
ins v20.d[0], v0.d[1]
ins v24.d[0], v1.d[0]
ins v28.d[0], v1.d[1]

ldp q2, q3, [A02], #32
ins v16.d[1], v2.d[0]
ins v20.d[1], v2.d[1]
ins v24.d[1], v3.d[0]
ins v28.d[1], v3.d[1]

ldp q4, q5, [A03], #32
ins v17.d[0], v4.d[0]
ins v21.d[0], v4.d[1]
ins v25.d[0], v5.d[0]
ins v29.d[0], v5.d[1]

ldp q6, q7, [A04], #32
ins v17.d[1], v6.d[0]
ins v21.d[1], v6.d[1]
ins v25.d[1], v7.d[0]
ins v29.d[1], v7.d[1]

ldp q8, q9, [A05], #32
ins v18.d[0], v8.d[0]
ins v22.d[0], v8.d[1]
ins v26.d[0], v9.d[0]
ins v30.d[0], v9.d[1]

ldp q10, q11, [A06], #32
ins v18.d[1], v10.d[0]
ins v22.d[1], v10.d[1]
ins v26.d[1], v11.d[0]
ins v30.d[1], v11.d[1]

ldp q12, q13, [A07], #32
ins v19.d[0], v12.d[0]
ins v23.d[0], v12.d[1]
ins v27.d[0], v13.d[0]
ins v31.d[0], v13.d[1]

ldp q14, q15, [A08], #32
ins v19.d[1], v14.d[0]
ins v23.d[1], v14.d[1]
ins v27.d[1], v15.d[0]
ins v31.d[1], v15.d[1]

st1 {v16.2d, v17.2d, v18.2d, v19.2d}, [B00]
add B00, B00, #64

st1 {v20.2d, v21.2d, v22.2d, v23.2d}, [B00]
add B00, B00, #64

st1 {v24.2d, v25.2d, v26.2d, v27.2d}, [B00]
add B00, B00, #64

st1 {v28.2d, v29.2d, v30.2d, v31.2d}, [B00]
add B00, B00, #64
.endm

.macro COPY1x8
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]
//prfm PLDL1KEEP, [A05, #A_PREFETCH]
//prfm PLDL1KEEP, [A06, #A_PREFETCH]
//prfm PLDL1KEEP, [A07, #A_PREFETCH]
//prfm PLDL1KEEP, [A08, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8
ldr d2, [A03], #8
ldr d3, [A04], #8
ldr d4, [A05], #8
ldr d5, [A06], #8
ldr d6, [A07], #8
ldr d7, [A08], #8

st1 {v0.1d, v1.1d, v2.1d, v3.1d}, [B00]
add B00, B00, #32
st1 {v4.1d, v5.1d, v6.1d, v7.1d}, [B00]
add B00, B00, #32

.endm


/*************************************************************************************/

.macro COPY8x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldp q0, q1, [A01], #32
ins v8.d[0], v0.d[0]
ins v10.d[0], v0.d[1]
ins v12.d[0], v1.d[0]
ins v14.d[0], v1.d[1]

ldp q2, q3, [A02], #32
ins v8.d[1], v2.d[0]
ins v10.d[1], v2.d[1]
ins v12.d[1], v3.d[0]
ins v14.d[1], v3.d[1]

ldp q4, q5, [A03], #32
ins v9.d[0], v4.d[0]
ins v11.d[0], v4.d[1]
ins v13.d[0], v5.d[0]
ins v15.d[0], v5.d[1]

ldp q6, q7, [A04], #32
ins v9.d[1], v6.d[0]
ins v11.d[1], v6.d[1]
ins v13.d[1], v7.d[0]
ins v15.d[1], v7.d[1]

st1 {v8.2d, v9.2d, v10.2d, v11.2d}, [B00]
add B00, B00, #64

st1 {v12.2d, v13.2d, v14.2d, v15.2d}, [B00]
add B00, B00, #64

ldp q16, q17, [A01], #32
ins v24.d[0], v16.d[0]
ins v26.d[0], v16.d[1]
ins v28.d[0], v17.d[0]
ins v30.d[0], v17.d[1]

ldp q18, q19, [A02], #32
ins v24.d[1], v18.d[0]
ins v26.d[1], v18.d[1]
ins v28.d[1], v19.d[0]
ins v30.d[1], v19.d[1]

ldp q20, q21, [A03], #32
ins v25.d[0], v20.d[0]
ins v27.d[0], v20.d[1]
ins v29.d[0], v21.d[0]
ins v31.d[0], v21.d[1]

ldp q22, q23, [A04], #32
ins v25.d[1], v22.d[0]
ins v27.d[1], v22.d[1]
ins v29.d[1], v23.d[0]
ins v31.d[1], v23.d[1]

st1 {v24.2d, v25.2d, v26.2d, v27.2d}, [B00]
add B00, B00, #64

st1 {v28.2d, v29.2d, v30.2d, v31.2d}, [B00]
add B00, B00, #64
.endm

.macro COPY1x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8
ldr d2, [A03], #8
ldr d3, [A04], #8

st1 {v0.1d, v1.1d, v2.1d, v3.1d}, [B00]
add B00, B00, #32
.endm

/*************************************************************************************/

.macro COPY8x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A01], #32

ins v8.d[0], v0.d[0]
ins v9.d[0], v0.d[1]
ins v10.d[0], v1.d[0]
ins v11.d[0], v1.d[1]
ins v12.d[0], v2.d[0]
ins v13.d[0], v2.d[1]
ins v14.d[0], v3.d[0]
ins v15.d[0], v3.d[1]

ldp q4, q5, [A02], #32
ldp q6, q7, [A02], #32

ins v8.d[1], v4.d[0]
ins v9.d[1], v4.d[1]
ins v10.d[1], v5.d[0]
ins v11.d[1], v5.d[1]
ins v12.d[1], v6.d[0]
ins v13.d[1], v6.d[1]
ins v14.d[1], v7.d[0]
ins v15.d[1], v7.d[1]

st1 {v8.2d, v9.2d, v10.2d, v11.2d}, [B00]
add B00, B00, #64
st1 {v12.2d, v13.2d, v14.2d, v15.2d}, [B00]
add B00, B00, #64
.endm


.macro COPY1x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8

stp d0, d1, [B00]
add B00, B00, #16
.endm

/*************************************************************************************/

.macro COPY8x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A01], #32
stp q0, q1, [B00], #32
stp q2, q3, [B00], #32
.endm


.macro COPY1x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldr d0, [A01], #8
str d0, [B00], #8
.endm

/**************************************************************************************
* End of macro definitions
**************************************************************************************/

PROLOGUE

.align 5

SAVE_REGS

lsl LDA, LDA, #3 // LDA = LDA * SIZE

dgemm_ncopy_L8_BEGIN:

asr J, N, #3 // J = N / 8
cmp J, #0
ble dgemm_ncopy_L4_BEGIN

dgemm_ncopy_L8_M8_BEGIN:

mov A01, A00
add A02, A01, LDA
add A03, A02, LDA
add A04, A03, LDA
add A05, A04, LDA
add A06, A05, LDA
add A07, A06, LDA
add A08, A07, LDA
add A00, A08, LDA


asr I, M, #3 // I = M / 8
cmp I, #0
ble dgemm_ncopy_L8_M8_40

dgemm_ncopy_L8_M8_20:

COPY8x8

subs I , I , #1
bne dgemm_ncopy_L8_M8_20


dgemm_ncopy_L8_M8_40:

and I, M , #7
cmp I, #0
ble dgemm_ncopy_L8_M8_END

dgemm_ncopy_L8_M8_60:

COPY1x8

subs I , I , #1
bne dgemm_ncopy_L8_M8_60


dgemm_ncopy_L8_M8_END:

subs J , J, #1 // j--
bne dgemm_ncopy_L8_M8_BEGIN

/*********************************************************************************************/

dgemm_ncopy_L4_BEGIN:

tst N, #7
ble dgemm_ncopy_L999

tst N, #4
ble dgemm_ncopy_L2_BEGIN

dgemm_ncopy_L4_M8_BEGIN:

mov A01, A00
add A02, A01, LDA
add A03, A02, LDA
add A04, A03, LDA
add A00, A04, LDA

asr I, M, #3 // I = M / 8
cmp I, #0
ble dgemm_ncopy_L4_M8_40

dgemm_ncopy_L4_M8_20:

COPY8x4

subs I , I , #1
bne dgemm_ncopy_L4_M8_20


dgemm_ncopy_L4_M8_40:

and I, M , #7
cmp I, #0
ble dgemm_ncopy_L4_M8_END

dgemm_ncopy_L4_M8_60:

COPY1x4

subs I , I , #1
bne dgemm_ncopy_L4_M8_60


dgemm_ncopy_L4_M8_END:


/*********************************************************************************************/

dgemm_ncopy_L2_BEGIN:

tst N, #3
ble dgemm_ncopy_L999

tst N, #2
ble dgemm_ncopy_L1_BEGIN

dgemm_ncopy_L2_M8_BEGIN:
mov A01, A00
add A02, A01, LDA
add A00, A02, LDA

asr I, M, #3 // I = M / 8
cmp I, #0
ble dgemm_ncopy_L2_M8_40

dgemm_ncopy_L2_M8_20:

COPY8x2

subs I , I , #1
bne dgemm_ncopy_L2_M8_20


dgemm_ncopy_L2_M8_40:

and I, M , #7
cmp I, #0
ble dgemm_ncopy_L2_M8_END

dgemm_ncopy_L2_M8_60:

COPY1x2

subs I , I , #1
bne dgemm_ncopy_L2_M8_60


dgemm_ncopy_L2_M8_END:


/*********************************************************************************************/

dgemm_ncopy_L1_BEGIN:

tst N, #1
ble dgemm_ncopy_L999


dgemm_ncopy_L1_M8_BEGIN:

mov A01, A00

asr I, M, #3 // I = M / 8
cmp I, #0
ble dgemm_ncopy_L1_M8_40

dgemm_ncopy_L1_M8_20:

COPY8x1

subs I , I , #1
bne dgemm_ncopy_L1_M8_20


dgemm_ncopy_L1_M8_40:

and I, M , #7
cmp I, #0
ble dgemm_ncopy_L1_M8_END

dgemm_ncopy_L1_M8_60:

COPY1x1

subs I , I , #1
bne dgemm_ncopy_L1_M8_60


dgemm_ncopy_L1_M8_END:

dgemm_ncopy_L999:

mov x0, #0
RESTORE_REGS
ret

EPILOGUE


+ 402
- 0
kernel/arm64/dgemm_tcopy_4.S View File

@@ -0,0 +1,402 @@
/***************************************************************************
Copyright (c) 2016, 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"

#define M x0
#define N x1
#define A x2
#define LDA x3
#define B x4

#define M4 x5

#define A01 x6
#define A02 x7
#define A03 x8
#define A04 x9

#define B01 x10
#define B02 x11
#define B03 x12
#define B04 x13

#define I x14
#define J x15

#define TEMP1 x16
#define TEMP2 x17

#define A_PREFETCH 2560
#define B_PREFETCH 256

/**************************************************************************************
* Macro definitions
**************************************************************************************/
.macro SAVE_REGS
add sp, sp, #-(11 * 16)
stp d8, d9, [sp, #(0 * 16)]
stp d10, d11, [sp, #(1 * 16)]
stp d12, d13, [sp, #(2 * 16)]
stp d14, d15, [sp, #(3 * 16)]
stp d16, d17, [sp, #(4 * 16)]
stp x18, x19, [sp, #(5 * 16)]
stp x20, x21, [sp, #(6 * 16)]
stp x22, x23, [sp, #(7 * 16)]
stp x24, x25, [sp, #(8 * 16)]
stp x26, x27, [sp, #(9 * 16)]
str x28, [sp, #(10 * 16)]
.endm

.macro RESTORE_REGS
ldp d8, d9, [sp, #(0 * 16)]
ldp d10, d11, [sp, #(1 * 16)]
ldp d12, d13, [sp, #(2 * 16)]
ldp d14, d15, [sp, #(3 * 16)]
ldp d16, d17, [sp, #(4 * 16)]
ldp x18, x19, [sp, #(5 * 16)]
ldp x20, x21, [sp, #(6 * 16)]
ldp x22, x23, [sp, #(7 * 16)]
ldp x24, x25, [sp, #(8 * 16)]
ldp x26, x27, [sp, #(9 * 16)]
ldr x28, [sp, #(10 * 16)]
add sp, sp, #(11*16)
.endm

.macro COPY4x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A02], #32

////prfm PLDL1KEEP, [B01, #B_PREFETCH]
st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B01]
add TEMP1, B01, #64

ldp q4, q5, [A03], #32
ldp q6, q7, [A04], #32

////prfm PLDL1KEEP, [B01, #B_PREFETCH]
st1 {v4.2d, v5.2d, v6.2d, v7.2d}, [TEMP1]

add B01, B01, M4
.endm

.macro COPY2x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldr q0, [A01], #16
ldr q1, [A02], #16
ldr q2, [A03], #16
ldr q3, [A04], #16

////prfm PLDL1KEEP, [B02, #B_PREFETCH]
st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B02]

add B02, B02, #64
.endm

.macro COPY1x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8
ldr d2, [A03], #8
ldr d3, [A04], #8

////prfm PLDL1KEEP, [B03, #B_PREFETCH]
st1 {v0.1d, v1.1d, v2.1d, v3.1d}, [B03]

add B03, B03, #32
.endm

/*************************************************************************************************************************/

.macro COPY4x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A02], #32

////prfm PLDL1KEEP, [B01, #B_PREFETCH]

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B01]
add B01, B01, M4
.endm

.macro COPY2x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldr q0, [A01], #16
ldr q1, [A02], #16

////prfm PLDL1KEEP, [B02, #B_PREFETCH]
stp q0, q1, [B02]

add B02, B02, #32
.endm

.macro COPY1x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8

////prfm PLDL1KEEP, [B03, #B_PREFETCH]
stp d0, d1, [B03]

add B03, B03, #16
.endm

/*************************************************************************************************************************/

.macro COPY4x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldp q0, q1, [A01], #32

////prfm PLDL1KEEP, [B01, #B_PREFETCH]
stp q0, q1, [B01]

add B01, B01, M4
.endm

.macro COPY2x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldr q0, [A01], #16

////prfm PLDL1KEEP, [B02, #B_PREFETCH]
str q0, [B02]

add B02, B02, #16
.endm

.macro COPY1x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldr d0, [A01], #8

////prfm PLDL1KEEP, [B03, #B_PREFETCH]
str d0, [B03]

add B03, B03, #8
.endm



/**************************************************************************************
* End of macro definitions
**************************************************************************************/

PROLOGUE

.align 5

SAVE_REGS

lsl LDA, LDA, #3 // LDA = LDA * SIZE

lsl TEMP1, M, #3 // x12 = M * SIZE

and B02 , N , #-4
and B03 , N , #-2

mul B02, B02, TEMP1
mul B03, B03, TEMP1

add B02 , B02, B
add B03 , B03, B

lsl M4, M, #5 // M4 = M * 4 * SIZE

dgemm_tcopy_L4_BEGIN:
asr J, M, #2 // J = M / 4
cmp J, #0
ble dgemm_tcopy_L2_BEGIN

.align 5
dgemm_tcopy_L4_M4_BEGIN:

mov A01, A
add A02, A01, LDA
add A03, A02, LDA
add A04, A03, LDA
add A, A04, LDA

mov B01, B
add B, B01, #128 // B = B + 16 * SIZE

asr I, N, #2 // I = N / 4
cmp I, #0
ble dgemm_tcopy_L4_M4_40

.align 5
dgemm_tcopy_L4_M4_20:

COPY4x4

subs I , I , #1
bne dgemm_tcopy_L4_M4_20


dgemm_tcopy_L4_M4_40:

tst N , #2
ble dgemm_tcopy_L4_M4_60

COPY2x4


dgemm_tcopy_L4_M4_60:

tst N, #1
ble dgemm_tcopy_L4_M4_END

COPY1x4


dgemm_tcopy_L4_M4_END:

subs J , J, #1 // j--
bne dgemm_tcopy_L4_M4_BEGIN



/*********************************************************************************************/

dgemm_tcopy_L2_BEGIN:

tst M, #3
ble dgemm_tcopy_L999

tst M, #2
ble dgemm_tcopy_L1_BEGIN

dgemm_tcopy_L2_M4_BEGIN:
mov A01, A
add A02, A01, LDA
add A, A02, LDA

mov B01, B
add B, B01, #64 // B = B + 8 * SIZE

asr I, N, #2 // I = N / 4
cmp I, #0
ble dgemm_tcopy_L2_M4_40

.align 5
dgemm_tcopy_L2_M4_20:

COPY4x2

subs I , I , #1
bne dgemm_tcopy_L2_M4_20


dgemm_tcopy_L2_M4_40:

tst N , #2
ble dgemm_tcopy_L2_M4_60

COPY2x2

dgemm_tcopy_L2_M4_60:

tst N , #1
ble dgemm_tcopy_L2_M4_END

COPY1x2


dgemm_tcopy_L2_M4_END:


/*********************************************************************************************/

dgemm_tcopy_L1_BEGIN:

tst M, #1
ble dgemm_tcopy_L999


dgemm_tcopy_L1_M4_BEGIN:

mov A01, A // A01 = A
mov B01, B

asr I, N, #2 // I = M / 4
cmp I, #0
ble dgemm_tcopy_L1_M4_40

.align 5
dgemm_tcopy_L1_M4_20:

COPY4x1

subs I , I , #1
bne dgemm_tcopy_L1_M4_20


dgemm_tcopy_L1_M4_40:

tst N , #2
ble dgemm_tcopy_L1_M4_60

COPY2x1

dgemm_tcopy_L1_M4_60:

tst N , #1
ble dgemm_tcopy_L1_M4_END

COPY1x1


dgemm_tcopy_L1_M4_END:


dgemm_tcopy_L999:
mov x0, #0 // set return value
RESTORE_REGS
ret

EPILOGUE


+ 682
- 0
kernel/arm64/dgemm_tcopy_8.S View File

@@ -0,0 +1,682 @@
/***************************************************************************
Copyright (c) 2016, 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"

#define M x0
#define N x1
#define A x2
#define LDA x3
#define B x4

#define M8 x5

#define A01 x6
#define A02 x7
#define A03 x8
#define A04 x9
#define A05 x10
#define A06 x11
#define A07 x12
#define A08 x13

#define B01 x14
#define B02 x15
#define B03 x16
#define B04 x17

#define I x18
#define J x19

#define TEMP1 x20
#define TEMP2 x21

#define A_PREFETCH 2560
#define B_PREFETCH 256

/**************************************************************************************
* Macro definitions
**************************************************************************************/
.macro SAVE_REGS
add sp, sp, #-(11 * 16)
stp d8, d9, [sp, #(0 * 16)]
stp d10, d11, [sp, #(1 * 16)]
stp d12, d13, [sp, #(2 * 16)]
stp d14, d15, [sp, #(3 * 16)]
stp d16, d17, [sp, #(4 * 16)]
stp x18, x19, [sp, #(5 * 16)]
stp x20, x21, [sp, #(6 * 16)]
stp x22, x23, [sp, #(7 * 16)]
stp x24, x25, [sp, #(8 * 16)]
stp x26, x27, [sp, #(9 * 16)]
str x28, [sp, #(10 * 16)]
.endm

.macro RESTORE_REGS
ldp d8, d9, [sp, #(0 * 16)]
ldp d10, d11, [sp, #(1 * 16)]
ldp d12, d13, [sp, #(2 * 16)]
ldp d14, d15, [sp, #(3 * 16)]
ldp d16, d17, [sp, #(4 * 16)]
ldp x18, x19, [sp, #(5 * 16)]
ldp x20, x21, [sp, #(6 * 16)]
ldp x22, x23, [sp, #(7 * 16)]
ldp x24, x25, [sp, #(8 * 16)]
ldp x26, x27, [sp, #(9 * 16)]
ldr x28, [sp, #(10 * 16)]
add sp, sp, #(11*16)
.endm

/*************************************************************************************************************************/

.macro COPY8x8
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]
//prfm PLDL1KEEP, [A05, #A_PREFETCH]
//prfm PLDL1KEEP, [A06, #A_PREFETCH]
//prfm PLDL1KEEP, [A07, #A_PREFETCH]
//prfm PLDL1KEEP, [A08, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A01], #32

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B01]
add TEMP1, B01, #64

ldp q4, q5, [A02], #32
ldp q6, q7, [A02], #32

st1 {v4.2d, v5.2d, v6.2d, v7.2d}, [TEMP1]
add TEMP1, TEMP1, #64

ldp q8, q9, [A03], #32
ldp q10, q11, [A03], #32

st1 {v8.2d, v9.2d, v10.2d, v11.2d}, [TEMP1]
add TEMP1, TEMP1, #64

ldp q12, q13, [A04], #32
ldp q14, q15, [A04], #32

st1 {v12.2d, v13.2d, v14.2d, v15.2d}, [TEMP1]
add TEMP1, TEMP1, #64

ldp q16, q17, [A05], #32
ldp q18, q19, [A05], #32

st1 {v16.2d, v17.2d, v18.2d, v19.2d}, [TEMP1]
add TEMP1, TEMP1, #64

ldp q20, q21, [A06], #32
ldp q22, q23, [A06], #32

st1 {v20.2d, v21.2d, v22.2d, v23.2d}, [TEMP1]
add TEMP1, TEMP1, #64

ldp q24, q25, [A07], #32
ldp q26, q27, [A07], #32

st1 {v24.2d, v25.2d, v26.2d, v27.2d}, [TEMP1]
add TEMP1, TEMP1, #64

ldp q28, q29, [A08], #32
ldp q30, q31, [A08], #32

st1 {v28.2d, v29.2d, v30.2d, v31.2d}, [TEMP1]
add TEMP1, TEMP1, #64

add B01, B01, M8
.endm

.macro COPY4x8
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]
//prfm PLDL1KEEP, [A05, #A_PREFETCH]
//prfm PLDL1KEEP, [A06, #A_PREFETCH]
//prfm PLDL1KEEP, [A07, #A_PREFETCH]
//prfm PLDL1KEEP, [A08, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A02], #32

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B02]
add B02, B02, #64

ldp q4, q5, [A03], #32
ldp q6, q7, [A04], #32

st1 {v4.2d, v5.2d, v6.2d, v7.2d}, [B02]
add B02, B02, #64

ldp q8, q9, [A05], #32
ldp q10, q11, [A06], #32

st1 {v8.2d, v9.2d, v10.2d, v11.2d}, [B02]
add B02, B02, #64

ldp q12, q13, [A07], #32
ldp q14, q15, [A08], #32

st1 {v12.2d, v13.2d, v14.2d, v15.2d}, [B02]
add B02, B02, #64
.endm

.macro COPY2x8
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]
//prfm PLDL1KEEP, [A05, #A_PREFETCH]
//prfm PLDL1KEEP, [A06, #A_PREFETCH]
//prfm PLDL1KEEP, [A07, #A_PREFETCH]
//prfm PLDL1KEEP, [A08, #A_PREFETCH]

ldr q0, [A01], #16
ldr q1, [A02], #16
ldr q2, [A03], #16
ldr q3, [A04], #16

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B03]
add B03, B03, #64

ldr q4, [A05], #16
ldr q5, [A06], #16
ldr q6, [A07], #16
ldr q7, [A08], #16

st1 {v4.2d, v5.2d, v6.2d, v7.2d}, [B03]
add B03, B03, #64
.endm

.macro COPY1x8
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]
//prfm PLDL1KEEP, [A05, #A_PREFETCH]
//prfm PLDL1KEEP, [A06, #A_PREFETCH]
//prfm PLDL1KEEP, [A07, #A_PREFETCH]
//prfm PLDL1KEEP, [A08, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8
ldr d2, [A03], #8
ldr d3, [A04], #8

st1 {v0.1d, v1.1d, v2.1d, v3.1d}, [B04]
add B04, B04, #32

ldr d4, [A05], #8
ldr d5, [A06], #8
ldr d6, [A07], #8
ldr d7, [A08], #8

st1 {v4.1d, v5.1d, v6.1d, v7.1d}, [B04]

add B04, B04, #32
.endm

/*************************************************************************************************************************/

.macro COPY8x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A01], #32

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B01]
add TEMP1, B01, #64

ldp q4, q5, [A02], #32
ldp q6, q7, [A02], #32

st1 {v4.2d, v5.2d, v6.2d, v7.2d}, [TEMP1]
add TEMP1, TEMP1, #64

ldp q8, q9, [A03], #32
ldp q10, q11, [A03], #32

st1 {v8.2d, v9.2d, v10.2d, v11.2d}, [TEMP1]
add TEMP1, TEMP1, #64

ldp q12, q13, [A04], #32
ldp q14, q15, [A04], #32

st1 {v12.2d, v13.2d, v14.2d, v15.2d}, [TEMP1]
add TEMP1, TEMP1, #64

add B01, B01, M8
.endm

.macro COPY4x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A02], #32

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B02]
add B02, B02, #64

ldp q4, q5, [A03], #32
ldp q6, q7, [A04], #32

st1 {v4.2d, v5.2d, v6.2d, v7.2d}, [B02]
add B02, B02, #64
.endm

.macro COPY2x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldr q0, [A01], #16
ldr q1, [A02], #16
ldr q2, [A03], #16
ldr q3, [A04], #16

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B03]

add B03, B03, #64
.endm

.macro COPY1x4
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]
//prfm PLDL1KEEP, [A03, #A_PREFETCH]
//prfm PLDL1KEEP, [A04, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8
ldr d2, [A03], #8
ldr d3, [A04], #8

st1 {v0.1d, v1.1d, v2.1d, v3.1d}, [B04]

add B04, B04, #32
.endm

/*************************************************************************************************************************/

.macro COPY8x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A01], #32
ldp q4, q5, [A02], #32
ldp q6, q7, [A02], #32

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B01]
add TEMP1, B01, #64
st1 {v4.2d, v5.2d, v6.2d, v7.2d}, [TEMP1]
add B01, B01, M8
.endm

.macro COPY4x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A02], #32

st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [B02]
add B02, B02, #64
.endm

.macro COPY2x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldr q0, [A01], #16
ldr q1, [A02], #16

stp q0, q1, [B03]

add B03, B03, #32
.endm

.macro COPY1x2
//prfm PLDL1KEEP, [A01, #A_PREFETCH]
//prfm PLDL1KEEP, [A02, #A_PREFETCH]

ldr d0, [A01], #8
ldr d1, [A02], #8

stp d0, d1, [B04]

add B04, B04, #16
.endm

/*************************************************************************************************************************/

.macro COPY8x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldp q0, q1, [A01], #32
ldp q2, q3, [A01], #32

stp q0, q1, [B01]
add TEMP1, B01, #32
stp q2, q3, [TEMP1]
add B01, B01, M8
.endm

.macro COPY4x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldp q0, q1, [A01], #32
stp q0, q1, [B02]

add B02, B02, #32
.endm

.macro COPY2x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldr q0, [A01], #16
str q0, [B03]

add B03, B03, #16
.endm

.macro COPY1x1
//prfm PLDL1KEEP, [A01, #A_PREFETCH]

ldr d0, [A01], #8
str d0, [B04]

add B04, B04, #8
.endm



/**************************************************************************************
* End of macro definitions
**************************************************************************************/

PROLOGUE

.align 5

SAVE_REGS

lsl LDA, LDA, #3 // LDA = LDA * SIZE

lsl TEMP1, M, #3 // TEMP1 = M * SIZE

and B02 , N , #-8
and B03 , N , #-4
and B04 , N , #-2

mul B02, B02, TEMP1
mul B03, B03, TEMP1
mul B04, B04, TEMP1

add B02 , B02, B
add B03 , B03, B
add B04 , B04, B

lsl M8, M, #6 // M8 = M * 8 * SIZE

dgemm_tcopy_L8_BEGIN:
asr J, M, #3 // J = M / 4
cmp J, #0
ble dgemm_tcopy_L4_BEGIN

.align 5
dgemm_tcopy_L8_M8_BEGIN:

mov A01, A
add A02, A01, LDA
add A03, A02, LDA
add A04, A03, LDA
add A05, A04, LDA
add A06, A05, LDA
add A07, A06, LDA
add A08, A07, LDA
add A, A08, LDA

mov B01, B
add B, B01, #512 // B = B + 64 * SIZE

asr I, N, #3 // I = N / 8
cmp I, #0
ble dgemm_tcopy_L8_M8_40

.align 5
dgemm_tcopy_L8_M8_20:

COPY8x8

subs I , I , #1
bne dgemm_tcopy_L8_M8_20

dgemm_tcopy_L8_M8_40:
tst N , #4
ble dgemm_tcopy_L8_M8_60

COPY4x8

dgemm_tcopy_L8_M8_60:

tst N , #2
ble dgemm_tcopy_L8_M8_80

COPY2x8


dgemm_tcopy_L8_M8_80:

tst N, #1
ble dgemm_tcopy_L8_M8_END

COPY1x8


dgemm_tcopy_L8_M8_END:

subs J , J, #1 // j--
bne dgemm_tcopy_L8_M8_BEGIN

/*********************************************************************************************/

dgemm_tcopy_L4_BEGIN:
tst M, #7
ble dgemm_tcopy_L999

tst M, #4
ble dgemm_tcopy_L2_BEGIN

dgemm_tcopy_L4_M8_BEGIN:

mov A01, A
add A02, A01, LDA
add A03, A02, LDA
add A04, A03, LDA
add A, A04, LDA

mov B01, B
add B, B01, #256 // B = B + 32 * SIZE

asr I, N, #3 // I = N / 8
cmp I, #0
ble dgemm_tcopy_L4_M8_40

.align 5
dgemm_tcopy_L4_M8_20:

COPY8x4

subs I , I , #1
bne dgemm_tcopy_L4_M8_20

dgemm_tcopy_L4_M8_40:
tst N , #4
ble dgemm_tcopy_L4_M8_60

COPY4x4

dgemm_tcopy_L4_M8_60:

tst N , #2
ble dgemm_tcopy_L4_M8_80

COPY2x4


dgemm_tcopy_L4_M8_80:

tst N, #1
ble dgemm_tcopy_L4_M8_END

COPY1x4


dgemm_tcopy_L4_M8_END:

/*********************************************************************************************/

dgemm_tcopy_L2_BEGIN:

tst M, #3
ble dgemm_tcopy_L999

tst M, #2
ble dgemm_tcopy_L1_BEGIN

dgemm_tcopy_L2_M8_BEGIN:
mov A01, A
add A02, A01, LDA
add A, A02, LDA

mov B01, B
add B, B01, #128 // B = B + 16 * SIZE

asr I, N, #3 // I = N / 8
cmp I, #0
ble dgemm_tcopy_L2_M8_40

.align 5
dgemm_tcopy_L2_M8_20:

COPY8x2

subs I , I , #1
bne dgemm_tcopy_L2_M8_20

dgemm_tcopy_L2_M8_40:
tst N , #4
ble dgemm_tcopy_L2_M8_60

COPY4x2

dgemm_tcopy_L2_M8_60:

tst N , #2
ble dgemm_tcopy_L2_M8_80

COPY2x2

dgemm_tcopy_L2_M8_80:

tst N , #1
ble dgemm_tcopy_L2_M8_END

COPY1x2


dgemm_tcopy_L2_M8_END:


/*********************************************************************************************/

dgemm_tcopy_L1_BEGIN:

tst M, #1
ble dgemm_tcopy_L999


dgemm_tcopy_L1_M8_BEGIN:

mov A01, A // A01 = A
mov B01, B

asr I, N, #3 // I = M / 8
cmp I, #0
ble dgemm_tcopy_L1_M8_40

.align 5
dgemm_tcopy_L1_M8_20:

COPY8x1

subs I , I , #1
bne dgemm_tcopy_L1_M8_20

dgemm_tcopy_L1_M8_40:
tst N , #4
ble dgemm_tcopy_L1_M8_60

COPY4x1

dgemm_tcopy_L1_M8_60:

tst N , #2
ble dgemm_tcopy_L1_M8_80

COPY2x1

dgemm_tcopy_L1_M8_80:

tst N , #1
ble dgemm_tcopy_L1_M8_END

COPY1x1


dgemm_tcopy_L1_M8_END:


dgemm_tcopy_L999:
mov x0, #0 // set return value
RESTORE_REGS
ret

EPILOGUE


+ 104
- 0
kernel/arm64/dot-thunderx.c View File

@@ -0,0 +1,104 @@
/***************************************************************************
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"

#if defined(DSDOT)
double CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
#else
FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
#endif
{
BLASLONG i=0;
BLASLONG ix=0,iy=0;

#if defined(DSDOT)
double dot = 0.0 ;
#else
FLOAT dot = 0.0 ;
#endif

if ( n < 0 ) return(dot);

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

int n1 = n & -4;

while(i < n1)
{

#if defined(DSDOT)
dot += (double) y[i] * (double) x[i]
+ (double) y[i+1] * (double) x[i+1]
+ (double) y[i+2] * (double) x[i+2]
+ (double) y[i+3] * (double) x[i+3] ;
#else
dot += y[i] * x[i]
+ y[i+1] * x[i+1]
+ y[i+2] * x[i+2]
+ y[i+3] * x[i+3] ;
#endif
i+=4 ;

}

while(i < n)
{

#if defined(DSDOT)
dot += (double) y[i] * (double) x[i] ;
#else
dot += y[i] * x[i] ;
#endif
i++ ;

}
return(dot);


}

while(i < n)
{

#if defined(DSDOT)
dot += (double) y[iy] * (double) x[ix] ;
#else
dot += y[iy] * x[ix] ;
#endif
ix += inc_x ;
iy += inc_y ;
i++ ;

}
return(dot);

}



+ 114
- 0
param.h View File

@@ -2303,6 +2303,44 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ZGEMM_DEFAULT_R 4096 #define ZGEMM_DEFAULT_R 4096




#define SYMV_P 16
#endif

#if defined(VULCAN)
#define SNUMOPT 2
#define DNUMOPT 2

#define GEMM_DEFAULT_OFFSET_A 0
#define GEMM_DEFAULT_OFFSET_B 0
#define GEMM_DEFAULT_ALIGN 0x03fffUL

#define SGEMM_DEFAULT_UNROLL_M 16
#define SGEMM_DEFAULT_UNROLL_N 4

#define DGEMM_DEFAULT_UNROLL_M 8
#define DGEMM_DEFAULT_UNROLL_N 4

#define CGEMM_DEFAULT_UNROLL_M 8
#define CGEMM_DEFAULT_UNROLL_N 4

#define ZGEMM_DEFAULT_UNROLL_M 4
#define ZGEMM_DEFAULT_UNROLL_N 4

#define SGEMM_DEFAULT_P 512
#define DGEMM_DEFAULT_P dgemm_p
#define CGEMM_DEFAULT_P 256
#define ZGEMM_DEFAULT_P 128

#define SGEMM_DEFAULT_Q 1024
#define DGEMM_DEFAULT_Q dgemm_q
#define CGEMM_DEFAULT_Q 512
#define ZGEMM_DEFAULT_Q 512

#define SGEMM_DEFAULT_R 4096
#define DGEMM_DEFAULT_R dgemm_r
#define CGEMM_DEFAULT_R 4096
#define ZGEMM_DEFAULT_R 2048

#define SYMV_P 16 #define SYMV_P 16
#endif #endif


@@ -2385,6 +2423,82 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define SYMV_P 16 #define SYMV_P 16
#endif #endif


#if defined(THUNDERX)
#define SNUMOPT 2
#define DNUMOPT 2

#define GEMM_DEFAULT_OFFSET_A 0
#define GEMM_DEFAULT_OFFSET_B 0
#define GEMM_DEFAULT_ALIGN 0x03fffUL

#define SGEMM_DEFAULT_UNROLL_M 4
#define SGEMM_DEFAULT_UNROLL_N 4

#define DGEMM_DEFAULT_UNROLL_M 2
#define DGEMM_DEFAULT_UNROLL_N 2

#define CGEMM_DEFAULT_UNROLL_M 2
#define CGEMM_DEFAULT_UNROLL_N 2

#define ZGEMM_DEFAULT_UNROLL_M 2
#define ZGEMM_DEFAULT_UNROLL_N 2

#define SGEMM_DEFAULT_P 128
#define DGEMM_DEFAULT_P 128
#define CGEMM_DEFAULT_P 96
#define ZGEMM_DEFAULT_P 64

#define SGEMM_DEFAULT_Q 240
#define DGEMM_DEFAULT_Q 120
#define CGEMM_DEFAULT_Q 120
#define ZGEMM_DEFAULT_Q 120

#define SGEMM_DEFAULT_R 12288
#define DGEMM_DEFAULT_R 8192
#define CGEMM_DEFAULT_R 4096
#define ZGEMM_DEFAULT_R 4096


#define SYMV_P 16
#endif

#if defined(THUNDERX2T99)
#define SNUMOPT 2
#define DNUMOPT 2

#define GEMM_DEFAULT_OFFSET_A 0
#define GEMM_DEFAULT_OFFSET_B 0
#define GEMM_DEFAULT_ALIGN 0x03fffUL

#define SGEMM_DEFAULT_UNROLL_M 16
#define SGEMM_DEFAULT_UNROLL_N 4

#define DGEMM_DEFAULT_UNROLL_M 8
#define DGEMM_DEFAULT_UNROLL_N 4

#define CGEMM_DEFAULT_UNROLL_M 8
#define CGEMM_DEFAULT_UNROLL_N 4

#define ZGEMM_DEFAULT_UNROLL_M 4
#define ZGEMM_DEFAULT_UNROLL_N 4

#define SGEMM_DEFAULT_P 512
#define DGEMM_DEFAULT_P dgemm_p
#define CGEMM_DEFAULT_P 256
#define ZGEMM_DEFAULT_P 128

#define SGEMM_DEFAULT_Q 1024
#define DGEMM_DEFAULT_Q dgemm_q
#define CGEMM_DEFAULT_Q 512
#define ZGEMM_DEFAULT_Q 512

#define SGEMM_DEFAULT_R 4096
#define DGEMM_DEFAULT_R dgemm_r
#define CGEMM_DEFAULT_R 4096
#define ZGEMM_DEFAULT_R 2048

#define SYMV_P 16
#endif


#if defined(ARMV5) #if defined(ARMV5)
#define SNUMOPT 2 #define SNUMOPT 2


Write Preview
Loading…
Cancel
Save