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

Remove redundant files

tags/v0.3.27
Heller Zheng 3 years ago
parent
bef47917bd
commit
5d0d1c5551
10 changed files with 2 additions and 2543 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +0
    -5
      Makefile.install
  2. +2
    -34
      kernel/riscv64/KERNEL.x280
  3. +0
    -92
      kernel/riscv64/gemm_ncopy_2_rvv.c
  4. +0
    -123
      kernel/riscv64/gemm_ncopy_4_rvv.c
  5. +0
    -108
      kernel/riscv64/gemm_tcopy_2_rvv.c
  6. +0
    -236
      kernel/riscv64/gemm_tcopy_4_rvv.c
  7. +0
    -214
      kernel/riscv64/gemmkernel_2x2_rvv.c
  8. +0
    -508
      kernel/riscv64/gemmkernel_4x4_rvv.c
  9. +0
    -342
      kernel/riscv64/trmmkernel_2x2_rvv.c
  10. +0
    -881
      kernel/riscv64/trmmkernel_4x4_rvv.c

+ 0
- 5
Makefile.install View File

@@ -8,7 +8,6 @@ PREFIX ?= /opt/OpenBLAS
OPENBLAS_INCLUDE_DIR := $(PREFIX)/include
OPENBLAS_LIBRARY_DIR := $(PREFIX)/lib
OPENBLAS_BINARY_DIR := $(PREFIX)/bin
OPENBLAS_RELEASE_DIR := $(PREFIX)/release
OPENBLAS_BUILD_DIR := $(CURDIR)
OPENBLAS_CMAKE_DIR := $(OPENBLAS_LIBRARY_DIR)/cmake/$(LIBSONAMEBASE)
OPENBLAS_CMAKE_CONFIG := OpenBLASConfig.cmake
@@ -39,7 +38,6 @@ install : lib.grd
@-mkdir -p "$(DESTDIR)$(OPENBLAS_INCLUDE_DIR)"
@-mkdir -p "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
@-mkdir -p "$(DESTDIR)$(OPENBLAS_BINARY_DIR)"
@-mkdir -p "$(DESTDIR)$(OPENBLAS_RELEASE_DIR)"
@-mkdir -p "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)"
@-mkdir -p "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)"
@echo Generating openblas_config.h in $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)
@@ -204,8 +202,5 @@ endif
@echo " endif ()" >> "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)"
@echo "endif ()" >> "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)"
@echo Install OK!
#Generating release tar
@echo Generating $(OPENBLAS_RELEASE_DIR)/$(basename $(LIBNAME)).tar.gz
@tar -cvz --file=$(OPENBLAS_RELEASE_DIR)/$(basename $(LIBNAME)).tar.gz --directory=$(PREFIX) --exclude=release .


+ 2
- 34
kernel/riscv64/KERNEL.x280 View File

@@ -122,23 +122,7 @@ CTRMMKERNEL = ztrmmkernel_2x2_rvv.c
ZTRMMKERNEL = ztrmmkernel_2x2_rvv.c

# SGEMM_UNROLL_N set in params.h
ifeq ($(SGEMM_UNROLL_N), 2)
SGEMMKERNEL = gemmkernel_2x2_rvv.c
SGEMMONCOPY = gemm_ncopy_2_rvv.c
SGEMMOTCOPY = gemm_tcopy_2_rvv.c
SGEMMONCOPYOBJ = sgemm_oncopy.o
SGEMMOTCOPYOBJ = sgemm_otcopy.o

STRMMKERNEL = trmmkernel_2x2_rvv.c
else ifeq ($(SGEMM_UNROLL_N), 4)
SGEMMKERNEL = gemmkernel_4x4_rvv.c
SGEMMONCOPY = gemm_ncopy_4_rvv.c
SGEMMOTCOPY = ../generic/gemm_tcopy_4.c
SGEMMONCOPYOBJ = sgemm_oncopy.o
SGEMMOTCOPYOBJ = sgemm_otcopy.o

STRMMKERNEL = trmmkernel_4x4_rvv.c
else ifeq ($(SGEMM_UNROLL_N), 8)
ifeq ($(SGEMM_UNROLL_N), 8)
# UNROLL_M is VLMAX
SGEMMKERNEL = gemmkernel_rvv_v1x8.c
SGEMMINCOPY = gemm_ncopy_rvv_v1.c
@@ -162,23 +146,7 @@ SSYMMLCOPY_M = symm_lcopy_rvv_v1.c
endif

# SGEMM_UNROLL_N set in params.h
ifeq ($(DGEMM_UNROLL_N), 2)
DGEMMKERNEL = gemmkernel_2x2_rvv.c
DGEMMONCOPY = gemm_ncopy_2_rvv.c
DGEMMOTCOPY = gemm_tcopy_2_rvv.c
DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o

DTRMMKERNEL = trmmkernel_2x2_rvv.c
else ifeq ($(DGEMM_UNROLL_N), 4)
DGEMMKERNEL = gemmkernel_4x4_rvv.c
DGEMMONCOPY = gemm_ncopy_4_rvv.c
DGEMMOTCOPY = ../generic/gemm_tcopy_4.c
DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o

DTRMMKERNEL = trmmkernel_4x4_rvv.c
else ifeq ($(DGEMM_UNROLL_N), 8)
ifeq ($(DGEMM_UNROLL_N), 8)
# UNROLL_M is VLMAX
DGEMMKERNEL = gemmkernel_rvv_v1x8.c
DGEMMINCOPY = gemm_ncopy_rvv_v1.c


+ 0
- 92
kernel/riscv64/gemm_ncopy_2_rvv.c View File

@@ -1,92 +0,0 @@
/***************************************************************************
Copyright (c) 2022, 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(DOUBLE)
#define VSETVL(n) vsetvl_e32m4(n)
#define FLOAT_V_T vfloat32m4_t
#define VLEV_FLOAT vle32_v_f32m4
#define VSEV_FLOAT vse32_v_f32m4
#define VSSEG2_FLOAT vsseg2e32_v_f32m4
#else
#define VSETVL(n) vsetvl_e64m4(n)
#define FLOAT_V_T vfloat64m4_t
#define VLEV_FLOAT vle64_v_f64m4
#define VSEV_FLOAT vse64_v_f64m4
#define VSSEG2_FLOAT vsseg2e64_v_f64m4
#endif

// Optimizes the implementation in ../generic/gemm_ncopy_2.c

int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
{
BLASLONG i, j;
IFLOAT *a_offset, *a_offset1, *a_offset2;
IFLOAT *b_offset;
FLOAT_V_T v1, v2;
size_t vl;

//fprintf(stderr, "gemm_ncopy_2 m=%ld n=%ld lda=%ld\n", m, n, lda); // KU

a_offset = a;
b_offset = b;

for(j = (n >> 1); j > 0; j--) {

a_offset1 = a_offset;
a_offset2 = a_offset + lda;
a_offset += 2 * lda;

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset1, vl);
v2 = VLEV_FLOAT(a_offset2, vl);
VSSEG2_FLOAT(b_offset, v1, v2, vl);

a_offset1 += vl;
a_offset2 += vl;
b_offset += vl*2;
}
}

if (n & 1) {

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset, vl);
VSEV_FLOAT(b_offset, v1, vl);

a_offset += vl;
b_offset += vl;
}
}

return 0;
}

+ 0
- 123
kernel/riscv64/gemm_ncopy_4_rvv.c View File

@@ -1,123 +0,0 @@
/***************************************************************************
Copyright (c) 2022, 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(DOUBLE)
#define VSETVL(n) vsetvl_e32m2(n)
#define FLOAT_V_T vfloat32m2_t
#define VLEV_FLOAT vle32_v_f32m2
#define VSEV_FLOAT vse32_v_f32m2
#define VSSEG2_FLOAT vsseg2e32_v_f32m2
#define VSSEG4_FLOAT vsseg4e32_v_f32m2
#else
#define VSETVL(n) vsetvl_e64m2(n)
#define FLOAT_V_T vfloat64m2_t
#define VLEV_FLOAT vle64_v_f64m2
#define VSEV_FLOAT vse64_v_f64m2
#define VSSEG2_FLOAT vsseg2e64_v_f64m2
#define VSSEG4_FLOAT vsseg4e64_v_f64m2
#endif

// Optimizes the implementation in ../generic/gemm_ncopy_4.c

int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
{
BLASLONG i, j;

FLOAT *a_offset, *a_offset1, *a_offset2, *a_offset3, *a_offset4;
FLOAT *b_offset;

FLOAT_V_T v1, v2, v3, v4;
size_t vl;

//fprintf(stderr, "gemm_ncopy_4 m=%ld n=%ld lda=%ld\n", m, n, lda);

a_offset = a;
b_offset = b;

for(j = (n >> 2); j > 0; j--) {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset3 = a_offset2 + lda;
a_offset4 = a_offset3 + lda;
a_offset += 4 * lda;

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset1, vl);
v2 = VLEV_FLOAT(a_offset2, vl);
v3 = VLEV_FLOAT(a_offset3, vl);
v4 = VLEV_FLOAT(a_offset4, vl);

VSSEG4_FLOAT(b_offset, v1, v2, v3, v4, vl);

a_offset1 += vl;
a_offset2 += vl;
a_offset3 += vl;
a_offset4 += vl;
b_offset += vl*4;
}
}

if (n & 2) {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset += 2 * lda;

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset1, vl);
v2 = VLEV_FLOAT(a_offset2, vl);

VSSEG2_FLOAT(b_offset, v1, v2, vl);

a_offset1 += vl;
a_offset2 += vl;
b_offset += vl*2;
}
}

if (n & 1) {
a_offset1 = a_offset;

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset1, vl);

VSEV_FLOAT(b_offset, v1, vl);

a_offset1 += vl;
b_offset += vl;
}
}

return 0;
}

+ 0
- 108
kernel/riscv64/gemm_tcopy_2_rvv.c View File

@@ -1,108 +0,0 @@
/***************************************************************************
Copyright (c) 2022, 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(DOUBLE)
#define VSETVL(n) vsetvl_e32m2(n)
#define FLOAT_V_T vfloat32m2_t
#define VLSEG2_FLOAT vlseg2e32_v_f32m2
#define VSSSEG2_FLOAT vssseg2e32_v_f32m2
#define VSSSEG4_FLOAT vssseg4e32_v_f32m2
#else
#define VSETVL(n) vsetvl_e64m2(n)
#define FLOAT_V_T vfloat64m2_t
#define VLSEG2_FLOAT vlseg2e64_v_f64m2
#define VSSSEG2_FLOAT vssseg2e64_v_f64m2
#define VSSSEG4_FLOAT vssseg4e64_v_f64m2
#endif

// Optimizes the implementation in ../generic/gemm_tcopy_2.c

int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
{
BLASLONG i, j;
IFLOAT *a_offset, *a_offset1, *a_offset2;
IFLOAT *b_offset, *b_offset1, *b_offset2;
FLOAT_V_T v1a, v1b, v2a, v2b;
size_t vl;

//fprintf(stderr, "gemm_tcopy_2 m=%ld n=%ld lda=%ld\n", m, n, lda); // KU

a_offset = a;
b_offset = b;
b_offset2 = b + m * (n & ~1);

for(i = (m >> 1); i > 0; i--) {

a_offset1 = a_offset;
a_offset2 = a_offset + lda;
a_offset += 2 * lda;

b_offset1 = b_offset;
b_offset += 4;

for(j = (n >> 1); j > 0; j -= vl) {
vl = VSETVL(j);

VLSEG2_FLOAT(&v1a, &v1b, a_offset1, vl);
VLSEG2_FLOAT(&v2a, &v2b, a_offset2, vl);

VSSSEG4_FLOAT(b_offset1, m*2*sizeof(FLOAT), v1a, v1b, v2a, v2b, vl);

a_offset1 += vl * 2;
a_offset2 += vl * 2;
b_offset1 += vl * m * 2;
}

if (n & 1) {
*(b_offset2 + 0) = *(a_offset1 + 0);
*(b_offset2 + 1) = *(a_offset2 + 0);
b_offset2 += 2;
}
}

if (m & 1) {

for(j = (n >> 1); j > 0; j -= vl) {
vl = VSETVL(j);

VLSEG2_FLOAT(&v1a, &v1b, a_offset, vl);

VSSSEG2_FLOAT(b_offset, m*2*sizeof(FLOAT), v1a, v1b, vl);

a_offset += vl * 2;
b_offset += vl * m * 2;
}

if (n & 1){
*(b_offset2 + 0) = *(a_offset + 0);
}
}

return 0;
}

+ 0
- 236
kernel/riscv64/gemm_tcopy_4_rvv.c View File

@@ -1,236 +0,0 @@
/***************************************************************************
Copyright (c) 2022, 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(DOUBLE)
#define VSETVL(n) vsetvl_e32m2(n)
#define FLOAT_V_T vfloat32m2_t
#define VLSEG2_FLOAT vlseg2e32_v_f32m2
#define VSSSEG2_FLOAT vssseg2e32_v_f32m2
#define VSSSEG4_FLOAT vssseg4e32_v_f32m2
#else
#define VSETVL(n) vsetvl_e64m2(n)
#define FLOAT_V_T vfloat64m2_t
#define VLSEG2_FLOAT vlseg2e64_v_f64m2
#define VSSSEG2_FLOAT vssseg2e64_v_f64m2
#define VSSSEG4_FLOAT vssseg4e64_v_f64m2
#endif

// Optimizes the implementation in ../generic/gemm_tcopy_4.c

int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
{
BLASLONG i, j;

FLOAT *a_offset, *a_offset1, *a_offset2, *a_offset3, *a_offset4;
FLOAT *b_offset, *b_offset1, *b_offset2, *b_offset3;
FLOAT ctemp1, ctemp2, ctemp3, ctemp4;
FLOAT ctemp5, ctemp6, ctemp7, ctemp8;
FLOAT ctemp9, ctemp10, ctemp11, ctemp12;
FLOAT ctemp13, ctemp14, ctemp15, ctemp16;

//fprintf(stderr, "gemm_tcopy_4 m=%ld n=%ld lda=%ld\n", m, n, lda);

a_offset = a;
b_offset = b;

b_offset2 = b + m * (n & ~3);
b_offset3 = b + m * (n & ~1);

for(j = (m >> 2); j > 0; j--) {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset3 = a_offset2 + lda;
a_offset4 = a_offset3 + lda;
a_offset += 4 * lda;

b_offset1 = b_offset;
b_offset += 16;

for(i = (n >> 2); i > 0; i--) {
v1 = VLEV_FLOAT(a_offset1, 4);
v2 = VLEV_FLOAT(a_offset2, 4);
v3 = VLEV_FLOAT(a_offset3, 4);
v4 = VLEV_FLOAT(a_offset4, 4);

a_offset1 += 4;
a_offset2 += 4;
a_offset3 += 4;
a_offset4 += 4;

VSEV_FLOAT(b_offset1, v1, 4);
VSEV_FLOAT(b_offset2+4, v2, 4);
VSEV_FLOAT(b_offset2+8, v3, 4);
VSEV_FLOAT(b_offset2+12, v4, 4);

b_offset1 += m * 4;
}

if (n & 2) {
v1 = VLEV_FLOAT(a_offset1, 2);
v2 = VLEV_FLOAT(a_offset2, 2);
v3 = VLEV_FLOAT(a_offset3, 2);
v4 = VLEV_FLOAT(a_offset4, 2);

a_offset1 += 2;
a_offset2 += 2;
a_offset3 += 2;
a_offset4 += 2;

VSEV_FLOAT(b_offset2, v1, 2);
VSEV_FLOAT(b_offset2+2, v2, 2);
VSEV_FLOAT(b_offset2+4, v3, 2);
VSEV_FLOAT(b_offset2+6, v4, 2);

b_offset2 += 8;
}

if (n & 1) {
v1 = VLEV_FLOAT(a_offset1, 1);
v2 = VLEV_FLOAT(a_offset2, 1);
v3 = VLEV_FLOAT(a_offset3, 1);
v4 = VLEV_FLOAT(a_offset4, 1);

VSSEG4_FLOAT(b_offset3, v1, v2, v3, v4, 1);

b_offset3 += 4;
}

}

// TODO cleanup

if (m & 2){
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset += 2 * lda;

b_offset1 = b_offset;
b_offset += 8;

i = (n >> 2);
if (i > 0){
do{
ctemp1 = *(a_offset1 + 0);
ctemp2 = *(a_offset1 + 1);
ctemp3 = *(a_offset1 + 2);
ctemp4 = *(a_offset1 + 3);

ctemp5 = *(a_offset2 + 0);
ctemp6 = *(a_offset2 + 1);
ctemp7 = *(a_offset2 + 2);
ctemp8 = *(a_offset2 + 3);

a_offset1 += 4;
a_offset2 += 4;

*(b_offset1 + 0) = ctemp1;
*(b_offset1 + 1) = ctemp2;
*(b_offset1 + 2) = ctemp3;
*(b_offset1 + 3) = ctemp4;

*(b_offset1 + 4) = ctemp5;
*(b_offset1 + 5) = ctemp6;
*(b_offset1 + 6) = ctemp7;
*(b_offset1 + 7) = ctemp8;

b_offset1 += m * 4;
i --;
}while(i > 0);
}

if (n & 2) {
ctemp1 = *(a_offset1 + 0);
ctemp2 = *(a_offset1 + 1);

ctemp3 = *(a_offset2 + 0);
ctemp4 = *(a_offset2 + 1);

a_offset1 += 2;
a_offset2 += 2;

*(b_offset2 + 0) = ctemp1;
*(b_offset2 + 1) = ctemp2;
*(b_offset2 + 2) = ctemp3;
*(b_offset2 + 3) = ctemp4;

b_offset2 += 4;
}

if (n & 1) {
ctemp1 = *(a_offset1 + 0);
ctemp2 = *(a_offset2 + 0);

*(b_offset3 + 0) = ctemp1;
*(b_offset3 + 1) = ctemp2;
b_offset3 += 2;
}
}

if (m & 1){
a_offset1 = a_offset;
b_offset1 = b_offset;

i = (n >> 2);
if (i > 0){
do{
ctemp1 = *(a_offset1 + 0);
ctemp2 = *(a_offset1 + 1);
ctemp3 = *(a_offset1 + 2);
ctemp4 = *(a_offset1 + 3);

a_offset1 += 4;

*(b_offset1 + 0) = ctemp1;
*(b_offset1 + 1) = ctemp2;
*(b_offset1 + 2) = ctemp3;
*(b_offset1 + 3) = ctemp4;

b_offset1 += 4 * m;

i --;
}while(i > 0);
}

if (n & 2) {
ctemp1 = *(a_offset1 + 0);
ctemp2 = *(a_offset1 + 1);
a_offset1 += 2;

*(b_offset2 + 0) = ctemp1;
*(b_offset2 + 1) = ctemp2;
}

if (n & 1) {
ctemp1 = *(a_offset1 + 0);
*(b_offset3 + 0) = ctemp1;
}
}

return 0;
}

+ 0
- 214
kernel/riscv64/gemmkernel_2x2_rvv.c View File

@@ -1,214 +0,0 @@
/***************************************************************************
Copyright (c) 2022, 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(DOUBLE)
#define VSETVL(n) vsetvl_e32m4(n)
#define VSETVL_MAX vsetvlmax_e32m4()
#define VSETVL_MAX_M1 vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VLEV_FLOAT vle32_v_f32m4
#define VLSEG2_FLOAT vlseg2e32_v_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
#define VFMACCVF_FLOAT vfmacc_vf_f32m4
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFREDSUMVS_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
#else
#define VSETVL(n) vsetvl_e64m4(n)
#define VSETVL_MAX vsetvlmax_e64m4()
#define VSETVL_MAX_M1 vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VLEV_FLOAT vle64_v_f64m4
#define VLSEG2_FLOAT vlseg2e64_v_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
#define VFMACCVF_FLOAT vfmacc_vf_f64m4
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFREDSUMVS_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
#endif

// Optimizes the implementation in ../generic/gemm_kernel_2x2.c

int CNAME(BLASLONG bm, BLASLONG bn, BLASLONG bk, FLOAT alpha, IFLOAT* ba, IFLOAT* bb, FLOAT* C, BLASLONG ldc
#ifdef TRMMKERNEL
,BLASLONG offset
#endif
)
{
BLASLONG i,j,k;
FLOAT *C0,*C1;
IFLOAT *ptrba,*ptrbb;

//fprintf(stderr, "gemm_kernel_2x2 bm=%ld bn=%ld bk=%ld alpha=%f ldc=%ld\n", bm, bn, bk, alpha, ldc);

FLOAT_V_T va0, va1, vb0, vb1;
FLOAT_V_T vres0, vres1, vres2, vres3;
FLOAT_V_T_M1 vsum0, vsum1, vsum2, vsum3;
FLOAT_V_T_M1 v_z0;

v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
size_t vlmax = VSETVL_MAX;
size_t vl;

for (j = bn/2; j > 0; j--) {
C0 = C;
C1 = C0 + ldc;
ptrba = ba;

for (i = bm/2; i > 0; i--) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
vres2 = VFMVVF_FLOAT(0.0, vlmax);
vres3 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);

ptrba += vl*2;
ptrbb += vl*2;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C1[0] += alpha * VFMVFS_FLOAT_M1(vsum2);
C1[1] += alpha * VFMVFS_FLOAT_M1(vsum3);

C0 += 2;
C1 += 2;
}

if(bm & 1) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

va0 = VLEV_FLOAT(ptrba, vl);
VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);

ptrba += vl;
ptrbb += vl*2;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C1[0] += alpha * VFMVFS_FLOAT_M1(vsum1);

C0 += 1;
C1 += 1;
}

bb += (bk<<1);
C += (ldc<<1);
}

if(bn & 1) {
C0 = C;
ptrba = ba;
for (i = bm/2; i > 0; i--) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);

ptrba += vl*2;
ptrbb += vl;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);

C0 += 2;
}

if(bm & 1) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

va0 = VLEV_FLOAT(ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);

ptrba += vl;
ptrbb += vl;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);

C0 += 1;
}

bb += (bk<<0);
C += ldc;
}

return 0;
}

+ 0
- 508
kernel/riscv64/gemmkernel_4x4_rvv.c View File

@@ -1,508 +0,0 @@
/***************************************************************************
Copyright (c) 2022, 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(DOUBLE)
#define VSETVL(n) vsetvl_e32m1(n)
#define VSETVL_MAX vsetvlmax_e32m1()
#define VSETVL_MAX_M1 vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m1_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VLEV_FLOAT vle32_v_f32m1
#define VLSEG2_FLOAT vlseg2e32_v_f32m1
#define VLSEG4_FLOAT vlseg4e32_v_f32m1
#define VFMVVF_FLOAT vfmv_v_f_f32m1
#define VFMACCVF_FLOAT vfmacc_vf_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m1
#define VFREDSUMVS_FLOAT vfredusum_vs_f32m1_f32m1
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
#else
#define VSETVL(n) vsetvl_e64m1(n)
#define VSETVL_MAX vsetvlmax_e64m1()
#define VSETVL_MAX_M1 vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m1_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VLEV_FLOAT vle64_v_f64m1
#define VLSEG2_FLOAT vlseg2e64_v_f64m1
#define VLSEG4_FLOAT vlseg4e64_v_f64m1
#define VFMVVF_FLOAT vfmv_v_f_f64m1
#define VFMACCVF_FLOAT vfmacc_vf_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m1
#define VFREDSUMVS_FLOAT vfredusum_vs_f64m1_f64m1
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
#endif

// Optimizes the implementation in ../generic/gemm_kernel_2x2.c

int CNAME(BLASLONG bm, BLASLONG bn, BLASLONG bk, FLOAT alpha, IFLOAT* ba, IFLOAT* bb, FLOAT* C, BLASLONG ldc
#ifdef TRMMKERNEL
,BLASLONG offset
#endif
)
{
BLASLONG i,j,k;
FLOAT *C0,*C1,*C2,*C3;
IFLOAT *ptrba,*ptrbb;

//fprintf(stderr, "gemm_kernel_4x4 bm=%ld bn=%ld bk=%ld alpha=%f ldc=%ld\n", bm, bn, bk, alpha, ldc); // KU

FLOAT_V_T va0, va1, va2, va3;
FLOAT_V_T vb0, vb1, vb2, vb3;
FLOAT_V_T vres0, vres1, vres2, vres3, vres4, vres5, vres6, vres7;
FLOAT_V_T vres8, vres9, vres10, vres11, vres12, vres13, vres14, vres15;
FLOAT_V_T_M1 vsum0, vsum1, vsum2, vsum3;
FLOAT_V_T_M1 v_z0;

v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
size_t vlmax = VSETVL_MAX;
size_t vl;

for (j = bn/4; j > 0; j--) {
C0 = C;
C1 = C0 + ldc;
C2 = C1 + ldc;
C3 = C2 + ldc;
ptrba = ba;

for (i = bm/4; i > 0; i--) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
vres2 = VFMVVF_FLOAT(0.0, vlmax);
vres3 = VFMVVF_FLOAT(0.0, vlmax);
vres4 = VFMVVF_FLOAT(0.0, vlmax);
vres5 = VFMVVF_FLOAT(0.0, vlmax);
vres6 = VFMVVF_FLOAT(0.0, vlmax);
vres7 = VFMVVF_FLOAT(0.0, vlmax);
vres8 = VFMVVF_FLOAT(0.0, vlmax);
vres9 = VFMVVF_FLOAT(0.0, vlmax);
vres10 = VFMVVF_FLOAT(0.0, vlmax);
vres11 = VFMVVF_FLOAT(0.0, vlmax);
vres12 = VFMVVF_FLOAT(0.0, vlmax);
vres13 = VFMVVF_FLOAT(0.0, vlmax);
vres14 = VFMVVF_FLOAT(0.0, vlmax);
vres15 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);

vres4 = VFMACCVV_FLOAT(vres4, va0, vb2, vl);
vres5 = VFMACCVV_FLOAT(vres5, va1, vb2, vl);
vres6 = VFMACCVV_FLOAT(vres6, va0, vb3, vl);
vres7 = VFMACCVV_FLOAT(vres7, va1, vb3, vl);

vres8 = VFMACCVV_FLOAT(vres8, va2, vb0, vl);
vres9 = VFMACCVV_FLOAT(vres9, va3, vb0, vl);
vres10 = VFMACCVV_FLOAT(vres10, va2, vb1, vl);
vres11 = VFMACCVV_FLOAT(vres11, va3, vb1, vl);

vres12 = VFMACCVV_FLOAT(vres12, va2, vb2, vl);
vres13 = VFMACCVV_FLOAT(vres13, va3, vb2, vl);
vres14 = VFMACCVV_FLOAT(vres14, va2, vb3, vl);
vres15 = VFMACCVV_FLOAT(vres15, va3, vb3, vl);

ptrba += vl*4;
ptrbb += vl*4;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres8, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres9, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C0[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
C0[3] += alpha * VFMVFS_FLOAT_M1(vsum3);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres2, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres3, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres10, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres11, v_z0, vlmax);
C1[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C1[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C1[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
C1[3] += alpha * VFMVFS_FLOAT_M1(vsum3);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres4, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres5, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres12, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres13, v_z0, vlmax);
C2[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C2[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C2[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
C2[3] += alpha * VFMVFS_FLOAT_M1(vsum3);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres6, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres7, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres14, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres15, v_z0, vlmax);
C3[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C3[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C3[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
C3[3] += alpha * VFMVFS_FLOAT_M1(vsum3);

C0 += 4;
C1 += 4;
C2 += 4;
C3 += 4;
}

if(bm & 2) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
vres2 = VFMVVF_FLOAT(0.0, vlmax);
vres3 = VFMVVF_FLOAT(0.0, vlmax);
vres4 = VFMVVF_FLOAT(0.0, vlmax);
vres5 = VFMVVF_FLOAT(0.0, vlmax);
vres6 = VFMVVF_FLOAT(0.0, vlmax);
vres7 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);

vres4 = VFMACCVV_FLOAT(vres4, va0, vb2, vl);
vres5 = VFMACCVV_FLOAT(vres5, va1, vb2, vl);
vres6 = VFMACCVV_FLOAT(vres6, va0, vb3, vl);
vres7 = VFMACCVV_FLOAT(vres7, va1, vb3, vl);

ptrba += vl*2;
ptrbb += vl*4;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
vsum0 = VFREDSUMVS_FLOAT(vsum0, vres2, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres3, v_z0, vlmax);
C1[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C1[1] += alpha * VFMVFS_FLOAT_M1(vsum1);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres4, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres5, v_z0, vlmax);
C2[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C2[1] += alpha * VFMVFS_FLOAT_M1(vsum1);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres6, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres7, v_z0, vlmax);
C3[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C3[1] += alpha * VFMVFS_FLOAT_M1(vsum1);

C0 += 2;
C1 += 2;
C2 += 2;
C3 += 2;
}

if(bm & 1) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
vres2 = VFMVVF_FLOAT(0.0, vlmax);
vres3 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

va0 = VLEV_FLOAT(ptrba, vl);
VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
vres2 = VFMACCVV_FLOAT(vres2, va0, vb2, vl);
vres3 = VFMACCVV_FLOAT(vres3, va0, vb3, vl);

ptrba += vl;
ptrbb += vl*4;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C1[0] += alpha * VFMVFS_FLOAT_M1(vsum1);
C2[0] += alpha * VFMVFS_FLOAT_M1(vsum2);
C3[0] += alpha * VFMVFS_FLOAT_M1(vsum3);

C0 += 1;
C1 += 1;
C2 += 1;
C3 += 1;
}

bb += (bk<<2);
C += (ldc<<2);
}

if(bn & 2) {

C0 = C;
C1 = C0 + ldc;
ptrba = ba;

for (i = bm/4; i > 0; i--) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
vres2 = VFMVVF_FLOAT(0.0, vlmax);
vres3 = VFMVVF_FLOAT(0.0, vlmax);

vres4 = VFMVVF_FLOAT(0.0, vlmax);
vres5 = VFMVVF_FLOAT(0.0, vlmax);
vres6 = VFMVVF_FLOAT(0.0, vlmax);
vres7 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
vres3 = VFMACCVV_FLOAT(vres3, va3, vb0, vl);

vres4 = VFMACCVV_FLOAT(vres4, va0, vb1, vl);
vres5 = VFMACCVV_FLOAT(vres5, va1, vb1, vl);
vres6 = VFMACCVV_FLOAT(vres6, va2, vb1, vl);
vres7 = VFMACCVV_FLOAT(vres7, va3, vb1, vl);

ptrba += vl*4;
ptrbb += vl*2;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C0[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
C0[3] += alpha * VFMVFS_FLOAT_M1(vsum3);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres4, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres5, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres6, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres7, v_z0, vlmax);
C1[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C1[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C1[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
C1[3] += alpha * VFMVFS_FLOAT_M1(vsum3);

C0 += 4;
C1 += 4;
}

if(bm & 2) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
vres2 = VFMVVF_FLOAT(0.0, vlmax);
vres3 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);

ptrba += vl*2;
ptrbb += vl*2;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C1[0] += alpha * VFMVFS_FLOAT_M1(vsum2);
C1[1] += alpha * VFMVFS_FLOAT_M1(vsum3);

C0 += 2;
C1 += 2;
}

if(bm & 1) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

va0 = VLEV_FLOAT(ptrba, vl);
VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);

ptrba += vl;
ptrbb += vl*2;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C1[0] += alpha * VFMVFS_FLOAT_M1(vsum1);

C0 += 1;
C1 += 1;
}

bb += (bk<<1);
C += (ldc<<1);
}

if(bn & 1) {
C0 = C;
ptrba = ba;
for (i = bm/4; i > 0; i--) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
vres2 = VFMVVF_FLOAT(0.0, vlmax);
vres3 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
vres3 = VFMACCVV_FLOAT(vres3, va3, vb0, vl);

ptrba += vl*4;
ptrbb += vl;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
C0[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
C0[3] += alpha * VFMVFS_FLOAT_M1(vsum3);

C0 += 4;
}

if(bm & 2) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);

ptrba += vl*2;
ptrbb += vl;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);

C0 += 2;
}

if(bm & 1) {
ptrbb = bb;

vres0 = VFMVVF_FLOAT(0.0, vlmax);

for (k = bk; k > 0; k -= vl) {
vl = VSETVL(k);

va0 = VLEV_FLOAT(ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);

ptrba += vl;
ptrbb += vl;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);

C0 += 1;
}

bb += (bk<<0);
C += ldc;
}

return 0;
}

+ 0
- 342
kernel/riscv64/trmmkernel_2x2_rvv.c View File

@@ -1,342 +0,0 @@
/***************************************************************************
Copyright (c) 2022, 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(DOUBLE)
#define VSETVL(n) vsetvl_e32m4(n)
#define VSETVL_MAX vsetvlmax_e32m4()
#define VSETVL_MAX_M1 vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VLEV_FLOAT vle32_v_f32m4
#define VLSEG_FLOAT vlseg2e32_v_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
#define VFMACCVF_FLOAT vfmacc_vf_f32m4
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFREDSUMVS_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
#else
#define VSETVL(n) vsetvl_e64m4(n)
#define VSETVL_MAX vsetvlmax_e64m4()
#define VSETVL_MAX_M1 vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VLEV_FLOAT vle64_v_f64m4
#define VLSEG_FLOAT vlseg2e64_v_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
#define VFMACCVF_FLOAT vfmacc_vf_f64m4
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFREDSUMVS_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
#endif


// Optimizes the implementation in ../generic/trmmkernel_2x2.c


int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc
#ifdef TRMMKERNEL
,BLASLONG offset
#endif
)
{
BLASLONG i,j,k;
FLOAT *C0,*C1,*ptrba,*ptrbb;
BLASLONG off, temp;

FLOAT_V_T va0, va1, vb0, vb1;
FLOAT_V_T vres0, vres1, vres2, vres3;
FLOAT_V_T_M1 v_res, v_z0;
v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
size_t vl;
size_t vlmax = VSETVL_MAX;
#if defined(TRMMKERNEL) && !defined(LEFT)
off = -offset;
#else
off = 0;
#endif

for (j = bn/2; j > 0; j--)
{
C0 = C;
C1 = C0+ldc;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
ptrba = ba;

for (i = bm/2; i > 0; i--)
{
#if (defined(LEFT) && defined(TRANSA)) || \
(!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*2;
ptrbb = bb + off*2;
#endif

#if (defined(LEFT) && !defined(TRANSA)) || \
(!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+2;
#else
temp = off+2;
#endif
vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
vres2 = VFMVVF_FLOAT(0.0, vlmax);
vres3 = VFMVVF_FLOAT(0.0, vlmax);
for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
VLSEG_FLOAT(&va0, &va1, ptrba, vl);
VLSEG_FLOAT(&vb0, &vb1, ptrbb, vl);

vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);

ptrba += vl * 2;
ptrbb += vl * 2;
}
v_res = VFREDSUMVS_FLOAT(v_res, vres0, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(v_res);
v_res = VFREDSUMVS_FLOAT(v_res, vres1, v_z0, vlmax);
C0[1] = alpha * VFMVFS_FLOAT_M1(v_res);
v_res = VFREDSUMVS_FLOAT(v_res, vres2, v_z0, vlmax);
C1[0] = alpha * VFMVFS_FLOAT_M1(v_res);
v_res = VFREDSUMVS_FLOAT(v_res, vres3, v_z0, vlmax);
C1[1] = alpha * VFMVFS_FLOAT_M1(v_res);

#if ( defined(LEFT) && defined(TRANSA)) || \
(!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 2;
#else
temp -= 2;
#endif
ptrba += temp*2;
ptrbb += temp*2;
#endif
#ifdef LEFT
off += 2;
#endif
C0 = C0+2;
C1 = C1+2;
}

if (bm & 1)
{
#if (defined(LEFT) && defined(TRANSA)) ||(!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off;
ptrbb = bb+off*2;
#endif

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+1;
#else
temp = off+2;
#endif
vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);

for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
va0 = VLEV_FLOAT(ptrba, vl);
VLSEG_FLOAT(&vb0, &vb1, ptrbb, vl);
vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);

ptrba += vl;
ptrbb += vl * 2;

}
v_res = VFREDSUMVS_FLOAT(v_res, vres0, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(v_res);
v_res = VFREDSUMVS_FLOAT(v_res, vres1, v_z0, vlmax);
C1[0] = alpha * VFMVFS_FLOAT_M1(v_res);

#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk-off;
#ifdef LEFT
temp -= 1;
#else
temp -= 2;
#endif
ptrba += temp;
ptrbb += temp*2;
#endif
#ifdef LEFT
off += 1;
#endif
C0 = C0+1;
C1 = C1+1;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 2;
#endif
k = (bk<<1);
bb = bb+k;
i = (ldc<<1);
C = C+i;
}

if (bn & 1)
{
C0 = C;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
ptrba = ba;

for (i = bm/2; i > 0; i--)
{
#if (defined(LEFT) && defined(TRANSA)) || \
(!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*2;
ptrbb = bb + off;
#endif


#if (defined(LEFT) && !defined(TRANSA)) || \
(!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+2;
#else
temp = off+1;
#endif
vres0 = VFMVVF_FLOAT(0.0, vlmax);
vres1 = VFMVVF_FLOAT(0.0, vlmax);
for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
vb0 = VLEV_FLOAT(ptrbb, vl);
VLSEG_FLOAT(&va0, &va1, ptrba, vl);
vres0 = VFMACCVV_FLOAT(vres0, vb0, va0, vl);
vres1 = VFMACCVV_FLOAT(vres1, vb0, va1, vl);
ptrba += vl * 2;
ptrbb += vl;
}
v_res = VFREDSUMVS_FLOAT(v_res, vres0, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(v_res);
v_res = VFREDSUMVS_FLOAT(v_res, vres1, v_z0, vlmax);
C0[1] = alpha * VFMVFS_FLOAT_M1(v_res);

#if ( defined(LEFT) && defined(TRANSA)) || \
(!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 2;
#else
temp -= 1;
#endif
ptrba += temp*2;
ptrbb += temp;
#endif
#ifdef LEFT
off += 2;
#endif

C0 = C0+2;
}

if (bm & 1)
{
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off;
ptrbb = bb+off;
#endif

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off + 1;
#else
temp = off + 1;
#endif
vres0 = VFMVVF_FLOAT(0.0, vlmax);
for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
va0 = VLEV_FLOAT(ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);
vres0 = VFMACCVV_FLOAT(vres0, vb0, va0, vl);
ptrba += vl;
ptrbb += vl;
}
v_res = VFREDSUMVS_FLOAT(v_res, vres0, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(v_res);

#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk-off;
#ifdef LEFT
temp -= 1;
#else
temp -= 1;
#endif
ptrba += temp;
ptrbb += temp;
#endif
#ifdef LEFT
off += 1;
#endif
C0 = C0+1;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 1;
#endif
k = (bk<<0);
bb = bb+k;
C = C+ldc;
}
return 0;
}


+ 0
- 881
kernel/riscv64/trmmkernel_4x4_rvv.c View File

@@ -1,881 +0,0 @@
/***************************************************************************
Copyright (c) 2022, 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 <stdbool.h>

#if !defined(DOUBLE)
#define VSETVL(n) vsetvl_e32m2(n)
#define VSETVL_MAX vsetvlmax_e32m2()
#define VSETVL_MAX_M1 vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m2_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VLEV_FLOAT vle32_v_f32m2
#define VLSEG4_FLOAT vlseg4e32_v_f32m2
#define VLSEG2_FLOAT vlseg2e32_v_f32m2
#define VFMVVF_FLOAT vfmv_v_f_f32m2
#define VFMUL_FLOAT vfmul_vv_f32m2
#define VFMACCVF_FLOAT vfmacc_vf_f32m2
#define VFMACCVV_FLOAT vfmacc_vv_f32m2
#define VFREDSUMVS_FLOAT vfredusum_vs_f32m2_f32m1
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
#else
#define VSETVL(n) vsetvl_e64m2(n)
#define VSETVL_MAX vsetvlmax_e64m2()
#define VSETVL_MAX_M1 vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m2_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VLEV_FLOAT vle64_v_f64m2
#define VLSEG4_FLOAT vlseg4e64_v_f64m2
#define VLSEG2_FLOAT vlseg2e64_v_f64m2
#define VFMVVF_FLOAT vfmv_v_f_f64m2
#define VFMUL_FLOAT vfmul_vv_f64m2
#define VFMACCVF_FLOAT vfmacc_vf_f64m2
#define VFMACCVV_FLOAT vfmacc_vv_f64m2
#define VFREDSUMVS_FLOAT vfredusum_vs_f64m2_f64m1
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
#endif


// Optimizes the implementation in ../generic/trmmkernel_4x4.c

int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc ,BLASLONG offset)
{

BLASLONG i,j,k;
FLOAT *C0,*C1,*C2,*C3,*ptrba,*ptrbb;

FLOAT_V_T va0, va1, va2, va3, vb0, vb1, vb2, vb3;
FLOAT_V_T_M1 vsum0, vsum1, vsum2, vsum3, v_z0;
v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
size_t vl;
size_t vlmax = VSETVL_MAX;

FLOAT_V_T vres0_0;
FLOAT_V_T vres0_1;
FLOAT_V_T vres0_2;
FLOAT_V_T vres0_3;

FLOAT_V_T vres1_0;
FLOAT_V_T vres1_1;
FLOAT_V_T vres1_2;
FLOAT_V_T vres1_3;

FLOAT_V_T vres2_0;
FLOAT_V_T vres2_1;
FLOAT_V_T vres2_2;
FLOAT_V_T vres2_3;

FLOAT_V_T vres3_0;
FLOAT_V_T vres3_1;
FLOAT_V_T vres3_2;
FLOAT_V_T vres3_3;

BLASLONG off, temp;

bool left;
bool transposed;
bool backwards;

#ifdef LEFT
left = true;
#else
left = false;
#endif

#ifdef TRANSA
transposed = true;
#else
transposed = false;
#endif

backwards = left != transposed;

if (!left) {
off = -offset;
}


for (j=0; j<bn/4; j+=1) // do blocks of the Mx4 loops
{
C0 = C;
C1 = C0+ldc;
C2 = C1+ldc;
C3 = C2+ldc;


if (left) {
off = offset;
}

ptrba = ba;

for (i=0; i<bm/4; i+=1) // do blocks of 4x4
{

ptrbb = bb;
if (backwards)
{
ptrba += off*4; // number of values in A
ptrbb += off*4; // number of values in B
}

vres0_0 = VFMVVF_FLOAT(0, vlmax);
vres0_1 = VFMVVF_FLOAT(0, vlmax);
vres0_2 = VFMVVF_FLOAT(0, vlmax);
vres0_3 = VFMVVF_FLOAT(0, vlmax);
vres1_0 = VFMVVF_FLOAT(0, vlmax);
vres1_1 = VFMVVF_FLOAT(0, vlmax);
vres1_2 = VFMVVF_FLOAT(0, vlmax);
vres1_3 = VFMVVF_FLOAT(0, vlmax);
vres2_0 = VFMVVF_FLOAT(0, vlmax);
vres2_1 = VFMVVF_FLOAT(0, vlmax);
vres2_2 = VFMVVF_FLOAT(0, vlmax);
vres2_3 = VFMVVF_FLOAT(0, vlmax);
vres3_0 = VFMVVF_FLOAT(0, vlmax);
vres3_1 = VFMVVF_FLOAT(0, vlmax);
vres3_2 = VFMVVF_FLOAT(0, vlmax);
vres3_3 = VFMVVF_FLOAT(0, vlmax);

temp = backwards ? bk-off :
left ? off + 4 : // number of values in A
off + 4; // number of values in B

for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
vres2_0 = VFMACCVV_FLOAT(vres2_0, va0, vb2, vl);
vres3_0 = VFMACCVV_FLOAT(vres3_0, va0, vb3, vl);
vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
vres1_1 = VFMACCVV_FLOAT(vres1_1, va1, vb1, vl);
vres2_1 = VFMACCVV_FLOAT(vres2_1, va1, vb2, vl);
vres3_1 = VFMACCVV_FLOAT(vres3_1, va1, vb3, vl);

vres0_2 = VFMACCVV_FLOAT(vres0_2, va2, vb0, vl);
vres1_2 = VFMACCVV_FLOAT(vres1_2, va2, vb1, vl);
vres2_2 = VFMACCVV_FLOAT(vres2_2, va2, vb2, vl);
vres3_2 = VFMACCVV_FLOAT(vres3_2, va2, vb3, vl);

vres0_3 = VFMACCVV_FLOAT(vres0_3, va3, vb0, vl);
vres1_3 = VFMACCVV_FLOAT(vres1_3, va3, vb1, vl);
vres2_3 = VFMACCVV_FLOAT(vres2_3, va3, vb2, vl);
vres3_3 = VFMACCVV_FLOAT(vres3_3, va3, vb3, vl);

ptrba += vl * 4;
ptrbb += vl * 4;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres0_2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres0_3, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C0[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
C0[3] = alpha * VFMVFS_FLOAT_M1(vsum3);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres1_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres1_2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres1_3, v_z0, vlmax);
C1[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C1[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C1[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
C1[3] = alpha * VFMVFS_FLOAT_M1(vsum3);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres2_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres2_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2_2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres2_3, v_z0, vlmax);
C2[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C2[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C2[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
C2[3] = alpha * VFMVFS_FLOAT_M1(vsum3);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres3_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres3_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres3_2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3_3, v_z0, vlmax);
C3[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C3[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C3[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
C3[3] = alpha * VFMVFS_FLOAT_M1(vsum3);

if (!backwards) {
temp = bk-off;
temp = left ? temp - 4 : // number of values in A
temp - 4; // number of values in B

ptrba += temp*4; // number of values in A
ptrbb += temp*4; // number of values in B
}
#ifdef LEFT
off += 4; // number of values in A
#endif

C0 = C0+4;
C1 = C1+4;
C2 = C2+4;
C3 = C3+4;

}

if ( bm & 2 ) // do any 2x4 loop
{

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*2;
ptrbb = bb + off*4;
#endif

vres0_0 = VFMVVF_FLOAT(0, vlmax);
vres0_1 = VFMVVF_FLOAT(0, vlmax);
vres1_0 = VFMVVF_FLOAT(0, vlmax);
vres1_1 = VFMVVF_FLOAT(0, vlmax);
vres2_0 = VFMVVF_FLOAT(0, vlmax);
vres2_1 = VFMVVF_FLOAT(0, vlmax);
vres3_0 = VFMVVF_FLOAT(0, vlmax);
vres3_1 = VFMVVF_FLOAT(0, vlmax);


#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+2; // number of values in A
#else
temp = off+4; // number of values in B
#endif
for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
vres2_0 = VFMACCVV_FLOAT(vres2_0, va0, vb2, vl);
vres3_0 = VFMACCVV_FLOAT(vres3_0, va0, vb3, vl);

vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
vres1_1 = VFMACCVV_FLOAT(vres1_1, va1, vb1, vl);
vres2_1 = VFMACCVV_FLOAT(vres2_1, va1, vb2, vl);
vres3_1 = VFMACCVV_FLOAT(vres3_1, va1, vb3, vl);

ptrba += vl * 2;
ptrbb += vl * 4;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres1_0, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres1_1, v_z0, vlmax);

C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C1[0] = alpha * VFMVFS_FLOAT_M1(vsum2);
C1[1] = alpha * VFMVFS_FLOAT_M1(vsum3);
vsum0 = VFREDSUMVS_FLOAT(vsum0, vres2_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres2_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres3_0, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3_1, v_z0, vlmax);

C2[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C2[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C3[0] = alpha * VFMVFS_FLOAT_M1(vsum2);
C3[1] = alpha * VFMVFS_FLOAT_M1(vsum3);


#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 2; // number of values in A
#else
temp -= 4; // number of values in B
#endif
ptrba += temp*2;
ptrbb += temp*4;
#endif

#ifdef LEFT
off += 2; // number of values in A
#endif

C0 = C0+2;
C1 = C1+2;
C2 = C2+2;
C3 = C3+2;

}

if ( bm & 1 ) // do any 1x4 loop
{

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*1;
ptrbb = bb + off*4;
#endif

vres0_0 = VFMVVF_FLOAT(0, vlmax);
vres1_0 = VFMVVF_FLOAT(0, vlmax);
vres2_0 = VFMVVF_FLOAT(0, vlmax);
vres3_0 = VFMVVF_FLOAT(0, vlmax);


#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+1; // number of values in A
#else
temp = off+4; // number of values in B
#endif

for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
va0 = VLEV_FLOAT(ptrba, vl);
VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
vres2_0 = VFMACCVV_FLOAT(vres2_0, va0, vb2, vl);
vres3_0 = VFMACCVV_FLOAT(vres3_0, va0, vb3, vl);

ptrba += vl;
ptrbb += vl * 4;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1_0, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2_0, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3_0, v_z0, vlmax);

C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C1[0] = alpha * VFMVFS_FLOAT_M1(vsum1);
C2[0] = alpha * VFMVFS_FLOAT_M1(vsum2);
C3[0] = alpha * VFMVFS_FLOAT_M1(vsum3);

#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 1; // number of values in A
#else
temp -= 4; // number of values in B
#endif
ptrba += temp*1;
ptrbb += temp*4;
#endif

#ifdef LEFT
off += 1; // number of values in A
#endif

C0 = C0+1;
C1 = C1+1;
C2 = C2+1;
C3 = C3+1;

}


#if defined(TRMMKERNEL) && !defined(LEFT)
off += 4;
#endif

k = (bk<<2);
bb = bb+k;
i = (ldc<<2);
C = C+i;
}

for (j=0; j<(bn&2); j+=2) // do the Mx2 loops
{
C0 = C;
C1 = C0+ldc;

#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif

ptrba = ba;

for (i=0; i<bm/4; i+=1) // do blocks of 4x2
{

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*4;
ptrbb = bb + off*2;
#endif
vres0_0 = VFMVVF_FLOAT(0, vlmax);
vres0_1 = VFMVVF_FLOAT(0, vlmax);
vres0_2 = VFMVVF_FLOAT(0, vlmax);
vres0_3 = VFMVVF_FLOAT(0, vlmax);
vres1_0 = VFMVVF_FLOAT(0, vlmax);
vres1_1 = VFMVVF_FLOAT(0, vlmax);
vres1_2 = VFMVVF_FLOAT(0, vlmax);
vres1_3 = VFMVVF_FLOAT(0, vlmax);

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+4; // number of values in A
#else
temp = off+2; // number of values in B
#endif

for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
vres1_1 = VFMACCVV_FLOAT(vres1_1, va1, vb1, vl);
vres0_2 = VFMACCVV_FLOAT(vres0_2, va2, vb0, vl);
vres1_2 = VFMACCVV_FLOAT(vres1_2, va2, vb1, vl);
vres0_3 = VFMACCVV_FLOAT(vres0_3, va3, vb0, vl);
vres1_3 = VFMACCVV_FLOAT(vres1_3, va3, vb1, vl);
ptrba += vl * 4;
ptrbb += vl * 2;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres0_2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres0_3, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C0[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
C0[3] = alpha * VFMVFS_FLOAT_M1(vsum3);

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres1_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres1_2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres1_3, v_z0, vlmax);
C1[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C1[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C1[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
C1[3] = alpha * VFMVFS_FLOAT_M1(vsum3);

#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 4; // number of values in A
#else
temp -= 2; // number of values in B
#endif
ptrba += temp*4;
ptrbb += temp*2;
#endif

#ifdef LEFT
off += 4; // number of values in A
#endif

C0 = C0+4;
C1 = C1+4;

}

if ( bm & 2 ) // do any 2x2 loop
{

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*2;
ptrbb = bb + off*2;
#endif

vres0_0 = VFMVVF_FLOAT(0, vlmax);
vres0_1 = VFMVVF_FLOAT(0, vlmax);
vres1_0 = VFMVVF_FLOAT(0, vlmax);
vres1_1 = VFMVVF_FLOAT(0, vlmax);


#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+2; // number of values in A
#else
temp = off+2; // number of values in B
#endif
for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
vres1_1 = VFMACCVV_FLOAT(vres1_1, va1, vb1, vl);
ptrba += vl * 2;
ptrbb += vl * 2;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres1_0, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres1_1, v_z0, vlmax);

C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C1[0] = alpha * VFMVFS_FLOAT_M1(vsum2);
C1[1] = alpha * VFMVFS_FLOAT_M1(vsum3);


#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 2; // number of values in A
#else
temp -= 2; // number of values in B
#endif
ptrba += temp*2;
ptrbb += temp*2;
#endif

#ifdef LEFT
off += 2; // number of values in A
#endif

C0 = C0+2;
C1 = C1+2;

}

if ( bm & 1 ) // do any 1x2 loop
{

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*1;
ptrbb = bb + off*2;
#endif


vres0_0 = VFMVVF_FLOAT(0, vlmax);
vres1_0 = VFMVVF_FLOAT(0, vlmax);


#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+1; // number of values in A
#else
temp = off+2; // number of values in B
#endif

for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
va0 = VLEV_FLOAT(ptrba, vl);
VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
ptrba += vl;
ptrbb += vl * 2;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1_0, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C1[0] = alpha * VFMVFS_FLOAT_M1(vsum1);


#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 1; // number of values in A
#else
temp -= 2; // number of values in B
#endif
ptrba += temp*1;
ptrbb += temp*2;
#endif

#ifdef LEFT
off += 1; // number of values in A
#endif

C0 = C0+1;
C1 = C1+1;

}


#if defined(TRMMKERNEL) && !defined(LEFT)
off += 2;
#endif

k = (bk<<1);
bb = bb+k;
i = (ldc<<1);
C = C+i;
}

for (j=0; j<(bn&1); j+=1) // do the Mx1 loops
{
C0 = C;

#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif

ptrba = ba;

for (i=0; i<bm/4; i+=1) // do blocks of 4x1 loops
{

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*4;
ptrbb = bb + off*1;
#endif

vres0_0 = VFMVVF_FLOAT(0, vlmax);
vres0_1 = VFMVVF_FLOAT(0, vlmax);
vres0_2 = VFMVVF_FLOAT(0, vlmax);
vres0_3 = VFMVVF_FLOAT(0, vlmax);


#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+4; // number of values in A
#else
temp = off+1; // number of values in B
#endif

for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
vres0_2 = VFMACCVV_FLOAT(vres0_2, va2, vb0, vl);
vres0_3 = VFMACCVV_FLOAT(vres0_3, va3, vb0, vl);
ptrba += vl * 4;
ptrbb += vl;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
vsum2 = VFREDSUMVS_FLOAT(vsum2, vres0_2, v_z0, vlmax);
vsum3 = VFREDSUMVS_FLOAT(vsum3, vres0_3, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
C0[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
C0[3] = alpha * VFMVFS_FLOAT_M1(vsum3);


#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 4; // number of values in A
#else
temp -= 1; // number of values in B
#endif
ptrba += temp*4;
ptrbb += temp*1;
#endif

#ifdef LEFT
off += 4; // number of values in A
#endif

C0 = C0+4;

}

if ( bm & 2 ) // do any 2x1 loop
{

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*2;
ptrbb = bb + off*1;
#endif

vres0_0 = VFMVVF_FLOAT(0, vlmax);
vres0_1 = VFMVVF_FLOAT(0, vlmax);

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+2; // number of values in A
#else
temp = off+1; // number of values in B
#endif

for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
ptrba += vl * 2;
ptrbb += vl;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);


#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 2; // number of values in A
#else
temp -= 1; // number of values in B
#endif
ptrba += temp*2;
ptrbb += temp*1;
#endif

#ifdef LEFT
off += 2; // number of values in A
#endif

C0 = C0+2;

}

if ( bm & 1 ) // do any 1x1 loop
{

#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
ptrbb = bb;
#else
ptrba += off*1;
ptrbb = bb + off*1;
#endif

vres0_0 = VFMVVF_FLOAT(0, vlmax);

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
temp = bk-off;
#elif defined(LEFT)
temp = off+1; // number of values in A
#else
temp = off+1; // number of values in B
#endif

for (k = temp; k > 0; k -= vl)
{
vl = VSETVL(k);
va0 = VLEV_FLOAT(ptrba, vl);
vb0 = VLEV_FLOAT(ptrbb, vl);
vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
ptrba += vl;
ptrbb += vl;
}

vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);


#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = bk - off;
#ifdef LEFT
temp -= 1; // number of values in A
#else
temp -= 1; // number of values in B
#endif
ptrba += temp*1;
ptrbb += temp*1;
#endif

#ifdef LEFT
off += 1; // number of values in A
#endif

C0 = C0+1;

}

#if defined(TRMMKERNEL) && !defined(LEFT)
off += 1;
#endif

k = (bk<<0);
bb = bb+k;
C = C+ldc;
}
return 0;
}

Write Preview
Loading…
Cancel
Save