floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Change file name to match the norm and delete useless code.

This commit is contained in:
Honglin Zhu 2022-10-28 17:09:39 +08:00
parent 4989e039a5
commit 79066b6bf3
8 changed files with 404 additions and 1296 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
kernel/arm64/KERNEL.NEOVERSEN2
View File

@ -189,12 +189,11 @@ ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX)
ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX)
SBGEMM_BETA = sbgemm_beta_neoversen2.c
# SBGEMMKERNEL = sbgemm_kernel_$(SBGEMM_UNROLL_M)x$(SBGEMM_UNROLL_N)_neoversen2.c
SBGEMMKERNEL = sbgemm_kernel_neoversen2_newbf16.c
SBGEMMINCOPY = sbgemm_ncopy_4_neoversen2.c
SBGEMMITCOPY = sbgemm_tcopy_8_neoversen2.c
SBGEMMONCOPY = sbgemm_ncopy_4_neoversen2.c
SBGEMMOTCOPY = sbgemm_tcopy_8_neoversen2.c
SBGEMMKERNEL = sbgemm_kernel_$(SBGEMM_UNROLL_M)x$(SBGEMM_UNROLL_N)_neoversen2.c
SBGEMMINCOPY = sbgemm_ncopy_$(SBGEMM_UNROLL_N)_neoversen2.c
SBGEMMITCOPY = sbgemm_tcopy_$(SBGEMM_UNROLL_M)_neoversen2.c
SBGEMMONCOPY = sbgemm_ncopy_$(SBGEMM_UNROLL_N)_neoversen2.c
SBGEMMOTCOPY = sbgemm_tcopy_$(SBGEMM_UNROLL_M)_neoversen2.c
SBGEMMINCOPYOBJ = sbgemm_incopy$(TSUFFIX).$(SUFFIX)
SBGEMMITCOPYOBJ = sbgemm_itcopy$(TSUFFIX).$(SUFFIX)
SBGEMMONCOPYOBJ = sbgemm_oncopy$(TSUFFIX).$(SUFFIX)

10
kernel/arm64/sbgemm_kernel_8x4_neoversen2.c
View File

@ -37,9 +37,9 @@
int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
FLOAT *C, BLASLONG ldc) {
if (alpha == 1.0f)
return sbgemm_kernel_neoversen2_alpha_one(m, n, k, alpha, A, B, C, ldc);
else
return sbgemm_kernel_neoversen2_alpha(m, n, k, alpha, A, B, C, ldc);
return 0;
if (alpha == 1.0f)
return sbgemm_kernel_neoversen2_alpha_one(m, n, k, alpha, A, B, C, ldc);
else
return sbgemm_kernel_neoversen2_alpha(m, n, k, alpha, A, B, C, ldc);
return 0;
}

982
kernel/arm64/sbgemm_kernel_8x4_neoversen2_impl.c
View File

File diff suppressed because it is too large Load Diff

467
kernel/arm64/sbgemm_kernel_neoversen2_newbf16.c
View File

@ -1,467 +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 <arm_sve.h>
#include "common.h"
#define LOAD_C(M, N) mc##M##N = svdup_f32(0);
#define MATMUL(M, N) mc##M##N = svbfmmla(mc##M##N, ma##M, mb##N);
#define LOAD_C_8x4 \
do { \
LOAD_C(0, 0); \
LOAD_C(0, 1); \
LOAD_C(1, 0); \
LOAD_C(1, 1); \
LOAD_C(2, 0); \
LOAD_C(2, 1); \
LOAD_C(3, 0); \
LOAD_C(3, 1); \
} while (0);
#define STORE_C(PG, PTR, SRC, DST) \
do { \
SRC = svld1_f32((PG), (PTR)); \
DST = svmad_z((PG), svalpha, DST, SRC); \
svst1_f32((PG), (PTR), DST); \
} while (0);
int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
FLOAT *C, BLASLONG ldc) {
BLASLONG pad_k = (k + 3) & ~3;
svbfloat16_t ma0, ma1, ma2, ma3, mb0, mb1;
svfloat32_t mc00, mc01, mc10, mc11, mc20, mc21, mc30, mc31,
vc0, vc1, vc2, vc3, vc4, vc5, vc6, vc7,
oc0, oc1, oc2, oc3, oc4, oc5, oc6, oc7;
svfloat32_t svalpha = svdup_f32(alpha);
svbool_t pg16 = svptrue_b16();
svbool_t pg16_low = svdupq_b16(1, 1, 1, 1, 0, 0, 0, 0);
svbool_t pg32 = svptrue_b32();
svbool_t pg32_low = svdupq_b32(1, 1, 0, 0);
svbool_t pg32_first = svdupq_b32(1, 0, 0, 0);
bfloat16_t *ptr_a = (bfloat16_t *)A;
bfloat16_t *ptr_b = (bfloat16_t *)B;
FLOAT *ptr_c = C;
bfloat16_t *ptr_a0, *ptr_a1, *ptr_a2, *ptr_a3;
bfloat16_t *ptr_b0, *ptr_b1;
FLOAT *ptr_c0, *ptr_c1, *ptr_c2, *ptr_c3;
for (BLASLONG j = 0; j < n / 4; j++) {
ptr_c0 = ptr_c;
ptr_c1 = ptr_c0 + ldc;
ptr_c2 = ptr_c1 + ldc;
ptr_c3 = ptr_c2 + ldc;
ptr_c += 4 * ldc;
ptr_a = (bfloat16_t *)A;
for (BLASLONG i = 0; i < m / 8; i++) {
ptr_a0 = ptr_a;
ptr_a += 8 * pad_k;
ptr_b0 = ptr_b;
LOAD_C_8x4;
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
ma1 = svld1_bf16(pg16, ptr_a0 + 8);
ma2 = svld1_bf16(pg16, ptr_a0 + 16);
ma3 = svld1_bf16(pg16, ptr_a0 + 24);
mb0 = svld1_bf16(pg16, ptr_b0);
mb1 = svld1_bf16(pg16, ptr_b0 + 8);
#if 0
for (int q = 0; q < 8; q++) {
float tmp = 0;
*((bfloat16_t *)(&tmp) + 1) = ptr_b0[8+q];
printf("%.1f ", tmp);
}
printf("\n");
#endif
MATMUL(0, 0); MATMUL(0, 1);
MATMUL(1, 0); MATMUL(1, 1);
MATMUL(2, 0); MATMUL(2, 1);
MATMUL(3, 0); MATMUL(3, 1);
ptr_a0 += 32;
ptr_b0 += 16;
}
vc0 = svuzp1(mc00, mc10);
vc1 = svuzp1(mc20, mc30);
vc2 = svuzp2(mc00, mc10);
vc3 = svuzp2(mc20, mc30);
vc4 = svuzp1(mc01, mc11);
vc5 = svuzp1(mc21, mc31);
vc6 = svuzp2(mc01, mc11);
vc7 = svuzp2(mc21, mc31);
STORE_C(pg32, ptr_c0, oc0, vc0);
STORE_C(pg32, ptr_c0+4, oc1, vc1);
STORE_C(pg32, ptr_c1, oc2, vc2);
STORE_C(pg32, ptr_c1+4, oc3, vc3);
STORE_C(pg32, ptr_c2, oc4, vc4)
STORE_C(pg32, ptr_c2+4, oc5, vc5);
STORE_C(pg32, ptr_c3, oc6, vc6)
STORE_C(pg32, ptr_c3+4, oc7, vc7);
ptr_c0 += 8;
ptr_c1 += 8;
ptr_c2 += 8;
ptr_c3 += 8;
}
if (m & 4) {
ptr_a0 = ptr_a;
ptr_a += 4 * pad_k;
ptr_b0 = ptr_b;
LOAD_C(0, 0); LOAD_C(0, 1);
LOAD_C(1, 0); LOAD_C(1, 1);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
ma1 = svld1_bf16(pg16, ptr_a0 + 8);
mb0 = svld1_bf16(pg16, ptr_b0);
mb1 = svld1_bf16(pg16, ptr_b0 + 8);
MATMUL(0, 0); MATMUL(0, 1);
MATMUL(1, 0); MATMUL(1, 1);
ptr_a0 += 16;
ptr_b0 += 16;
}
vc0 = svuzp1(mc00, mc10);
vc1 = svuzp2(mc00, mc10);
vc2 = svuzp1(mc01, mc11);
vc3 = svuzp2(mc01, mc11);
STORE_C(pg32, ptr_c0, oc0, vc0);
STORE_C(pg32, ptr_c1, oc1, vc1);
STORE_C(pg32, ptr_c2, oc2, vc2);
STORE_C(pg32, ptr_c3, oc3, vc3);
ptr_c0 += 4;
ptr_c1 += 4;
ptr_c2 += 4;
ptr_c3 += 4;
}
if (m & 2) {
ptr_a0 = ptr_a;
ptr_a += 2 * pad_k;
ptr_b0 = ptr_b;
LOAD_C(0, 0); LOAD_C(0, 1);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
mb0 = svld1_bf16(pg16, ptr_b0);
mb1 = svld1_bf16(pg16, ptr_b0 + 8);
MATMUL(0, 0); MATMUL(0, 1);
ptr_a0 += 8;
ptr_b0 += 16;
}
vc0 = svuzp1(mc00, mc00);
vc1 = svuzp2(mc00, mc00);
vc2 = svuzp1(mc01, mc01);
vc3 = svuzp2(mc01, mc01);
STORE_C(pg32_low, ptr_c0, oc0, vc0);
STORE_C(pg32_low, ptr_c1, oc1, vc1);
STORE_C(pg32_low, ptr_c2, oc2, vc2);
STORE_C(pg32_low, ptr_c3, oc3, vc3);
ptr_c0 += 2;
ptr_c1 += 2;
ptr_c2 += 2;
ptr_c3 += 2;
}
if (m & 1) {
ptr_a0 = ptr_a;
ptr_b0 = ptr_b;
LOAD_C(0, 0); LOAD_C(0, 1);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16_low, ptr_a0);
mb0 = svld1_bf16(pg16, ptr_b0);
mb1 = svld1_bf16(pg16, ptr_b0 + 8);
MATMUL(0, 0); MATMUL(0, 1);
ptr_a0 += 4;
ptr_b0 += 16;
}
vc1 = svuzp2(mc00, mc00);
vc3 = svuzp2(mc01, mc01);
STORE_C(pg32_first, ptr_c0, oc0, mc00);
STORE_C(pg32_first, ptr_c1, oc1, vc1);
STORE_C(pg32_first, ptr_c2, oc2, mc01);
STORE_C(pg32_first, ptr_c3, oc3, vc3);
}
ptr_b += 4 * pad_k;
}
if (n & 2) {
ptr_c0 = ptr_c;
ptr_c1 = ptr_c0 + ldc;
ptr_c += 2 * ldc;
ptr_a = (bfloat16_t *)A;
for (BLASLONG i = 0; i < m / 8; i++) {
ptr_a0 = ptr_a;
ptr_a += 8 * pad_k;
ptr_b0 = ptr_b;
LOAD_C(0, 0);
LOAD_C(1, 0);
LOAD_C(2, 0);
LOAD_C(3, 0);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
ma1 = svld1_bf16(pg16, ptr_a0 + 8);
ma2 = svld1_bf16(pg16, ptr_a0 + 16);
ma3 = svld1_bf16(pg16, ptr_a0 + 24);
mb0 = svld1_bf16(pg16, ptr_b0);
MATMUL(0, 0);
MATMUL(1, 0);
MATMUL(2, 0);
MATMUL(3, 0);
ptr_a0 += 32;
ptr_b0 += 8;
}
vc0 = svuzp1(mc00, mc10);
vc1 = svuzp1(mc20, mc30);
vc2 = svuzp2(mc00, mc10);
vc3 = svuzp2(mc20, mc30);
STORE_C(pg32, ptr_c0, oc0, vc0);
STORE_C(pg32, ptr_c0 + 4, oc1, vc1);
STORE_C(pg32, ptr_c1, oc2, vc2);
STORE_C(pg32, ptr_c1 + 4, oc3, vc3);
ptr_c0 += 8;
ptr_c1 += 8;
}
if (m & 4) {
ptr_a0 = ptr_a;
ptr_a += 4 * pad_k;
ptr_b0 = ptr_b;
LOAD_C(0, 0);
LOAD_C(1, 0);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
ma1 = svld1_bf16(pg16, ptr_a0 + 8);
mb0 = svld1_bf16(pg16, ptr_b0);
MATMUL(0, 0);
MATMUL(1, 0);
ptr_a0 += 16;
ptr_b0 += 8;
}
vc0 = svuzp1(mc00, mc10);
vc1 = svuzp2(mc00, mc10);
STORE_C(pg32, ptr_c0, oc0, vc0);
STORE_C(pg32, ptr_c1, oc1, vc1);
ptr_c0 += 4;
ptr_c1 += 4;
}
if (m & 2) {
ptr_a0 = ptr_a;
ptr_a += 2 * pad_k;
ptr_b0 = ptr_b;
LOAD_C(0, 0);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
mb0 = svld1_bf16(pg16, ptr_b0);
MATMUL(0, 0);
ptr_a0 += 8;
ptr_b0 += 8;
}
vc0 = svuzp1(mc00, mc00);
vc1 = svuzp2(mc00, mc00);
STORE_C(pg32_low, ptr_c0, oc0, vc0);
STORE_C(pg32_low, ptr_c1, oc1, vc1);
ptr_c0 += 2;
ptr_c1 += 2;
}
if (m & 1) {
ptr_a0 = ptr_a;
ptr_b0 = ptr_b;
LOAD_C(0, 0);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16_low, ptr_a0);
mb0 = svld1_bf16(pg16, ptr_b0);
MATMUL(0, 0);
ptr_a0 += 4;
ptr_b0 += 8;
}
vc1 = svuzp2(mc00, mc00);
STORE_C(pg32_first, ptr_c0, oc0, mc00);
STORE_C(pg32_first, ptr_c1, oc1, vc1);
}
ptr_b += 2 * pad_k;
}
if (n & 1) {
ptr_c0 = ptr_c;
ptr_a = (bfloat16_t *)A;
for (BLASLONG i = 0; i < m / 8; i++) {
ptr_a0 = ptr_a;
ptr_a += 8 * pad_k;
ptr_b0 = ptr_b;
LOAD_C(0, 0);
LOAD_C(1, 0);
LOAD_C(2, 0);
LOAD_C(3, 0);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
ma1 = svld1_bf16(pg16, ptr_a0 + 8);
ma2 = svld1_bf16(pg16, ptr_a0 + 16);
ma3 = svld1_bf16(pg16, ptr_a0 + 24);
mb0 = svld1_bf16(pg16_low, ptr_b0);
MATMUL(0, 0);
MATMUL(1, 0);
MATMUL(2, 0);
MATMUL(3, 0);
ptr_a0 += 32;
ptr_b0 += 4;
}
vc0 = svuzp1(mc00, mc10);
vc1 = svuzp1(mc20, mc30);
STORE_C(pg32, ptr_c0, oc0, vc0);
STORE_C(pg32, ptr_c0 + 4, oc1, vc1);
ptr_c0 += 8;
}
if (m & 4) {
ptr_a0 = ptr_a;
ptr_a += 4 * pad_k;
ptr_b0 = ptr_b;
LOAD_C(0, 0);
LOAD_C(1, 0);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
ma1 = svld1_bf16(pg16, ptr_a0 + 8);
mb0 = svld1_bf16(pg16_low, ptr_b0);
MATMUL(0, 0);
MATMUL(1, 0);
ptr_a0 += 16;
ptr_b0 += 4;
}
vc0 = svuzp1(mc00, mc10);
STORE_C(pg32, ptr_c0, oc0, vc0);
ptr_c0 += 4;
}
if (m & 2) {
ptr_a0 = ptr_a;
ptr_a += 2 * pad_k;
ptr_b0 = ptr_b;
LOAD_C(0, 0);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16, ptr_a0);
mb0 = svld1_bf16(pg16_low, ptr_b0);
MATMUL(0, 0);
ptr_a0 += 8;
ptr_b0 += 4;
}
vc0 = svuzp1(mc00, mc00);
STORE_C(pg32_low, ptr_c0, oc0, vc0);
ptr_c0 += 2;
}
if (m & 1) {
ptr_a0 = ptr_a;
ptr_b0 = ptr_b;
LOAD_C(0, 0);
for (BLASLONG p = 0; p < pad_k; p += 4) {
ma0 = svld1_bf16(pg16_low, ptr_a0);
mb0 = svld1_bf16(pg16_low, ptr_b0);
MATMUL(0, 0);
ptr_a0 += 4;
ptr_b0 += 4;
}
STORE_C(pg32_first, ptr_c0, oc0, mc00);
}
}
return 0;
}

11
kernel/arm64/sbgemm_ncopy_4_neoversen2.c
View File

@ -58,17 +58,6 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
svst1_bf16(pg16, (bfloat16_t *)b_offset + 8, v2);
svst1_bf16(pg16, (bfloat16_t *)b_offset + 12, v3);
#if 0
for (int line = 0; line < 4; line++) {
for (int p = 0; p < 4; p++) {
float tmp = 0;
*((bfloat16 *)(&tmp) + 1) = b_offset[line * 4 + p];
printf("%f ", tmp);
}
printf("\n");
}
#endif
b_offset += 16;
a_offsetx[0] += 4;
a_offsetx[1] += 4;

101
kernel/arm64/sbgemm_ncopy_neoversen2.c
View File

@ -1,101 +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"
int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
IFLOAT *a_offset, *a_offset1, *a_offset2;
IFLOAT *b_offset;
a_offset = a;
b_offset = b;
for (BLASLONG j = 0; j < n / 2; j++) {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset += 2 * lda;
for (BLASLONG i = 0; i < m / 4; i++) {
*(b_offset + 0) = *(a_offset1 + 0);
*(b_offset + 1) = *(a_offset1 + 1);
*(b_offset + 2) = *(a_offset1 + 2);
*(b_offset + 3) = *(a_offset1 + 3);
*(b_offset + 4) = *(a_offset2 + 0);
*(b_offset + 5) = *(a_offset2 + 1);
*(b_offset + 6) = *(a_offset2 + 2);
*(b_offset + 7) = *(a_offset2 + 3);
a_offset1 += 4;
a_offset2 += 4;
b_offset += 8;
}
BLASLONG rest = m & 3;
if (rest == 3) {
*(b_offset + 0) = *(a_offset1 + 0);
*(b_offset + 1) = *(a_offset1 + 1);
*(b_offset + 2) = *(a_offset1 + 2);
*(b_offset + 3) = *(a_offset2 + 0);
*(b_offset + 4) = *(a_offset2 + 1);
*(b_offset + 5) = *(a_offset2 + 2);
b_offset += 6;
} else if (rest == 2) {
*(b_offset + 0) = *(a_offset1 + 0);
*(b_offset + 1) = *(a_offset1 + 1);
*(b_offset + 2) = *(a_offset2 + 0);
*(b_offset + 3) = *(a_offset2 + 1);
b_offset += 4;
} else if (rest == 1) {
*(b_offset + 0) = *(a_offset1 + 0);
*(b_offset + 1) = *(a_offset2 + 0);
b_offset += 2;
}
}
if (n & 1) {
for (BLASLONG i = 0; i < m / 4; i++) {
*(b_offset + 0) = *(a_offset + 0);
*(b_offset + 1) = *(a_offset + 1);
*(b_offset + 2) = *(a_offset + 2);
*(b_offset + 3) = *(a_offset + 3);
b_offset += 4;
a_offset += 4;
}
BLASLONG rest = m & 3;
if (rest == 3) {
*(b_offset + 0) = *(a_offset + 0);
*(b_offset + 1) = *(a_offset + 1);
*(b_offset + 2) = *(a_offset + 2);
} else if (rest == 2) {
*(b_offset + 0) = *(a_offset + 0);
*(b_offset + 1) = *(a_offset + 1);
} else if (rest == 1) {
*(b_offset + 0) = *(a_offset + 0);
}
}
return 0;
}

9
kernel/arm64/sbgemm_tcopy_8_neoversen2.c
View File

@ -43,15 +43,6 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
for (BLASLONG i = 0; i < m / 4; i++) {
for (BLASLONG line = 0; line < 8; line++) {
#if 0
float fv0 = 0, fv1 = 0, fv2 = 0, fv3 = 0;
*((bfloat16 *)(&fv0) + 1) = a_offset0[line];
*((bfloat16 *)(&fv1) + 1) = a_offset1[line];
*((bfloat16 *)(&fv2) + 1) = a_offset2[line];
*((bfloat16 *)(&fv3) + 1) = a_offset3[line];
printf("%f %f %f %f\n", fv0, fv1, fv2, fv3);
#endif
b_offset[line * 4] = a_offset0[line];
b_offset[line * 4 + 1] = a_offset1[line];
b_offset[line * 4 + 2] = a_offset2[line];

109
kernel/arm64/sbgemm_tcopy_neoversen2.c
View File

@ -1,109 +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"
int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
IFLOAT *a_offset, *a_offset1, *a_offset2, *a_offset3, *a_offset4;
IFLOAT *b_offset;
a_offset = a;
b_offset = b;
for (BLASLONG j = 0; j < n / 2; j++) {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset3 = a_offset2 + lda;
a_offset4 = a_offset3 + lda;
a_offset += 2;
for (BLASLONG i = 0; i < m / 4; i++) {
*(b_offset + 0) = *(a_offset1 + 0);
*(b_offset + 1) = *(a_offset2 + 0);
*(b_offset + 2) = *(a_offset3 + 0);
*(b_offset + 3) = *(a_offset4 + 0);
*(b_offset + 4) = *(a_offset1 + 1);
*(b_offset + 5) = *(a_offset2 + 1);
*(b_offset + 6) = *(a_offset3 + 1);
*(b_offset + 7) = *(a_offset4 + 1);
b_offset += 8;
a_offset1 += 4 * lda;
a_offset2 += 4 * lda;
a_offset3 += 4 * lda;
a_offset4 += 4 * lda;
}
if (m & 3) {
BLASLONG rest = m & 3;
if (rest == 3) {
*(b_offset + 0) = *(a_offset1 + 0);
*(b_offset + 1) = *(a_offset2 + 0);
*(b_offset + 2) = *(a_offset3 + 0);
*(b_offset + 3) = *(a_offset1 + 1);
*(b_offset + 4) = *(a_offset2 + 1);
*(b_offset + 5) = *(a_offset3 + 1);
b_offset += 6;
} else if (rest == 2) {
*(b_offset + 0) = *(a_offset1 + 0);
*(b_offset + 1) = *(a_offset2 + 0);
*(b_offset + 2) = *(a_offset1 + 1);
*(b_offset + 3) = *(a_offset2 + 1);
b_offset += 4;
} else if (rest == 1) {
*(b_offset + 0) = *(a_offset1 + 0);
*(b_offset + 1) = *(a_offset1 + 1);
b_offset += 2;
}
}
}
if (n & 1) {
for (BLASLONG i = 0; i < m / 4; i++) {
*(b_offset + 0) = *(a_offset);
*(b_offset + 1) = *(a_offset + lda);
*(b_offset + 2) = *(a_offset + lda * 2);
*(b_offset + 3) = *(a_offset + lda * 3);
b_offset += 4;
a_offset += 4 * lda;
}
BLASLONG rest = m & 3;
if (rest == 3) {
*(b_offset + 0) = *(a_offset);
*(b_offset + 1) = *(a_offset + lda);
*(b_offset + 2) = *(a_offset + lda * 2);
} else if (rest == 2) {
*(b_offset + 0) = *(a_offset);
*(b_offset + 1) = *(a_offset + lda);
} else if (rest == 1) {
*(b_offset + 0) = *(a_offset);
}
}
return 0;
}