1131 lines
32 KiB
C
1131 lines
32 KiB
C
/*******************************************************************************
|
|
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.
|
|
*******************************************************************************/
|
|
|
|
#include "common.h"
|
|
#include "macros_msa.h"
|
|
|
|
#define ENABLE_PREFETCH
|
|
|
|
#ifdef ENABLE_PREFETCH
|
|
inline static void prefetch_load_lf(unsigned char *src) {
|
|
__asm__ __volatile__("pref 0, 0(%[src]) \n\t" : : [src] "r"(src));
|
|
}
|
|
#endif
|
|
|
|
static void __attribute__ ((noinline))
|
|
dgemmkernel_8x4_core_msa(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
|
|
FLOAT *C, BLASLONG ldc)
|
|
{
|
|
BLASLONG i, j, l;
|
|
FLOAT *pc0, *pc1, *pc2, *pc3, *pa0, *pb0;
|
|
v2f64 v_alpha = {alpha, alpha};
|
|
v2f64 src_a0, src_a1, src_a2, src_a3, src_b, src_b0, src_b1;
|
|
v2f64 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
|
|
v2f64 res0, res1, res2, res3, res4, res5, res6, res7;
|
|
v2f64 res8, res9, res10, res11, res12, res13, res14, res15;
|
|
|
|
for (j = (n >> 2); j--;)
|
|
{
|
|
pc0 = C;
|
|
pc1 = pc0 + ldc;
|
|
pc2 = pc1 + ldc;
|
|
pc3 = pc2 + ldc;
|
|
|
|
pa0 = A;
|
|
|
|
for (i = (m >> 3); i--;)
|
|
{
|
|
pb0 = B;
|
|
|
|
#ifdef ENABLE_PREFETCH
|
|
prefetch_load_lf((unsigned char *)(pa0 + 8));
|
|
prefetch_load_lf((unsigned char *)(pa0 + 12));
|
|
prefetch_load_lf((unsigned char *)(pa0 + 16));
|
|
prefetch_load_lf((unsigned char *)(pa0 + 20));
|
|
|
|
prefetch_load_lf((unsigned char *)(pb0 + 4));
|
|
prefetch_load_lf((unsigned char *)(pb0 + 8));
|
|
prefetch_load_lf((unsigned char *)(pb0 + 12));
|
|
#endif
|
|
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 = src_a0 * src_b;
|
|
res1 = src_a1 * src_b;
|
|
res2 = src_a2 * src_b;
|
|
res3 = src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res4 = src_a0 * src_b;
|
|
res5 = src_a1 * src_b;
|
|
res6 = src_a2 * src_b;
|
|
res7 = src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res8 = src_a0 * src_b;
|
|
res9 = src_a1 * src_b;
|
|
res10 = src_a2 * src_b;
|
|
res11 = src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res12 = src_a0 * src_b;
|
|
res13 = src_a1 * src_b;
|
|
res14 = src_a2 * src_b;
|
|
res15 = src_a3 * src_b;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
#ifdef ENABLE_PREFETCH
|
|
prefetch_load_lf((unsigned char *)(pa0 + 16));
|
|
prefetch_load_lf((unsigned char *)(pa0 + 20));
|
|
prefetch_load_lf((unsigned char *)(pa0 + 24));
|
|
prefetch_load_lf((unsigned char *)(pa0 + 28));
|
|
#endif
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
res6 += src_a2 * src_b;
|
|
res7 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res8 += src_a0 * src_b;
|
|
res9 += src_a1 * src_b;
|
|
res10 += src_a2 * src_b;
|
|
res11 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res12 += src_a0 * src_b;
|
|
res13 += src_a1 * src_b;
|
|
res14 += src_a2 * src_b;
|
|
res15 += src_a3 * src_b;
|
|
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
#ifdef ENABLE_PREFETCH
|
|
prefetch_load_lf((unsigned char *)(pb0 + 8));
|
|
prefetch_load_lf((unsigned char *)(pb0 + 12));
|
|
#endif
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
res6 += src_a2 * src_b;
|
|
res7 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res8 += src_a0 * src_b;
|
|
res9 += src_a1 * src_b;
|
|
res10 += src_a2 * src_b;
|
|
res11 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res12 += src_a0 * src_b;
|
|
res13 += src_a1 * src_b;
|
|
res14 += src_a2 * src_b;
|
|
res15 += src_a3 * src_b;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
res6 += src_a2 * src_b;
|
|
res7 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res8 += src_a0 * src_b;
|
|
res9 += src_a1 * src_b;
|
|
res10 += src_a2 * src_b;
|
|
res11 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res12 += src_a0 * src_b;
|
|
res13 += src_a1 * src_b;
|
|
res14 += src_a2 * src_b;
|
|
res15 += src_a3 * src_b;
|
|
}
|
|
|
|
#ifdef ENABLE_PREFETCH
|
|
prefetch_load_lf((unsigned char *)(pc0 + 8));
|
|
prefetch_load_lf((unsigned char *)(pc1 + 8));
|
|
prefetch_load_lf((unsigned char *)(pc2 + 8));
|
|
prefetch_load_lf((unsigned char *)(pc3 + 8));
|
|
#endif
|
|
LD_DP4(pc0, 2, dst0, dst1, dst2, dst3);
|
|
LD_DP4(pc1, 2, dst4, dst5, dst6, dst7);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
dst1 += res1 * v_alpha;
|
|
dst2 += res2 * v_alpha;
|
|
dst3 += res3 * v_alpha;
|
|
dst4 += res4 * v_alpha;
|
|
dst5 += res5 * v_alpha;
|
|
dst6 += res6 * v_alpha;
|
|
dst7 += res7 * v_alpha;
|
|
|
|
ST_DP4_INC(dst0, dst1, dst2, dst3, pc0, 2);
|
|
ST_DP4_INC(dst4, dst5, dst6, dst7, pc1, 2);
|
|
|
|
LD_DP4(pc2, 2, dst0, dst1, dst2, dst3);
|
|
LD_DP4(pc3, 2, dst4, dst5, dst6, dst7);
|
|
|
|
dst0 += res8 * v_alpha;
|
|
dst1 += res9 * v_alpha;
|
|
dst2 += res10 * v_alpha;
|
|
dst3 += res11 * v_alpha;
|
|
dst4 += res12 * v_alpha;
|
|
dst5 += res13 * v_alpha;
|
|
dst6 += res14 * v_alpha;
|
|
dst7 += res15 * v_alpha;
|
|
|
|
ST_DP4_INC(dst0, dst1, dst2, dst3, pc2, 2);
|
|
ST_DP4_INC(dst4, dst5, dst6, dst7, pc3, 2);
|
|
}
|
|
|
|
l = (k << 2);
|
|
B = B + l;
|
|
i = (ldc << 2);
|
|
C = C + i;
|
|
}
|
|
}
|
|
|
|
static void __attribute__ ((noinline))
|
|
dgemmkernel_7x4_core_msa(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
|
|
FLOAT *C, BLASLONG ldc)
|
|
{
|
|
BLASLONG i, j, l;
|
|
FLOAT *pc0, *pc1, *pc2, *pc3, *pa0, *pb0;
|
|
FLOAT tmp0, tmp1, tmp2, tmp3;
|
|
FLOAT a0, b0, b1, b2, b3;
|
|
v2f64 v_alpha = {alpha, alpha};
|
|
v2f64 src_a0, src_a1, src_b, src_b0, src_b1;
|
|
v2f64 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
|
|
v2f64 res0, res1, res2, res3, res4, res5, res6, res7;
|
|
|
|
for (j = (n >> 2); j--;)
|
|
{
|
|
|
|
pc0 = C + 8 * (m >> 3);
|
|
pc1 = pc0 + ldc;
|
|
pc2 = pc1 + ldc;
|
|
pc3 = pc2 + ldc;
|
|
|
|
pa0 = A + k * 8 * (m >> 3);
|
|
|
|
if (m & 4)
|
|
{
|
|
pb0 = B;
|
|
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 = src_a0 * src_b;
|
|
res1 = src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res2 = src_a0 * src_b;
|
|
res3 = src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res4 = src_a0 * src_b;
|
|
res5 = src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res6 = src_a0 * src_b;
|
|
res7 = src_a1 * src_b;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res2 += src_a0 * src_b;
|
|
res3 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res6 += src_a0 * src_b;
|
|
res7 += src_a1 * src_b;
|
|
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res2 += src_a0 * src_b;
|
|
res3 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res6 += src_a0 * src_b;
|
|
res7 += src_a1 * src_b;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res2 += src_a0 * src_b;
|
|
res3 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res6 += src_a0 * src_b;
|
|
res7 += src_a1 * src_b;
|
|
}
|
|
|
|
LD_DP2(pc0, 2, dst0, dst1);
|
|
LD_DP2(pc1, 2, dst2, dst3);
|
|
LD_DP2(pc2, 2, dst4, dst5);
|
|
LD_DP2(pc3, 2, dst6, dst7);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
dst1 += res1 * v_alpha;
|
|
dst2 += res2 * v_alpha;
|
|
dst3 += res3 * v_alpha;
|
|
dst4 += res4 * v_alpha;
|
|
dst5 += res5 * v_alpha;
|
|
dst6 += res6 * v_alpha;
|
|
dst7 += res7 * v_alpha;
|
|
|
|
ST_DP2_INC(dst0, dst1, pc0, 2);
|
|
ST_DP2_INC(dst2, dst3, pc1, 2);
|
|
ST_DP2_INC(dst4, dst5, pc2, 2);
|
|
ST_DP2_INC(dst6, dst7, pc3, 2);
|
|
}
|
|
|
|
if (m & 2)
|
|
{
|
|
pb0 = B;
|
|
|
|
src_a0 = LD_DP(pa0);
|
|
pa0 += 2;
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 = src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res1 = src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res2 = src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res3 = src_a0 * src_b;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
src_a0 = LD_DP(pa0);
|
|
pa0 += 2;
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res1 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res2 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res3 += src_a0 * src_b;
|
|
|
|
src_a0 = LD_DP(pa0);
|
|
pa0 += 2;
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res1 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res2 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res3 += src_a0 * src_b;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
src_a0 = LD_DP(pa0);
|
|
pa0 += 2;
|
|
LD_DP2_INC(pb0, 2, src_b0, src_b1);
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res1 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res2 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1);
|
|
res3 += src_a0 * src_b;
|
|
}
|
|
|
|
dst0 = LD_DP(pc0);
|
|
dst1 = LD_DP(pc1);
|
|
dst2 = LD_DP(pc2);
|
|
dst3 = LD_DP(pc3);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
dst1 += res1 * v_alpha;
|
|
dst2 += res2 * v_alpha;
|
|
dst3 += res3 * v_alpha;
|
|
|
|
ST_DP(dst0, pc0);
|
|
ST_DP(dst1, pc1);
|
|
ST_DP(dst2, pc2);
|
|
ST_DP(dst3, pc3);
|
|
|
|
pc0 += 2;
|
|
pc1 += 2;
|
|
pc2 += 2;
|
|
pc3 += 2;
|
|
}
|
|
|
|
if (m & 1)
|
|
{
|
|
pb0 = B;
|
|
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 = a0 * b0;
|
|
|
|
b1 = pb0[1];
|
|
tmp1 = a0 * b1;
|
|
|
|
b2 = pb0[2];
|
|
tmp2 = a0 * b2;
|
|
|
|
b3 = pb0[3];
|
|
tmp3 = a0 * b3;
|
|
|
|
pa0 += 1;
|
|
pb0 += 4;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
b1 = pb0[1];
|
|
tmp1 += a0 * b1;
|
|
|
|
b2 = pb0[2];
|
|
tmp2 += a0 * b2;
|
|
|
|
b3 = pb0[3];
|
|
tmp3 += a0 * b3;
|
|
|
|
pa0 += 1;
|
|
pb0 += 4;
|
|
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
b1 = pb0[1];
|
|
tmp1 += a0 * b1;
|
|
|
|
b2 = pb0[2];
|
|
tmp2 += a0 * b2;
|
|
|
|
b3 = pb0[3];
|
|
tmp3 += a0 * b3;
|
|
|
|
pa0 += 1;
|
|
pb0 += 4;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
b1 = pb0[1];
|
|
tmp1 += a0 * b1;
|
|
|
|
b2 = pb0[2];
|
|
tmp2 += a0 * b2;
|
|
|
|
b3 = pb0[3];
|
|
tmp3 += a0 * b3;
|
|
|
|
pa0 += 1;
|
|
pb0 += 4;
|
|
}
|
|
|
|
tmp0 = alpha * tmp0;
|
|
tmp1 = alpha * tmp1;
|
|
tmp2 = alpha * tmp2;
|
|
tmp3 = alpha * tmp3;
|
|
|
|
pc0[0] += tmp0;
|
|
pc1[0] += tmp1;
|
|
pc2[0] += tmp2;
|
|
pc3[0] += tmp3;
|
|
|
|
pc0 += 1;
|
|
pc1 += 1;
|
|
pc2 += 1;
|
|
pc3 += 1;
|
|
}
|
|
|
|
l = (k << 2);
|
|
B = B + l;
|
|
i = (ldc << 2);
|
|
C = C + i;
|
|
}
|
|
}
|
|
|
|
static void __attribute__ ((noinline))
|
|
dgemmkernel_8x4_non_core_msa(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
|
|
FLOAT *C, BLASLONG ldc)
|
|
{
|
|
BLASLONG i, l;
|
|
FLOAT *pc0, *pc1, *pa0, *pb0;
|
|
FLOAT tmp0, tmp1;
|
|
FLOAT a0, b0, b1;
|
|
v2f64 v_alpha = {alpha, alpha};
|
|
v2f64 src_a0, src_a1, src_a2, src_a3, src_b, src_b0;
|
|
v2f64 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
|
|
v2f64 res0, res1, res2, res3, res4, res5, res6, res7;
|
|
|
|
if (n & 2)
|
|
{
|
|
pc0 = C;
|
|
pc1 = pc0 + ldc;
|
|
|
|
pa0 = A;
|
|
|
|
for (i = (m >> 3); i--;)
|
|
{
|
|
pb0 = B;
|
|
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 = src_a0 * src_b;
|
|
res1 = src_a1 * src_b;
|
|
res2 = src_a2 * src_b;
|
|
res3 = src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res4 = src_a0 * src_b;
|
|
res5 = src_a1 * src_b;
|
|
res6 = src_a2 * src_b;
|
|
res7 = src_a3 * src_b;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
res6 += src_a2 * src_b;
|
|
res7 += src_a3 * src_b;
|
|
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
res6 += src_a2 * src_b;
|
|
res7 += src_a3 * src_b;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res4 += src_a0 * src_b;
|
|
res5 += src_a1 * src_b;
|
|
res6 += src_a2 * src_b;
|
|
res7 += src_a3 * src_b;
|
|
}
|
|
|
|
LD_DP4(pc0, 2, dst0, dst1, dst2, dst3);
|
|
LD_DP4(pc1, 2, dst4, dst5, dst6, dst7);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
dst1 += res1 * v_alpha;
|
|
dst2 += res2 * v_alpha;
|
|
dst3 += res3 * v_alpha;
|
|
dst4 += res4 * v_alpha;
|
|
dst5 += res5 * v_alpha;
|
|
dst6 += res6 * v_alpha;
|
|
dst7 += res7 * v_alpha;
|
|
|
|
ST_DP4_INC(dst0, dst1, dst2, dst3, pc0, 2);
|
|
ST_DP4_INC(dst4, dst5, dst6, dst7, pc1, 2);
|
|
}
|
|
|
|
if (m & 4)
|
|
{
|
|
pb0 = B;
|
|
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 = src_a0 * src_b;
|
|
res1 = src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res2 = src_a0 * src_b;
|
|
res3 = src_a1 * src_b;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res2 += src_a0 * src_b;
|
|
res3 += src_a1 * src_b;
|
|
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res2 += src_a0 * src_b;
|
|
res3 += src_a1 * src_b;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res2 += src_a0 * src_b;
|
|
res3 += src_a1 * src_b;
|
|
}
|
|
|
|
LD_DP2(pc0, 2, dst0, dst1);
|
|
LD_DP2(pc1, 2, dst2, dst3);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
dst1 += res1 * v_alpha;
|
|
dst2 += res2 * v_alpha;
|
|
dst3 += res3 * v_alpha;
|
|
|
|
ST_DP2_INC(dst0, dst1, pc0, 2);
|
|
ST_DP2_INC(dst2, dst3, pc1, 2);
|
|
}
|
|
|
|
if (m & 2)
|
|
{
|
|
pb0 = B;
|
|
|
|
src_a0 = LD_DP(pa0);
|
|
pa0 += 2;
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 = src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res1 = src_a0 * src_b;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
src_a0 = LD_DP(pa0);
|
|
pa0 += 2;
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res1 += src_a0 * src_b;
|
|
|
|
src_a0 = LD_DP(pa0);
|
|
pa0 += 2;
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res1 += src_a0 * src_b;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
src_a0 = LD_DP(pa0);
|
|
pa0 += 2;
|
|
src_b0 = LD_DP(pb0);
|
|
pb0 += 2;
|
|
|
|
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res0 += src_a0 * src_b;
|
|
|
|
src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0);
|
|
res1 += src_a0 * src_b;
|
|
}
|
|
|
|
dst0 = LD_DP(pc0);
|
|
dst1 = LD_DP(pc1);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
dst1 += res1 * v_alpha;
|
|
|
|
ST_DP(dst0, pc0);
|
|
ST_DP(dst1, pc1);
|
|
|
|
pc0 += 2;
|
|
pc1 += 2;
|
|
}
|
|
|
|
if (m & 1)
|
|
{
|
|
pb0 = B;
|
|
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 = a0 * b0;
|
|
|
|
b1 = pb0[1];
|
|
tmp1 = a0 * b1;
|
|
|
|
pa0 += 1;
|
|
pb0 += 2;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
b1 = pb0[1];
|
|
tmp1 += a0 * b1;
|
|
|
|
pa0 += 1;
|
|
pb0 += 2;
|
|
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
b1 = pb0[1];
|
|
tmp1 += a0 * b1;
|
|
|
|
pa0 += 1;
|
|
pb0 += 2;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
b1 = pb0[1];
|
|
tmp1 += a0 * b1;
|
|
|
|
pa0 += 1;
|
|
pb0 += 2;
|
|
}
|
|
|
|
tmp0 = alpha * tmp0;
|
|
tmp1 = alpha * tmp1;
|
|
|
|
pc0[0] += tmp0;
|
|
pc1[0] += tmp1;
|
|
|
|
pc0 += 1;
|
|
pc1 += 1;
|
|
}
|
|
|
|
l = (k << 1);
|
|
B = B + l;
|
|
i = (ldc << 1);
|
|
C = C + i;
|
|
}
|
|
|
|
if (n & 1)
|
|
{
|
|
pc0 = C;
|
|
pa0 = A;
|
|
|
|
for (i = (m >> 3); i--;)
|
|
{
|
|
pb0 = B;
|
|
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 = src_a0 * src_b;
|
|
res1 = src_a1 * src_b;
|
|
res2 = src_a2 * src_b;
|
|
res3 = src_a3 * src_b;
|
|
|
|
pb0 += 1;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
pb0 += 1;
|
|
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
pb0 += 1;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
res2 += src_a2 * src_b;
|
|
res3 += src_a3 * src_b;
|
|
|
|
pb0 += 1;
|
|
}
|
|
|
|
LD_DP4(pc0, 2, dst0, dst1, dst2, dst3);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
dst1 += res1 * v_alpha;
|
|
dst2 += res2 * v_alpha;
|
|
dst3 += res3 * v_alpha;
|
|
|
|
ST_DP4_INC(dst0, dst1, dst2, dst3, pc0, 2);
|
|
}
|
|
|
|
if (m & 4)
|
|
{
|
|
pb0 = B;
|
|
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 = src_a0 * src_b;
|
|
res1 = src_a1 * src_b;
|
|
|
|
pb0 += 1;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
pb0 += 1;
|
|
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
pb0 += 1;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
LD_DP2_INC(pa0, 2, src_a0, src_a1);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
res1 += src_a1 * src_b;
|
|
|
|
pb0 += 1;
|
|
}
|
|
|
|
LD_DP2(pc0, 2, dst0, dst1);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
dst1 += res1 * v_alpha;
|
|
|
|
ST_DP2_INC(dst0, dst1, pc0, 2);
|
|
}
|
|
|
|
if (m & 2)
|
|
{
|
|
pb0 = B;
|
|
|
|
src_a0 = LD_DP(pa0);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 = src_a0 * src_b;
|
|
|
|
pa0 += 2;
|
|
pb0 += 1;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
src_a0 = LD_DP(pa0);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
|
|
pa0 += 2;
|
|
pb0 += 1;
|
|
|
|
src_a0 = LD_DP(pa0);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
|
|
pa0 += 2;
|
|
pb0 += 1;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
src_a0 = LD_DP(pa0);
|
|
src_b[0] = pb0[0];
|
|
src_b[1] = pb0[0];
|
|
|
|
res0 += src_a0 * src_b;
|
|
|
|
pa0 += 2;
|
|
pb0 += 1;
|
|
}
|
|
|
|
dst0 = LD_DP(pc0);
|
|
|
|
dst0 += res0 * v_alpha;
|
|
|
|
ST_DP(dst0, pc0);
|
|
|
|
pc0 += 2;
|
|
}
|
|
|
|
if (m & 1)
|
|
{
|
|
pb0 = B;
|
|
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 = a0 * b0;
|
|
|
|
pa0 += 1;
|
|
pb0 += 1;
|
|
|
|
for (l = ((k - 1) >> 1); l--;)
|
|
{
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
pa0 += 1;
|
|
pb0 += 1;
|
|
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
pa0 += 1;
|
|
pb0 += 1;
|
|
}
|
|
|
|
if ((k - 1) & 1)
|
|
{
|
|
a0 = pa0[0];
|
|
b0 = pb0[0];
|
|
tmp0 += a0 * b0;
|
|
|
|
pa0 += 1;
|
|
pb0 += 1;
|
|
}
|
|
|
|
pc0[0] += alpha * tmp0;
|
|
|
|
pc0 += 1;
|
|
}
|
|
|
|
l = (k << 0);
|
|
B = B + l;
|
|
i = (ldc << 0);
|
|
C = C + i;
|
|
}
|
|
}
|
|
|
|
int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
|
|
FLOAT *C, BLASLONG ldc
|
|
#ifdef TRMMKERNEL
|
|
, BLASLONG offset
|
|
#endif
|
|
)
|
|
{
|
|
if (n >> 2)
|
|
{
|
|
if (m >> 3)
|
|
dgemmkernel_8x4_core_msa(m, n, k, alpha, A, B, C, ldc);
|
|
|
|
if (m & 7)
|
|
dgemmkernel_7x4_core_msa(m, n, k, alpha, A, B, C, ldc);
|
|
}
|
|
|
|
if (n & 3)
|
|
{
|
|
B = B + (k << 2) * (n >> 2);
|
|
C = C + (ldc << 2) * (n >> 2);
|
|
|
|
dgemmkernel_8x4_non_core_msa(m, n, k, alpha, A, B, C, ldc);
|
|
}
|
|
|
|
return 0;
|
|
}
|