Skip to content

Commit acf6cab

Browse files
XiWeiGuXiWeiGu
committed
LoongArch64: Opt somatcopy_rn with LASX
1 parent 15edb44 commit acf6cab

File tree

2 files changed

+194
-0
lines changed

2 files changed

+194
-0
lines changed

‎kernel/loongarch64/KERNEL.LA464

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -174,4 +174,5 @@ DGEMM_SMALL_K_TT = dgemm_small_kernel_tt_lasx.S
174174
DGEMM_SMALL_K_B0_TT = dgemm_small_kernel_tt_lasx.S
175175

176176
SOMATCOPY_RT = somatcopy_rt_lasx.c
177+
SOMATCOPY_RN = somatcopy_rn_lasx.c
177178
endif

‎kernel/loongarch64/somatcopy_rn_lasx.c

Lines changed: 193 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,193 @@
1+
/***************************************************************************
2+
Copyright (c) 2024, The OpenBLAS Project
3+
All rights reserved.
4+
Redistribution and use in source and binary forms, with or without
5+
modification, are permitted provided that the following conditions are
6+
met:
7+
1. Redistributions of source code must retain the above copyright
8+
notice, this list of conditions and the following disclaimer.
9+
2. Redistributions in binary form must reproduce the above copyright
10+
notice, this list of conditions and the following disclaimer in
11+
the documentation and/or other materials provided with the
12+
distribution.
13+
3. Neither the name of the OpenBLAS project nor the names of
14+
its contributors may be used to endorse or promote products
15+
derived from this software without specific prior written permission.
16+
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
17+
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18+
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19+
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
20+
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21+
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22+
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23+
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24+
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
25+
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26+
*****************************************************************************/
27+
28+
#include "common.h"
29+
#define SAVE_4x1(b1, b2, b3, b4) \
30+
"vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" \
31+
"vstelm.w $vr"#b3",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b4",%1,0,0;add.d %1,%1,%3;\n\t"
32+
#define SAVE_4x2(b1, b2, b3, b4) \
33+
"vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" \
34+
"vstelm.d $vr"#b3",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b4",%1,0,0;add.d %1,%1,%3;\n\t"
35+
#define SAVE_4x4(b1, b2, b3, b4) \
36+
"vst $vr"#b1",%1,0;add.d %1,%1,%3;vst $vr"#b2",%1,0;add.d %1,%1,%3;\n\t" \
37+
"vst $vr"#b3",%1,0;add.d %1,%1,%3;vst $vr"#b4",%1,0;add.d %1,%1,%3;\n\t"
38+
#define SAVE_4x8(b1, b2, b3, b4) \
39+
"xvst $xr"#b1",%1,0;add.d %1,%1,%3;xvst $xr"#b2",%1,0;add.d %1,%1,%3;\n\t" \
40+
"xvst $xr"#b3",%1,0;add.d %1,%1,%3;xvst $xr"#b4",%1,0;add.d %1,%1,%3;\n\t"
41+
#define SAVE_4x16(b1, b2, b3, b4, b5, b6, b7, b8) \
42+
"xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t" \
43+
"xvst $xr"#b3",%1,0;xvst $xr"#b4",%1,32;add.d %1,%1,%3;\n\t" \
44+
"xvst $xr"#b5",%1,0;xvst $xr"#b6",%1,32;add.d %1,%1,%3;\n\t" \
45+
"xvst $xr"#b7",%1,0;xvst $xr"#b8",%1,32;add.d %1,%1,%3;\n\t"
46+
#define SAVE_2x16(b1, b2, b3, b4) \
47+
"xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t" \
48+
"xvst $xr"#b3",%1,0;xvst $xr"#b4",%1,32;add.d %1,%1,%3;\n\t"
49+
#define SAVE_2x8(b1, b2) \
50+
"xvst $xr"#b1",%1,0;add.d %1,%1,%3;xvst $xr"#b2",%1,0;add.d %1,%1,%3;\n\t"
51+
#define SAVE_2x4(b1, b2) \
52+
"vst $vr"#b1",%1,0;add.d %1,%1,%3;vst $vr"#b2",%1,0;add.d %1,%1,%3;\n\t"
53+
#define SAVE_2x2(b1, b2) \
54+
"vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t"
55+
#define SAVE_2x1(b1, b2) \
56+
"vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t"
57+
#define SAVE_1x16(b1, b2) \
58+
"xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t"
59+
#define SAVE_1x8(b1) \
60+
"xvst $xr"#b1",%1,0;add.d %1,%1,%3;\n\t"
61+
#define SAVE_1x4(b1) \
62+
"vst $vr"#b1",%1,0;add.d %1,%1,%3;\n\t"
63+
#define SAVE_1x2(b1) \
64+
"vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;\n\t"
65+
#define SAVE_1x1(b1) \
66+
"vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;\n\t"
67+
#define COPY_4x16 \
68+
"xvld $xr0,%0,0; xvld $xr4,%0,32; add.d %0,%0,%2; \n\t" \
69+
"xvld $xr1,%0,0; xvld $xr5,%0,32; add.d %0,%0,%2; \n\t" \
70+
"xvld $xr2,%0,0; xvld $xr6,%0,32; add.d %0,%0,%2; \n\t" \
71+
"xvld $xr3,%0,0; xvld $xr7,%0,32; add.d %0,%0,%2; \n\t" \
72+
"xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \
73+
"xvfmul.s $xr4,$xr4,$xr15;xvfmul.s $xr5,$xr5,$xr15;xvfmul.s $xr6,$xr6,$xr15;xvfmul.s $xr7,$xr7,$xr15 \n\t" \
74+
SAVE_4x16(0,4,1,5,2,6,3,7)
75+
#define COPY_4x8 \
76+
"xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \
77+
"xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \
78+
"xvld $xr2,%0,0; add.d %0,%0,%2; \n\t" \
79+
"xvld $xr3,%0,0; add.d %0,%0,%2; \n\t" \
80+
"xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \
81+
SAVE_4x8(0,1,2,3)
82+
#define COPY_4x4 \
83+
"vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \
84+
"vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \
85+
"vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \
86+
"vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \
87+
"vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \
88+
SAVE_4x4(0,1,2,3)
89+
#define COPY_4x2 \
90+
"vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \
91+
"vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \
92+
"vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \
93+
"vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \
94+
"vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \
95+
SAVE_4x2(0,1,2,3)
96+
#define COPY_4x1 \
97+
"fld.s $f0,%0,0; add.d %0,%0,%2; \n\t" \
98+
"fld.s $f1,%0,0; add.d %0,%0,%2; \n\t" \
99+
"fld.s $f2,%0,0; add.d %0,%0,%2; \n\t" \
100+
"fld.s $f3,%0,0; add.d %0,%0,%2; \n\t" \
101+
"fmul.s $f0,$f0,$f15;fmul.s $f1,$f1,$f15;fmul.s $f2,$f2,$f15;fmul.s $f3,$f3,$f15 \n\t" \
102+
SAVE_4x1(0,1,2,3)
103+
#define COPY_2x16 \
104+
"xvld $xr0,%0,0; xvld $xr2,%0,32; add.d %0,%0,%2; \n\t" \
105+
"xvld $xr1,%0,0; xvld $xr3,%0,32; add.d %0,%0,%2; \n\t" \
106+
"xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \
107+
SAVE_2x16(0,2,1,3)
108+
#define COPY_2x8 \
109+
"xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \
110+
"xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \
111+
"xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15; \n\t" \
112+
SAVE_2x8(0,1)
113+
#define COPY_2x4 \
114+
"vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \
115+
"vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \
116+
"vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15; \n\t" \
117+
SAVE_2x4(0,1)
118+
#define COPY_2x2 \
119+
"fld.d $f0,%0,0;add.d %0,%0,%2; \n\t" \
120+
"fld.d $f1,%0,0;add.d %0,%0,%2; \n\t" \
121+
"vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15; \n\t" \
122+
SAVE_2x2(0,1)
123+
#define COPY_2x1 \
124+
"fld.s $f0,%0,0;add.d %0,%0,%2; \n\t" \
125+
"fld.s $f1,%0,0;add.d %0,%0,%2; \n\t" \
126+
"fmul.s $f0,$f0,$f15;fmul.s $f1,$f1,$f15; \n\t" \
127+
SAVE_2x1(0,1)
128+
#define COPY_1x16 \
129+
"xvld $xr0,%0,0; xvld $xr1,%0,32; add.d %0,%0,%2; \n\t" \
130+
"xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15; \n\t" \
131+
SAVE_1x16(0,1)
132+
#define COPY_1x8 \
133+
"xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \
134+
"xvfmul.s $xr0,$xr0,$xr15; \n\t" \
135+
SAVE_1x8(0)
136+
#define COPY_1x4 \
137+
"vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \
138+
"vfmul.s $vr0,$vr0,$vr15; \n\t" \
139+
SAVE_1x4(0)
140+
#define COPY_1x2 \
141+
"fld.d $f0,%0,0;add.d %0,%0,%2; \n\t" \
142+
"vfmul.s $vr0,$vr0,$vr15; \n\t" \
143+
SAVE_1x2(0)
144+
#define COPY_1x1 \
145+
"fld.s $f0,%0,0;add.d %0,%0,%2; \n\t" \
146+
"fmul.s $f0,$f0,$f15; \n\t" \
147+
SAVE_1x1(0)
148+
#define ROWS_OF_BLOCK 128
149+
#define COMPUTE(ndim) \
150+
src = src_base; dst = dst_base; \
151+
__asm__ __volatile__( \
152+
"xvldrepl.w $xr15, %6, 0 \n\t" \
153+
"srli.d $r6, %5, 2 \n\t" \
154+
"beqz $r6, "#ndim"3f \n\t" \
155+
#ndim"4: \n\t" \
156+
COPY_4x##ndim \
157+
"addi.d $r6, $r6, -1 \n\t" \
158+
"bnez $r6, "#ndim"4b \n\t" \
159+
#ndim"3: \n\t" \
160+
"andi $r6, %5, 2 \n\t" \
161+
"beqz $r6, "#ndim"1f \n\t" \
162+
#ndim"2: \n\t" \
163+
COPY_2x##ndim \
164+
#ndim"1: \n\t" \
165+
"andi $r6, %5, 1 \n\t" \
166+
"beqz $r6, "#ndim"0f \n\t" \
167+
COPY_1x##ndim \
168+
#ndim"0: \n\t" \
169+
:"+r"(src),"+r"(dst),"+r"(src_ld_bytes),"+r"(dst_ld_bytes),"+r"(dst_tmp) \
170+
:"r"(num_rows),"r"(&ALPHA) \
171+
:"memory", "$r6", "$f0", "$f1", "$f2", "$f3", "$f4", "$f5", "$f6", "$f7", "$f8", "$f9", "$f10", "$f11", "$f15" \
172+
);
173+
174+
int CNAME(BLASLONG rows, BLASLONG cols, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG ldb){
175+
float *src, *dst, *dst_tmp=0, *src_base, *dst_base;
176+
uint64_t src_ld_bytes = (uint64_t)lda * sizeof(float), dst_ld_bytes = (uint64_t)ldb * sizeof(float), num_rows = 0;
177+
BLASLONG cols_left, rows_done; float ALPHA = alpha;
178+
if (ALPHA == 0.0) {
179+
dst_base = b;
180+
for (cols_left = cols; cols_left > 0; cols_left--) {memset(dst_base, 0, rows * sizeof(float)); dst_base += ldb;}
181+
return 0;
182+
}
183+
for (rows_done = 0; rows_done < rows; rows_done += num_rows) {
184+
num_rows = rows - rows_done;
185+
if (num_rows > ROWS_OF_BLOCK) num_rows = ROWS_OF_BLOCK;
186+
cols_left = cols; src_base = a + (int64_t)lda * (int64_t)rows_done; dst_base = b + rows_done;
187+
for (;cols_left > 15; cols_left -= 16) {COMPUTE(16) src_base += 16; dst_base += 16;}
188+
for (;cols_left > 7; cols_left -= 8) {COMPUTE(8) src_base += 8; dst_base += 8;}
189+
for (;cols_left > 3; cols_left -= 4) {COMPUTE(4) src_base += 4; dst_base += 4;}
190+
for (;cols_left > 1; cols_left -= 2) {COMPUTE(2) src_base += 2; dst_base += 2;}
191+
if (cols_left > 0) {COMPUTE(1) src_base ++; dst_base ++;}
192+
}
193+
}

0 commit comments

Comments
 (0)