Skip to content

Commit 9d6df1d

Browse files
martin-frbgmartin-frbg
authored
Merge pull request #5422 from ChipKerchner/addRVVVectorizedPacking
Add and use vectorized packing in ZVL128B and ZVL256B for RISCV
2 parents f3b2a15 + 64401b4 commit 9d6df1d

File tree

6 files changed

+709
-1
lines changed

6 files changed

+709
-1
lines changed

‎kernel/riscv64/KERNEL.RISCV64_ZVL128B

Lines changed: 11 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -96,14 +96,24 @@ CGEMVTKERNEL = zgemv_t_rvv.c
9696
ZGEMVTKERNEL = zgemv_t_rvv.c
9797

9898
SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_zvl128b.c
99+
ifneq ($(filter $(SGEMM_UNROLL_N),4 8 16),)
100+
SGEMMONCOPY = gemm_ncopy_$(SGEMM_UNROLL_N)_rvv.c
101+
SGEMMOTCOPY = gemm_tcopy_$(SGEMM_UNROLL_N)_rvv.c
102+
else
99103
SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
100104
SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
105+
endif
101106
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
102107
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
103108

104109
ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
110+
ifneq ($(filter $(SGEMM_UNROLL_M),4 8 16),)
111+
SGEMMINCOPY = gemm_ncopy_$(SGEMM_UNROLL_M)_rvv.c
112+
SGEMMITCOPY = gemm_tcopy_$(SGEMM_UNROLL_M)_rvv.c
113+
else
105114
SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
106115
SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
116+
endif
107117
SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
108118
SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
109119
endif
@@ -255,4 +265,4 @@ SHGEMMOTCOPYOBJ = shgemm_otcopy$(TSUFFIX).$(SUFFIX)
255265
ifndef SHGEMM_BETA
256266
SHGEMM_BETA = gemm_beta_rvv.c
257267
endif
258-
endif
268+
endif

‎kernel/riscv64/KERNEL.RISCV64_ZVL256B

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -96,13 +96,23 @@ CTRMMKERNEL = ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N)_zvl256b.c
9696
ZTRMMKERNEL = ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N)_zvl256b.c
9797

9898
SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_zvl256b.c
99+
ifneq ($(filter $(SGEMM_UNROLL_N),4 8 16),)
100+
SGEMMONCOPY = gemm_ncopy_$(SGEMM_UNROLL_N)_rvv.c
101+
SGEMMOTCOPY = gemm_tcopy_$(SGEMM_UNROLL_N)_rvv.c
102+
else
99103
SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
100104
SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
105+
endif
101106
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
102107
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
103108
ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
109+
ifneq ($(filter $(SGEMM_UNROLL_M),4 8 16),)
110+
SGEMMINCOPY = gemm_ncopy_$(SGEMM_UNROLL_M)_rvv.c
111+
SGEMMITCOPY = gemm_tcopy_$(SGEMM_UNROLL_M)_rvv.c
112+
else
104113
SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
105114
SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
115+
endif
106116
SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
107117
SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
108118
endif

‎kernel/riscv64/gemm_ncopy_16_rvv.c

Lines changed: 325 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,325 @@
1+
/***************************************************************************
2+
Copyright (c) 2025, 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+
30+
#if !defined(DOUBLE)
31+
#define VSETVL(n) __riscv_vsetvl_e32m1(n)
32+
#define FLOAT_V_T vfloat32m1_t
33+
#define FLOAT_VX2_T vfloat32m1x2_t
34+
#define FLOAT_VX4_T vfloat32m1x4_t
35+
#define FLOAT_VX8_T vfloat32m1x8_t
36+
#define VLSEG_FLOAT __riscv_vlse32_v_f32m1
37+
#define VLSSEG2_FLOAT __riscv_vlsseg2e32_v_f32m1x2
38+
#define VLSSEG4_FLOAT __riscv_vlsseg4e32_v_f32m1x4
39+
#define VLSSEG8_FLOAT __riscv_vlsseg8e32_v_f32m1x8
40+
#define VGET_VX2 __riscv_vget_v_f32m1x2_f32m1
41+
#define VGET_VX4 __riscv_vget_v_f32m1x4_f32m1
42+
#define VGET_VX8 __riscv_vget_v_f32m1x8_f32m1
43+
#define VSET_VX2 __riscv_vset_v_f32m1_f32m1x2
44+
#define VSET_VX4 __riscv_vset_v_f32m1_f32m1x4
45+
#define VSET_VX8 __riscv_vset_v_f32m1_f32m1x8
46+
#define VLEV_FLOAT __riscv_vle32_v_f32m1
47+
#define VSEV_FLOAT __riscv_vse32_v_f32m1
48+
#define VSSEG2_FLOAT __riscv_vsseg2e32_v_f32m1x2
49+
#define VSSEG4_FLOAT __riscv_vsseg4e32_v_f32m1x4
50+
#define VSSEG8_FLOAT __riscv_vsseg8e32_v_f32m1x8
51+
#else
52+
#define VSETVL(n) __riscv_vsetvl_e64m1(n)
53+
#define FLOAT_V_T vfloat64m1_t
54+
#define FLOAT_VX2_T vfloat64m1x2_t
55+
#define FLOAT_VX4_T vfloat64m1x4_t
56+
#define FLOAT_VX8_T vfloat64m1x8_t
57+
#define VLSEG_FLOAT __riscv_vlse64_v_f64m1
58+
#define VLSSEG2_FLOAT __riscv_vlsseg2e64_v_f64m1x2
59+
#define VLSSEG4_FLOAT __riscv_vlsseg4e64_v_f64m1x4
60+
#define VLSSEG8_FLOAT __riscv_vlsseg8e64_v_f64m1x8
61+
#define VGET_VX2 __riscv_vget_v_f64m1x2_f64m1
62+
#define VGET_VX4 __riscv_vget_v_f64m1x4_f64m1
63+
#define VGET_VX8 __riscv_vget_v_f64m1x8_f64m1
64+
#define VSET_VX2 __riscv_vset_v_f64m1_f64m1x2
65+
#define VSET_VX4 __riscv_vset_v_f64m1_f64m1x4
66+
#define VSET_VX8 __riscv_vset_v_f64m1_f64m1x8
67+
#define VLEV_FLOAT __riscv_vle64_v_f64m1
68+
#define VSEV_FLOAT __riscv_vse64_v_f64m1
69+
#define VSSEG2_FLOAT __riscv_vsseg2e64_v_f64m1x2
70+
#define VSSEG4_FLOAT __riscv_vsseg4e64_v_f64m1x4
71+
#define VSSEG8_FLOAT __riscv_vsseg8e64_v_f64m1x8
72+
#endif
73+
74+
// Optimizes the implementation in ../generic/gemm_ncopy_16.c
75+
76+
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
77+
{
78+
BLASLONG i, j;
79+
80+
FLOAT *a_offset;
81+
FLOAT *a_offset1, *a_offset2, *a_offset3, *a_offset4;
82+
FLOAT *a_offset5, *a_offset6, *a_offset7, *a_offset8;
83+
FLOAT *b_offset;
84+
85+
FLOAT_V_T v1, v2, v3, v4, v5, v6, v7, v8;
86+
FLOAT_V_T v9, v10, v11, v12, v13, v14, v15, v16;
87+
FLOAT_VX2_T vx2, vx21;
88+
FLOAT_VX4_T vx4, vx41;
89+
FLOAT_VX8_T vx8, vx81;
90+
91+
size_t vl;
92+
93+
//fprintf(stderr, "gemm_ncopy_16 m=%ld n=%ld lda=%ld\n", m, n, lda);
94+
95+
a_offset = a;
96+
b_offset = b;
97+
98+
j = (n >> 4);
99+
if (j) {
100+
vl = VSETVL(8);
101+
102+
do {
103+
a_offset1 = a_offset;
104+
a_offset2 = a_offset1 + lda * 8;
105+
a_offset += 16 * lda;
106+
107+
i = m >> 3;
108+
if (i) {
109+
do {
110+
vx8 = VLSSEG8_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
111+
vx81 = VLSSEG8_FLOAT(a_offset2, lda * sizeof(FLOAT), vl);
112+
113+
v1 = VGET_VX8(vx8, 0);
114+
v2 = VGET_VX8(vx8, 1);
115+
v3 = VGET_VX8(vx8, 2);
116+
v4 = VGET_VX8(vx8, 3);
117+
v5 = VGET_VX8(vx8, 4);
118+
v6 = VGET_VX8(vx8, 5);
119+
v7 = VGET_VX8(vx8, 6);
120+
v8 = VGET_VX8(vx8, 7);
121+
v9 = VGET_VX8(vx81, 0);
122+
v10 = VGET_VX8(vx81, 1);
123+
v11 = VGET_VX8(vx81, 2);
124+
v12 = VGET_VX8(vx81, 3);
125+
v13 = VGET_VX8(vx81, 4);
126+
v14 = VGET_VX8(vx81, 5);
127+
v15 = VGET_VX8(vx81, 6);
128+
v16 = VGET_VX8(vx81, 7);
129+
130+
VSEV_FLOAT(b_offset, v1, vl);
131+
VSEV_FLOAT(b_offset + 8, v9, vl);
132+
VSEV_FLOAT(b_offset + 16, v2, vl);
133+
VSEV_FLOAT(b_offset + 24, v10, vl);
134+
VSEV_FLOAT(b_offset + 32, v3, vl);
135+
VSEV_FLOAT(b_offset + 40, v11, vl);
136+
VSEV_FLOAT(b_offset + 48, v4, vl);
137+
VSEV_FLOAT(b_offset + 56, v12, vl);
138+
VSEV_FLOAT(b_offset + 64, v5, vl);
139+
VSEV_FLOAT(b_offset + 72, v13, vl);
140+
VSEV_FLOAT(b_offset + 80, v6, vl);
141+
VSEV_FLOAT(b_offset + 88, v14, vl);
142+
VSEV_FLOAT(b_offset + 96, v7, vl);
143+
VSEV_FLOAT(b_offset + 104, v15, vl);
144+
VSEV_FLOAT(b_offset + 112, v8, vl);
145+
VSEV_FLOAT(b_offset + 120, v16, vl);
146+
147+
a_offset1 += 8;
148+
a_offset2 += 8;
149+
b_offset += 128;
150+
} while (--i);
151+
}
152+
153+
if (m & 4) {
154+
vx4 = VLSSEG4_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
155+
vx41 = VLSSEG4_FLOAT(a_offset2, lda * sizeof(FLOAT), vl);
156+
157+
v1 = VGET_VX4(vx4, 0);
158+
v2 = VGET_VX4(vx4, 1);
159+
v3 = VGET_VX4(vx4, 2);
160+
v4 = VGET_VX4(vx4, 3);
161+
v5 = VGET_VX4(vx41, 0);
162+
v6 = VGET_VX4(vx41, 1);
163+
v7 = VGET_VX4(vx41, 2);
164+
v8 = VGET_VX4(vx41, 3);
165+
166+
VSEV_FLOAT(b_offset, v1, vl);
167+
VSEV_FLOAT(b_offset + 8, v5, vl);
168+
VSEV_FLOAT(b_offset + 16, v2, vl);
169+
VSEV_FLOAT(b_offset + 24, v6, vl);
170+
VSEV_FLOAT(b_offset + 32, v3, vl);
171+
VSEV_FLOAT(b_offset + 40, v7, vl);
172+
VSEV_FLOAT(b_offset + 48, v4, vl);
173+
VSEV_FLOAT(b_offset + 56, v8, vl);
174+
175+
a_offset1 += 4;
176+
a_offset2 += 4;
177+
b_offset += 64;
178+
}
179+
180+
if (m & 2) {
181+
vx2 = VLSSEG2_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
182+
vx21 = VLSSEG2_FLOAT(a_offset2, lda * sizeof(FLOAT), vl);
183+
184+
v1 = VGET_VX2(vx2, 0);
185+
v2 = VGET_VX2(vx2, 1);
186+
v3 = VGET_VX2(vx21, 0);
187+
v4 = VGET_VX2(vx21, 1);
188+
189+
VSEV_FLOAT(b_offset, v1, vl);
190+
VSEV_FLOAT(b_offset + 8, v3, vl);
191+
VSEV_FLOAT(b_offset + 16, v2, vl);
192+
VSEV_FLOAT(b_offset + 24, v4, vl);
193+
194+
a_offset1 += 2;
195+
a_offset2 += 2;
196+
b_offset += 32;
197+
}
198+
199+
if (m & 1) {
200+
v1 = VLSEG_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
201+
v2 = VLSEG_FLOAT(a_offset2, lda * sizeof(FLOAT), vl);
202+
203+
VSEV_FLOAT(b_offset, v1, vl);
204+
VSEV_FLOAT(b_offset + 8, v2, vl);
205+
206+
b_offset += 16;
207+
}
208+
} while (--j);
209+
}
210+
211+
if (n & 8) {
212+
a_offset1 = a_offset;
213+
a_offset2 = a_offset1 + lda;
214+
a_offset3 = a_offset2 + lda;
215+
a_offset4 = a_offset3 + lda;
216+
a_offset5 = a_offset4 + lda;
217+
a_offset6 = a_offset5 + lda;
218+
a_offset7 = a_offset6 + lda;
219+
a_offset8 = a_offset7 + lda;
220+
a_offset += 8 * lda;
221+
222+
for(i = m; i > 0; i -= vl) {
223+
vl = VSETVL(i);
224+
225+
v1 = VLEV_FLOAT(a_offset1, vl);
226+
v2 = VLEV_FLOAT(a_offset2, vl);
227+
v3 = VLEV_FLOAT(a_offset3, vl);
228+
v4 = VLEV_FLOAT(a_offset4, vl);
229+
v5 = VLEV_FLOAT(a_offset5, vl);
230+
v6 = VLEV_FLOAT(a_offset6, vl);
231+
v7 = VLEV_FLOAT(a_offset7, vl);
232+
v8 = VLEV_FLOAT(a_offset8, vl);
233+
234+
vx8 = VSET_VX8(vx8, 0, v1);
235+
vx8 = VSET_VX8(vx8, 1, v2);
236+
vx8 = VSET_VX8(vx8, 2, v3);
237+
vx8 = VSET_VX8(vx8, 3, v4);
238+
vx8 = VSET_VX8(vx8, 4, v5);
239+
vx8 = VSET_VX8(vx8, 5, v6);
240+
vx8 = VSET_VX8(vx8, 6, v7);
241+
vx8 = VSET_VX8(vx8, 7, v8);
242+
243+
VSSEG8_FLOAT(b_offset, vx8, vl);
244+
245+
a_offset1 += vl;
246+
a_offset2 += vl;
247+
a_offset3 += vl;
248+
a_offset4 += vl;
249+
a_offset5 += vl;
250+
a_offset6 += vl;
251+
a_offset7 += vl;
252+
a_offset8 += vl;
253+
b_offset += vl*8;
254+
}
255+
}
256+
257+
if (n & 4) {
258+
a_offset1 = a_offset;
259+
a_offset2 = a_offset1 + lda;
260+
a_offset3 = a_offset2 + lda;
261+
a_offset4 = a_offset3 + lda;
262+
a_offset += 4 * lda;
263+
264+
for(i = m; i > 0; i -= vl) {
265+
vl = VSETVL(i);
266+
267+
v1 = VLEV_FLOAT(a_offset1, vl);
268+
v2 = VLEV_FLOAT(a_offset2, vl);
269+
v3 = VLEV_FLOAT(a_offset3, vl);
270+
v4 = VLEV_FLOAT(a_offset4, vl);
271+
272+
vx4 = VSET_VX4(vx4, 0, v1);
273+
vx4 = VSET_VX4(vx4, 1, v2);
274+
vx4 = VSET_VX4(vx4, 2, v3);
275+
vx4 = VSET_VX4(vx4, 3, v4);
276+
277+
VSSEG4_FLOAT(b_offset, vx4, vl);
278+
279+
a_offset1 += vl;
280+
a_offset2 += vl;
281+
a_offset3 += vl;
282+
a_offset4 += vl;
283+
b_offset += vl*4;
284+
}
285+
}
286+
287+
if (n & 2) {
288+
a_offset1 = a_offset;
289+
a_offset2 = a_offset1 + lda;
290+
a_offset += 2 * lda;
291+
292+
for(i = m; i > 0; i -= vl) {
293+
vl = VSETVL(i);
294+
295+
v1 = VLEV_FLOAT(a_offset1, vl);
296+
v2 = VLEV_FLOAT(a_offset2, vl);
297+
298+
vx2 = VSET_VX2(vx2, 0, v1);
299+
vx2 = VSET_VX2(vx2, 1, v2);
300+
301+
VSSEG2_FLOAT(b_offset, vx2, vl);
302+
303+
a_offset1 += vl;
304+
a_offset2 += vl;
305+
b_offset += vl*2;
306+
}
307+
}
308+
309+
if (n & 1) {
310+
a_offset1 = a_offset;
311+
312+
for(i = m; i > 0; i -= vl) {
313+
vl = VSETVL(i);
314+
315+
v1 = VLEV_FLOAT(a_offset1, vl);
316+
317+
VSEV_FLOAT(b_offset, v1, vl);
318+
319+
a_offset1 += vl;
320+
b_offset += vl;
321+
}
322+
}
323+
324+
return 0;
325+
}

0 commit comments

Comments
 (0)