Skip to content

Commit 3356043

Browse files
ChipKerchnerChipKerchner
committed
Convert inputs from BF16 to FP32 and use FP32 vector madds. 18% faster.
1 parent e3cb067 commit 3356043

2 files changed

Lines changed: 214 additions & 0 deletions

File tree

‎kernel/riscv64/sbgemm_kernel_16x8_zvl256b.c‎

Lines changed: 128 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,8 @@
11
#include "common.h"
22
#include <riscv_vector.h>
33

4+
#define BF16_WIDEN_ONE
5+
46
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B, FLOAT *C, BLASLONG ldc)
57
{
68
BLASLONG gvl = 0;
@@ -28,6 +30,30 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
2830
vfloat32m2_t result7 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
2931

3032
for (BLASLONG k=0; k<K; k++) {
33+
#ifdef BF16_WIDEN_ONE
34+
float B0 = (float)(BB[bi+0]);
35+
float B1 = (float)(BB[bi+1]);
36+
float B2 = (float)(BB[bi+2]);
37+
float B3 = (float)(BB[bi+3]);
38+
float B4 = (float)(BB[bi+4]);
39+
float B5 = (float)(BB[bi+5]);
40+
float B6 = (float)(BB[bi+6]);
41+
float B7 = (float)(BB[bi+7]);
42+
bi += 8;
43+
44+
vbfloat16m1_t A00 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
45+
vfloat32m2_t A0 = __riscv_vfwcvtbf16_f_f_v_f32m2(A00, gvl);
46+
ai += 16;
47+
48+
result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
49+
result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
50+
result2 = __riscv_vfmacc_vf_f32m2(result2, B2, A0, gvl);
51+
result3 = __riscv_vfmacc_vf_f32m2(result3, B3, A0, gvl);
52+
result4 = __riscv_vfmacc_vf_f32m2(result4, B4, A0, gvl);
53+
result5 = __riscv_vfmacc_vf_f32m2(result5, B5, A0, gvl);
54+
result6 = __riscv_vfmacc_vf_f32m2(result6, B6, A0, gvl);
55+
result7 = __riscv_vfmacc_vf_f32m2(result7, B7, A0, gvl);
56+
#else
3157
__bf16 B0 = BB[bi+0];
3258
__bf16 B1 = BB[bi+1];
3359
__bf16 B2 = BB[bi+2];
@@ -49,6 +75,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
4975
result5 = __riscv_vfwmaccbf16_vf_f32m2(result5, B5, A0, gvl);
5076
result6 = __riscv_vfwmaccbf16_vf_f32m2(result6, B6, A0, gvl);
5177
result7 = __riscv_vfwmaccbf16_vf_f32m2(result7, B7, A0, gvl);
78+
#endif
5279
}
5380

5481
BLASLONG ci=n_top*ldc+m_top;
@@ -102,6 +129,30 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
102129
vfloat32m1_t result7 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
103130

104131
for (BLASLONG k=0; k<K; k++) {
132+
#ifdef BF16_WIDEN_ONE
133+
float B0 = (float)(BB[bi+0]);
134+
float B1 = (float)(BB[bi+1]);
135+
float B2 = (float)(BB[bi+2]);
136+
float B3 = (float)(BB[bi+3]);
137+
float B4 = (float)(BB[bi+4]);
138+
float B5 = (float)(BB[bi+5]);
139+
float B6 = (float)(BB[bi+6]);
140+
float B7 = (float)(BB[bi+7]);
141+
bi += 8;
142+
143+
vbfloat16mf2_t A00 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
144+
vfloat32m1_t A0 = __riscv_vfwcvtbf16_f_f_v_f32m1(A00, gvl);
145+
ai += 8;
146+
147+
result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
148+
result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
149+
result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
150+
result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
151+
result4 = __riscv_vfmacc_vf_f32m1(result4, B4, A0, gvl);
152+
result5 = __riscv_vfmacc_vf_f32m1(result5, B5, A0, gvl);
153+
result6 = __riscv_vfmacc_vf_f32m1(result6, B6, A0, gvl);
154+
result7 = __riscv_vfmacc_vf_f32m1(result7, B7, A0, gvl);
155+
#else
105156
__bf16 B0 = BB[bi+0];
106157
__bf16 B1 = BB[bi+1];
107158
__bf16 B2 = BB[bi+2];
@@ -123,6 +174,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
123174
result5 = __riscv_vfwmaccbf16_vf_f32m1(result5, B5, A0, gvl);
124175
result6 = __riscv_vfwmaccbf16_vf_f32m1(result6, B6, A0, gvl);
125176
result7 = __riscv_vfwmaccbf16_vf_f32m1(result7, B7, A0, gvl);
177+
#endif
126178
}
127179

128180
BLASLONG ci=n_top*ldc+m_top;
@@ -174,6 +226,30 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
174226
vfloat32m1_t result7 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
175227

176228
for (BLASLONG k=0; k < K; ++k) {
229+
#ifdef BF16_WIDEN_ONE
230+
float B0 = (float)(BB[bi+0]);
231+
float B1 = (float)(BB[bi+1]);
232+
float B2 = (float)(BB[bi+2]);
233+
float B3 = (float)(BB[bi+3]);
234+
float B4 = (float)(BB[bi+4]);
235+
float B5 = (float)(BB[bi+5]);
236+
float B6 = (float)(BB[bi+6]);
237+
float B7 = (float)(BB[bi+7]);
238+
bi += 8;
239+
240+
vbfloat16mf4_t A00 = __riscv_vle16_v_bf16mf4( &AA[ai+0*gvl], gvl );
241+
vfloat32m1_t A0 = __riscv_vlmul_ext_v_f32mf2_f32m1(__riscv_vfwcvtbf16_f_f_v_f32mf2(A00, gvl));
242+
ai += 4;
243+
244+
result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
245+
result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
246+
result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
247+
result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
248+
result4 = __riscv_vfmacc_vf_f32m1(result4, B4, A0, gvl);
249+
result5 = __riscv_vfmacc_vf_f32m1(result5, B5, A0, gvl);
250+
result6 = __riscv_vfmacc_vf_f32m1(result6, B6, A0, gvl);
251+
result7 = __riscv_vfmacc_vf_f32m1(result7, B7, A0, gvl);
252+
#else
177253
__bf16 B0 = BB[bi+0];
178254
__bf16 B1 = BB[bi+1];
179255
__bf16 B2 = BB[bi+2];
@@ -195,6 +271,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
195271
result5 = __riscv_vfwmaccbf16_vf_f32m1(result5, B5, A0, gvl);
196272
result6 = __riscv_vfwmaccbf16_vf_f32m1(result6, B6, A0, gvl);
197273
result7 = __riscv_vfwmaccbf16_vf_f32m1(result7, B7, A0, gvl);
274+
#endif
198275
}
199276

200277
BLASLONG ci = n_top * ldc + m_top;
@@ -356,6 +433,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
356433
vfloat32m2_t result3 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
357434

358435
for (BLASLONG k=0; k<K; k++) {
436+
#ifdef BF16_WIDEN_ONE
437+
float B0 = (float)(BB[bi+0]);
438+
float B1 = (float)(BB[bi+1]);
439+
float B2 = (float)(BB[bi+2]);
440+
float B3 = (float)(BB[bi+3]);
441+
bi += 4;
442+
443+
vbfloat16m1_t A00 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
444+
vfloat32m2_t A0 = __riscv_vfwcvtbf16_f_f_v_f32m2(A00, gvl);
445+
ai += 16;
446+
447+
result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
448+
result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
449+
result2 = __riscv_vfmacc_vf_f32m2(result2, B2, A0, gvl);
450+
result3 = __riscv_vfmacc_vf_f32m2(result3, B3, A0, gvl);
451+
#else
359452
__bf16 B0 = BB[bi+0];
360453
__bf16 B1 = BB[bi+1];
361454
__bf16 B2 = BB[bi+2];
@@ -369,6 +462,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
369462
result1 = __riscv_vfwmaccbf16_vf_f32m2(result1, B1, A0, gvl);
370463
result2 = __riscv_vfwmaccbf16_vf_f32m2(result2, B2, A0, gvl);
371464
result3 = __riscv_vfwmaccbf16_vf_f32m2(result3, B3, A0, gvl);
465+
#endif
372466
}
373467

374468
BLASLONG ci=n_top*ldc+m_top;
@@ -403,6 +497,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
403497
vfloat32m1_t result3 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
404498

405499
for (BLASLONG k=0; k<K; k++) {
500+
#ifdef BF16_WIDEN_ONE
501+
float B0 = (float)(BB[bi+0]);
502+
float B1 = (float)(BB[bi+1]);
503+
float B2 = (float)(BB[bi+2]);
504+
float B3 = (float)(BB[bi+3]);
505+
bi += 4;
506+
507+
vbfloat16mf2_t A00 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
508+
vfloat32m1_t A0 = __riscv_vfwcvtbf16_f_f_v_f32m1(A00, gvl);
509+
ai += 8;
510+
511+
result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
512+
result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
513+
result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
514+
result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
515+
#else
406516
__bf16 B0 = BB[bi+0];
407517
__bf16 B1 = BB[bi+1];
408518
__bf16 B2 = BB[bi+2];
@@ -416,6 +526,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
416526
result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
417527
result2 = __riscv_vfwmaccbf16_vf_f32m1(result2, B2, A0, gvl);
418528
result3 = __riscv_vfwmaccbf16_vf_f32m1(result3, B3, A0, gvl);
529+
#endif
419530
}
420531

421532
BLASLONG ci=n_top*ldc+m_top;
@@ -451,6 +562,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
451562
vfloat32m1_t result3 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
452563

453564
for (BLASLONG k=0; k < K; ++k) {
565+
#ifdef BF16_WIDEN_ONE
566+
float B0 = (float)(BB[bi+0]);
567+
float B1 = (float)(BB[bi+1]);
568+
float B2 = (float)(BB[bi+2]);
569+
float B3 = (float)(BB[bi+3]);
570+
bi += 4;
571+
572+
vbfloat16mf4_t A00 = __riscv_vle16_v_bf16mf4( &AA[ai+0*gvl], gvl );
573+
vfloat32m1_t A0 = __riscv_vlmul_ext_v_f32mf2_f32m1(__riscv_vfwcvtbf16_f_f_v_f32mf2(A00, gvl));
574+
ai += 4;
575+
576+
result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
577+
result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
578+
result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
579+
result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
580+
#else
454581
__bf16 B0 = BB[bi+0];
455582
__bf16 B1 = BB[bi+1];
456583
__bf16 B2 = BB[bi+2];
@@ -464,6 +591,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
464591
result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
465592
result2 = __riscv_vfwmaccbf16_vf_f32m1(result2, B2, A0, gvl);
466593
result3 = __riscv_vfwmaccbf16_vf_f32m1(result3, B3, A0, gvl);
594+
#endif
467595
}
468596

469597
BLASLONG ci = n_top * ldc + m_top;

‎kernel/riscv64/sbgemm_kernel_8x8_zvl128b.c‎

Lines changed: 86 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,8 @@
11
#include "common.h"
22
#include <riscv_vector.h>
33

4+
#define BF16_WIDEN_ONE
5+
46
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B, FLOAT *C, BLASLONG ldc)
57
{
68
BLASLONG gvl = 0;
@@ -28,6 +30,30 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
2830
vfloat32m2_t result7 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
2931

3032
for (BLASLONG k=0; k<K; k++) {
33+
#ifdef BF16_WIDEN_ONE
34+
float B0 = (float)(BB[bi+0]);
35+
float B1 = (float)(BB[bi+1]);
36+
float B2 = (float)(BB[bi+2]);
37+
float B3 = (float)(BB[bi+3]);
38+
float B4 = (float)(BB[bi+4]);
39+
float B5 = (float)(BB[bi+5]);
40+
float B6 = (float)(BB[bi+6]);
41+
float B7 = (float)(BB[bi+7]);
42+
bi += 8;
43+
44+
vbfloat16m1_t A00 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
45+
vfloat32m2_t A0 = __riscv_vfwcvtbf16_f_f_v_f32m2(A00, gvl);
46+
ai += 8;
47+
48+
result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
49+
result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
50+
result2 = __riscv_vfmacc_vf_f32m2(result2, B2, A0, gvl);
51+
result3 = __riscv_vfmacc_vf_f32m2(result3, B3, A0, gvl);
52+
result4 = __riscv_vfmacc_vf_f32m2(result4, B4, A0, gvl);
53+
result5 = __riscv_vfmacc_vf_f32m2(result5, B5, A0, gvl);
54+
result6 = __riscv_vfmacc_vf_f32m2(result6, B6, A0, gvl);
55+
result7 = __riscv_vfmacc_vf_f32m2(result7, B7, A0, gvl);
56+
#else
3157
__bf16 B0 = BB[bi+0];
3258
__bf16 B1 = BB[bi+1];
3359
__bf16 B2 = BB[bi+2];
@@ -49,6 +75,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
4975
result5 = __riscv_vfwmaccbf16_vf_f32m2(result5, B5, A0, gvl);
5076
result6 = __riscv_vfwmaccbf16_vf_f32m2(result6, B6, A0, gvl);
5177
result7 = __riscv_vfwmaccbf16_vf_f32m2(result7, B7, A0, gvl);
78+
#endif
5279
}
5380

5481
BLASLONG ci=n_top*ldc+m_top;
@@ -102,6 +129,30 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
102129
vfloat32m1_t result7 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
103130

104131
for (BLASLONG k=0; k < K; ++k) {
132+
#ifdef BF16_WIDEN_ONE
133+
float B0 = (float)(BB[bi+0]);
134+
float B1 = (float)(BB[bi+1]);
135+
float B2 = (float)(BB[bi+2]);
136+
float B3 = (float)(BB[bi+3]);
137+
float B4 = (float)(BB[bi+4]);
138+
float B5 = (float)(BB[bi+5]);
139+
float B6 = (float)(BB[bi+6]);
140+
float B7 = (float)(BB[bi+7]);
141+
bi += 8;
142+
143+
vbfloat16mf2_t A00 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
144+
vfloat32m1_t A0 = __riscv_vfwcvtbf16_f_f_v_f32m1(A00, gvl);
145+
ai += 4;
146+
147+
result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
148+
result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
149+
result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
150+
result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
151+
result4 = __riscv_vfmacc_vf_f32m1(result4, B4, A0, gvl);
152+
result5 = __riscv_vfmacc_vf_f32m1(result5, B5, A0, gvl);
153+
result6 = __riscv_vfmacc_vf_f32m1(result6, B6, A0, gvl);
154+
result7 = __riscv_vfmacc_vf_f32m1(result7, B7, A0, gvl);
155+
#else
105156
__bf16 B0 = BB[bi+0];
106157
__bf16 B1 = BB[bi+1];
107158
__bf16 B2 = BB[bi+2];
@@ -123,6 +174,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
123174
result5 = __riscv_vfwmaccbf16_vf_f32m1(result5, B5, A0, gvl);
124175
result6 = __riscv_vfwmaccbf16_vf_f32m1(result6, B6, A0, gvl);
125176
result7 = __riscv_vfwmaccbf16_vf_f32m1(result7, B7, A0, gvl);
177+
#endif
126178
}
127179

128180
BLASLONG ci = n_top * ldc + m_top;
@@ -286,6 +338,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
286338
vfloat32m2_t result3 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
287339

288340
for (BLASLONG k=0; k<K; k++) {
341+
#ifdef BF16_WIDEN_ONE
342+
float B0 = (float)(BB[bi+0]);
343+
float B1 = (float)(BB[bi+1]);
344+
float B2 = (float)(BB[bi+2]);
345+
float B3 = (float)(BB[bi+3]);
346+
bi += 4;
347+
348+
vbfloat16m1_t A00 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
349+
vfloat32m2_t A0 = __riscv_vfwcvtbf16_f_f_v_f32m2(A00, gvl);
350+
ai += 8;
351+
352+
result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
353+
result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
354+
result2 = __riscv_vfmacc_vf_f32m2(result2, B2, A0, gvl);
355+
result3 = __riscv_vfmacc_vf_f32m2(result3, B3, A0, gvl);
356+
#else
289357
__bf16 B0 = BB[bi+0];
290358
__bf16 B1 = BB[bi+1];
291359
__bf16 B2 = BB[bi+2];
@@ -299,6 +367,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
299367
result1 = __riscv_vfwmaccbf16_vf_f32m2(result1, B1, A0, gvl);
300368
result2 = __riscv_vfwmaccbf16_vf_f32m2(result2, B2, A0, gvl);
301369
result3 = __riscv_vfwmaccbf16_vf_f32m2(result3, B3, A0, gvl);
370+
#endif
302371
}
303372

304373
BLASLONG ci=n_top*ldc+m_top;
@@ -334,6 +403,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
334403
vfloat32m1_t result3 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
335404

336405
for (BLASLONG k=0; k < K; ++k) {
406+
#ifdef BF16_WIDEN_ONE
407+
float B0 = (float)(BB[bi+0]);
408+
float B1 = (float)(BB[bi+1]);
409+
float B2 = (float)(BB[bi+2]);
410+
float B3 = (float)(BB[bi+3]);
411+
bi += 4;
412+
413+
vbfloat16mf4_t A00 = __riscv_vle16_v_bf16mf4( &AA[ai+0*gvl], gvl );
414+
vfloat32m1_t A0 = __riscv_vlmul_ext_v_f32mf2_f32m1(__riscv_vfwcvtbf16_f_f_v_f32mf2(A00, gvl));
415+
ai += 4;
416+
417+
result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
418+
result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
419+
result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
420+
result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
421+
#else
337422
__bf16 B0 = BB[bi+0];
338423
__bf16 B1 = BB[bi+1];
339424
__bf16 B2 = BB[bi+2];
@@ -347,6 +432,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
347432
result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
348433
result2 = __riscv_vfwmaccbf16_vf_f32m1(result2, B2, A0, gvl);
349434
result3 = __riscv_vfwmaccbf16_vf_f32m1(result3, B3, A0, gvl);
435+
#endif
350436
}
351437

352438
BLASLONG ci = n_top * ldc + m_top;

0 commit comments

Comments
 (0)