Skip to content

Commit 123c25c

Browse files
martin-frbgmartin-frbg
authored
Merge pull request #5527 from ChipKerchner/fixbfloat16Tobf16conversions
Prevent possible conversion from bfloat16 to __bf16 - RISC-V
2 parents 3a9da52 + 00a7336 commit 123c25c

File tree

4 files changed

+223
-197
lines changed

4 files changed

+223
-197
lines changed

‎kernel/riscv64/sbgemm_kernel_16x8_zvl256b.c‎

Lines changed: 104 additions & 102 deletions
Original file line numberDiff line numberDiff line change
@@ -6,6 +6,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
66
BLASLONG gvl = 0;
77
BLASLONG m_top = 0;
88
BLASLONG n_top = 0;
9+
__bf16 *BB = (__bf16 *)(B);
10+
__bf16 *AA = (__bf16 *)(A);
911

1012
// -- MAIN PASS
1113
for (BLASLONG j=0; j<N/8; j+=1) {
@@ -26,17 +28,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
2628
vfloat32m2_t result7 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
2729

2830
for (BLASLONG k=0; k<K; k++) {
29-
__bf16 B0 = B[bi+0];
30-
__bf16 B1 = B[bi+1];
31-
__bf16 B2 = B[bi+2];
32-
__bf16 B3 = B[bi+3];
33-
__bf16 B4 = B[bi+4];
34-
__bf16 B5 = B[bi+5];
35-
__bf16 B6 = B[bi+6];
36-
__bf16 B7 = B[bi+7];
31+
__bf16 B0 = BB[bi+0];
32+
__bf16 B1 = BB[bi+1];
33+
__bf16 B2 = BB[bi+2];
34+
__bf16 B3 = BB[bi+3];
35+
__bf16 B4 = BB[bi+4];
36+
__bf16 B5 = BB[bi+5];
37+
__bf16 B6 = BB[bi+6];
38+
__bf16 B7 = BB[bi+7];
3739
bi += 8;
3840

39-
vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &A[ai+0*gvl], gvl );
41+
vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
4042
ai += 16;
4143

4244
result0 = __riscv_vfwmaccbf16_vf_f32m2(result0, B0, A0, gvl);
@@ -100,17 +102,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
100102
vfloat32m1_t result7 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
101103

102104
for (BLASLONG k=0; k<K; k++) {
103-
__bf16 B0 = B[bi+0];
104-
__bf16 B1 = B[bi+1];
105-
__bf16 B2 = B[bi+2];
106-
__bf16 B3 = B[bi+3];
107-
__bf16 B4 = B[bi+4];
108-
__bf16 B5 = B[bi+5];
109-
__bf16 B6 = B[bi+6];
110-
__bf16 B7 = B[bi+7];
105+
__bf16 B0 = BB[bi+0];
106+
__bf16 B1 = BB[bi+1];
107+
__bf16 B2 = BB[bi+2];
108+
__bf16 B3 = BB[bi+3];
109+
__bf16 B4 = BB[bi+4];
110+
__bf16 B5 = BB[bi+5];
111+
__bf16 B6 = BB[bi+6];
112+
__bf16 B7 = BB[bi+7];
111113
bi += 8;
112114

113-
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &A[ai+0*gvl], gvl );
115+
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
114116
ai += 8;
115117

116118
result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
@@ -172,17 +174,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
172174
vfloat32m1_t result7 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
173175

174176
for (BLASLONG k=0; k < K; ++k) {
175-
__bf16 B0 = B[bi+0];
176-
__bf16 B1 = B[bi+1];
177-
__bf16 B2 = B[bi+2];
178-
__bf16 B3 = B[bi+3];
179-
__bf16 B4 = B[bi+4];
180-
__bf16 B5 = B[bi+5];
181-
__bf16 B6 = B[bi+6];
182-
__bf16 B7 = B[bi+7];
177+
__bf16 B0 = BB[bi+0];
178+
__bf16 B1 = BB[bi+1];
179+
__bf16 B2 = BB[bi+2];
180+
__bf16 B3 = BB[bi+3];
181+
__bf16 B4 = BB[bi+4];
182+
__bf16 B5 = BB[bi+5];
183+
__bf16 B6 = BB[bi+6];
184+
__bf16 B7 = BB[bi+7];
183185
bi += 8;
184186

185-
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &A[ai+0*gvl], gvl );
187+
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
186188
ai += 4;
187189

188190
result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
@@ -256,22 +258,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
256258
BLASLONG bi = n_top * K;
257259

258260
for (BLASLONG k=0; k<K; k++) {
259-
result0+=(float)(A[ai+0])*(float)(B[bi+0]);
260-
result1+=(float)(A[ai+1])*(float)(B[bi+0]);
261-
result2+=(float)(A[ai+0])*(float)(B[bi+1]);
262-
result3+=(float)(A[ai+1])*(float)(B[bi+1]);
263-
result4+=(float)(A[ai+0])*(float)(B[bi+2]);
264-
result5+=(float)(A[ai+1])*(float)(B[bi+2]);
265-
result6+=(float)(A[ai+0])*(float)(B[bi+3]);
266-
result7+=(float)(A[ai+1])*(float)(B[bi+3]);
267-
result8+=(float)(A[ai+0])*(float)(B[bi+4]);
268-
result9+=(float)(A[ai+1])*(float)(B[bi+4]);
269-
result10+=(float)(A[ai+0])*(float)(B[bi+5]);
270-
result11+=(float)(A[ai+1])*(float)(B[bi+5]);
271-
result12+=(float)(A[ai+0])*(float)(B[bi+6]);
272-
result13+=(float)(A[ai+1])*(float)(B[bi+6]);
273-
result14+=(float)(A[ai+0])*(float)(B[bi+7]);
274-
result15+=(float)(A[ai+1])*(float)(B[bi+7]);
261+
result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
262+
result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
263+
result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
264+
result3+=(float)(AA[ai+1])*(float)(BB[bi+1]);
265+
result4+=(float)(AA[ai+0])*(float)(BB[bi+2]);
266+
result5+=(float)(AA[ai+1])*(float)(BB[bi+2]);
267+
result6+=(float)(AA[ai+0])*(float)(BB[bi+3]);
268+
result7+=(float)(AA[ai+1])*(float)(BB[bi+3]);
269+
result8+=(float)(AA[ai+0])*(float)(BB[bi+4]);
270+
result9+=(float)(AA[ai+1])*(float)(BB[bi+4]);
271+
result10+=(float)(AA[ai+0])*(float)(BB[bi+5]);
272+
result11+=(float)(AA[ai+1])*(float)(BB[bi+5]);
273+
result12+=(float)(AA[ai+0])*(float)(BB[bi+6]);
274+
result13+=(float)(AA[ai+1])*(float)(BB[bi+6]);
275+
result14+=(float)(AA[ai+0])*(float)(BB[bi+7]);
276+
result15+=(float)(AA[ai+1])*(float)(BB[bi+7]);
275277
ai+=2;
276278
bi+=8;
277279
}
@@ -314,14 +316,14 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
314316
BLASLONG bi = n_top * K;
315317

316318
for (BLASLONG k=0; k<K; k++) {
317-
result0+=(float)(A[ai+0])*(float)(B[bi+0]);
318-
result1+=(float)(A[ai+0])*(float)(B[bi+1]);
319-
result2+=(float)(A[ai+0])*(float)(B[bi+2]);
320-
result3+=(float)(A[ai+0])*(float)(B[bi+3]);
321-
result4+=(float)(A[ai+0])*(float)(B[bi+4]);
322-
result5+=(float)(A[ai+0])*(float)(B[bi+5]);
323-
result6+=(float)(A[ai+0])*(float)(B[bi+6]);
324-
result7+=(float)(A[ai+0])*(float)(B[bi+7]);
319+
result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
320+
result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
321+
result2+=(float)(AA[ai+0])*(float)(BB[bi+2]);
322+
result3+=(float)(AA[ai+0])*(float)(BB[bi+3]);
323+
result4+=(float)(AA[ai+0])*(float)(BB[bi+4]);
324+
result5+=(float)(AA[ai+0])*(float)(BB[bi+5]);
325+
result6+=(float)(AA[ai+0])*(float)(BB[bi+6]);
326+
result7+=(float)(AA[ai+0])*(float)(BB[bi+7]);
325327
ai+=1;
326328
bi+=8;
327329
}
@@ -354,13 +356,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
354356
vfloat32m2_t result3 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
355357

356358
for (BLASLONG k=0; k<K; k++) {
357-
__bf16 B0 = B[bi+0];
358-
__bf16 B1 = B[bi+1];
359-
__bf16 B2 = B[bi+2];
360-
__bf16 B3 = B[bi+3];
359+
__bf16 B0 = BB[bi+0];
360+
__bf16 B1 = BB[bi+1];
361+
__bf16 B2 = BB[bi+2];
362+
__bf16 B3 = BB[bi+3];
361363
bi += 4;
362364

363-
vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &A[ai+0*gvl], gvl );
365+
vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
364366
ai += 16;
365367

366368
result0 = __riscv_vfwmaccbf16_vf_f32m2(result0, B0, A0, gvl);
@@ -401,13 +403,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
401403
vfloat32m1_t result3 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
402404

403405
for (BLASLONG k=0; k<K; k++) {
404-
__bf16 B0 = B[bi+0];
405-
__bf16 B1 = B[bi+1];
406-
__bf16 B2 = B[bi+2];
407-
__bf16 B3 = B[bi+3];
406+
__bf16 B0 = BB[bi+0];
407+
__bf16 B1 = BB[bi+1];
408+
__bf16 B2 = BB[bi+2];
409+
__bf16 B3 = BB[bi+3];
408410
bi += 4;
409411

410-
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &A[ai+0*gvl], gvl );
412+
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
411413
ai += 8;
412414

413415
result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
@@ -449,13 +451,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
449451
vfloat32m1_t result3 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
450452

451453
for (BLASLONG k=0; k < K; ++k) {
452-
__bf16 B0 = B[bi+0];
453-
__bf16 B1 = B[bi+1];
454-
__bf16 B2 = B[bi+2];
455-
__bf16 B3 = B[bi+3];
454+
__bf16 B0 = BB[bi+0];
455+
__bf16 B1 = BB[bi+1];
456+
__bf16 B2 = BB[bi+2];
457+
__bf16 B3 = BB[bi+3];
456458
bi += 4;
457459

458-
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &A[ai+0*gvl], gvl );
460+
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
459461
ai += 4;
460462

461463
result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
@@ -501,14 +503,14 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
501503
BLASLONG bi = n_top * K;
502504

503505
for (BLASLONG k=0; k<K; k++) {
504-
result0+=(float)(A[ai+0])*(float)(B[bi+0]);
505-
result1+=(float)(A[ai+1])*(float)(B[bi+0]);
506-
result2+=(float)(A[ai+0])*(float)(B[bi+1]);
507-
result3+=(float)(A[ai+1])*(float)(B[bi+1]);
508-
result4+=(float)(A[ai+0])*(float)(B[bi+2]);
509-
result5+=(float)(A[ai+1])*(float)(B[bi+2]);
510-
result6+=(float)(A[ai+0])*(float)(B[bi+3]);
511-
result7+=(float)(A[ai+1])*(float)(B[bi+3]);
506+
result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
507+
result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
508+
result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
509+
result3+=(float)(AA[ai+1])*(float)(BB[bi+1]);
510+
result4+=(float)(AA[ai+0])*(float)(BB[bi+2]);
511+
result5+=(float)(AA[ai+1])*(float)(BB[bi+2]);
512+
result6+=(float)(AA[ai+0])*(float)(BB[bi+3]);
513+
result7+=(float)(AA[ai+1])*(float)(BB[bi+3]);
512514
ai+=2;
513515
bi+=4;
514516
}
@@ -537,10 +539,10 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
537539
BLASLONG bi = n_top * K;
538540

539541
for (BLASLONG k=0; k<K; k++) {
540-
result0+=(float)(A[ai+0])*(float)(B[bi+0]);
541-
result1+=(float)(A[ai+0])*(float)(B[bi+1]);
542-
result2+=(float)(A[ai+0])*(float)(B[bi+2]);
543-
result3+=(float)(A[ai+0])*(float)(B[bi+3]);
542+
result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
543+
result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
544+
result2+=(float)(AA[ai+0])*(float)(BB[bi+2]);
545+
result3+=(float)(AA[ai+0])*(float)(BB[bi+3]);
544546
ai+=1;
545547
bi+=4;
546548
}
@@ -569,11 +571,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
569571
vfloat32m2_t result1 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
570572

571573
for (BLASLONG k=0; k<K; k++) {
572-
__bf16 B0 = B[bi+0];
573-
__bf16 B1 = B[bi+1];
574+
__bf16 B0 = BB[bi+0];
575+
__bf16 B1 = BB[bi+1];
574576
bi += 2;
575577

576-
vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &A[ai+0*gvl], gvl );
578+
vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
577579
ai += 16;
578580

579581
result0 = __riscv_vfwmaccbf16_vf_f32m2(result0, B0, A0, gvl);
@@ -603,11 +605,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
603605
vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
604606

605607
for (BLASLONG k=0; k<K; k++) {
606-
__bf16 B0 = B[bi+0];
607-
__bf16 B1 = B[bi+1];
608+
__bf16 B0 = BB[bi+0];
609+
__bf16 B1 = BB[bi+1];
608610
bi += 2;
609611

610-
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &A[ai+0*gvl], gvl );
612+
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
611613
ai += 8;
612614

613615
result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
@@ -639,11 +641,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
639641
vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
640642

641643
for (BLASLONG k=0; k < K; ++k) {
642-
__bf16 B0 = B[bi+0];
643-
__bf16 B1 = B[bi+1];
644+
__bf16 B0 = BB[bi+0];
645+
__bf16 B1 = BB[bi+1];
644646
bi += 2;
645647

646-
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &A[ai+0*gvl], gvl );
648+
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
647649
ai += 4;
648650

649651
result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
@@ -675,10 +677,10 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
675677
BLASLONG bi = n_top * K;
676678

677679
for (BLASLONG k=0; k<K; k++) {
678-
result0+=(float)(A[ai+0])*(float)(B[bi+0]);
679-
result1+=(float)(A[ai+1])*(float)(B[bi+0]);
680-
result2+=(float)(A[ai+0])*(float)(B[bi+1]);
681-
result3+=(float)(A[ai+1])*(float)(B[bi+1]);
680+
result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
681+
result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
682+
result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
683+
result3+=(float)(AA[ai+1])*(float)(BB[bi+1]);
682684
ai+=2;
683685
bi+=2;
684686
}
@@ -701,8 +703,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
701703
BLASLONG bi = n_top * K;
702704

703705
for (BLASLONG k=0; k<K; k++) {
704-
result0+=(float)(A[ai+0])*(float)(B[bi+0]);
705-
result1+=(float)(A[ai+0])*(float)(B[bi+1]);
706+
result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
707+
result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
706708
ai+=1;
707709
bi+=2;
708710
}
@@ -728,10 +730,10 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
728730
vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
729731

730732
for (BLASLONG k=0; k<K; k++) {
731-
__bf16 B0 = B[bi+0];
733+
__bf16 B0 = BB[bi+0];
732734
bi += 1;
733735

734-
vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &A[ai+0*gvl], gvl );
736+
vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
735737
ai += 16;
736738

737739
result0 = __riscv_vfwmaccbf16_vf_f32m2(result0, B0, A0, gvl);
@@ -757,10 +759,10 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
757759
vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
758760

759761
for (BLASLONG k=0; k<K; k++) {
760-
__bf16 B0 = B[bi+0];
762+
__bf16 B0 = BB[bi+0];
761763
bi += 1;
762764

763-
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &A[ai+0*gvl], gvl );
765+
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
764766
ai += 8;
765767

766768
result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
@@ -787,10 +789,10 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
787789
vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
788790

789791
for (BLASLONG k=0; k < K; ++k) {
790-
__bf16 B0 = B[bi+0];
792+
__bf16 B0 = BB[bi+0];
791793
bi += 1;
792794

793-
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &A[ai+0*gvl], gvl );
795+
vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
794796
ai += 4;
795797

796798
result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
@@ -814,8 +816,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
814816
BLASLONG bi = n_top * K;
815817

816818
for (BLASLONG k=0; k<K; k++) {
817-
result0+=(float)(A[ai+0])*(float)(B[bi+0]);
818-
result1+=(float)(A[ai+1])*(float)(B[bi+0]);
819+
result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
820+
result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
819821
ai+=2;
820822
bi+=1;
821823
}
@@ -835,7 +837,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
835837
BLASLONG bi = n_top * K;
836838

837839
for (BLASLONG k=0; k<K; k++) {
838-
result0+=(float)(A[ai+0])*(float)(B[bi+0]);
840+
result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
839841
ai+=1;
840842
bi+=1;
841843
}

0 commit comments

Comments
 (0)