OpenMathLib/OpenBLAS
1
0
Fork 0
mirror of https://github.com/OpenMathLib/OpenBLAS synced 2026-05-31 00:45:48 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Missing one gemv conversion.

Browse Source
This commit is contained in:
Chip Kerchner
2025-11-04 22:27:53 +00:00
parent edf2e5900c
commit 00a7336fc9
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
kernel/riscv64/sbgemv_t_vector.c
View File

@@ -60,8 +60,10 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *
BLASLONG ix = 0, iy = 0;
#if defined(HFLOAT16)
_Float16 *a_ptr = (_Float16 *)(a);
_Float16 *x_ptr = (_Float16 *)(x);
#else
__bf16 *a_ptr = (__bf16 *)(a);
__bf16 *x_ptr = (__bf16 *)(x);
#endif
FLOAT temp;
@@ -83,7 +85,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *
#endif
for (k = 0; k < m/gvl; k++) {
va = VLEV_IFLOAT(&a_ptr[j], gvl);
vx = VLEV_IFLOAT(&x[j], gvl);
vx = VLEV_IFLOAT(&x_ptr[j], gvl);
vr = VFMACCVV_FLOAT(vz, va, vx, gvl); // could vfmacc here and reduce outside loop
v_res = VFREDSUM_FLOAT(vr, v_res, gvl); // but that reordering diverges far enough from scalar path to make tests fail
j += gvl;
@@ -91,7 +93,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *
if (j < m) {
gvl = VSETVL(m-j);
va = VLEV_IFLOAT(&a_ptr[j], gvl);
vx = VLEV_IFLOAT(&x[j], gvl);
vx = VLEV_IFLOAT(&x_ptr[j], gvl);
vr = VFMACCVV_FLOAT(vz, va, vx, gvl);
v_res = VFREDSUM_FLOAT(vr, v_res, gvl);
}
@@ -113,7 +115,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *
#endif
for (k = 0; k < m/gvl; k++) {
va = VLEV_IFLOAT(&a_ptr[j], gvl);
vx = VLSEV_IFLOAT(&x[ix], stride_x, gvl);
vx = VLSEV_IFLOAT(&x_ptr[ix], stride_x, gvl);
vr = VFMACCVV_FLOAT(vz, va, vx, gvl);
v_res = VFREDSUM_FLOAT(vr, v_res, gvl);
j += gvl;
@@ -122,7 +124,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *
if (j < m) {
gvl = VSETVL(m-j);
va = VLEV_IFLOAT(&a_ptr[j], gvl);
vx = VLSEV_IFLOAT(&x[ix], stride_x, gvl);
vx = VLSEV_IFLOAT(&x_ptr[ix], stride_x, gvl);
vr = VFMACCVV_FLOAT(vz, va, vx, gvl);
v_res = VFREDSUM_FLOAT(vr, v_res, gvl);
}