Skip to content

Tranverse matrix data in a cache friendly manner for GEMV_N (RISCV) #5476

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Oct 1, 2025
Merged

Tranverse matrix data in a cache friendly manner for GEMV_N (RISCV) #5476

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
dd7a1d6
Merge branch 'develop' of https://github.com/OpenMathLib/OpenBLAS int…
ChipKerchner Sep 30, 2025
2d82d14
Tranverse matrix data in a cache friendly manner for GEMV_N (RISCV).
ChipKerchner Sep 30, 2025
36f9cb8
Fix pre-RVV 1.0.
ChipKerchner Sep 30, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
257 changes: 47 additions & 210 deletions kernel/riscv64/gemv_n_vector.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,230 +26,67 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#include "common.h"

#if !defined(DOUBLE)
#define VSETVL(n) RISCV_RVV(vsetvl_e32m8)(n)
#define FLOAT_V_T vfloat32m8_t
#define VLEV_FLOAT RISCV_RVV(vle32_v_f32m8)
#define VLSEV_FLOAT RISCV_RVV(vlse32_v_f32m8)
#define VSEV_FLOAT RISCV_RVV(vse32_v_f32m8)
#define VSSEV_FLOAT RISCV_RVV(vsse32_v_f32m8)
#define VFMACCVF_FLOAT RISCV_RVV(vfmacc_vf_f32m8)
#define VFMUL_VF_FLOAT RISCV_RVV(vfmul_vf_f32m8)
#define VREINTERPRET_FLOAT RISCV_RVV(vreinterpret_v_i32m8_f32m8)
#define VFILL_INT RISCV_RVV(vmv_v_x_i32m8)
#define VSETVL(n) RISCV_RVV(vsetvl_e32m8)(n)
#define FLOAT_V_T vfloat32m8_t
#define VLEV_FLOAT RISCV_RVV(vle32_v_f32m8)
#define VLSEV_FLOAT RISCV_RVV(vlse32_v_f32m8)
#define VSEV_FLOAT RISCV_RVV(vse32_v_f32m8)
#define VSSEV_FLOAT RISCV_RVV(vsse32_v_f32m8)
#define VFMACCVF_FLOAT RISCV_RVV(vfmacc_vf_f32m8)
#else
#define VSETVL(n) RISCV_RVV(vsetvl_e64m4)(n)
#define FLOAT_V_T vfloat64m4_t
#define VLEV_FLOAT RISCV_RVV(vle64_v_f64m4)
#define VLSEV_FLOAT RISCV_RVV(vlse64_v_f64m4)
#define VSEV_FLOAT RISCV_RVV(vse64_v_f64m4)
#define VSSEV_FLOAT RISCV_RVV(vsse64_v_f64m4)
#define VFMACCVF_FLOAT RISCV_RVV(vfmacc_vf_f64m4)
#define VFMUL_VF_FLOAT RISCV_RVV(vfmul_vf_f64m4)
#define VREINTERPRET_FLOAT RISCV_RVV(vreinterpret_v_i64m4_f64m4)
#define VFILL_INT RISCV_RVV(vmv_v_x_i64m4)
#define VSETVL(n) RISCV_RVV(vsetvl_e64m8)(n)
#define FLOAT_V_T vfloat64m8_t
#define VLEV_FLOAT RISCV_RVV(vle64_v_f64m8)
#define VLSEV_FLOAT RISCV_RVV(vlse64_v_f64m8)
#define VSEV_FLOAT RISCV_RVV(vse64_v_f64m8)
#define VSSEV_FLOAT RISCV_RVV(vsse64_v_f64m8)
#define VFMACCVF_FLOAT RISCV_RVV(vfmacc_vf_f64m8)
#endif

int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
{
BLASLONG i = 0, j = 0, k = 0;
BLASLONG ix = 0, iy = 0;

if(n < 0) return(0);
FLOAT *a_ptr = a;
FLOAT temp[4];
FLOAT_V_T va0, va1, vy0, vy1,vy0_temp, vy1_temp ,va0_0 , va0_1 , va1_0 ,va1_1 ,va2_0 ,va2_1 ,va3_0 ,va3_1 ;
unsigned int gvl = 0;
if(inc_y == 1 && inc_x == 1){
gvl = VSETVL(m);
if(gvl <= m/2){
for(k=0,j=0; k<m/(2*gvl); k++){
a_ptr = a;
ix = 0;
vy0_temp = VLEV_FLOAT(&y[j], gvl);
vy1_temp = VLEV_FLOAT(&y[j+gvl], gvl);
vy0 = VREINTERPRET_FLOAT(VFILL_INT(0, gvl));
vy1 = VREINTERPRET_FLOAT(VFILL_INT(0, gvl));
int i;

int remainder = n % 4;
for(i = 0; i < remainder; i++){
temp[0] = x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp[0], va0, gvl);

va1 = VLEV_FLOAT(&a_ptr[j+gvl], gvl);
vy1 = VFMACCVF_FLOAT(vy1, temp[0], va1, gvl);
a_ptr += lda;
ix ++;
}

for(i = remainder; i < n; i += 4){
va0_0 = VLEV_FLOAT(&(a_ptr)[j], gvl);
va0_1 = VLEV_FLOAT(&(a_ptr)[j+gvl], gvl);
va1_0 = VLEV_FLOAT(&(a_ptr+lda * 1)[j], gvl);
va1_1 = VLEV_FLOAT(&(a_ptr+lda * 1)[j+gvl], gvl);
va2_0 = VLEV_FLOAT(&(a_ptr+lda * 2)[j], gvl);
va2_1 = VLEV_FLOAT(&(a_ptr+lda * 2)[j+gvl], gvl);
va3_0 = VLEV_FLOAT(&(a_ptr+lda * 3)[j], gvl);
va3_1 = VLEV_FLOAT(&(a_ptr+lda * 3)[j+gvl], gvl);

vy0 = VFMACCVF_FLOAT(vy0, x[ix], va0_0, gvl);
vy1 = VFMACCVF_FLOAT(vy1, x[ix], va0_1, gvl);

vy0 = VFMACCVF_FLOAT(vy0, x[ix+1], va1_0, gvl);
vy1 = VFMACCVF_FLOAT(vy1, x[ix+1], va1_1, gvl);
if (n < 0) return(0);

vy0 = VFMACCVF_FLOAT(vy0, x[ix+2], va2_0, gvl);
vy1 = VFMACCVF_FLOAT(vy1, x[ix+2], va2_1, gvl);

vy0 = VFMACCVF_FLOAT(vy0, x[ix+3], va3_0, gvl);
vy1 = VFMACCVF_FLOAT(vy1, x[ix+3], va3_1, gvl);
a_ptr += 4 * lda;
ix +=4;
}
vy0 = VFMACCVF_FLOAT(vy0_temp, alpha, vy0, gvl);
vy1 = VFMACCVF_FLOAT(vy1_temp, alpha, vy1, gvl);
VSEV_FLOAT(&y[j], vy0, gvl);
VSEV_FLOAT(&y[j+gvl], vy1, gvl);
j += gvl * 2;
}
}
//tail
if(gvl <= m - j ){
a_ptr = a;
ix = 0;
vy0_temp = VLEV_FLOAT(&y[j], gvl);
vy0 = VREINTERPRET_FLOAT(VFILL_INT(0, gvl));
int i;

int remainder = n % 4;
for(i = 0; i < remainder; i++){
temp[0] = x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp[0], va0, gvl);
a_ptr += lda;
ix ++;
}

for(i = remainder; i < n; i += 4){
va0_0 = VLEV_FLOAT(&(a_ptr)[j], gvl);
va1_0 = VLEV_FLOAT(&(a_ptr+lda * 1)[j], gvl);
va2_0 = VLEV_FLOAT(&(a_ptr+lda * 2)[j], gvl);
va3_0 = VLEV_FLOAT(&(a_ptr+lda * 3)[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, x[ix], va0_0, gvl);
vy0 = VFMACCVF_FLOAT(vy0, x[ix+1], va1_0, gvl);
vy0 = VFMACCVF_FLOAT(vy0, x[ix+2], va2_0, gvl);
vy0 = VFMACCVF_FLOAT(vy0, x[ix+3], va3_0, gvl);
a_ptr += 4 * lda;
ix +=4;
}
vy0 = VFMACCVF_FLOAT(vy0_temp, alpha, vy0, gvl);

VSEV_FLOAT(&y[j], vy0, gvl);

j += gvl ;
}

FLOAT *a_ptr, *y_ptr, temp;
BLASLONG i, j, vl;
FLOAT_V_T va, vy;

for(;j < m;){
gvl = VSETVL(m-j);
if (inc_y == 1) {
for (j = 0; j < n; j++) {
temp = alpha * x[0];
y_ptr = y;
a_ptr = a;
ix = 0;
vy0 = VLEV_FLOAT(&y[j], gvl);
for(i = 0; i < n; i++){
temp[0] = alpha * x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp[0], va0, gvl);

a_ptr += lda;
ix += inc_x;
}
VSEV_FLOAT(&y[j], vy0, gvl);
j += gvl;
}
}else if (inc_y == 1 && inc_x !=1) {
gvl = VSETVL(m);
if(gvl <= m/2){
for(k=0,j=0; k<m/(2*gvl); k++){
a_ptr = a;
ix = 0;
vy0 = VLEV_FLOAT(&y[j], gvl);
vy1 = VLEV_FLOAT(&y[j+gvl], gvl);
for(i = 0; i < n; i++){
temp[0] = alpha * x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp[0], va0, gvl);

va1 = VLEV_FLOAT(&a_ptr[j+gvl], gvl);
vy1 = VFMACCVF_FLOAT(vy1, temp[0], va1, gvl);
a_ptr += lda;
ix += inc_x;
}
VSEV_FLOAT(&y[j], vy0, gvl);
VSEV_FLOAT(&y[j+gvl], vy1, gvl);
j += gvl * 2;
}
}
//tail
for(;j < m;){
gvl = VSETVL(m-j);
a_ptr = a;
ix = 0;
vy0 = VLEV_FLOAT(&y[j], gvl);
for(i = 0; i < n; i++){
temp[0] = alpha * x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp[0], va0, gvl);

a_ptr += lda;
ix += inc_x;
for (i = m; i > 0; i -= vl) {
vl = VSETVL(i);
vy = VLEV_FLOAT(y_ptr, vl);
va = VLEV_FLOAT(a_ptr, vl);
vy = VFMACCVF_FLOAT(vy, temp, va, vl);
VSEV_FLOAT(y_ptr, vy, vl);
y_ptr += vl;
a_ptr += vl;
}
VSEV_FLOAT(&y[j], vy0, gvl);
j += gvl;
x += inc_x;
a += lda;
}
}else{
} else {
BLASLONG stride_y = inc_y * sizeof(FLOAT);
gvl = VSETVL(m);
if(gvl <= m/2){
BLASLONG inc_yv = inc_y * gvl;
for(k=0,j=0; k<m/(2*gvl); k++){
a_ptr = a;
ix = 0;
vy0 = VLSEV_FLOAT(&y[iy], stride_y, gvl);
vy1 = VLSEV_FLOAT(&y[iy+inc_yv], stride_y, gvl);
for(i = 0; i < n; i++){
temp[0] = alpha * x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp[0], va0, gvl);

va1 = VLEV_FLOAT(&a_ptr[j+gvl], gvl);
vy1 = VFMACCVF_FLOAT(vy1, temp[0], va1, gvl);
a_ptr += lda;
ix += inc_x;
}
VSSEV_FLOAT(&y[iy], stride_y, vy0, gvl);
VSSEV_FLOAT(&y[iy+inc_yv], stride_y, vy1, gvl);
j += gvl * 2;
iy += inc_yv * 2;
}
}
//tail
for(;j < m;){
gvl = VSETVL(m-j);
for (j = 0; j < n; j++) {
temp = alpha * x[0];
y_ptr = y;
a_ptr = a;
ix = 0;
vy0 = VLSEV_FLOAT(&y[j*inc_y], stride_y, gvl);
for(i = 0; i < n; i++){
temp[0] = alpha * x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp[0], va0, gvl);

a_ptr += lda;
ix += inc_x;
for (i = m; i > 0; i -= vl) {
vl = VSETVL(i);
vy = VLSEV_FLOAT(y_ptr, stride_y, vl);
va = VLEV_FLOAT(a_ptr, vl);
vy = VFMACCVF_FLOAT(vy, temp, va, vl);
VSSEV_FLOAT(y_ptr, stride_y, vy, vl);
y_ptr += vl * inc_y;
a_ptr += vl;
}
VSSEV_FLOAT(&y[j*inc_y], stride_y, vy0, gvl);
j += gvl;
x += inc_x;
a += lda;
}
}
return(0);
Expand Down
Loading