OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Optimizing the Implementation of GEMV on the RISC-V V Extension 

Specialized some scenarios, performed loop unrolling, and reduced the
number of multiplications.
tags/v0.3.30
lglglglgy 5 months ago
parent
67c5bdd639
commit
1ff303f36e
1 changed files with 207 additions and 97 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +207
    -97
      kernel/riscv64/gemv_n_vector.c

+ 207
- 97
kernel/riscv64/gemv_n_vector.c View File

@@ -27,13 +27,15 @@ 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_e32m4)(n)
#define FLOAT_V_T vfloat32m4_t
#define VLEV_FLOAT RISCV_RVV(vle32_v_f32m4)
#define VLSEV_FLOAT RISCV_RVV(vlse32_v_f32m4)
#define VSEV_FLOAT RISCV_RVV(vse32_v_f32m4)
#define VSSEV_FLOAT RISCV_RVV(vsse32_v_f32m4)
#define VFMACCVF_FLOAT RISCV_RVV(vfmacc_vf_f32m4)
#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 VFILL_ZERO_FLOAT RISCV_RVV(vfsub_vv_f32m8)
#else
#define VSETVL(n) RISCV_RVV(vsetvl_e64m4)(n)
#define FLOAT_V_T vfloat64m4_t
@@ -42,103 +44,211 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#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 VFILL_ZERO_FLOAT RISCV_RVV(vfsub_vv_f64m4)
#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 = 0.0;
FLOAT_V_T va0, va1, vy0, vy1;
unsigned int gvl = 0;
if(inc_y == 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 = alpha * x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp, va0, gvl);

va1 = VLEV_FLOAT(&a_ptr[j+gvl], gvl);
vy1 = VFMACCVF_FLOAT(vy1, temp, 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 = alpha * x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp, va0, gvl);

a_ptr += lda;
ix += inc_x;
}
VSEV_FLOAT(&y[j], vy0, gvl);
j += gvl;
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 , temp_v ,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 = VFILL_ZERO_FLOAT(vy0 , vy0 , gvl);
vy1 = VFILL_ZERO_FLOAT(vy1 , vy1 , 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 ++;
}
}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 = alpha * x[ix];
va0 = VLEV_FLOAT(&a_ptr[j], gvl);
vy0 = VFMACCVF_FLOAT(vy0, temp, va0, gvl);

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

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);

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

a_ptr += lda;
ix += inc_x;
}
VSSEV_FLOAT(&y[j*inc_y], stride_y, vy0, gvl);
j += gvl;
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 = VFILL_ZERO_FLOAT(vy0 , vy0 , 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 ;
}

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;
}
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;
}
}
return(0);
}
//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;
}
VSEV_FLOAT(&y[j], vy0, gvl);
j += gvl;
}
}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);
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;
}
VSSEV_FLOAT(&y[j*inc_y], stride_y, vy0, gvl);
j += gvl;
}
}
return(0);
}

Write Preview
Loading…
Cancel
Save