format code
This commit is contained in:
parent
55d686d41e
commit
bc3728475f
|
@ -27,10 +27,11 @@
|
||||||
* *****************************************************************************/
|
* *****************************************************************************/
|
||||||
|
|
||||||
#include <arm_sve.h>
|
#include <arm_sve.h>
|
||||||
|
|
||||||
#include "common.h"
|
#include "common.h"
|
||||||
|
|
||||||
int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
|
int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B, FLOAT *C,
|
||||||
FLOAT *C, BLASLONG ldc) {
|
BLASLONG ldc) {
|
||||||
// printf("m: %d, n: %d, k: %d\n", m, n, k);
|
// printf("m: %d, n: %d, k: %d\n", m, n, k);
|
||||||
BLASLONG padk = (k + 3) & ~3;
|
BLASLONG padk = (k + 3) & ~3;
|
||||||
BLASLONG padm = (m + 1) & ~1;
|
BLASLONG padm = (m + 1) & ~1;
|
||||||
|
@ -78,10 +79,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
|
||||||
ptr_b0 = ptr_b;
|
ptr_b0 = ptr_b;
|
||||||
ptr_b1 = ptr_b0 + 2 * padk;
|
ptr_b1 = ptr_b0 + 2 * padk;
|
||||||
|
|
||||||
mc00 = svdup_f32(0); mc01 = svdup_f32(0);
|
mc00 = svdup_f32(0);
|
||||||
mc10 = svdup_f32(0); mc11 = svdup_f32(0);
|
mc01 = svdup_f32(0);
|
||||||
mc20 = svdup_f32(0); mc21 = svdup_f32(0);
|
mc10 = svdup_f32(0);
|
||||||
mc30 = svdup_f32(0); mc31 = svdup_f32(0);
|
mc11 = svdup_f32(0);
|
||||||
|
mc20 = svdup_f32(0);
|
||||||
|
mc21 = svdup_f32(0);
|
||||||
|
mc30 = svdup_f32(0);
|
||||||
|
mc31 = svdup_f32(0);
|
||||||
|
|
||||||
for (BLASLONG p = 0; p < padk / 4; p++) {
|
for (BLASLONG p = 0; p < padk / 4; p++) {
|
||||||
ma0 = svld1_bf16(pg16, (bfloat16_t *)ptr_a0);
|
ma0 = svld1_bf16(pg16, (bfloat16_t *)ptr_a0);
|
||||||
|
@ -135,8 +140,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
|
||||||
ptr_b0 = ptr_b;
|
ptr_b0 = ptr_b;
|
||||||
ptr_b1 = ptr_b0 + 2 * padk;
|
ptr_b1 = ptr_b0 + 2 * padk;
|
||||||
|
|
||||||
mc00 = svdup_f32(0); mc01 = svdup_f32(0);
|
mc00 = svdup_f32(0);
|
||||||
mc10 = svdup_f32(0); mc11 = svdup_f32(0);
|
mc01 = svdup_f32(0);
|
||||||
|
mc10 = svdup_f32(0);
|
||||||
|
mc11 = svdup_f32(0);
|
||||||
for (BLASLONG p = 0; p < padk / 4; p++) {
|
for (BLASLONG p = 0; p < padk / 4; p++) {
|
||||||
ma0 = svld1_bf16(pg16, (bfloat16_t *)ptr_a0);
|
ma0 = svld1_bf16(pg16, (bfloat16_t *)ptr_a0);
|
||||||
ma1 = svld1_bf16(pg16, (bfloat16_t *)ptr_a1);
|
ma1 = svld1_bf16(pg16, (bfloat16_t *)ptr_a1);
|
||||||
|
@ -171,7 +178,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
|
||||||
ptr_b0 = ptr_b;
|
ptr_b0 = ptr_b;
|
||||||
ptr_b1 = ptr_b0 + 2 * padk;
|
ptr_b1 = ptr_b0 + 2 * padk;
|
||||||
|
|
||||||
mc00 = svdup_f32(0); mc01 = svdup_f32(0);
|
mc00 = svdup_f32(0);
|
||||||
|
mc01 = svdup_f32(0);
|
||||||
for (BLASLONG p = 0; p < padk / 4; p++) {
|
for (BLASLONG p = 0; p < padk / 4; p++) {
|
||||||
ma0 = svld1_bf16(pg16, (bfloat16_t *)ptr_a0);
|
ma0 = svld1_bf16(pg16, (bfloat16_t *)ptr_a0);
|
||||||
mb0 = svld1_bf16(pg16, (bfloat16_t *)ptr_b0);
|
mb0 = svld1_bf16(pg16, (bfloat16_t *)ptr_b0);
|
||||||
|
@ -189,7 +197,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
|
||||||
}
|
}
|
||||||
|
|
||||||
ptr_b += 4 * padk;
|
ptr_b += 4 * padk;
|
||||||
|
|
||||||
}
|
}
|
||||||
|
|
||||||
if (padn & 2) {
|
if (padn & 2) {
|
||||||
|
@ -299,7 +306,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
|
||||||
for (i = 0; i < m / 4; i++) {
|
for (i = 0; i < m / 4; i++) {
|
||||||
org_vc0 = svld1_f32(pg32, org_c0);
|
org_vc0 = svld1_f32(pg32, org_c0);
|
||||||
raw_vc0 = svld1_f32(pg32, raw_c0);
|
raw_vc0 = svld1_f32(pg32, raw_c0);
|
||||||
org_vc0 = svmad_z(pg32, svalpha, raw_vc0, org_vc0); // alpha * raw + org, raw -> a * b
|
org_vc0 = svmad_z(pg32, svalpha, raw_vc0,
|
||||||
|
org_vc0); // alpha * raw + org, raw -> a * b
|
||||||
svst1_f32(pg32, org_c0, org_vc0);
|
svst1_f32(pg32, org_c0, org_vc0);
|
||||||
org_c0 += 4;
|
org_c0 += 4;
|
||||||
raw_c0 += 4;
|
raw_c0 += 4;
|
||||||
|
@ -310,5 +318,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT *A, IFLOAT *B,
|
||||||
raw_c0++;
|
raw_c0++;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
return 0;
|
return 0;
|
||||||
}
|
}
|
||||||
|
|
|
@ -97,7 +97,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
|
||||||
*(b_offset + 6) = 0;
|
*(b_offset + 6) = 0;
|
||||||
*(b_offset + 7) = 0;
|
*(b_offset + 7) = 0;
|
||||||
a_offset += 4;
|
a_offset += 4;
|
||||||
b_offset += 8;
|
b_offset += 4;
|
||||||
}
|
}
|
||||||
if (i < m) {
|
if (i < m) {
|
||||||
*(b_offset + 4) = 0;
|
*(b_offset + 4) = 0;
|
||||||
|
|
|
@ -87,7 +87,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
|
||||||
b_offset += 8;
|
b_offset += 8;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
if (j < n) { // padding 2
|
if (j < n) { // rest 1
|
||||||
BLASLONG i = 0;
|
BLASLONG i = 0;
|
||||||
for (; i < m4; i += 4) {
|
for (; i < m4; i += 4) {
|
||||||
*(b_offset + 0) = *(a_offset + 0);
|
*(b_offset + 0) = *(a_offset + 0);
|
||||||
|
@ -98,7 +98,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
|
||||||
*(b_offset + 5) = 0;
|
*(b_offset + 5) = 0;
|
||||||
*(b_offset + 6) = 0;
|
*(b_offset + 6) = 0;
|
||||||
*(b_offset + 7) = 0;
|
*(b_offset + 7) = 0;
|
||||||
b_offset += 8;
|
b_offset += 4;
|
||||||
a_offset += 4 * lda;
|
a_offset += 4 * lda;
|
||||||
}
|
}
|
||||||
if (i < m) {
|
if (i < m) {
|
||||||
|
@ -113,5 +113,6 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
|
||||||
*(b_offset + 3) = 0;
|
*(b_offset + 3) = 0;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
return 0;
|
return 0;
|
||||||
}
|
}
|
||||||
|
|
Loading…
Reference in New Issue