bugfix for sbgemm_ncopy_8_neoversen2

lilh9598 · lilh9598 · commit 729af6406fd9 · 2022-12-05T05:10:18.000-05:00
diff --git a/kernel/arm64/sbgemm_ncopy_8_neoversen2.c b/kernel/arm64/sbgemm_ncopy_8_neoversen2.c
@@ -83,13 +83,13 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
 
     if (m & 3) {
       BLASLONG rest = m & 3;
-      for (BLASLONG col = 0; col < 4; col++) {
+      for (BLASLONG col = 0; col < 8; col++) {
         b_offset[4 * col] = a_offsetx[col][0];
         b_offset[4 * col + 1] = rest == 1 ? 0 : a_offsetx[col][1];
         b_offset[4 * col + 2] = rest <= 2 ? 0 : a_offsetx[col][2];
         b_offset[4 * col + 3] = rest <= 3 ? 0 : a_offsetx[col][3];
       }
-      b_offset += 16;
+      b_offset += 32;
     }
   }
 

Original file line number	Diff line number	Diff line change
`@@ -83,13 +83,13 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT a, BLASLONG lda, IFLOAT b) {`
`83`	`83`
`84`	`84`	`if (m & 3) {`
`85`	`85`	`BLASLONG rest = m & 3;`
`86`		`- for (BLASLONG col = 0; col < 4; col++) {`
	`86`	`+ for (BLASLONG col = 0; col < 8; col++) {`
`87`	`87`	`b_offset[4 * col] = a_offsetx[col][0];`
`88`	`88`	`b_offset[4 * col + 1] = rest == 1 ? 0 : a_offsetx[col][1];`
`89`	`89`	`b_offset[4 * col + 2] = rest <= 2 ? 0 : a_offsetx[col][2];`
`90`	`90`	`b_offset[4 * col + 3] = rest <= 3 ? 0 : a_offsetx[col][3];`
`91`	`91`	`}`
`92`		`- b_offset += 16;`
	`92`	`+ b_offset += 32;`
`93`	`93`	`}`
`94`	`94`	`}`
`95`	`95`