@@ -126,7 +126,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
126
126
127
127
for (jjs = js ; jjs < js + min_j ; jjs += min_jj ){
128
128
min_jj = min_j + js - jjs ;
129
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
129
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
130
130
else
131
131
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
132
132
@@ -182,7 +182,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
182
182
183
183
for (jjs = 0 ; jjs < min_j - min_l - ls + js ; jjs += min_jj ){
184
184
min_jj = min_j - min_l - ls + js - jjs ;
185
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
185
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
186
186
else
187
187
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
188
188
@@ -243,7 +243,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
243
243
244
244
for (jjs = js ; jjs < js + min_j ; jjs += min_jj ){
245
245
min_jj = min_j + js - jjs ;
246
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
246
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
247
247
else
248
248
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
249
249
@@ -304,7 +304,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
304
304
305
305
for (jjs = 0 ; jjs < min_j - js + ls ; jjs += min_jj ){
306
306
min_jj = min_j - js + ls - jjs ;
307
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
307
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
308
308
else
309
309
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
310
310
0 commit comments