@@ -126,7 +126,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
126
126
/* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */
127
127
if (min_jj >= 6 * GEMM_UNROLL_N ) min_jj = 6 * GEMM_UNROLL_N ;
128
128
#else
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
#endif
@@ -150,7 +150,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
150
150
/* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */
151
151
if (min_jj >= 6 * GEMM_UNROLL_N ) min_jj = 6 * GEMM_UNROLL_N ;
152
152
#else
153
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
153
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
154
154
else
155
155
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
156
156
#endif
@@ -207,7 +207,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
207
207
/* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */
208
208
if (min_jj >= 6 * GEMM_UNROLL_N ) min_jj = 6 * GEMM_UNROLL_N ;
209
209
#else
210
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
210
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
211
211
else
212
212
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
213
213
#endif
@@ -262,7 +262,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
262
262
/* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */
263
263
if (min_jj >= 6 * GEMM_UNROLL_N ) min_jj = 6 * GEMM_UNROLL_N ;
264
264
#else
265
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
265
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
266
266
else
267
267
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
268
268
#endif
@@ -287,7 +287,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
287
287
/* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */
288
288
if (min_jj >= 6 * GEMM_UNROLL_N ) min_jj = 6 * GEMM_UNROLL_N ;
289
289
#else
290
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
290
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
291
291
else
292
292
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
293
293
#endif
@@ -348,7 +348,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
348
348
/* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */
349
349
if (min_jj >= 6 * GEMM_UNROLL_N ) min_jj = 6 * GEMM_UNROLL_N ;
350
350
#else
351
- if (min_jj > GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
351
+ if (min_jj >= GEMM_UNROLL_N * 3 ) min_jj = GEMM_UNROLL_N * 3 ;
352
352
else
353
353
if (min_jj > GEMM_UNROLL_N ) min_jj = GEMM_UNROLL_N ;
354
354
#endif
0 commit comments