Refs #2587 fix small matrix c/zgemm bug.

This commit is contained in:
Xianyi Zhang 2020-08-28 22:36:36 +08:00
parent 255b6dd0fa
commit 4130d1732e
10 changed files with 105 additions and 109 deletions

View File

@ -536,85 +536,85 @@ int dgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLA
int dgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int dgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int cgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_nr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_nc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_nr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_nc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_tr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_tc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_tr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_tc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_rn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_rt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_rr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_rc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_rn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_rt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_rr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_rc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_cn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_ct(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_cr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_cc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * beta, float * C, BLASLONG ldc);
int cgemm_small_kernel_cn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_ct(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_cr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int cgemm_small_kernel_cc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
int zgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_nr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_nc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_nr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_nc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_tr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_tc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_tr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_tc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_rn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_rt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_rr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_rc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_rn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_rt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_rr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_rc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_cn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_ct(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_cr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_cc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * beta, double * C, BLASLONG ldc);
int zgemm_small_kernel_cn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_ct(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_cr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int zgemm_small_kernel_cc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
int cgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_nr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_nc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_nr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_nc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_tr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_tc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_tr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_tc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_rn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_rt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_rr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_rc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_rn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_rt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_rr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_rc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_cn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_ct(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_cr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_cc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_cn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_ct(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_cr(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int cgemm_small_kernel_b0_cc(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
int zgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_nr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_nc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_nr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_nc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_tr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_tc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_tr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_tc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_rn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_rt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_rr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_rc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_rn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_rt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_rr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_rc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_cn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_ct(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_cr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_cc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_cn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_ct(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_cr(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
int zgemm_small_kernel_b0_cc(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
#endif

View File

@ -104,47 +104,43 @@ static int (*gemm[])(blas_arg_t *, BLASLONG *, BLASLONG *, IFLOAT *, IFLOAT *, B
};
#ifdef SMALL_MATRIX_OPT
//Only support s/dgemm small matrix optimiztion so far.
#ifndef COMPLEX
static int (*gemm_small_kernel[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT ,FLOAT *, BLASLONG, FLOAT, FLOAT *, BLASLONG) = {
#ifndef GEMM3M
#ifndef COMPLEX
GEMM_SMALL_KERNEL_NN, GEMM_SMALL_KERNEL_TN, NULL, NULL,
GEMM_SMALL_KERNEL_NT, GEMM_SMALL_KERNEL_TT, NULL, NULL,
#endif
#endif
};
static int (*gemm_small_kernel_b0[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG) = {
#ifndef GEMM3M
#ifndef COMPLEX
GEMM_SMALL_KERNEL_B0_NN, GEMM_SMALL_KERNEL_B0_TN, NULL, NULL,
GEMM_SMALL_KERNEL_B0_NT, GEMM_SMALL_KERNEL_B0_TT, NULL, NULL,
#endif
#endif
};
static int (*zgemm_small_kernel[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT *,FLOAT *, BLASLONG, FLOAT *, FLOAT *, BLASLONG) = {
#else
static int (*zgemm_small_kernel[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG) = {
#ifndef GEMM3M
#ifdef COMPLEX
GEMM_SMALL_KERNEL_NN, GEMM_SMALL_KERNEL_TN, GEMM_SMALL_KERNEL_RN, GEMM_SMALL_KERNEL_CN,
GEMM_SMALL_KERNEL_NT, GEMM_SMALL_KERNEL_TT, GEMM_SMALL_KERNEL_RT, GEMM_SMALL_KERNEL_CT,
GEMM_SMALL_KERNEL_NR, GEMM_SMALL_KERNEL_TR, GEMM_SMALL_KERNEL_RR, GEMM_SMALL_KERNEL_CR,
GEMM_SMALL_KERNEL_NC, GEMM_SMALL_KERNEL_TC, GEMM_SMALL_KERNEL_RC, GEMM_SMALL_KERNEL_CC,
#endif
#endif
};
static int (*zgemm_small_kernel_b0[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT *, FLOAT *, BLASLONG, FLOAT *, BLASLONG) = {
static int (*zgemm_small_kernel_b0[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG) = {
#ifndef GEMM3M
#ifdef COMPLEX
GEMM_SMALL_KERNEL_B0_NN, GEMM_SMALL_KERNEL_B0_TN, GEMM_SMALL_KERNEL_B0_RN, GEMM_SMALL_KERNEL_B0_CN,
GEMM_SMALL_KERNEL_B0_NT, GEMM_SMALL_KERNEL_B0_TT, GEMM_SMALL_KERNEL_B0_RT, GEMM_SMALL_KERNEL_B0_CT,
GEMM_SMALL_KERNEL_B0_NR, GEMM_SMALL_KERNEL_B0_TR, GEMM_SMALL_KERNEL_B0_RR, GEMM_SMALL_KERNEL_B0_CR,
GEMM_SMALL_KERNEL_B0_NC, GEMM_SMALL_KERNEL_B0_TC, GEMM_SMALL_KERNEL_B0_RC, GEMM_SMALL_KERNEL_B0_CC,
#endif
#endif
};
#endif
#endif
#ifndef CBLAS
@ -476,9 +472,9 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
}
#else
if(beta[0] == 0.0 && beta[1] == 0.0){
(zgemm_small_kernel_b0[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda, (FLOAT *)(args.alpha), args.b, args.ldb, args.c, args.ldc);
(zgemm_small_kernel_b0[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda, alpha[0], alpha[1], args.b, args.ldb, args.c, args.ldc);
}else{
(zgemm_small_kernel[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda, (FLOAT *)(args.alpha), args.b, args.ldb, (FLOAT *)(args.beta), args.c, args.ldc);
(zgemm_small_kernel[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda, alpha[0], alpha[1], args.b, args.ldb, beta[0], beta[1], args.c, args.ldc);
}
#endif
return;

View File

@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* alpha, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha0, FLOAT alpha1, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
{
FLOAT real, imag;
@ -65,8 +65,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* al
#endif
}
C[j*2*ldc + 2*i] = alpha[0]*real - alpha[1]*imag;
C[j*2*ldc+ 2*i + 1] = alpha[0]*imag + real*alpha[1];
C[j*2*ldc + 2*i] = alpha0*real - alpha1*imag;
C[j*2*ldc+ 2*i + 1] = alpha0*imag + real*alpha1;
}
}

View File

@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* alpha, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha0, FLOAT alpha1, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
{
FLOAT real, imag;
int i, j, l;
@ -68,8 +68,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* al
#endif
}
C[j*2*ldc + 2*i] = alpha[0]*real - alpha[1]*imag;
C[j*2*ldc+ 2*i + 1] = alpha[0]*imag + real*alpha[1];
C[j*2*ldc + 2*i] = alpha0*real - alpha1*imag;
C[j*2*ldc+ 2*i + 1] = alpha0*imag + real*alpha1;
}
}

View File

@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* alpha, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha0, FLOAT alpha1, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
{
FLOAT real, imag;
int i, j, l;
@ -68,8 +68,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* al
#endif
}
C[j*2*ldc + 2*i] = alpha[0]*real - alpha[1]*imag;
C[j*2*ldc+ 2*i + 1] = alpha[0]*imag + real*alpha[1];
C[j*2*ldc + 2*i] = alpha0*real - alpha1*imag;
C[j*2*ldc+ 2*i + 1] = alpha0*imag + real*alpha1;
}
}

View File

@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* alpha, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha0, FLOAT alpha1, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
{
FLOAT real, imag;
int i, j, l;
@ -68,8 +68,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* al
#endif
}
C[j*2*ldc + 2*i] = alpha[0]*real - alpha[1]*imag;
C[j*2*ldc+ 2*i + 1] = alpha[0]*imag + real*alpha[1];
C[j*2*ldc + 2*i] = alpha0*real - alpha1*imag;
C[j*2*ldc+ 2*i + 1] = alpha0*imag + real*alpha1;
}
}

View File

@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* alpha, FLOAT * B, BLASLONG ldb, FLOAT* beta, FLOAT * C, BLASLONG ldc)
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha0, FLOAT alpha1, FLOAT * B, BLASLONG ldb, FLOAT beta0, FLOAT beta1, FLOAT * C, BLASLONG ldc)
{
FLOAT real, imag;
FLOAT tmp0, tmp1;
@ -65,12 +65,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* al
#endif
}
tmp0 = beta[0]*C[j*2*ldc + 2*i] - beta[1]*C[j*2*ldc+ 2*i + 1];
tmp1 = beta[0]*C[j*2*ldc+ 2*i + 1] + beta[1]*C[j*2*ldc + 2*i];
tmp0 = beta0*C[j*2*ldc + 2*i] - beta1*C[j*2*ldc+ 2*i + 1];
tmp1 = beta0*C[j*2*ldc+ 2*i + 1] + beta1*C[j*2*ldc + 2*i];
C[j*2*ldc + 2*i] =tmp0+ alpha[0]*real - alpha[1]*imag;
C[j*2*ldc+ 2*i + 1] = tmp1+ alpha[0]*imag + real*alpha[1];
C[j*2*ldc + 2*i] =tmp0+ alpha0*real - alpha1*imag;
C[j*2*ldc+ 2*i + 1] = tmp1+ alpha0*imag + real*alpha1;
}
}

View File

@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* alpha, FLOAT * B, BLASLONG ldb, FLOAT* beta, FLOAT * C, BLASLONG ldc)
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha0, FLOAT alpha1, FLOAT * B, BLASLONG ldb, FLOAT beta0, FLOAT beta1, FLOAT * C, BLASLONG ldc)
{
FLOAT real, imag;
FLOAT tmp0, tmp1;
@ -69,12 +69,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* al
#endif
}
tmp0 = beta[0]*C[j*2*ldc + 2*i] - beta[1]*C[j*2*ldc+ 2*i + 1];
tmp1 = beta[0]*C[j*2*ldc+ 2*i + 1] + beta[1]*C[j*2*ldc + 2*i];
tmp0 = beta0*C[j*2*ldc + 2*i] - beta1*C[j*2*ldc+ 2*i + 1];
tmp1 = beta0*C[j*2*ldc+ 2*i + 1] + beta1*C[j*2*ldc + 2*i];
C[j*2*ldc + 2*i] =tmp0+ alpha[0]*real - alpha[1]*imag;
C[j*2*ldc+ 2*i + 1] = tmp1+ alpha[0]*imag + real*alpha[1];
C[j*2*ldc + 2*i] =tmp0+ alpha0*real - alpha1*imag;
C[j*2*ldc+ 2*i + 1] = tmp1+ alpha0*imag + real*alpha1;
}
}

View File

@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* alpha, FLOAT * B, BLASLONG ldb, FLOAT* beta, FLOAT * C, BLASLONG ldc)
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha0, FLOAT alpha1, FLOAT * B, BLASLONG ldb, FLOAT beta0, FLOAT beta1, FLOAT * C, BLASLONG ldc)
{
FLOAT real, imag;
FLOAT tmp0, tmp1;
@ -69,12 +69,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* al
#endif
}
tmp0 = beta[0]*C[j*2*ldc + 2*i] - beta[1]*C[j*2*ldc+ 2*i + 1];
tmp1 = beta[0]*C[j*2*ldc+ 2*i + 1] + beta[1]*C[j*2*ldc + 2*i];
tmp0 = beta0*C[j*2*ldc + 2*i] - beta1*C[j*2*ldc+ 2*i + 1];
tmp1 = beta0*C[j*2*ldc+ 2*i + 1] + beta1*C[j*2*ldc + 2*i];
C[j*2*ldc + 2*i] =tmp0+ alpha[0]*real - alpha[1]*imag;
C[j*2*ldc+ 2*i + 1] = tmp1+ alpha[0]*imag + real*alpha[1];
C[j*2*ldc + 2*i] =tmp0+ alpha0*real - alpha1*imag;
C[j*2*ldc+ 2*i + 1] = tmp1+ alpha0*imag + real*alpha1;
}
}

View File

@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* alpha, FLOAT * B, BLASLONG ldb, FLOAT* beta, FLOAT * C, BLASLONG ldc)
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha0, FLOAT alpha1, FLOAT * B, BLASLONG ldb, FLOAT beta0, FLOAT beta1, FLOAT * C, BLASLONG ldc)
{
FLOAT real, imag;
FLOAT tmp0, tmp1;
@ -69,12 +69,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT* al
#endif
}
tmp0 = beta[0]*C[j*2*ldc + 2*i] - beta[1]*C[j*2*ldc+ 2*i + 1];
tmp1 = beta[0]*C[j*2*ldc+ 2*i + 1] + beta[1]*C[j*2*ldc + 2*i];
tmp0 = beta0*C[j*2*ldc + 2*i] - beta1*C[j*2*ldc+ 2*i + 1];
tmp1 = beta0*C[j*2*ldc+ 2*i + 1] + beta1*C[j*2*ldc + 2*i];
C[j*2*ldc + 2*i] =tmp0+ alpha[0]*real - alpha[1]*imag;
C[j*2*ldc+ 2*i + 1] = tmp1+ alpha[0]*imag + real*alpha[1];
C[j*2*ldc + 2*i] =tmp0+ alpha0*real - alpha1*imag;
C[j*2*ldc+ 2*i + 1] = tmp1+ alpha0*imag + real*alpha1;
}
}