|
libflame
12600
|
Functions | |
| FLA_Error | FLA_Gemm_internal (FLA_Trans transa, FLA_Trans transb, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl) |
Variables | |
| fla_gemm_t * | flash_gemm_cntl_blas |
| fla_gemm_t * | flash_gemm_cntl_mm_op |
| FLA_Error FLA_Gemm_internal | ( | FLA_Trans | transa, |
| FLA_Trans | transb, | ||
| FLA_Obj | alpha, | ||
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C, | ||
| fla_gemm_t * | cntl | ||
| ) |
References FLA_Check_error_level(), FLA_Gemm_cc(), FLA_Gemm_ch(), FLA_Gemm_cn(), FLA_Gemm_ct(), FLA_Gemm_hc(), FLA_Gemm_hh(), FLA_Gemm_hn(), FLA_Gemm_ht(), FLA_Gemm_internal(), FLA_Gemm_internal_check(), FLA_Gemm_nc(), FLA_Gemm_nh(), FLA_Gemm_nn(), FLA_Gemm_nt(), FLA_Gemm_tc(), FLA_Gemm_th(), FLA_Gemm_tn(), FLA_Gemm_tt(), FLA_Obj_elemtype(), flash_gemm_cntl_blas, and FLASH_Queue_get_enabled().
Referenced by FLA_Apply_CAQ2_UT_lhfc_blk_var1(), FLA_Apply_Q2_UT_lhfc_blk_var1(), FLA_Apply_Q2_UT_lnfc_blk_var1(), FLA_Apply_Q_UT_lhbc_blk_var1(), FLA_Apply_Q_UT_lhbc_blk_var3(), FLA_Apply_Q_UT_lhbr_blk_var1(), FLA_Apply_Q_UT_lhbr_blk_var3(), FLA_Apply_Q_UT_lhfc_blk_var1(), FLA_Apply_Q_UT_lhfc_blk_var3(), FLA_Apply_Q_UT_lhfr_blk_var1(), FLA_Apply_Q_UT_lhfr_blk_var3(), FLA_Apply_Q_UT_lnbc_blk_var1(), FLA_Apply_Q_UT_lnbc_blk_var3(), FLA_Apply_Q_UT_lnbr_blk_var1(), FLA_Apply_Q_UT_lnbr_blk_var3(), FLA_Apply_Q_UT_lnfc_blk_var1(), FLA_Apply_Q_UT_lnfc_blk_var3(), FLA_Apply_Q_UT_lnfr_blk_var1(), FLA_Apply_Q_UT_lnfr_blk_var3(), FLA_Apply_Q_UT_rhbc_blk_var1(), FLA_Apply_Q_UT_rhbc_blk_var3(), FLA_Apply_Q_UT_rhbr_blk_var1(), FLA_Apply_Q_UT_rhbr_blk_var3(), FLA_Apply_Q_UT_rhfc_blk_var1(), FLA_Apply_Q_UT_rhfc_blk_var3(), FLA_Apply_Q_UT_rhfr_blk_var1(), FLA_Apply_Q_UT_rhfr_blk_var3(), FLA_Apply_Q_UT_rnbc_blk_var1(), FLA_Apply_Q_UT_rnbc_blk_var3(), FLA_Apply_Q_UT_rnbr_blk_var1(), FLA_Apply_Q_UT_rnbr_blk_var3(), FLA_Apply_Q_UT_rnfc_blk_var1(), FLA_Apply_Q_UT_rnfc_blk_var3(), FLA_Apply_Q_UT_rnfr_blk_var1(), FLA_Apply_Q_UT_rnfr_blk_var3(), FLA_Apply_QUD_UT_lhfc_blk_var1(), FLA_CAQR2_UT_blk_var1(), FLA_Chol_l_blk_var2(), FLA_Chol_u_blk_var2(), FLA_Eig_gest_il_blk_var2(), FLA_Eig_gest_il_blk_var3(), FLA_Eig_gest_il_blk_var4(), FLA_Eig_gest_iu_blk_var2(), FLA_Eig_gest_iu_blk_var3(), FLA_Eig_gest_iu_blk_var4(), FLA_Eig_gest_nl_blk_var2(), FLA_Eig_gest_nl_blk_var4(), FLA_Eig_gest_nu_blk_var2(), FLA_Eig_gest_nu_blk_var4(), FLA_Gemm(), FLA_Gemm_cc_blk_var1(), FLA_Gemm_cc_blk_var2(), FLA_Gemm_cc_blk_var3(), FLA_Gemm_cc_blk_var4(), FLA_Gemm_cc_blk_var5(), FLA_Gemm_cc_blk_var6(), FLA_Gemm_ch_blk_var1(), FLA_Gemm_ch_blk_var2(), FLA_Gemm_ch_blk_var3(), FLA_Gemm_ch_blk_var4(), FLA_Gemm_ch_blk_var5(), FLA_Gemm_ch_blk_var6(), FLA_Gemm_cn_blk_var1(), FLA_Gemm_cn_blk_var2(), FLA_Gemm_cn_blk_var3(), FLA_Gemm_cn_blk_var4(), FLA_Gemm_cn_blk_var5(), FLA_Gemm_cn_blk_var6(), FLA_Gemm_ct_blk_var1(), FLA_Gemm_ct_blk_var2(), FLA_Gemm_ct_blk_var3(), FLA_Gemm_ct_blk_var4(), FLA_Gemm_ct_blk_var5(), FLA_Gemm_ct_blk_var6(), FLA_Gemm_hc_blk_var1(), FLA_Gemm_hc_blk_var2(), FLA_Gemm_hc_blk_var3(), FLA_Gemm_hc_blk_var4(), FLA_Gemm_hc_blk_var5(), FLA_Gemm_hc_blk_var6(), FLA_Gemm_hh_blk_var1(), FLA_Gemm_hh_blk_var2(), FLA_Gemm_hh_blk_var3(), FLA_Gemm_hh_blk_var4(), FLA_Gemm_hh_blk_var5(), FLA_Gemm_hh_blk_var6(), FLA_Gemm_hn_blk_var1(), FLA_Gemm_hn_blk_var2(), FLA_Gemm_hn_blk_var3(), FLA_Gemm_hn_blk_var4(), FLA_Gemm_hn_blk_var5(), FLA_Gemm_hn_blk_var6(), FLA_Gemm_ht_blk_var1(), FLA_Gemm_ht_blk_var2(), FLA_Gemm_ht_blk_var3(), FLA_Gemm_ht_blk_var4(), FLA_Gemm_ht_blk_var5(), FLA_Gemm_ht_blk_var6(), FLA_Gemm_internal(), FLA_Gemm_nc_blk_var1(), FLA_Gemm_nc_blk_var2(), FLA_Gemm_nc_blk_var3(), FLA_Gemm_nc_blk_var4(), FLA_Gemm_nc_blk_var5(), FLA_Gemm_nc_blk_var6(), FLA_Gemm_nh_blk_var1(), FLA_Gemm_nh_blk_var2(), FLA_Gemm_nh_blk_var3(), FLA_Gemm_nh_blk_var4(), FLA_Gemm_nh_blk_var5(), FLA_Gemm_nh_blk_var6(), FLA_Gemm_nn_blk_var1(), FLA_Gemm_nn_blk_var2(), FLA_Gemm_nn_blk_var3(), FLA_Gemm_nn_blk_var4(), FLA_Gemm_nn_blk_var5(), FLA_Gemm_nn_blk_var6(), FLA_Gemm_nt_blk_var1(), FLA_Gemm_nt_blk_var2(), FLA_Gemm_nt_blk_var3(), FLA_Gemm_nt_blk_var4(), FLA_Gemm_nt_blk_var5(), FLA_Gemm_nt_blk_var6(), FLA_Gemm_tc_blk_var1(), FLA_Gemm_tc_blk_var2(), FLA_Gemm_tc_blk_var3(), FLA_Gemm_tc_blk_var4(), FLA_Gemm_tc_blk_var5(), FLA_Gemm_tc_blk_var6(), FLA_Gemm_th_blk_var1(), FLA_Gemm_th_blk_var2(), FLA_Gemm_th_blk_var3(), FLA_Gemm_th_blk_var4(), FLA_Gemm_th_blk_var5(), FLA_Gemm_th_blk_var6(), FLA_Gemm_tn_blk_var1(), FLA_Gemm_tn_blk_var2(), FLA_Gemm_tn_blk_var3(), FLA_Gemm_tn_blk_var4(), FLA_Gemm_tn_blk_var5(), FLA_Gemm_tn_blk_var6(), FLA_Gemm_tt_blk_var1(), FLA_Gemm_tt_blk_var2(), FLA_Gemm_tt_blk_var3(), FLA_Gemm_tt_blk_var4(), FLA_Gemm_tt_blk_var5(), FLA_Gemm_tt_blk_var6(), FLA_Hemm_ll_blk_var1(), FLA_Hemm_ll_blk_var2(), FLA_Hemm_ll_blk_var3(), FLA_Hemm_ll_blk_var4(), FLA_Hemm_ll_blk_var5(), FLA_Hemm_ll_blk_var6(), FLA_Hemm_ll_blk_var7(), FLA_Hemm_ll_blk_var8(), FLA_Hemm_lu_blk_var1(), FLA_Hemm_lu_blk_var2(), FLA_Hemm_lu_blk_var3(), FLA_Hemm_lu_blk_var4(), FLA_Hemm_lu_blk_var5(), FLA_Hemm_lu_blk_var6(), FLA_Hemm_lu_blk_var7(), FLA_Hemm_lu_blk_var8(), FLA_Hemm_rl_blk_var1(), FLA_Hemm_rl_blk_var2(), FLA_Hemm_rl_blk_var3(), FLA_Hemm_rl_blk_var4(), FLA_Hemm_rl_blk_var5(), FLA_Hemm_rl_blk_var6(), FLA_Hemm_rl_blk_var7(), FLA_Hemm_rl_blk_var8(), FLA_Hemm_ru_blk_var1(), FLA_Hemm_ru_blk_var2(), FLA_Hemm_ru_blk_var3(), FLA_Hemm_ru_blk_var4(), FLA_Hemm_ru_blk_var5(), FLA_Hemm_ru_blk_var6(), FLA_Hemm_ru_blk_var7(), FLA_Hemm_ru_blk_var8(), FLA_Her2k_lh_blk_var1(), FLA_Her2k_lh_blk_var2(), FLA_Her2k_lh_blk_var3(), FLA_Her2k_lh_blk_var4(), FLA_Her2k_lh_blk_var5(), FLA_Her2k_lh_blk_var6(), FLA_Her2k_lh_blk_var7(), FLA_Her2k_lh_blk_var8(), FLA_Her2k_ln_blk_var1(), FLA_Her2k_ln_blk_var2(), FLA_Her2k_ln_blk_var3(), FLA_Her2k_ln_blk_var4(), FLA_Her2k_ln_blk_var5(), FLA_Her2k_ln_blk_var6(), FLA_Her2k_ln_blk_var7(), FLA_Her2k_ln_blk_var8(), FLA_Her2k_uh_blk_var1(), FLA_Her2k_uh_blk_var2(), FLA_Her2k_uh_blk_var3(), FLA_Her2k_uh_blk_var4(), FLA_Her2k_uh_blk_var5(), FLA_Her2k_uh_blk_var6(), FLA_Her2k_uh_blk_var7(), FLA_Her2k_uh_blk_var8(), FLA_Her2k_un_blk_var1(), FLA_Her2k_un_blk_var2(), FLA_Her2k_un_blk_var3(), FLA_Her2k_un_blk_var4(), FLA_Her2k_un_blk_var5(), FLA_Her2k_un_blk_var6(), FLA_Her2k_un_blk_var7(), FLA_Her2k_un_blk_var8(), FLA_Herk_lh_blk_var1(), FLA_Herk_lh_blk_var2(), FLA_Herk_lh_blk_var3(), FLA_Herk_lh_blk_var4(), FLA_Herk_ln_blk_var1(), FLA_Herk_ln_blk_var2(), FLA_Herk_ln_blk_var3(), FLA_Herk_ln_blk_var4(), FLA_Herk_uh_blk_var1(), FLA_Herk_uh_blk_var2(), FLA_Herk_uh_blk_var3(), FLA_Herk_uh_blk_var4(), FLA_Herk_un_blk_var1(), FLA_Herk_un_blk_var2(), FLA_Herk_un_blk_var3(), FLA_Herk_un_blk_var4(), FLA_LU_nopiv_blk_var1(), FLA_LU_nopiv_blk_var2(), FLA_LU_nopiv_blk_var3(), FLA_LU_nopiv_blk_var4(), FLA_LU_nopiv_blk_var5(), FLA_LU_piv_blk_var3(), FLA_LU_piv_blk_var4(), FLA_LU_piv_blk_var5(), FLA_Lyap_h_blk_var2(), FLA_Lyap_h_blk_var3(), FLA_Lyap_n_blk_var2(), FLA_Lyap_n_blk_var3(), FLA_QR2_UT_blk_var1(), FLA_Sylv_hh_blk_var1(), FLA_Sylv_hh_blk_var10(), FLA_Sylv_hh_blk_var11(), FLA_Sylv_hh_blk_var12(), FLA_Sylv_hh_blk_var13(), FLA_Sylv_hh_blk_var14(), FLA_Sylv_hh_blk_var15(), FLA_Sylv_hh_blk_var16(), FLA_Sylv_hh_blk_var17(), FLA_Sylv_hh_blk_var18(), FLA_Sylv_hh_blk_var2(), FLA_Sylv_hh_blk_var3(), FLA_Sylv_hh_blk_var4(), FLA_Sylv_hh_blk_var5(), FLA_Sylv_hh_blk_var6(), FLA_Sylv_hh_blk_var7(), FLA_Sylv_hh_blk_var8(), FLA_Sylv_hh_blk_var9(), FLA_Sylv_hn_blk_var1(), FLA_Sylv_hn_blk_var10(), FLA_Sylv_hn_blk_var11(), FLA_Sylv_hn_blk_var12(), FLA_Sylv_hn_blk_var13(), FLA_Sylv_hn_blk_var14(), FLA_Sylv_hn_blk_var15(), FLA_Sylv_hn_blk_var16(), FLA_Sylv_hn_blk_var17(), FLA_Sylv_hn_blk_var18(), FLA_Sylv_hn_blk_var2(), FLA_Sylv_hn_blk_var3(), FLA_Sylv_hn_blk_var4(), FLA_Sylv_hn_blk_var5(), FLA_Sylv_hn_blk_var6(), FLA_Sylv_hn_blk_var7(), FLA_Sylv_hn_blk_var8(), FLA_Sylv_hn_blk_var9(), FLA_Sylv_nh_blk_var1(), FLA_Sylv_nh_blk_var10(), FLA_Sylv_nh_blk_var11(), FLA_Sylv_nh_blk_var12(), FLA_Sylv_nh_blk_var13(), FLA_Sylv_nh_blk_var14(), FLA_Sylv_nh_blk_var15(), FLA_Sylv_nh_blk_var16(), FLA_Sylv_nh_blk_var17(), FLA_Sylv_nh_blk_var18(), FLA_Sylv_nh_blk_var2(), FLA_Sylv_nh_blk_var3(), FLA_Sylv_nh_blk_var4(), FLA_Sylv_nh_blk_var5(), FLA_Sylv_nh_blk_var6(), FLA_Sylv_nh_blk_var7(), FLA_Sylv_nh_blk_var8(), FLA_Sylv_nh_blk_var9(), FLA_Sylv_nn_blk_var1(), FLA_Sylv_nn_blk_var10(), FLA_Sylv_nn_blk_var11(), FLA_Sylv_nn_blk_var12(), FLA_Sylv_nn_blk_var13(), FLA_Sylv_nn_blk_var14(), FLA_Sylv_nn_blk_var15(), FLA_Sylv_nn_blk_var16(), FLA_Sylv_nn_blk_var17(), FLA_Sylv_nn_blk_var18(), FLA_Sylv_nn_blk_var2(), FLA_Sylv_nn_blk_var3(), FLA_Sylv_nn_blk_var4(), FLA_Sylv_nn_blk_var5(), FLA_Sylv_nn_blk_var6(), FLA_Sylv_nn_blk_var7(), FLA_Sylv_nn_blk_var8(), FLA_Sylv_nn_blk_var9(), FLA_Symm_ll_blk_var1(), FLA_Symm_ll_blk_var2(), FLA_Symm_ll_blk_var3(), FLA_Symm_ll_blk_var4(), FLA_Symm_ll_blk_var5(), FLA_Symm_ll_blk_var6(), FLA_Symm_ll_blk_var7(), FLA_Symm_ll_blk_var8(), FLA_Symm_lu_blk_var1(), FLA_Symm_lu_blk_var2(), FLA_Symm_lu_blk_var3(), FLA_Symm_lu_blk_var4(), FLA_Symm_lu_blk_var5(), FLA_Symm_lu_blk_var6(), FLA_Symm_lu_blk_var7(), FLA_Symm_lu_blk_var8(), FLA_Symm_rl_blk_var1(), FLA_Symm_rl_blk_var2(), FLA_Symm_rl_blk_var3(), FLA_Symm_rl_blk_var4(), FLA_Symm_rl_blk_var5(), FLA_Symm_rl_blk_var6(), FLA_Symm_rl_blk_var7(), FLA_Symm_rl_blk_var8(), FLA_Symm_ru_blk_var1(), FLA_Symm_ru_blk_var2(), FLA_Symm_ru_blk_var3(), FLA_Symm_ru_blk_var4(), FLA_Symm_ru_blk_var5(), FLA_Symm_ru_blk_var6(), FLA_Symm_ru_blk_var7(), FLA_Symm_ru_blk_var8(), FLA_Syr2k_ln_blk_var1(), FLA_Syr2k_ln_blk_var2(), FLA_Syr2k_ln_blk_var3(), FLA_Syr2k_ln_blk_var4(), FLA_Syr2k_ln_blk_var5(), FLA_Syr2k_ln_blk_var6(), FLA_Syr2k_ln_blk_var7(), FLA_Syr2k_ln_blk_var8(), FLA_Syr2k_lt_blk_var1(), FLA_Syr2k_lt_blk_var2(), FLA_Syr2k_lt_blk_var3(), FLA_Syr2k_lt_blk_var4(), FLA_Syr2k_lt_blk_var5(), FLA_Syr2k_lt_blk_var6(), FLA_Syr2k_lt_blk_var7(), FLA_Syr2k_lt_blk_var8(), FLA_Syr2k_un_blk_var1(), FLA_Syr2k_un_blk_var2(), FLA_Syr2k_un_blk_var3(), FLA_Syr2k_un_blk_var4(), FLA_Syr2k_un_blk_var5(), FLA_Syr2k_un_blk_var6(), FLA_Syr2k_un_blk_var7(), FLA_Syr2k_un_blk_var8(), FLA_Syr2k_ut_blk_var1(), FLA_Syr2k_ut_blk_var2(), FLA_Syr2k_ut_blk_var3(), FLA_Syr2k_ut_blk_var4(), FLA_Syr2k_ut_blk_var5(), FLA_Syr2k_ut_blk_var6(), FLA_Syr2k_ut_blk_var7(), FLA_Syr2k_ut_blk_var8(), FLA_Syrk_ln_blk_var1(), FLA_Syrk_ln_blk_var2(), FLA_Syrk_ln_blk_var3(), FLA_Syrk_ln_blk_var4(), FLA_Syrk_lt_blk_var1(), FLA_Syrk_lt_blk_var2(), FLA_Syrk_lt_blk_var3(), FLA_Syrk_lt_blk_var4(), FLA_Syrk_un_blk_var1(), FLA_Syrk_un_blk_var2(), FLA_Syrk_un_blk_var3(), FLA_Syrk_un_blk_var4(), FLA_Syrk_ut_blk_var1(), FLA_Syrk_ut_blk_var2(), FLA_Syrk_ut_blk_var3(), FLA_Syrk_ut_blk_var4(), FLA_Trinv_ln_blk_var3(), FLA_Trinv_ln_blk_var4(), FLA_Trinv_lu_blk_var3(), FLA_Trinv_lu_blk_var4(), FLA_Trinv_un_blk_var3(), FLA_Trinv_un_blk_var4(), FLA_Trinv_uu_blk_var3(), FLA_Trinv_uu_blk_var4(), FLA_Trmm_llc_blk_var1(), FLA_Trmm_llc_blk_var2(), FLA_Trmm_llh_blk_var1(), FLA_Trmm_llh_blk_var2(), FLA_Trmm_lln_blk_var1(), FLA_Trmm_lln_blk_var2(), FLA_Trmm_llt_blk_var1(), FLA_Trmm_llt_blk_var2(), FLA_Trmm_luc_blk_var1(), FLA_Trmm_luc_blk_var2(), FLA_Trmm_luh_blk_var1(), FLA_Trmm_luh_blk_var2(), FLA_Trmm_lun_blk_var1(), FLA_Trmm_lun_blk_var2(), FLA_Trmm_lut_blk_var1(), FLA_Trmm_lut_blk_var2(), FLA_Trmm_rlc_blk_var1(), FLA_Trmm_rlc_blk_var2(), FLA_Trmm_rlh_blk_var1(), FLA_Trmm_rlh_blk_var2(), FLA_Trmm_rln_blk_var1(), FLA_Trmm_rln_blk_var2(), FLA_Trmm_rlt_blk_var1(), FLA_Trmm_rlt_blk_var2(), FLA_Trmm_ruc_blk_var1(), FLA_Trmm_ruc_blk_var2(), FLA_Trmm_ruh_blk_var1(), FLA_Trmm_ruh_blk_var2(), FLA_Trmm_run_blk_var1(), FLA_Trmm_run_blk_var2(), FLA_Trmm_rut_blk_var1(), FLA_Trmm_rut_blk_var2(), FLA_Trsm_llc_blk_var1(), FLA_Trsm_llc_blk_var2(), FLA_Trsm_llh_blk_var1(), FLA_Trsm_llh_blk_var2(), FLA_Trsm_lln_blk_var1(), FLA_Trsm_lln_blk_var2(), FLA_Trsm_llt_blk_var1(), FLA_Trsm_llt_blk_var2(), FLA_Trsm_luc_blk_var1(), FLA_Trsm_luc_blk_var2(), FLA_Trsm_luh_blk_var1(), FLA_Trsm_luh_blk_var2(), FLA_Trsm_lun_blk_var1(), FLA_Trsm_lun_blk_var2(), FLA_Trsm_lut_blk_var1(), FLA_Trsm_lut_blk_var2(), FLA_Trsm_rlc_blk_var1(), FLA_Trsm_rlc_blk_var2(), FLA_Trsm_rlh_blk_var1(), FLA_Trsm_rlh_blk_var2(), FLA_Trsm_rln_blk_var1(), FLA_Trsm_rln_blk_var2(), FLA_Trsm_rlt_blk_var1(), FLA_Trsm_rlt_blk_var2(), FLA_Trsm_ruc_blk_var1(), FLA_Trsm_ruc_blk_var2(), FLA_Trsm_ruh_blk_var1(), FLA_Trsm_ruh_blk_var2(), FLA_Trsm_run_blk_var1(), FLA_Trsm_run_blk_var2(), FLA_Trsm_rut_blk_var1(), FLA_Trsm_rut_blk_var2(), FLA_Ttmm_l_blk_var2(), FLA_Ttmm_u_blk_var2(), and FLASH_Gemm().
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Gemm_internal_check( transa, transb, alpha, A, B, beta, C, cntl );
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_MATRIX &&
FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
// Recurse
r_val = FLA_Gemm_internal( transa,
transb,
alpha,
*FLASH_OBJ_PTR_AT( A ),
*FLASH_OBJ_PTR_AT( B ),
beta,
*FLASH_OBJ_PTR_AT( C ),
flash_gemm_cntl_mm_op );
}
else if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
FLASH_Queue_get_enabled( ) )
{
// Enqueue
ENQUEUE_FLASH_Gemm( transa, transb, alpha, A, B, beta, C, cntl );
}
else
{
if ( FLA_Cntl_matrix_type( cntl ) == FLA_HIER &&
FLA_Obj_elemtype( A ) == FLA_SCALAR &&
!FLASH_Queue_get_enabled( ) )
{
// Execute leaf
cntl = flash_gemm_cntl_blas;
}
// Parameter combinations
if ( transa == FLA_NO_TRANSPOSE )
{
if ( transb == FLA_NO_TRANSPOSE )
r_val = FLA_Gemm_nn( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_TRANSPOSE )
r_val = FLA_Gemm_nt( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Gemm_nc( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_TRANSPOSE )
r_val = FLA_Gemm_nh( alpha, A, B, beta, C, cntl );
}
else if ( transa == FLA_TRANSPOSE )
{
if ( transb == FLA_NO_TRANSPOSE )
r_val = FLA_Gemm_tn( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_TRANSPOSE )
r_val = FLA_Gemm_tt( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Gemm_tc( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_TRANSPOSE )
r_val = FLA_Gemm_th( alpha, A, B, beta, C, cntl );
}
else if ( transa == FLA_CONJ_NO_TRANSPOSE )
{
if ( transb == FLA_NO_TRANSPOSE )
r_val = FLA_Gemm_cn( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_TRANSPOSE )
r_val = FLA_Gemm_ct( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Gemm_cc( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_TRANSPOSE )
r_val = FLA_Gemm_ch( alpha, A, B, beta, C, cntl );
}
else if ( transa == FLA_CONJ_TRANSPOSE )
{
if ( transb == FLA_NO_TRANSPOSE )
r_val = FLA_Gemm_hn( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_TRANSPOSE )
r_val = FLA_Gemm_ht( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_NO_TRANSPOSE )
r_val = FLA_Gemm_hc( alpha, A, B, beta, C, cntl );
else if ( transb == FLA_CONJ_TRANSPOSE )
r_val = FLA_Gemm_hh( alpha, A, B, beta, C, cntl );
}
}
return r_val;
}
Referenced by FLA_Gemm_internal().
1.7.6.1