|
646 | 646 |
|
647 | 647 | #elif defined(HALF)
|
648 | 648 |
|
| 649 | +#define AXPYU_K SAXPYU_K |
| 650 | +#define AXPYC_K SAXPYC_K |
| 651 | +#define SCAL_K SSCAL_K |
| 652 | +#define GEMV_N SGEMV_N |
| 653 | +#define GEMV_T SGEMV_T |
| 654 | +#define SYMV_U SSYMV_U |
| 655 | +#define SYMV_L SSYMV_L |
| 656 | +#define GERU_K SGERU_K |
| 657 | +#define GERC_K SGERC_K |
| 658 | +#define GERV_K SGERV_K |
| 659 | +#define GERD_K SGERD_K |
| 660 | +#define SYMV_THREAD_U SSYMV_THREAD_U |
| 661 | +#define SYMV_THREAD_L SSYMV_THREAD_L |
649 | 662 | #define GEMM_BETA SHGEMM_BETA
|
650 | 663 | #define GEMM_KERNEL_N SHGEMM_KERNEL
|
651 | 664 | #define GEMM_KERNEL_L SHGEMM_KERNEL
|
|
672 | 685 | #define GEMM_OTCOPY SHGEMM_OTCOPY
|
673 | 686 | #define GEMM_INCOPY SHGEMM_INCOPY
|
674 | 687 | #define GEMM_ITCOPY SHGEMM_ITCOPY
|
| 688 | +#define SYMM_THREAD_LU SSYMM_THREAD_LU |
| 689 | +#define SYMM_THREAD_LL SSYMM_THREAD_LL |
| 690 | +#define SYMM_THREAD_RU SSYMM_THREAD_RU |
| 691 | +#define SYMM_THREAD_RL SSYMM_THREAD_RL |
| 692 | +#define SYMM_LU SSYMM_LU |
| 693 | +#define SYMM_LL SSYMM_LL |
| 694 | +#define SYMM_RU SSYMM_RU |
| 695 | +#define SYMM_RL SSYMM_RL |
| 696 | + |
| 697 | + |
| 698 | +#define HEMM_THREAD_LU SHEMM_THREAD_LU |
| 699 | +#define HEMM_THREAD_LL SHEMM_THREAD_LL |
| 700 | +#define HEMM_THREAD_RU SHEMM_THREAD_RU |
| 701 | +#define HEMM_THREAD_RL SHEMM_THREAD_RL |
675 | 702 |
|
676 | 703 | #define GEMM_THREAD_NN SHGEMM_THREAD_NN
|
677 | 704 | #define GEMM_THREAD_CN SHGEMM_THREAD_TN
|
|
690 | 717 | #define GEMM_THREAD_RC SHGEMM_THREAD_NT
|
691 | 718 | #define GEMM_THREAD_RR SHGEMM_THREAD_NN
|
692 | 719 |
|
| 720 | +#ifdef UNIT |
| 721 | + |
| 722 | +#define TRMM_OUNCOPY STRMM_OUNUCOPY |
| 723 | +#define TRMM_OUTCOPY STRMM_OUTUCOPY |
| 724 | +#define TRMM_OLNCOPY STRMM_OLNUCOPY |
| 725 | +#define TRMM_OLTCOPY STRMM_OLTUCOPY |
| 726 | +#define TRSM_OUNCOPY STRSM_OUNUCOPY |
| 727 | +#define TRSM_OUTCOPY STRSM_OUTUCOPY |
| 728 | +#define TRSM_OLNCOPY STRSM_OLNUCOPY |
| 729 | +#define TRSM_OLTCOPY STRSM_OLTUCOPY |
| 730 | + |
| 731 | +#define TRMM_IUNCOPY STRMM_IUNUCOPY |
| 732 | +#define TRMM_IUTCOPY STRMM_IUTUCOPY |
| 733 | +#define TRMM_ILNCOPY STRMM_ILNUCOPY |
| 734 | +#define TRMM_ILTCOPY STRMM_ILTUCOPY |
| 735 | +#define TRSM_IUNCOPY STRSM_IUNUCOPY |
| 736 | +#define TRSM_IUTCOPY STRSM_IUTUCOPY |
| 737 | +#define TRSM_ILNCOPY STRSM_ILNUCOPY |
| 738 | +#define TRSM_ILTCOPY STRSM_ILTUCOPY |
| 739 | + |
| 740 | +#else |
| 741 | + |
| 742 | +#define TRMM_OUNCOPY STRMM_OUNNCOPY |
| 743 | +#define TRMM_OUTCOPY STRMM_OUTNCOPY |
| 744 | +#define TRMM_OLNCOPY STRMM_OLNNCOPY |
| 745 | +#define TRMM_OLTCOPY STRMM_OLTNCOPY |
| 746 | +#define TRSM_OUNCOPY STRSM_OUNNCOPY |
| 747 | +#define TRSM_OUTCOPY STRSM_OUTNCOPY |
| 748 | +#define TRSM_OLNCOPY STRSM_OLNNCOPY |
| 749 | +#define TRSM_OLTCOPY STRSM_OLTNCOPY |
| 750 | + |
| 751 | +#define TRMM_IUNCOPY STRMM_IUNNCOPY |
| 752 | +#define TRMM_IUTCOPY STRMM_IUTNCOPY |
| 753 | +#define TRMM_ILNCOPY STRMM_ILNNCOPY |
| 754 | +#define TRMM_ILTCOPY STRMM_ILTNCOPY |
| 755 | +#define TRSM_IUNCOPY STRSM_IUNNCOPY |
| 756 | +#define TRSM_IUTCOPY STRSM_IUTNCOPY |
| 757 | +#define TRSM_ILNCOPY STRSM_ILNNCOPY |
| 758 | +#define TRSM_ILTCOPY STRSM_ILTNCOPY |
| 759 | + |
| 760 | +#define TRMM_KERNEL_LN STRMM_KERNEL_LN |
| 761 | +#define TRMM_KERNEL_LT STRMM_KERNEL_LT |
| 762 | +#define TRMM_KERNEL_LR STRMM_KERNEL_LN |
| 763 | +#define TRMM_KERNEL_LC STRMM_KERNEL_LT |
| 764 | +#define TRMM_KERNEL_RN STRMM_KERNEL_RN |
| 765 | +#define TRMM_KERNEL_RT STRMM_KERNEL_RT |
| 766 | +#define TRMM_KERNEL_RR STRMM_KERNEL_RN |
| 767 | +#define TRMM_KERNEL_RC STRMM_KERNEL_RT |
| 768 | + |
| 769 | +#define TRSM_KERNEL_LN STRSM_KERNEL_LN |
| 770 | +#define TRSM_KERNEL_LT STRSM_KERNEL_LT |
| 771 | +#define TRSM_KERNEL_LR STRSM_KERNEL_LN |
| 772 | +#define TRSM_KERNEL_LC STRSM_KERNEL_LT |
| 773 | +#define TRSM_KERNEL_RN STRSM_KERNEL_RN |
| 774 | +#define TRSM_KERNEL_RT STRSM_KERNEL_RT |
| 775 | +#define TRSM_KERNEL_RR STRSM_KERNEL_RN |
| 776 | +#define TRSM_KERNEL_RC STRSM_KERNEL_RT |
| 777 | + |
| 778 | +#define SYMM_IUTCOPY SSYMM_IUTCOPY |
| 779 | +#define SYMM_ILTCOPY SSYMM_ILTCOPY |
| 780 | +#define SYMM_OUTCOPY SSYMM_OUTCOPY |
| 781 | +#define SYMM_OLTCOPY SSYMM_OLTCOPY |
| 782 | +#define TRMM_LNUU STRMM_LNUU |
| 783 | +#define TRMM_LNUN STRMM_LNUN |
| 784 | +#define TRMM_LNLU STRMM_LNLU |
| 785 | +#define TRMM_LNLN STRMM_LNLN |
| 786 | +#define TRMM_LTUU STRMM_LTUU |
| 787 | +#define TRMM_LTUN STRMM_LTUN |
| 788 | +#define TRMM_LTLU STRMM_LTLU |
| 789 | +#define TRMM_LTLN STRMM_LTLN |
| 790 | +#define TRMM_LRUU STRMM_LNUU |
| 791 | +#define TRMM_LRUN STRMM_LNUN |
| 792 | +#define TRMM_LRLU STRMM_LNLU |
| 793 | +#define TRMM_LRLN STRMM_LNLN |
| 794 | +#define TRMM_LCUU STRMM_LTUU |
| 795 | +#define TRMM_LCUN STRMM_LTUN |
| 796 | +#define TRMM_LCLU STRMM_LTLU |
| 797 | +#define TRMM_LCLN STRMM_LTLN |
| 798 | +#define TRMM_RNUU STRMM_RNUU |
| 799 | +#define TRMM_RNUN STRMM_RNUN |
| 800 | +#define TRMM_RNLU STRMM_RNLU |
| 801 | +#define TRMM_RNLN STRMM_RNLN |
| 802 | +#define TRMM_RTUU STRMM_RTUU |
| 803 | +#define TRMM_RTUN STRMM_RTUN |
| 804 | +#define TRMM_RTLU STRMM_RTLU |
| 805 | +#define TRMM_RTLN STRMM_RTLN |
| 806 | +#define TRMM_RRUU STRMM_RNUU |
| 807 | +#define TRMM_RRUN STRMM_RNUN |
| 808 | +#define TRMM_RRLU STRMM_RNLU |
| 809 | +#define TRMM_RRLN STRMM_RNLN |
| 810 | +#define TRMM_RCUU STRMM_RTUU |
| 811 | +#define TRMM_RCUN STRMM_RTUN |
| 812 | +#define TRMM_RCLU STRMM_RTLU |
| 813 | +#define TRMM_RCLN STRMM_RTLN |
| 814 | + |
| 815 | +#define TRSM_LNUU STRSM_LNUU |
| 816 | +#define TRSM_LNUN STRSM_LNUN |
| 817 | +#define TRSM_LNLU STRSM_LNLU |
| 818 | +#define TRSM_LNLN STRSM_LNLN |
| 819 | +#define TRSM_LTUU STRSM_LTUU |
| 820 | +#define TRSM_LTUN STRSM_LTUN |
| 821 | +#define TRSM_LTLU STRSM_LTLU |
| 822 | +#define TRSM_LTLN STRSM_LTLN |
| 823 | +#define TRSM_LRUU STRSM_LNUU |
| 824 | +#define TRSM_LRUN STRSM_LNUN |
| 825 | +#define TRSM_LRLU STRSM_LNLU |
| 826 | +#define TRSM_LRLN STRSM_LNLN |
| 827 | +#define TRSM_LCUU STRSM_LTUU |
| 828 | +#define TRSM_LCUN STRSM_LTUN |
| 829 | +#define TRSM_LCLU STRSM_LTLU |
| 830 | +#define TRSM_LCLN STRSM_LTLN |
| 831 | +#define TRSM_RNUU STRSM_RNUU |
| 832 | +#define TRSM_RNUN STRSM_RNUN |
| 833 | +#define TRSM_RNLU STRSM_RNLU |
| 834 | +#define TRSM_RNLN STRSM_RNLN |
| 835 | +#define TRSM_RTUU STRSM_RTUU |
| 836 | +#define TRSM_RTUN STRSM_RTUN |
| 837 | +#define TRSM_RTLU STRSM_RTLU |
| 838 | +#define TRSM_RTLN STRSM_RTLN |
| 839 | +#define TRSM_RRUU STRSM_RNUU |
| 840 | +#define TRSM_RRUN STRSM_RNUN |
| 841 | +#define TRSM_RRLU STRSM_RNLU |
| 842 | +#define TRSM_RRLN STRSM_RNLN |
| 843 | +#define TRSM_RCUU STRSM_RTUU |
| 844 | +#define TRSM_RCUN STRSM_RTUN |
| 845 | +#define TRSM_RCLU STRSM_RTLU |
| 846 | +#define TRSM_RCLN STRSM_RTLN |
| 847 | +#define SYRK_UN SSYRK_UN |
| 848 | +#define SYRK_UT SSYRK_UT |
| 849 | +#define SYRK_LN SSYRK_LN |
| 850 | +#define SYRK_LT SSYRK_LT |
| 851 | +#define SYRK_UR SSYRK_UN |
| 852 | +#define SYRK_UC SSYRK_UT |
| 853 | +#define SYRK_LR SSYRK_LN |
| 854 | +#define SYRK_LC SSYRK_LT |
| 855 | + |
| 856 | +#define SYRK_KERNEL_U SSYRK_KERNEL_U |
| 857 | +#define SYRK_KERNEL_L SSYRK_KERNEL_L |
| 858 | + |
| 859 | +#define HERK_UN SSYRK_UN |
| 860 | +#define HERK_LN SSYRK_LN |
| 861 | +#define HERK_UC SSYRK_UT |
| 862 | +#define HERK_LC SSYRK_LT |
| 863 | + |
| 864 | +#define HER2K_UN SSYR2K_UN |
| 865 | +#define HER2K_LN SSYR2K_LN |
| 866 | +#define HER2K_UC SSYR2K_UT |
| 867 | +#define HER2K_LC SSYR2K_LT |
| 868 | + |
| 869 | +#define SYR2K_UN SSYR2K_UN |
| 870 | +#define SYR2K_UT SSYR2K_UT |
| 871 | +#define SYR2K_LN SSYR2K_LN |
| 872 | +#define SYR2K_LT SSYR2K_LT |
| 873 | +#define SYR2K_UR SSYR2K_UN |
| 874 | +#define SYR2K_UC SSYR2K_UT |
| 875 | +#define SYR2K_LR SSYR2K_LN |
| 876 | +#define SYR2K_LC SSYR2K_LT |
| 877 | + |
| 878 | +#define SYR2K_KERNEL_U SSYR2K_KERNEL_U |
| 879 | +#define SYR2K_KERNEL_L SSYR2K_KERNEL_L |
| 880 | +#define SYRK_THREAD_UN SSYRK_THREAD_UN |
| 881 | +#define SYRK_THREAD_UT SSYRK_THREAD_UT |
| 882 | +#define SYRK_THREAD_LN SSYRK_THREAD_LN |
| 883 | +#define SYRK_THREAD_LT SSYRK_THREAD_LT |
| 884 | +#define SYRK_THREAD_UR SSYRK_THREAD_UR |
| 885 | +#define SYRK_THREAD_UC SSYRK_THREAD_UC |
| 886 | +#define SYRK_THREAD_LR SSYRK_THREAD_LN |
| 887 | +#define SYRK_THREAD_LC SSYRK_THREAD_LT |
| 888 | + |
| 889 | +#define HERK_THREAD_UN SSYRK_THREAD_UN |
| 890 | +#define HERK_THREAD_UT SSYRK_THREAD_UT |
| 891 | +#define HERK_THREAD_LN SSYRK_THREAD_LN |
| 892 | +#define HERK_THREAD_LT SSYRK_THREAD_LT |
| 893 | +#define HERK_THREAD_UR SSYRK_THREAD_UR |
| 894 | +#define HERK_THREAD_UC SSYRK_THREAD_UC |
| 895 | +#define HERK_THREAD_LR SSYRK_THREAD_LN |
| 896 | +#define HERK_THREAD_LC SSYRK_THREAD_LT |
| 897 | + |
| 898 | +#endif |
| 899 | + |
693 | 900 | #else
|
694 | 901 |
|
695 | 902 | #define AMAX_K SAMAX_K
|
|
721 | 928 | #define GEMV_S SGEMV_S
|
722 | 929 | #define GEMV_D SGEMV_D
|
723 | 930 |
|
| 931 | + |
| 932 | +#define SYMV_U SSYMV_U |
| 933 | +#define SYMV_L SSYMV_L |
724 | 934 | #define GERU_K SGERU_K
|
725 | 935 | #define GERC_K SGERC_K
|
726 | 936 | #define GERV_K SGERV_K
|
727 | 937 | #define GERD_K SGERD_K
|
728 | 938 |
|
729 |
| -#define SYMV_U SSYMV_U |
730 |
| -#define SYMV_L SSYMV_L |
731 |
| - |
732 | 939 | #define SYMV_THREAD_U SSYMV_THREAD_U
|
733 | 940 | #define SYMV_THREAD_L SSYMV_THREAD_L
|
734 | 941 |
|
|
0 commit comments