|
1 |
| -#ifndef COMMON_SH_H |
2 |
| -#define COMMON_SH_H |
| 1 | +#ifndef COMMON_SB_H |
| 2 | +#define COMMON_SB_H |
3 | 3 |
|
4 | 4 | #ifndef DYNAMIC_ARCH
|
5 | 5 |
|
6 |
| -#define SHDOT_K shdot_k |
7 |
| -#define SHSTOBF16_K shstobf16_k |
8 |
| -#define SHDTOBF16_K shdtobf16_k |
| 6 | +#define SBDOT_K sbdot_k |
| 7 | +#define SBSTOBF16_K sbstobf16_k |
| 8 | +#define SBDTOBF16_K sbdtobf16_k |
9 | 9 | #define SBF16TOS_K sbf16tos_k
|
10 | 10 | #define DBF16TOD_K dbf16tod_k
|
11 | 11 |
|
12 |
| -#define SHGEMM_ONCOPY shgemm_oncopy |
13 |
| -#define SHGEMM_OTCOPY shgemm_otcopy |
| 12 | +#define SBGEMM_ONCOPY sbgemm_oncopy |
| 13 | +#define SBGEMM_OTCOPY sbgemm_otcopy |
14 | 14 |
|
15 |
| -#if SHGEMM_DEFAULT_UNROLL_M == SHGEMM_DEFAULT_UNROLL_N |
16 |
| -#define SHGEMM_INCOPY shgemm_oncopy |
17 |
| -#define SHGEMM_ITCOPY shgemm_otcopy |
| 15 | +#if SBGEMM_DEFAULT_UNROLL_M == SBGEMM_DEFAULT_UNROLL_N |
| 16 | +#define SBGEMM_INCOPY sbgemm_oncopy |
| 17 | +#define SBGEMM_ITCOPY sbgemm_otcopy |
18 | 18 | #else
|
19 |
| -#define SHGEMM_INCOPY shgemm_incopy |
20 |
| -#define SHGEMM_ITCOPY shgemm_itcopy |
| 19 | +#define SBGEMM_INCOPY sbgemm_incopy |
| 20 | +#define SBGEMM_ITCOPY sbgemm_itcopy |
21 | 21 | #endif
|
22 |
| -#define SHGEMM_BETA shgemm_beta |
23 |
| -#define SHGEMM_KERNEL shgemm_kernel |
| 22 | +#define SBGEMM_BETA sbgemm_beta |
| 23 | +#define SBGEMM_KERNEL sbgemm_kernel |
24 | 24 |
|
25 | 25 | #else
|
26 | 26 |
|
27 |
| -#define SHDOT_K gotoblas -> shdot_k |
28 |
| -#define SHSTOBF16_K gotoblas -> shstobf16_k |
29 |
| -#define SHDTOBF16_K gotoblas -> shdtobf16_k |
| 27 | +#define SBDOT_K gotoblas -> sbdot_k |
| 28 | +#define SBSTOBF16_K gotoblas -> sbstobf16_k |
| 29 | +#define SBDTOBF16_K gotoblas -> sbdtobf16_k |
30 | 30 | #define SBF16TOS_K gotoblas -> sbf16tos_k
|
31 | 31 | #define DBF16TOD_K gotoblas -> dbf16tod_k
|
32 | 32 |
|
33 |
| -#define SHGEMM_ONCOPY gotoblas -> shgemm_oncopy |
34 |
| -#define SHGEMM_OTCOPY gotoblas -> shgemm_otcopy |
35 |
| -#define SHGEMM_INCOPY gotoblas -> shgemm_incopy |
36 |
| -#define SHGEMM_ITCOPY gotoblas -> shgemm_itcopy |
37 |
| -#define SHGEMM_BETA gotoblas -> shgemm_beta |
38 |
| -#define SHGEMM_KERNEL gotoblas -> shgemm_kernel |
| 33 | +#define SBGEMM_ONCOPY gotoblas -> sbgemm_oncopy |
| 34 | +#define SBGEMM_OTCOPY gotoblas -> sbgemm_otcopy |
| 35 | +#define SBGEMM_INCOPY gotoblas -> sbgemm_incopy |
| 36 | +#define SBGEMM_ITCOPY gotoblas -> sbgemm_itcopy |
| 37 | +#define SBGEMM_BETA gotoblas -> sbgemm_beta |
| 38 | +#define SBGEMM_KERNEL gotoblas -> sbgemm_kernel |
39 | 39 |
|
40 | 40 | #endif
|
41 | 41 |
|
42 |
| -#define SHGEMM_NN shgemm_nn |
43 |
| -#define SHGEMM_CN shgemm_tn |
44 |
| -#define SHGEMM_TN shgemm_tn |
45 |
| -#define SHGEMM_NC shgemm_nt |
46 |
| -#define SHGEMM_NT shgemm_nt |
47 |
| -#define SHGEMM_CC shgemm_tt |
48 |
| -#define SHGEMM_CT shgemm_tt |
49 |
| -#define SHGEMM_TC shgemm_tt |
50 |
| -#define SHGEMM_TT shgemm_tt |
51 |
| -#define SHGEMM_NR shgemm_nn |
52 |
| -#define SHGEMM_TR shgemm_tn |
53 |
| -#define SHGEMM_CR shgemm_tn |
54 |
| -#define SHGEMM_RN shgemm_nn |
55 |
| -#define SHGEMM_RT shgemm_nt |
56 |
| -#define SHGEMM_RC shgemm_nt |
57 |
| -#define SHGEMM_RR shgemm_nn |
| 42 | +#define SBGEMM_NN sbgemm_nn |
| 43 | +#define SBGEMM_CN sbgemm_tn |
| 44 | +#define SBGEMM_TN sbgemm_tn |
| 45 | +#define SBGEMM_NC sbgemm_nt |
| 46 | +#define SBGEMM_NT sbgemm_nt |
| 47 | +#define SBGEMM_CC sbgemm_tt |
| 48 | +#define SBGEMM_CT sbgemm_tt |
| 49 | +#define SBGEMM_TC sbgemm_tt |
| 50 | +#define SBGEMM_TT sbgemm_tt |
| 51 | +#define SBGEMM_NR sbgemm_nn |
| 52 | +#define SBGEMM_TR sbgemm_tn |
| 53 | +#define SBGEMM_CR sbgemm_tn |
| 54 | +#define SBGEMM_RN sbgemm_nn |
| 55 | +#define SBGEMM_RT sbgemm_nt |
| 56 | +#define SBGEMM_RC sbgemm_nt |
| 57 | +#define SBGEMM_RR sbgemm_nn |
58 | 58 |
|
59 |
| -#define SHGEMM_THREAD_NN shgemm_thread_nn |
60 |
| -#define SHGEMM_THREAD_CN shgemm_thread_tn |
61 |
| -#define SHGEMM_THREAD_TN shgemm_thread_tn |
62 |
| -#define SHGEMM_THREAD_NC shgemm_thread_nt |
63 |
| -#define SHGEMM_THREAD_NT shgemm_thread_nt |
64 |
| -#define SHGEMM_THREAD_CC shgemm_thread_tt |
65 |
| -#define SHGEMM_THREAD_CT shgemm_thread_tt |
66 |
| -#define SHGEMM_THREAD_TC shgemm_thread_tt |
67 |
| -#define SHGEMM_THREAD_TT shgemm_thread_tt |
68 |
| -#define SHGEMM_THREAD_NR shgemm_thread_nn |
69 |
| -#define SHGEMM_THREAD_TR shgemm_thread_tn |
70 |
| -#define SHGEMM_THREAD_CR shgemm_thread_tn |
71 |
| -#define SHGEMM_THREAD_RN shgemm_thread_nn |
72 |
| -#define SHGEMM_THREAD_RT shgemm_thread_nt |
73 |
| -#define SHGEMM_THREAD_RC shgemm_thread_nt |
74 |
| -#define SHGEMM_THREAD_RR shgemm_thread_nn |
| 59 | +#define SBGEMM_THREAD_NN sbgemm_thread_nn |
| 60 | +#define SBGEMM_THREAD_CN sbgemm_thread_tn |
| 61 | +#define SBGEMM_THREAD_TN sbgemm_thread_tn |
| 62 | +#define SBGEMM_THREAD_NC sbgemm_thread_nt |
| 63 | +#define SBGEMM_THREAD_NT sbgemm_thread_nt |
| 64 | +#define SBGEMM_THREAD_CC sbgemm_thread_tt |
| 65 | +#define SBGEMM_THREAD_CT sbgemm_thread_tt |
| 66 | +#define SBGEMM_THREAD_TC sbgemm_thread_tt |
| 67 | +#define SBGEMM_THREAD_TT sbgemm_thread_tt |
| 68 | +#define SBGEMM_THREAD_NR sbgemm_thread_nn |
| 69 | +#define SBGEMM_THREAD_TR sbgemm_thread_tn |
| 70 | +#define SBGEMM_THREAD_CR sbgemm_thread_tn |
| 71 | +#define SBGEMM_THREAD_RN sbgemm_thread_nn |
| 72 | +#define SBGEMM_THREAD_RT sbgemm_thread_nt |
| 73 | +#define SBGEMM_THREAD_RC sbgemm_thread_nt |
| 74 | +#define SBGEMM_THREAD_RR sbgemm_thread_nn |
75 | 75 |
|
76 | 76 | #endif
|
77 | 77 |
|
0 commit comments