@@ -51,39 +51,39 @@ typedef struct {
51
51
int sbgemm_p , sbgemm_q , sbgemm_r ;
52
52
int sbgemm_unroll_m , sbgemm_unroll_n , sbgemm_unroll_mn ;
53
53
54
- void (* shstobf16_k ) (BLASLONG , float * , BLASLONG , bfloat16 * , BLASLONG );
55
- void (* shdtobf16_k ) (BLASLONG , double * , BLASLONG , bfloat16 * , BLASLONG );
54
+ void (* sbstobf16_k ) (BLASLONG , float * , BLASLONG , bfloat16 * , BLASLONG );
55
+ void (* sbdtobf16_k ) (BLASLONG , double * , BLASLONG , bfloat16 * , BLASLONG );
56
56
void (* sbf16tos_k ) (BLASLONG , bfloat16 * , BLASLONG , float * , BLASLONG );
57
57
void (* dbf16tod_k ) (BLASLONG , bfloat16 * , BLASLONG , double * , BLASLONG );
58
58
59
- float (* shamax_k ) (BLASLONG , float * , BLASLONG );
60
- float (* shamin_k ) (BLASLONG , float * , BLASLONG );
61
- float (* shmax_k ) (BLASLONG , float * , BLASLONG );
62
- float (* shmin_k ) (BLASLONG , float * , BLASLONG );
63
- BLASLONG (* ishamax_k )(BLASLONG , float * , BLASLONG );
64
- BLASLONG (* ishamin_k )(BLASLONG , float * , BLASLONG );
65
- BLASLONG (* ishmax_k ) (BLASLONG , float * , BLASLONG );
66
- BLASLONG (* ishmin_k ) (BLASLONG , float * , BLASLONG );
67
-
68
- float (* shnrm2_k ) (BLASLONG , float * , BLASLONG );
69
- float (* shasum_k ) (BLASLONG , float * , BLASLONG );
70
- float (* shsum_k ) (BLASLONG , float * , BLASLONG );
71
- int (* shcopy_k ) (BLASLONG , float * , BLASLONG , float * , BLASLONG );
59
+ float (* sbamax_k ) (BLASLONG , float * , BLASLONG );
60
+ float (* sbamin_k ) (BLASLONG , float * , BLASLONG );
61
+ float (* sbmax_k ) (BLASLONG , float * , BLASLONG );
62
+ float (* sbmin_k ) (BLASLONG , float * , BLASLONG );
63
+ BLASLONG (* isbamax_k )(BLASLONG , float * , BLASLONG );
64
+ BLASLONG (* isbamin_k )(BLASLONG , float * , BLASLONG );
65
+ BLASLONG (* isbmax_k ) (BLASLONG , float * , BLASLONG );
66
+ BLASLONG (* isbmin_k ) (BLASLONG , float * , BLASLONG );
67
+
68
+ float (* sbnrm2_k ) (BLASLONG , float * , BLASLONG );
69
+ float (* sbasum_k ) (BLASLONG , float * , BLASLONG );
70
+ float (* sbsum_k ) (BLASLONG , float * , BLASLONG );
71
+ int (* sbcopy_k ) (BLASLONG , float * , BLASLONG , float * , BLASLONG );
72
72
float (* sbdot_k ) (BLASLONG , bfloat16 * , BLASLONG , bfloat16 * , BLASLONG );
73
73
double (* dsbdot_k ) (BLASLONG , float * , BLASLONG , float * , BLASLONG );
74
74
75
- int (* shrot_k ) (BLASLONG , float * , BLASLONG , float * , BLASLONG , float , float );
75
+ int (* sbrot_k ) (BLASLONG , float * , BLASLONG , float * , BLASLONG , float , float );
76
76
77
- int (* shaxpy_k ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG );
78
- int (* shscal_k ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG );
79
- int (* shswap_k ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG );
77
+ int (* sbaxpy_k ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG );
78
+ int (* sbscal_k ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG );
79
+ int (* sbswap_k ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG );
80
80
81
81
int (* sbgemv_n ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG , float * );
82
82
int (* sbgemv_t ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG , float * );
83
- int (* shger_k ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG , float * );
83
+ int (* sbger_k ) (BLASLONG , BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG , float * );
84
84
85
- int (* shsymv_L ) (BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG , float * );
86
- int (* shsymv_U ) (BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG , float * );
85
+ int (* sbsymv_L ) (BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG , float * );
86
+ int (* sbsymv_U ) (BLASLONG , BLASLONG , float , float * , BLASLONG , float * , BLASLONG , float * , BLASLONG , float * );
87
87
88
88
int (* sbgemm_kernel )(BLASLONG , BLASLONG , BLASLONG , float , bfloat16 * , bfloat16 * , float * , BLASLONG );
89
89
int (* sbgemm_beta )(BLASLONG , BLASLONG , BLASLONG , float , bfloat16 * , BLASLONG , bfloat16 * , BLASLONG , float * , BLASLONG );
@@ -93,57 +93,57 @@ BLASLONG (*ishmin_k) (BLASLONG, float *, BLASLONG);
93
93
int (* sbgemm_oncopy )(BLASLONG , BLASLONG , bfloat16 * , BLASLONG , bfloat16 * );
94
94
int (* sbgemm_otcopy )(BLASLONG , BLASLONG , bfloat16 * , BLASLONG , bfloat16 * );
95
95
96
- int (* shtrsm_kernel_LN )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
97
- int (* shtrsm_kernel_LT )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
98
- int (* shtrsm_kernel_RN )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
99
- int (* shtrsm_kernel_RT )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
100
-
101
- int (* shtrsm_iunucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
102
- int (* shtrsm_iunncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
103
- int (* shtrsm_iutucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
104
- int (* shtrsm_iutncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
105
- int (* shtrsm_ilnucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
106
- int (* shtrsm_ilnncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
107
- int (* shtrsm_iltucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
108
- int (* shtrsm_iltncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
109
- int (* shtrsm_ounucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
110
- int (* shtrsm_ounncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
111
- int (* shtrsm_outucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
112
- int (* shtrsm_outncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
113
- int (* shtrsm_olnucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
114
- int (* shtrsm_olnncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
115
- int (* shtrsm_oltucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
116
- int (* shtrsm_oltncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
117
-
118
- int (* shtrmm_kernel_RN )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
119
- int (* shtrmm_kernel_RT )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
120
- int (* shtrmm_kernel_LN )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
121
- int (* shtrmm_kernel_LT )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
122
-
123
- int (* shtrmm_iunucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
124
- int (* shtrmm_iunncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
125
- int (* shtrmm_iutucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
126
- int (* shtrmm_iutncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
127
- int (* shtrmm_ilnucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
128
- int (* shtrmm_ilnncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
129
- int (* shtrmm_iltucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
130
- int (* shtrmm_iltncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
131
- int (* shtrmm_ounucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
132
- int (* shtrmm_ounncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
133
- int (* shtrmm_outucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
134
- int (* shtrmm_outncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
135
- int (* shtrmm_olnucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
136
- int (* shtrmm_olnncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
137
- int (* shtrmm_oltucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
138
- int (* shtrmm_oltncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
139
-
140
- int (* shsymm_iutcopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
141
- int (* shsymm_iltcopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
142
- int (* shsymm_outcopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
143
- int (* shsymm_oltcopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
144
-
145
- int (* shneg_tcopy ) (BLASLONG , BLASLONG , float * , BLASLONG , float * );
146
- int (* shlaswp_ncopy ) (BLASLONG , BLASLONG , BLASLONG , float * , BLASLONG , blasint * , float * );
96
+ int (* sbtrsm_kernel_LN )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
97
+ int (* sbtrsm_kernel_LT )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
98
+ int (* sbtrsm_kernel_RN )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
99
+ int (* sbtrsm_kernel_RT )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
100
+
101
+ int (* sbtrsm_iunucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
102
+ int (* sbtrsm_iunncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
103
+ int (* sbtrsm_iutucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
104
+ int (* sbtrsm_iutncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
105
+ int (* sbtrsm_ilnucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
106
+ int (* sbtrsm_ilnncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
107
+ int (* sbtrsm_iltucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
108
+ int (* sbtrsm_iltncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
109
+ int (* sbtrsm_ounucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
110
+ int (* sbtrsm_ounncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
111
+ int (* sbtrsm_outucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
112
+ int (* sbtrsm_outncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
113
+ int (* sbtrsm_olnucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
114
+ int (* sbtrsm_olnncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
115
+ int (* sbtrsm_oltucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
116
+ int (* sbtrsm_oltncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , float * );
117
+
118
+ int (* sbtrmm_kernel_RN )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
119
+ int (* sbtrmm_kernel_RT )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
120
+ int (* sbtrmm_kernel_LN )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
121
+ int (* sbtrmm_kernel_LT )(BLASLONG , BLASLONG , BLASLONG , float , float * , float * , float * , BLASLONG , BLASLONG );
122
+
123
+ int (* sbtrmm_iunucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
124
+ int (* sbtrmm_iunncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
125
+ int (* sbtrmm_iutucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
126
+ int (* sbtrmm_iutncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
127
+ int (* sbtrmm_ilnucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
128
+ int (* sbtrmm_ilnncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
129
+ int (* sbtrmm_iltucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
130
+ int (* sbtrmm_iltncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
131
+ int (* sbtrmm_ounucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
132
+ int (* sbtrmm_ounncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
133
+ int (* sbtrmm_outucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
134
+ int (* sbtrmm_outncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
135
+ int (* sbtrmm_olnucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
136
+ int (* sbtrmm_olnncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
137
+ int (* sbtrmm_oltucopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
138
+ int (* sbtrmm_oltncopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
139
+
140
+ int (* sbsymm_iutcopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
141
+ int (* sbsymm_iltcopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
142
+ int (* sbsymm_outcopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
143
+ int (* sbsymm_oltcopy )(BLASLONG , BLASLONG , float * , BLASLONG , BLASLONG , BLASLONG , float * );
144
+
145
+ int (* sbneg_tcopy ) (BLASLONG , BLASLONG , float * , BLASLONG , float * );
146
+ int (* sblaswp_ncopy ) (BLASLONG , BLASLONG , BLASLONG , float * , BLASLONG , blasint * , float * );
147
147
148
148
#endif
149
149
0 commit comments