@@ -20,6 +20,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
20
20
f(in_T, out_T, W_T, narrow, 1152 ) \
21
21
f(in_T, out_T, W_T, narrow, 1280 ) \
22
22
f(in_T, out_T, W_T, narrow, 1536 ) \
23
+ f(in_T, out_T, W_T, narrow, 1664 ) \
23
24
f(in_T, out_T, W_T, narrow, 1728 ) \
24
25
f(in_T, out_T, W_T, narrow, 1792 ) \
25
26
f(in_T, out_T, W_T, narrow, 2048 ) \
@@ -36,6 +37,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
36
37
f(in_T, out_T, W_T, narrow, 5120 ) \
37
38
f(in_T, out_T, W_T, narrow, 5504 ) \
38
39
f(in_T, out_T, W_T, narrow, 5632 ) \
40
+ f(in_T, out_T, W_T, narrow, 5888 ) \
39
41
f(in_T, out_T, W_T, narrow, 6144 ) \
40
42
f(in_T, out_T, W_T, narrow, 6400 ) \
41
43
f(in_T, out_T, W_T, narrow, 6848 ) \
@@ -45,6 +47,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
45
47
f(in_T, out_T, W_T, narrow, 9216 ) \
46
48
f(in_T, out_T, W_T, narrow, 10240 ) \
47
49
f(in_T, out_T, W_T, narrow, 11008 ) \
50
+ f(in_T, out_T, W_T, narrow, 11264 ) \
48
51
f(in_T, out_T, W_T, narrow, 12288 ) \
49
52
f(in_T, out_T, W_T, narrow, 13696 ) \
50
53
f(in_T, out_T, W_T, narrow, 13824 ) \
@@ -53,6 +56,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
53
56
f(in_T, out_T, W_T, narrow, 16384 ) \
54
57
f(in_T, out_T, W_T, narrow, 20480 ) \
55
58
f(in_T, out_T, W_T, narrow, 22016 ) \
59
+ f(in_T, out_T, W_T, narrow, 22528 ) \
56
60
f(in_T, out_T, W_T, narrow, 24576 ) \
57
61
f(in_T, out_T, W_T, narrow, 27392 ) \
58
62
f(in_T, out_T, W_T, narrow, 27648 ) \
@@ -91,6 +95,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
91
95
f(in_T, out_T, W_T, 1152 , narrow) \
92
96
f(in_T, out_T, W_T, 1280 , narrow) \
93
97
f(in_T, out_T, W_T, 1536 , narrow) \
98
+ f(in_T, out_T, W_T, 1664 , narrow) \
94
99
f(in_T, out_T, W_T, 1728 , narrow) \
95
100
f(in_T, out_T, W_T, 1792 , narrow) \
96
101
f(in_T, out_T, W_T, 2048 , narrow) \
@@ -107,6 +112,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
107
112
f(in_T, out_T, W_T, 5120 , narrow) \
108
113
f(in_T, out_T, W_T, 5504 , narrow) \
109
114
f(in_T, out_T, W_T, 5632 , narrow) \
115
+ f(in_T, out_T, W_T, 5888 , narrow) \
110
116
f(in_T, out_T, W_T, 6144 , narrow) \
111
117
f(in_T, out_T, W_T, 6400 , narrow) \
112
118
f(in_T, out_T, W_T, 6848 , narrow) \
@@ -116,6 +122,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
116
122
f(in_T, out_T, W_T, 9216 , narrow) \
117
123
f(in_T, out_T, W_T, 10240 , narrow) \
118
124
f(in_T, out_T, W_T, 11008 , narrow) \
125
+ f(in_T, out_T, W_T, 11264 , narrow) \
119
126
f(in_T, out_T, W_T, 12288 , narrow) \
120
127
f(in_T, out_T, W_T, 13696 , narrow) \
121
128
f(in_T, out_T, W_T, 13824 , narrow) \
@@ -124,6 +131,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
124
131
f(in_T, out_T, W_T, 16384 , narrow) \
125
132
f(in_T, out_T, W_T, 20480 , narrow) \
126
133
f(in_T, out_T, W_T, 22016 , narrow) \
134
+ f(in_T, out_T, W_T, 22528 , narrow) \
127
135
f(in_T, out_T, W_T, 24576 , narrow) \
128
136
f(in_T, out_T, W_T, 27392 , narrow) \
129
137
f(in_T, out_T, W_T, 27648 , narrow) \
0 commit comments