blob: 2c64ba635757c6ef8fccee7697988ab720261718 [file] [log] [blame]
Marat Dukhan9b474cf2021-05-25 16:37:48 -07001# Copyright 2021 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5
Marat Dukhanc46e6712021-06-01 19:00:16 -07006- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64
7 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
8 k-block: 8
9- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64
10 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
11 k-block: 8
12- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64
13 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
14 k-block: 8
15- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64
16 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
17 k-block: 8
18- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__ssse3_ld64
19 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
20 k-block: 8
21- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__ssse3_ld64
22 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
23 k-block: 8
24- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__ssse3_ld64
25 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
26 k-block: 8
27- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__ssse3_ld64
28 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
29 k-block: 8
30- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64
31 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
32 k-block: 8
33- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64
34 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
35 k-block: 8
36- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64
37 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
38 k-block: 8
39- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64
40 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
41 k-block: 8
42- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64
43 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
44 k-block: 8
45- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64
46 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
47 k-block: 8
48- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64
49 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
50 k-block: 8
51- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64
52 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
53 k-block: 8
54- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64
55 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
56 k-block: 8
57- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64
58 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
59 k-block: 8
60- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64
61 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
62 k-block: 8
63- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64
64 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
65 k-block: 8
66- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128
67 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
68 k-block: 8
69- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128
70 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
71 k-block: 8
72- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128
73 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
74 k-block: 8
75- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128
76 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
77 k-block: 8
78- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__ssse3_ld128
79 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
80 k-block: 8
81- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__ssse3_ld128
82 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
83 k-block: 8
84- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__ssse3_ld128
85 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
86 k-block: 8
87- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__ssse3_ld128
88 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
89 k-block: 8
90- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128
91 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
92 k-block: 8
93- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128
94 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
95 k-block: 8
96- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128
97 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
98 k-block: 8
99- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128
100 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
101 k-block: 8
102- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128
103 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
104 k-block: 8
105- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128
106 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
107 k-block: 8
108- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128
109 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
110 k-block: 8
111- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128
112 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
113 k-block: 8
114- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128
115 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
116 k-block: 8
117- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128
118 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
119 k-block: 8
120- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128
121 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
122 k-block: 8
123- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128
124 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
125 k-block: 8
126- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64
127 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
128 k-block: 8
129- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64
130 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
131 k-block: 8
132- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64
133 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
134 k-block: 8
135- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__ssse3_ld64
136 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
137 k-block: 8
138- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__ssse3_ld64
139 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
140 k-block: 8
141- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld64
142 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
143 k-block: 8
144- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64
145 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
146 k-block: 8
147- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64
148 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
149 k-block: 8
150- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64
151 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
152 k-block: 8
153- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64
154 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
155 k-block: 8
156- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64
157 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
158 k-block: 8
159- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64
160 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
161 k-block: 8
162- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64
163 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
164 k-block: 8
165- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64
166 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
167 k-block: 8
168- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64
169 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
170 k-block: 8
171- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128
172 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
173 k-block: 8
174- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128
175 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
176 k-block: 8
177- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128
178 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
179 k-block: 8
180- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__ssse3_ld128
181 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
182 k-block: 8
183- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__ssse3_ld128
184 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
185 k-block: 8
186- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld128
187 init: xnn_init_qs8_conv_minmax_fp32_sse2_params
188 k-block: 8
189- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128
190 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
191 k-block: 8
192- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128
193 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
194 k-block: 8
195- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128
196 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
197 k-block: 8
198- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128
199 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
200 k-block: 8
201- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128
202 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
203 k-block: 8
204- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128
205 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
206 k-block: 8
207- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128
208 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
209 k-block: 8
210- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128
211 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
212 k-block: 8
213- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128
214 init: xnn_init_qs8_conv_minmax_fp32_sse4_params
215 k-block: 8
Marat Dukhan9b474cf2021-05-25 16:37:48 -0700216- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__avx2
217 init: xnn_init_qs8_conv_minmax_fp32_avx2_params
218 k-block: 8
219- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__avx2
220 init: xnn_init_qs8_conv_minmax_fp32_avx2_params
221 k-block: 8
222- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x8c8__avx2
223 init: xnn_init_qs8_conv_minmax_fp32_avx2_params
224 k-block: 8
Marat Dukhan71855ee2021-05-25 19:05:06 -0700225- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx
226 init: xnn_init_qs8_conv_minmax_fp32_avx512_params
227 k-block: 8
228- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx
229 init: xnn_init_qs8_conv_minmax_fp32_avx512_params
230 k-block: 8
231- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx
232 init: xnn_init_qs8_conv_minmax_fp32_avx512_params
233 k-block: 8
234- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx
235 init: xnn_init_qs8_conv_minmax_fp32_avx512_params
236 k-block: 8