blob: ccdd1d6b949b154505014184575dc8ce9d4772bd [file] [log] [blame]
Miao Wange9993472020-02-10 15:00:10 -08001// Copyright (C) 2020 The Android Open Source Project
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7// http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15OPERATOR_SRCS = [
16 "src/add-nc.c",
17 "src/argmax-pooling-nhwc.c",
18 "src/average-pooling-nhwc.c",
19 "src/binary-elementwise-nd.c",
20 "src/channel-pad-nc.c",
21 "src/channel-shuffle-nc.c",
22 "src/clamp-nc.c",
23 "src/convolution-nchw.c",
24 "src/convolution-nhwc.c",
25 "src/deconvolution-nhwc.c",
26 "src/fully-connected-nc.c",
27 "src/global-average-pooling-ncw.c",
28 "src/global-average-pooling-nwc.c",
29 "src/hardswish-nc.c",
30 "src/leaky-relu-nc.c",
31 "src/max-pooling-nhwc.c",
32 "src/prelu-nc.c",
33 "src/resize-bilinear-nhwc.c",
34 "src/sigmoid-nc.c",
35 "src/softmax-nc.c",
36 "src/unpooling-nhwc.c",
37]
38
39TABLE_SRCS = [
40 "src/tables/exp2-k-over-64.c",
41 "src/tables/exp2-k-over-2048.c",
42]
43
44SCALAR_UKERNELS = [
45 "src/f32-argmaxpool/4x-scalar-c1.c",
46 "src/f32-argmaxpool/9p8x-scalar-c1.c",
47 "src/f32-argmaxpool/9x-scalar-c1.c",
Miao Wang400e4042020-04-17 10:15:59 -070048 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
49 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Miao Wang2534c2f2020-03-16 11:58:04 -070050 "src/f32-clamp/gen/scalar-x1.c",
51 "src/f32-clamp/gen/scalar-x2.c",
52 "src/f32-clamp/gen/scalar-x4.c",
Miao Wange9993472020-02-10 15:00:10 -080053 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
54 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
55 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
56 "src/f32-dwconv-spchw/3x3p1-scalar.c",
57 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
58 "src/f32-dwconv-spchw/5x5p2-scalar.c",
59 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -080060 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
61 "src/f32-dwconv/gen/up1x4-scalar.c",
62 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
63 "src/f32-dwconv/gen/up1x9-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -070064 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
65 "src/f32-dwconv/gen/up1x25-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -080066 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
67 "src/f32-dwconv/gen/up2x4-scalar.c",
68 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
69 "src/f32-dwconv/gen/up2x9-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -070070 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
71 "src/f32-dwconv/gen/up2x25-scalar.c",
72 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
73 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
74 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
75 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
76 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
77 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
78 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
79 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
80 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
81 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
82 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
83 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -080084 "src/f32-gavgpool-spchw/scalar-x1.c",
Miao Wang400e4042020-04-17 10:15:59 -070085 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
86 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
87 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
88 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
89 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -080090 "src/f32-gemm/gen/1x4-scalar.c",
91 "src/f32-gemm/gen/2x4-scalar.c",
92 "src/f32-gemm/gen/4x2-scalar.c",
93 "src/f32-gemm/gen/4x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -070094 "src/f32-gemm/gen/1x4-minmax-scalar.c",
95 "src/f32-gemm/gen/2x4-minmax-scalar.c",
96 "src/f32-gemm/gen/4x2-minmax-scalar.c",
97 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -080098 "src/f32-hswish/gen/scalar-x1.c",
99 "src/f32-hswish/gen/scalar-x2.c",
100 "src/f32-hswish/gen/scalar-x4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700101 "src/f32-ibilinear/gen/scalar-c1.c",
102 "src/f32-ibilinear/gen/scalar-c2.c",
103 "src/f32-ibilinear/gen/scalar-c4.c",
Miao Wange9993472020-02-10 15:00:10 -0800104 "src/f32-igemm/gen/1x4-scalar.c",
105 "src/f32-igemm/gen/2x4-scalar.c",
106 "src/f32-igemm/gen/4x2-scalar.c",
107 "src/f32-igemm/gen/4x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700108 "src/f32-igemm/gen/1x4-minmax-scalar.c",
109 "src/f32-igemm/gen/2x4-minmax-scalar.c",
110 "src/f32-igemm/gen/4x2-minmax-scalar.c",
111 "src/f32-igemm/gen/4x4-minmax-scalar.c",
112 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
113 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
114 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
115 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
116 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
117 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
118 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -0800119 "src/f32-prelu/gen/scalar-2x1.c",
120 "src/f32-prelu/gen/scalar-2x4.c",
121 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
122 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
123 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
124 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
125 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
126 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
127 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
128 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
129 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
130 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
131 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
132 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
133 "src/f32-rmax/scalar.c",
134 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
135 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
136 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
137 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
138 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
139 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
140 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
141 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
142 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700143 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
144 "src/f32-spmm/gen/1x1-minmax-scalar.c",
145 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
146 "src/f32-spmm/gen/2x1-minmax-scalar.c",
147 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
148 "src/f32-spmm/gen/4x1-minmax-scalar.c",
149 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
150 "src/f32-spmm/gen/8x1-minmax-scalar.c",
151 "src/f32-spmm/gen/8x2-minmax-scalar.c",
152 "src/f32-spmm/gen/8x4-minmax-scalar.c",
153 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
154 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
155 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
156 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
157 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
158 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
159 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
160 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
161 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
162 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
163 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
164 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800165 "src/f32-vbinary/gen/vmax-scalar-x1.c",
166 "src/f32-vbinary/gen/vmax-scalar-x2.c",
167 "src/f32-vbinary/gen/vmax-scalar-x4.c",
168 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
169 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
170 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
171 "src/f32-vbinary/gen/vmin-scalar-x1.c",
172 "src/f32-vbinary/gen/vmin-scalar-x2.c",
173 "src/f32-vbinary/gen/vmin-scalar-x4.c",
174 "src/f32-vbinary/gen/vminc-scalar-x1.c",
175 "src/f32-vbinary/gen/vminc-scalar-x2.c",
176 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700177 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
178 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
179 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
180 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
181 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
182 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
183 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
184 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
185 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
186 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
187 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
188 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
189 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
190 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
191 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
192 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
193 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
194 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
195 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
196 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
197 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Miao Wange9993472020-02-10 15:00:10 -0800198 "src/math/expminus-scalar-lut2048-p1.c",
199 "src/math/expminus-scalar-lut64-p2.c",
200 "src/math/expminus-scalar-p5.c",
201 "src/math/sigmoid-scalar-lut2048-p1-div.c",
202 "src/math/sigmoid-scalar-lut64-p2-div.c",
203 "src/math/sigmoid-scalar-p5-div.c",
Miao Wang400e4042020-04-17 10:15:59 -0700204 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
205 "src/q8-avgpool/9x-minmax-scalar-c1.c",
206 "src/q8-dwconv/up1x9-minmax-scalar.c",
207 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
208 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
209 "src/q8-gemm/2x2-minmax-scalar.c",
210 "src/q8-igemm/2x2-minmax-scalar.c",
211 "src/q8-vadd/minmax-scalar.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700212 "src/u8-clamp/scalar-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800213 "src/u8-lut32norm/scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700214 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Miao Wange9993472020-02-10 15:00:10 -0800215 "src/u8-rmax/scalar.c",
216 "src/x32-packx/x2-scalar.c",
217 "src/x32-packx/x3-scalar.c",
218 "src/x32-packx/x4-scalar.c",
219 "src/x32-pad/x2-scalar.c",
220 "src/x32-unpool/scalar.c",
221 "src/x32-zip/x2-scalar.c",
222 "src/x32-zip/x3-scalar.c",
223 "src/x32-zip/x4-scalar.c",
224 "src/x32-zip/xm-scalar.c",
225 "src/x8-lut/scalar.c",
226 "src/x8-zip/x2-scalar.c",
227 "src/x8-zip/x3-scalar.c",
228 "src/x8-zip/x4-scalar.c",
229 "src/x8-zip/xm-scalar.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700230 "src/requantization/precise-scalar.c",
231 "src/requantization/fp32-scalar.c",
232 "src/requantization/q31-scalar.c",
233 "src/requantization/gemmlowp-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -0800234]
235
236PSIMD_FASTMATH_UKERNELS = [
237 "src/f32-argmaxpool/4x-psimd-c4.c",
238 "src/f32-argmaxpool/9p8x-psimd-c4.c",
239 "src/f32-argmaxpool/9x-psimd-c4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700240 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
241 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700242 "src/f32-clamp/gen/psimd-x4.c",
243 "src/f32-clamp/gen/psimd-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700244 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
245 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
246 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
247 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
248 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
249 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
250 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
251 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
252 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
253 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
254 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
255 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
256 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
257 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
258 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
259 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
260 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
261 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
262 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
263 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
264 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
265 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
266 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
267 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
268 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
269 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
270 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
271 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
272 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
273 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
274 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
275 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
276 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Miao Wange9993472020-02-10 15:00:10 -0800277 "src/f32-hswish/gen/psimd-x4.c",
278 "src/f32-hswish/gen/psimd-x8.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700279 "src/f32-ibilinear/gen/psimd-c4.c",
280 "src/f32-ibilinear/gen/psimd-c8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700281 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
282 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
283 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
284 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
285 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
286 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
287 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
288 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
289 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
290 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
291 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
292 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
293 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
294 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Miao Wange9993472020-02-10 15:00:10 -0800295 "src/f32-prelu/gen/psimd-2x4.c",
296 "src/f32-prelu/gen/psimd-2x8.c",
297 "src/f32-rmax/psimd.c",
Miao Wang400e4042020-04-17 10:15:59 -0700298 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
299 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
300 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
301 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
302 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
303 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
304 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
305 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800306 "src/f32-vbinary/gen/vmax-psimd-x4.c",
307 "src/f32-vbinary/gen/vmax-psimd-x8.c",
308 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
309 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
310 "src/f32-vbinary/gen/vmin-psimd-x4.c",
311 "src/f32-vbinary/gen/vmin-psimd-x8.c",
312 "src/f32-vbinary/gen/vminc-psimd-x4.c",
313 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700314 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
315 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
316 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
317 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
318 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
319 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
320 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
321 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
322 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
323 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
324 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
325 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
326 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
327 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Miao Wange9993472020-02-10 15:00:10 -0800328 "src/x32-packx/x4-psimd.c",
329 "src/x32-pad/x2-psimd.c",
330 "src/x32-unpool/psimd.c",
331 "src/x32-zip/x2-psimd.c",
332 "src/x32-zip/x3-psimd.c",
333 "src/x32-zip/x4-psimd.c",
334 "src/x32-zip/xm-psimd.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700335 "src/requantization/precise-psimd.c",
336 "src/requantization/fp32-psimd.c",
Miao Wange9993472020-02-10 15:00:10 -0800337]
338
339PSIMD_ACCMATH_UKERNELS = [
340 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
341 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
342 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
343 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
344 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
345 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
346 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
347 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
348 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
349 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
350 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
351 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
352 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
353 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
354 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
355 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
356 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
357 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
358 "src/math/sigmoid-psimd-p5-div.c",
359]
360
361// ISA-specific micro-kernels
362NEON_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -0700363 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
364 "src/f32-avgpool/9x-minmax-neon-c4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700365 "src/f32-clamp/gen/neon-x4.c",
366 "src/f32-clamp/gen/neon-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700367 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
368 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
369 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
370 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Miao Wange9993472020-02-10 15:00:10 -0800371 "src/f32-gavgpool-spchw/neon-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700372 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
373 "src/f32-gavgpool/7x-minmax-neon-c4.c",
374 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
375 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
376 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
377 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
378 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
379 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
380 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
381 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
382 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
383 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
384 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
385 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
386 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
387 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
388 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
389 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
390 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
391 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
392 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
393 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
394 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
395 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
396 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
397 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
398 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
399 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
400 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
401 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
402 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
403 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
404 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800405 "src/f32-hswish/gen/neon-x4.c",
406 "src/f32-hswish/gen/neon-x8.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700407 "src/f32-ibilinear/gen/neon-c4.c",
408 "src/f32-ibilinear/gen/neon-c8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700409 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
410 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
411 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
412 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
413 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
414 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
415 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
416 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
417 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
418 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
419 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
420 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
421 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
422 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
423 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
424 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
425 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
426 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
427 "src/f32-pavgpool/9x-minmax-neon-c4.c",
428 "src/f32-ppmm/gen/4x8-minmax-neon.c",
429 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800430 "src/f32-prelu/gen/neon-2x4.c",
431 "src/f32-prelu/gen/neon-2x8.c",
432 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
433 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
434 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
435 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
436 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
437 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
438 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
439 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
440 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
441 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
442 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
443 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
444 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
445 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
446 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
447 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
448 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
449 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
450 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
451 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
452 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
453 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
454 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
455 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
456 "src/f32-rmax/neon.c",
457 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
458 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
459 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
460 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
461 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
462 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
463 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
464 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
465 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
466 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
467 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
468 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
469 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
470 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
471 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
472 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
473 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
474 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
475 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Miao Wang400e4042020-04-17 10:15:59 -0700476 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
477 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
478 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
479 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800480 "src/f32-vbinary/gen/vmax-neon-x4.c",
481 "src/f32-vbinary/gen/vmax-neon-x8.c",
482 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
483 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
484 "src/f32-vbinary/gen/vmin-neon-x4.c",
485 "src/f32-vbinary/gen/vmin-neon-x8.c",
486 "src/f32-vbinary/gen/vminc-neon-x4.c",
487 "src/f32-vbinary/gen/vminc-neon-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700488 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
489 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
490 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
491 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
492 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
493 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
494 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
495 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
496 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
497 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
498 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
499 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
500 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
501 "src/q8-avgpool/9x-minmax-neon-c8.c",
502 "src/q8-dwconv/up8x9-minmax-neon.c",
503 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
504 "src/q8-gavgpool/7x-minmax-neon-c8.c",
505 "src/q8-gemm/4x8-minmax-neon.c",
506 "src/q8-gemm/8x8-minmax-neon.c",
507 "src/q8-igemm/4x8-minmax-neon.c",
508 "src/q8-igemm/8x8-minmax-neon.c",
509 "src/q8-vadd/minmax-neon.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700510 "src/u8-clamp/neon-x64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700511 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Miao Wange9993472020-02-10 15:00:10 -0800512 "src/u8-rmax/neon.c",
513 "src/x32-packx/x4-neon-st4.c",
514 "src/x32-pad/x2-neon.c",
515 "src/x32-zip/x2-neon.c",
516 "src/x32-zip/x3-neon.c",
517 "src/x32-zip/x4-neon.c",
518 "src/x32-zip/xm-neon.c",
519 "src/x8-zip/x2-neon.c",
520 "src/x8-zip/x3-neon.c",
521 "src/x8-zip/x4-neon.c",
522 "src/x8-zip/xm-neon.c",
523 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
524 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
525 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
526 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
527 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
528 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
529 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700530 "src/requantization/precise-neon.c",
531 "src/requantization/fp32-neon.c",
532 "src/requantization/q31-neon.c",
533 "src/requantization/gemmlowp-neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800534]
535
536NEONFMA_UKERNELS = [
Miao Wang2534c2f2020-03-16 11:58:04 -0700537 "src/f32-ibilinear/gen/neonfma-c4.c",
538 "src/f32-ibilinear/gen/neonfma-c8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700539 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
540 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
541 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
542 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
543 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
544 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
545 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
546 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
547 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
548 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
549 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
550 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
551 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
552 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
553 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
554 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
555 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
556 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
557 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
558 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
559 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
560 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
561 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
562 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
563 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
564 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
565 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
566 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
567 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
568 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
569 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Miao Wange9993472020-02-10 15:00:10 -0800570 "src/f32-hswish/gen/neonfma-x4.c",
571 "src/f32-hswish/gen/neonfma-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700572 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
573 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Miao Wange9993472020-02-10 15:00:10 -0800574 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
575 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
576 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
577 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
578 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
579 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
580 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
581 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
582 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
583 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
584 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
585 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
586 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
587 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
588 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
589 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
590 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
591 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
592 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
593 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
594 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
595 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
596 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
597 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
598 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
599 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
600 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
601 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
602 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
603 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
604 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
605 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
606 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
607 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
608 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
609 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
610 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
611 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
612 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
613 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
614 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
615 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
616 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
617 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
618 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
619 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
620 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
621 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
622 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
623 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
624 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
625 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
626 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
627 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
628 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
629 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
630 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
631 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
632 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
633 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
634 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
635 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
636 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
637 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
638 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
639 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
640 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
641 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
642 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
643 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
644 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
645 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
646 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
647 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
648 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
649 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
650 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
651 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Miao Wang400e4042020-04-17 10:15:59 -0700652 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
653 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Miao Wange9993472020-02-10 15:00:10 -0800654 "src/math/exp-neonfma-lut64-p2.c",
655 "src/math/exp-neonfma-p5.c",
656 "src/math/expminus-neonfma-lut2048-p1.c",
657 "src/math/expminus-neonfma-lut64-p2.c",
658 "src/math/expminus-neonfma-p5.c",
659 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
660 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
661 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
662 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
663 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
664 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
665 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
666 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
667 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
668 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
669 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
670 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
671 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
672 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
673 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
674 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
675 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
676 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
677]
678
679AARCH64_NEONFMA_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -0700680 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
681 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
682 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
683 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
684 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
685 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
686 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
687 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
688 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
689 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
690 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
691 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
692 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
693 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
694 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
695 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
696 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
697 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
698 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
699 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
700 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
701 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
702 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
703 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
704 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
705 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Miao Wange9993472020-02-10 15:00:10 -0800706 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
707 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
708 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
709 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
710 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
711 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
712 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
713 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
714 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
715 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
716 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
717 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
718 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
719 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
720 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
721 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
722 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
723 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
724 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
725 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
726 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
727 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
728 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
729 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
730 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Miao Wang400e4042020-04-17 10:15:59 -0700731 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
732 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
733 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
734 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
735 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
736 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
737 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
738 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
739 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
740 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
741 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
742 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
743 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
744 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
745 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
746 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
747 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
748 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Miao Wange9993472020-02-10 15:00:10 -0800749 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
750 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
751 "src/math/sigmoid-neonfma-rr1-p5-div.c",
752 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
753 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
754 "src/math/sigmoid-neonfma-rr2-p5-div.c",
755]
756
757AARCH64_NEONFP16ARITH_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -0700758 "src/f16-gemm/gen/1x8-neonfp16arith-ld64.c",
Miao Wange9993472020-02-10 15:00:10 -0800759 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
760 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
761 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700762 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
763 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
764 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
765 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
766 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
767 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
768 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
769 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Miao Wange9993472020-02-10 15:00:10 -0800770]
771
772SSE_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -0700773 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
774 "src/f32-avgpool/9x-minmax-sse-c4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700775 "src/f32-clamp/gen/sse-x4.c",
776 "src/f32-clamp/gen/sse-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800777 "src/f32-dwconv-spchw/3x3p1-sse.c",
778 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -0700779 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
780 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
781 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
782 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
783 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
784 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
785 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
786 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
787 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
788 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
789 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
790 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Miao Wange9993472020-02-10 15:00:10 -0800791 "src/f32-gavgpool-spchw/sse-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700792 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
793 "src/f32-gavgpool/7x-minmax-sse-c4.c",
794 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
795 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
796 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
797 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
798 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
799 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
800 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
801 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
802 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
803 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
804 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
805 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
806 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Miao Wange9993472020-02-10 15:00:10 -0800807 "src/f32-hswish/gen/sse-x4.c",
808 "src/f32-hswish/gen/sse-x8.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700809 "src/f32-ibilinear/gen/sse-c4.c",
810 "src/f32-ibilinear/gen/sse-c8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700811 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
812 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
813 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
814 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
815 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
816 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
817 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
818 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
819 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
820 "src/f32-pavgpool/9x-minmax-sse-c4.c",
821 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Miao Wange9993472020-02-10 15:00:10 -0800822 "src/f32-rmax/sse.c",
Miao Wang400e4042020-04-17 10:15:59 -0700823 "src/f32-spmm/gen/4x1-minmax-sse.c",
824 "src/f32-spmm/gen/8x1-minmax-sse.c",
825 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
826 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
827 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
828 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
829 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
830 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
831 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
832 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800833 "src/f32-vbinary/gen/vmax-sse-x4.c",
834 "src/f32-vbinary/gen/vmax-sse-x8.c",
835 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
836 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
837 "src/f32-vbinary/gen/vmin-sse-x4.c",
838 "src/f32-vbinary/gen/vmin-sse-x8.c",
839 "src/f32-vbinary/gen/vminc-sse-x4.c",
840 "src/f32-vbinary/gen/vminc-sse-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700841 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
842 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
843 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
844 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
845 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
846 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
847 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
848 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
849 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
850 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
851 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
852 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
853 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
854 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Miao Wange9993472020-02-10 15:00:10 -0800855 "src/x32-packx/x4-sse.c",
856]
857
858SSE2_UKERNELS = [
859 "src/f32-argmaxpool/9p8x-sse2-c4.c",
860 "src/f32-argmaxpool/4x-sse2-c4.c",
861 "src/f32-argmaxpool/9x-sse2-c4.c",
862 "src/f32-prelu/gen/sse2-2x4.c",
863 "src/f32-prelu/gen/sse2-2x8.c",
864 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
865 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
866 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
867 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
868 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
869 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
870 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
871 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
872 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
873 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
874 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
875 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
876 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
877 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
878 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
879 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
880 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
881 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Miao Wang400e4042020-04-17 10:15:59 -0700882 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
883 "src/q8-avgpool/9x-minmax-sse2-c8.c",
884 "src/q8-igemm/4x4c2-minmax-sse2.c",
885 "src/q8-dwconv/up8x9-minmax-sse2.c",
886 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
887 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
888 "src/q8-gemm/2x4c8-minmax-sse2.c",
889 "src/q8-gemm/4x4c2-minmax-sse2.c",
890 "src/q8-vadd/minmax-sse2.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700891 "src/u8-clamp/sse2-x64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700892 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Miao Wange9993472020-02-10 15:00:10 -0800893 "src/u8-rmax/sse2.c",
894 "src/x32-pad/x2-sse2.c",
895 "src/x32-zip/x2-sse2.c",
896 "src/x32-zip/x3-sse2.c",
897 "src/x32-zip/x4-sse2.c",
898 "src/x32-zip/xm-sse2.c",
899 "src/x8-zip/x2-sse2.c",
900 "src/x8-zip/x3-sse2.c",
901 "src/x8-zip/x4-sse2.c",
902 "src/x8-zip/xm-sse2.c",
903 "src/math/exp-sse2-p5.c",
904 "src/math/expminus-sse2-p5.c",
905 "src/math/sigmoid-sse2-p5-div.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700906 "src/requantization/precise-sse2.c",
907 "src/requantization/fp32-sse2.c",
908 "src/requantization/q31-sse2.c",
909 "src/requantization/gemmlowp-sse2.c",
910]
911
912SSSE3_UKERNELS = [
913 "src/requantization/precise-ssse3.c",
914 "src/requantization/q31-ssse3.c",
915 "src/requantization/gemmlowp-ssse3.c",
Miao Wange9993472020-02-10 15:00:10 -0800916]
917
918SSE41_UKERNELS = [
919 "src/f32-prelu/gen/sse41-2x4.c",
920 "src/f32-prelu/gen/sse41-2x8.c",
921 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
922 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
923 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
924 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
925 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
926 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700927 "src/requantization/precise-sse4.c",
928 "src/requantization/q31-sse4.c",
929 "src/requantization/gemmlowp-sse4.c",
Miao Wange9993472020-02-10 15:00:10 -0800930]
931
932AVX_UKERNELS = [
Miao Wang2534c2f2020-03-16 11:58:04 -0700933 "src/f32-clamp/gen/avx-x8.c",
934 "src/f32-clamp/gen/avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -0700935 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
936 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
937 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
938 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
939 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
940 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
941 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
942 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
943 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
944 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
945 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
946 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
947 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
948 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
949 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
950 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
951 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
952 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
953 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
954 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
955 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
956 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
957 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
958 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
959 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
960 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
961 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
962 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
963 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
964 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Miao Wange9993472020-02-10 15:00:10 -0800965 "src/f32-hswish/gen/avx-x8.c",
966 "src/f32-hswish/gen/avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -0700967 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
968 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
969 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
970 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
971 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
972 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
973 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
974 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
975 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700976 "src/f32-prelu/gen/avx-2x8.c",
977 "src/f32-prelu/gen/avx-2x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800978 "src/f32-rmax/avx.c",
Miao Wang400e4042020-04-17 10:15:59 -0700979 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
980 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
981 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
982 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
983 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
984 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
985 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
986 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800987 "src/f32-vbinary/gen/vmax-avx-x8.c",
988 "src/f32-vbinary/gen/vmax-avx-x16.c",
989 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
990 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
991 "src/f32-vbinary/gen/vmin-avx-x8.c",
992 "src/f32-vbinary/gen/vmin-avx-x16.c",
993 "src/f32-vbinary/gen/vminc-avx-x8.c",
994 "src/f32-vbinary/gen/vminc-avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -0700995 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
996 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
997 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
998 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
999 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1000 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1001 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1002 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1003 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1004 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1005 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1006 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Miao Wange9993472020-02-10 15:00:10 -08001007 "src/f32-vscale/avx-unroll32.c",
1008]
1009
1010FMA3_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -07001011 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1012 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1013 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1014 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1015 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1016 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1017 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1018 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1019 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1020 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1021 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1022 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1023 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1024 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1025 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1026 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1027 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1028 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1029 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1030 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1031 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1032 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1033 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1034 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1035 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1036 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1037 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1038 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1039 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1040 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1041 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1042 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1043 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1044 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1045 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1046 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1047 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1048 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1049 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1050 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Miao Wange9993472020-02-10 15:00:10 -08001051 "src/f32-hswish/gen/fma3-x8.c",
1052 "src/f32-hswish/gen/fma3-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001053 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1054 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1055 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1056 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1057 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1058 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1059 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1060 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1061 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1062 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1063 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1064 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1065 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1066 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Miao Wange9993472020-02-10 15:00:10 -08001067]
1068
1069AVX2_UKERNELS = [
1070 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1071 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1072 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1073 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1074 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1075 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1076 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1077 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1078 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1079 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1080 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1081 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1082 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1083 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1084 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1085 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1086 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1087 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1088 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1089 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1090 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1091 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1092 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1093 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1094 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1095 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1096 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1097 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1098 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1099 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1100 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1101 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1102 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1103 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1104 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1105 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
1106 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1107 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1108 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1109 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1110 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1111 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1112 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1113 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1114 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1115 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1116 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1117 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1118 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1119 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1120 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1121 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1122 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1123 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1124 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1125 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1126 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1127 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1128 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1129 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1130 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1131 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1132 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1133 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1134 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1135 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
1136 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1137 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1138 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1139 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1140 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1141 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1142 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1143 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1144 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1145 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1146 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1147 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1148 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1149 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1150 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1151 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1152 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1153 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1154 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1155 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1156 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1157 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1158 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1159 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
1160 "src/math/exp-avx2-p5.c",
1161 "src/math/exp-avx2-perm-p3.c",
1162 "src/math/exp-avx2-perm-p4.c",
1163 "src/math/expminus-avx2-p5.c",
1164 "src/math/extexp-avx2-p5.c",
1165 "src/math/sigmoid-avx2-rr2-p5-div.c",
1166 "src/math/sigmoid-avx2-rr1-p5-div.c",
1167 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1168 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1169 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1170 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
1171]
1172
1173AVX512F_UKERNELS = [
Miao Wang2534c2f2020-03-16 11:58:04 -07001174 "src/f32-clamp/gen/avx512f-x16.c",
1175 "src/f32-clamp/gen/avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07001176 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1177 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1178 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1179 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1180 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1181 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1182 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1183 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1184 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1185 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1186 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1187 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1188 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1189 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1190 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1191 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1192 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1193 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1194 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1195 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1196 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1197 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1198 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1199 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Miao Wange9993472020-02-10 15:00:10 -08001200 "src/f32-hswish/gen/avx512f-x16.c",
1201 "src/f32-hswish/gen/avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07001202 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1203 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1204 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1205 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1206 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1207 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001208 "src/f32-prelu/gen/avx512f-2x16.c",
1209 "src/f32-prelu/gen/avx512f-2x32.c",
Miao Wange9993472020-02-10 15:00:10 -08001210 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1211 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1212 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1213 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1214 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1215 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1216 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1217 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1218 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1219 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1220 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1221 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1222 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1223 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1224 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1225 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1226 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1227 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1228 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1229 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1230 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1231 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1232 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1233 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1234 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1235 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1236 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1237 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1238 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1239 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1240 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1241 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1242 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1243 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1244 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1245 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1246 "src/f32-rmax/avx512f.c",
Miao Wang400e4042020-04-17 10:15:59 -07001247 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1248 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1249 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1250 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1251 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1252 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1253 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1254 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08001255 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1256 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1257 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1258 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1259 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1260 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1261 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1262 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07001263 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1264 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1265 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1266 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1267 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1268 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1269 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1270 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1271 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1272 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1273 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1274 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08001275 "src/f32-vscale/avx512f-unroll64.c",
1276 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1277 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1278 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1279 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1280 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1281 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1282 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1283 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1284 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1285 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1286 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1287 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1288 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1289 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1290 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1291 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1292 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1293 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1294 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1295 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1296 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1297 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1298 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1299 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
1300 "src/math/exp-avx512f-p5-scalef.c",
1301 "src/math/exp-avx512f-p5.c",
1302 "src/math/exp-avx512f-perm-p3.c",
1303 "src/math/exp-avx512f-perm2-p2.c",
1304 "src/math/extexp-avx512f-p5.c",
1305]
1306
1307AARCH32_ASM_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -07001308 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
1309 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1310 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1311 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1312 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1313 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1314 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1315 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1316 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1317 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1318 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Miao Wange9993472020-02-10 15:00:10 -08001319]
1320
1321AARCH64_ASM_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -07001322 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
1323 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1324 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
1325 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1326 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1327 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1328 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1329 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
1330 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
1331 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1332 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1333 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1334 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1335 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1336 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1337 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1338 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1339 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1340 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1341 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1342 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1343 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1344 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1345 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1346 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1347 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1348 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1349 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1350 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1351 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
1352 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
1353 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1354 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1355 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1356 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1357 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1358 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1359 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1360 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1361 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1362 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1363 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1364 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1365 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1366 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1367 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1368 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1369 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1370 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1371 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1372 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1373 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1374 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1375 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1376 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1377 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1378 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1379 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1380 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1381 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1382 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1383 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1384 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1385 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1386 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1387 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1388 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1389 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1390 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Miao Wange9993472020-02-10 15:00:10 -08001391]
1392
1393cc_defaults {
1394 name: "xnnpack_internal_default",
1395 vendor_available: true,
1396 sdk_version: "current",
1397 local_include_dirs: [
1398 "include",
1399 "src",
1400 ],
1401 cflags: [
1402 "-std=c99",
1403 "-DXNN_LOG_LEVEL=2",
1404 "-Wno-unused-parameter",
1405 "-Wno-missing-field-initializers",
1406 "-Wno-pointer-arith",
1407 ],
1408 stl: "libc++_static",
1409}
1410
1411cc_library_static {
1412 name: "xnnpack_tables",
1413 defaults: ["xnnpack_internal_default"],
1414 srcs: TABLE_SRCS,
1415}
1416
1417cc_library_static {
1418 name: "xnnpack_im2col",
1419 defaults: ["xnnpack_internal_default"],
1420 srcs: [
1421 "src/im2col.c",
1422 ],
1423}
1424
1425cc_library_static {
1426 name: "xnnpack_indirection",
1427 defaults: ["xnnpack_internal_default"],
1428 srcs: [
1429 "src/indirection.c",
1430 ],
1431 header_libs: [
1432 "fp16_headers",
1433 "fxdiv_headers",
1434 ],
1435 static_libs: [
1436 "libpthreadpool",
1437 ],
1438}
1439
1440cc_library_static {
1441 name: "xnnpack_operator_run",
1442 defaults: ["xnnpack_internal_default"],
1443 srcs: [
1444 "src/operator-run.c",
1445 ],
1446 cflags: [
1447 "-Wno-vla",
1448 ],
1449 header_libs: [
1450 "fp16_headers",
1451 "fxdiv_headers",
1452 ],
1453 static_libs: [
1454 "libclog",
1455 "libpthreadpool",
1456 ],
1457}
1458
1459cc_library_static {
1460 name: "xnnpack_operators",
1461 defaults: ["xnnpack_internal_default"],
1462 srcs: OPERATOR_SRCS + [
1463 "src/memory.c",
1464 "src/operator-delete.c",
1465 ],
1466 header_libs: [
1467 "fp16_headers",
1468 "fxdiv_headers",
1469 ],
1470 static_libs: [
1471 "libclog",
1472 "libpthreadpool",
1473 ],
1474 whole_static_libs: [
1475 "xnnpack_indirection",
1476 ],
1477}
1478
1479cc_library_static {
1480 name: "xnnpack_scalar_ukernels",
1481 defaults: ["xnnpack_internal_default"],
1482 srcs: SCALAR_UKERNELS,
1483 header_libs: [
1484 "fp16_headers",
1485 "fxdiv_headers",
1486 ],
1487 static_libs: [
1488 "libpthreadpool",
1489 "xnnpack_tables",
1490 ],
1491}
1492
1493cc_library_static {
1494 name: "xnnpack_psimd_fastmath_ukernels",
1495 defaults: ["xnnpack_internal_default"],
1496 srcs: PSIMD_FASTMATH_UKERNELS,
1497 cflags: [
1498 "-O3",
1499 "-ffast-math",
1500 ],
1501 arch: {
1502 arm: {
1503 cflags: [
1504 "-marm",
1505 "-mfpu=neon",
1506 ],
1507 },
1508 },
1509 header_libs: [
1510 "fp16_headers",
1511 "psimd_headers",
1512 ],
1513 static_libs: [
1514 "libpthreadpool",
1515 "xnnpack_tables",
1516 ],
1517}
1518
1519cc_library_static {
1520 name: "xnnpack_psimd_accmath_ukernels",
1521 defaults: ["xnnpack_internal_default"],
1522 srcs: PSIMD_ACCMATH_UKERNELS,
1523 cflags: [
1524 "-O3",
1525 ],
1526 arch: {
1527 arm: {
1528 cflags: [
1529 "-marm",
1530 "-mfpu=neon",
1531 ],
1532 },
1533 },
1534 header_libs: [
1535 "fp16_headers",
1536 "psimd_headers",
1537 ],
1538 static_libs: [
1539 "libpthreadpool",
1540 "xnnpack_tables",
1541 ],
1542}
1543
1544cc_library_static {
1545 name: "xnnpack_neon_ukernels",
1546 defaults: ["xnnpack_internal_default"],
1547 arch: {
1548 arm: {
1549 srcs: NEON_UKERNELS,
1550 cflags: [
1551 "-marm",
1552 "-mfpu=neon",
1553 ],
1554 },
1555 arm64: {
1556 srcs: NEON_UKERNELS,
1557 },
1558 x86: { enabled: false, },
1559 x86_64: { enabled: false, },
1560 },
1561 header_libs: [
1562 "fp16_headers",
1563 ],
1564 static_libs: [
1565 "libpthreadpool",
1566 "xnnpack_tables",
1567 ],
1568}
1569
1570cc_library_static {
1571 name: "xnnpack_neonfma_ukernels",
1572 defaults: ["xnnpack_internal_default"],
1573 arch: {
1574 arm: {
1575 srcs: NEONFMA_UKERNELS,
1576 cflags: [
1577 "-marm",
1578 "-mfpu=neon-vfpv4",
1579 ],
1580 },
1581 arm64: {
1582 srcs: NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1583 },
1584 x86: { enabled: false, },
1585 x86_64: { enabled: false, },
1586 },
1587 header_libs: [
1588 "fp16_headers",
1589 ],
1590 static_libs: [
1591 "libpthreadpool",
1592 "xnnpack_tables",
1593 ],
1594}
1595
1596cc_library_static {
1597 name: "xnnpack_neonfp16arith_ukernels",
1598 defaults: ["xnnpack_internal_default"],
1599 arch: {
1600 arm: { enabled: false, },
1601 arm64: {
1602 srcs: AARCH64_NEONFP16ARITH_UKERNELS,
1603 cflags: [
1604 "-march=armv8.2-a+fp16",
1605 ],
1606 },
1607 x86: { enabled: false, },
1608 x86_64: { enabled: false, },
1609 },
1610 header_libs: [
1611 "fp16_headers",
1612 ],
1613 static_libs: [
1614 "libpthreadpool",
1615 "xnnpack_tables",
1616 ],
1617}
1618
1619cc_library_static {
1620 name: "xnnpack_asm_ukernels",
1621 defaults: ["xnnpack_internal_default"],
1622 arch: {
1623 arm: {
1624 srcs: AARCH32_ASM_UKERNELS,
1625 },
1626 arm64: {
1627 srcs: AARCH64_ASM_UKERNELS,
Miao Wang400e4042020-04-17 10:15:59 -07001628 clang_asflags: [
1629 "-march=armv8.2-a+fp16",
1630 ],
Miao Wange9993472020-02-10 15:00:10 -08001631 },
1632 x86: { enabled: false, },
1633 x86_64: { enabled: false, },
1634 },
1635}
1636
1637cc_library_static {
1638 name: "xnnpack_sse2_ukernels",
1639 defaults: ["xnnpack_internal_default"],
1640 arch: {
1641 arm: { enabled: false, },
1642 arm64: { enabled: false, },
1643 x86: {
1644 srcs: SSE_UKERNELS + SSE2_UKERNELS,
1645 cflags: [
1646 "-msse2",
1647 ],
1648 },
1649 x86_64: {
1650 srcs: SSE_UKERNELS + SSE2_UKERNELS,
1651 cflags: [
1652 "-msse2",
1653 ],
1654 },
1655 },
1656 header_libs: [
1657 "fp16_headers",
1658 ],
1659 static_libs: [
1660 "libpthreadpool",
1661 "xnnpack_tables",
1662 ],
1663}
1664
1665cc_library_static {
Miao Wang2534c2f2020-03-16 11:58:04 -07001666 name: "xnnpack_ssse3_ukernels",
1667 defaults: ["xnnpack_internal_default"],
1668 arch: {
1669 arm: { enabled: false, },
1670 arm64: { enabled: false, },
1671 x86: {
1672 srcs: SSSE3_UKERNELS,
1673 cflags: [
1674 "-mssse3",
1675 ],
1676 },
1677 x86_64: {
1678 srcs: SSSE3_UKERNELS,
1679 cflags: [
1680 "-mssse3",
1681 ],
1682 },
1683 },
1684 header_libs: [
1685 "fp16_headers",
1686 ],
1687 static_libs: [
1688 "libpthreadpool",
1689 "xnnpack_tables",
1690 ],
1691}
1692
1693cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08001694 name: "xnnpack_sse41_ukernels",
1695 defaults: ["xnnpack_internal_default"],
1696 arch: {
1697 arm: { enabled: false, },
1698 arm64: { enabled: false, },
1699 x86: {
1700 srcs: SSE41_UKERNELS,
1701 cflags: [
1702 "-msse4.1",
1703 ],
1704 },
1705 x86_64: {
1706 srcs: SSE41_UKERNELS,
1707 cflags: [
1708 "-msse4.1",
1709 ],
1710 },
1711 },
1712 header_libs: [
1713 "fp16_headers",
1714 ],
1715 static_libs: [
1716 "libpthreadpool",
1717 "xnnpack_tables",
1718 ],
1719}
1720
1721cc_library_static {
1722 name: "xnnpack_avx_ukernels",
1723 defaults: ["xnnpack_internal_default"],
1724 arch: {
1725 arm: { enabled: false, },
1726 arm64: { enabled: false, },
1727 x86: {
1728 srcs: AVX_UKERNELS,
1729 cflags: [
1730 "-mavx",
1731 ],
1732 },
1733 x86_64: {
1734 srcs: AVX_UKERNELS,
1735 cflags: [
1736 "-mavx",
1737 ],
1738 },
1739 },
1740 header_libs: [
1741 "fp16_headers",
1742 ],
1743 static_libs: [
1744 "libpthreadpool",
1745 "xnnpack_tables",
1746 ],
1747}
1748
1749cc_library_static {
1750 name: "xnnpack_fma3_ukernels",
1751 defaults: ["xnnpack_internal_default"],
1752 arch: {
1753 arm: { enabled: false, },
1754 arm64: { enabled: false, },
1755 x86: {
1756 srcs: FMA3_UKERNELS,
1757 cflags: [
1758 "-mfma",
1759 ],
1760 },
1761 x86_64: {
1762 srcs: FMA3_UKERNELS,
1763 cflags: [
1764 "-mfma",
1765 ],
1766 },
1767 },
1768 header_libs: [
1769 "fp16_headers",
1770 ],
1771 static_libs: [
1772 "libpthreadpool",
1773 "xnnpack_tables",
1774 ],
1775}
1776
1777cc_library_static {
1778 name: "xnnpack_avx2_ukernels",
1779 defaults: ["xnnpack_internal_default"],
1780 arch: {
1781 arm: { enabled: false, },
1782 arm64: { enabled: false, },
1783 x86: {
1784 srcs: AVX2_UKERNELS,
1785 cflags: [
1786 "-mfma",
1787 "-mavx2",
1788 ],
1789 },
1790 x86_64: {
1791 srcs: AVX2_UKERNELS,
1792 cflags: [
1793 "-mfma",
1794 "-mavx2",
1795 ],
1796 },
1797 },
1798 header_libs: [
1799 "fp16_headers",
1800 ],
1801 static_libs: [
1802 "libpthreadpool",
1803 "xnnpack_tables",
1804 ],
1805}
1806
1807cc_library_static {
1808 name: "xnnpack_avx512f_ukernels",
1809 defaults: ["xnnpack_internal_default"],
1810 arch: {
1811 arm: { enabled: false, },
1812 arm64: { enabled: false, },
1813 x86: {
1814 srcs: AVX512F_UKERNELS,
1815 cflags: [
1816 "-mavx512f",
1817 ],
1818 },
1819 x86_64: {
1820 srcs: AVX512F_UKERNELS,
1821 cflags: [
1822 "-mavx512f",
1823 ],
1824 },
1825 },
1826 header_libs: [
1827 "fp16_headers",
1828 ],
1829 static_libs: [
1830 "libpthreadpool",
1831 "xnnpack_tables",
1832 ],
1833}
1834
1835cc_library_static {
1836 name: "xnnpack_ukernels",
1837 defaults: ["xnnpack_internal_default"],
1838 arch: {
1839 arm: {
1840 whole_static_libs: [
1841 "xnnpack_psimd_fastmath_ukernels",
1842 "xnnpack_psimd_accmath_ukernels",
1843 "xnnpack_neon_ukernels",
1844 "xnnpack_neonfma_ukernels",
1845 "xnnpack_asm_ukernels",
1846 ],
1847 },
1848 arm64: {
1849 whole_static_libs: [
1850 "xnnpack_psimd_fastmath_ukernels",
1851 "xnnpack_psimd_accmath_ukernels",
1852 "xnnpack_neon_ukernels",
1853 "xnnpack_neonfma_ukernels",
1854 "xnnpack_neonfp16arith_ukernels",
1855 "xnnpack_asm_ukernels",
1856 ],
1857 },
1858 x86: {
1859 whole_static_libs: [
1860 "xnnpack_psimd_fastmath_ukernels",
1861 "xnnpack_psimd_accmath_ukernels",
1862 "xnnpack_sse2_ukernels",
Miao Wang2534c2f2020-03-16 11:58:04 -07001863 "xnnpack_ssse3_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08001864 "xnnpack_sse41_ukernels",
1865 "xnnpack_avx_ukernels",
1866 "xnnpack_fma3_ukernels",
1867 "xnnpack_avx2_ukernels",
1868 "xnnpack_avx512f_ukernels",
1869 ],
1870 },
1871 x86_64: {
1872 whole_static_libs: [
1873 "xnnpack_psimd_fastmath_ukernels",
1874 "xnnpack_psimd_accmath_ukernels",
1875 "xnnpack_sse2_ukernels",
Miao Wang2534c2f2020-03-16 11:58:04 -07001876 "xnnpack_ssse3_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08001877 "xnnpack_sse41_ukernels",
1878 "xnnpack_avx_ukernels",
1879 "xnnpack_fma3_ukernels",
1880 "xnnpack_avx2_ukernels",
1881 "xnnpack_avx512f_ukernels",
1882 ],
1883 },
1884 },
1885 whole_static_libs: [
1886 "xnnpack_scalar_ukernels",
1887 "xnnpack_tables",
1888 ],
1889}
1890
1891cc_library_static {
1892 name: "libXNNPACK",
1893 defaults: ["xnnpack_internal_default"],
1894 export_include_dirs: ["include"],
1895 srcs: [
1896 "src/init.c",
1897 "src/runtime.c",
1898 "src/subgraph.c",
1899 "src/tensor.c",
1900 ],
1901 whole_static_libs: [
1902 "libclog",
1903 "libcpuinfo",
1904 "libpthreadpool",
1905 "xnnpack_ukernels",
1906 "xnnpack_operator_run",
1907 "xnnpack_operators",
1908 ],
1909}
1910
1911// Tests and benchmarks
1912cc_defaults {
1913 name: "xnnpack_tests_default",
1914 vendor_available: true,
1915 stl: "libc++_static",
1916 local_include_dirs: [
1917 "bench",
1918 "models",
1919 "test",
1920 "src",
1921 ],
1922 cflags: [
Miao Wang400e4042020-04-17 10:15:59 -07001923 "-Wno-unused-function",
1924 "-Wno-unused-parameter",
Miao Wange9993472020-02-10 15:00:10 -08001925 ],
1926 header_libs: [
1927 "fp16_headers",
1928 ],
1929 static_libs: [
1930 "libXNNPACK",
1931 "libpthreadpool",
1932 "libgmock",
1933 ],
1934 shared_libs: [
1935 "liblog",
1936 ],
1937}
1938
1939cc_library_static {
1940 name: "xnnpack_mobilenet_v1",
1941 defaults: ["xnnpack_tests_default"],
1942 srcs: [
1943 "models/mobilenet-v1.cc",
1944 ],
1945}
1946
1947cc_library_static {
1948 name: "xnnpack_mobilenet_v2",
1949 defaults: ["xnnpack_tests_default"],
1950 srcs: [
1951 "models/mobilenet-v2.cc",
1952 ],
1953}
1954
1955cc_library_static {
1956 name: "xnnpack_mobilenet_v3_large",
1957 defaults: ["xnnpack_tests_default"],
1958 srcs: [
1959 "models/mobilenet-v3-large.cc",
1960 ],
1961}
1962
1963cc_library_static {
1964 name: "xnnpack_mobilenet_v3_small",
1965 defaults: ["xnnpack_tests_default"],
1966 srcs: [
1967 "models/mobilenet-v3-small.cc",
1968 ],
1969}
1970
1971cc_benchmark {
1972 name: "xnnpack_end2end_bench",
1973 defaults: ["xnnpack_tests_default"],
1974 srcs: [
1975 "bench/end2end.cc",
1976 "bench/utils.cc",
1977 ],
1978 cflags: [
1979 "-Wno-unused-result"
1980 ],
1981 static_libs: [
1982 "libcpuinfo",
1983 "libgoogle-benchmark",
1984 "xnnpack_mobilenet_v1",
1985 "xnnpack_mobilenet_v2",
1986 "xnnpack_mobilenet_v3_large",
1987 "xnnpack_mobilenet_v3_small",
1988 ],
1989}
1990
1991cc_test {
1992 name: "xnnpack_add_nc_test",
1993 defaults: ["xnnpack_tests_default"],
1994 srcs: [
1995 "test/add-nc.cc",
1996 ],
1997 test_suites: [
1998 "general-tests",
1999 ],
2000}
2001
2002cc_test {
2003 name: "xnnpack_add_nd_test",
2004 defaults: ["xnnpack_tests_default"],
2005 srcs: [
2006 "test/add-nd.cc",
2007 ],
2008 test_suites: [
2009 "general-tests",
2010 ],
2011}
2012
2013cc_test {
2014 name: "xnnpack_argmax_pooling_nhwc_test",
2015 defaults: ["xnnpack_tests_default"],
2016 srcs: [
2017 "test/argmax-pooling-nhwc.cc",
2018 ],
2019 test_suites: [
2020 "general-tests",
2021 ],
2022}
2023
2024cc_test {
2025 name: "xnnpack_average_pooling_nhwc_test",
2026 defaults: ["xnnpack_tests_default"],
2027 srcs: [
2028 "test/average-pooling-nhwc.cc",
2029 ],
2030 test_suites: [
2031 "general-tests",
2032 ],
2033}
2034
2035cc_test {
2036 name: "xnnpack_channel_pad_nc_test",
2037 defaults: ["xnnpack_tests_default"],
2038 srcs: [
2039 "test/channel-pad-nc.cc",
2040 ],
2041 test_suites: [
2042 "general-tests",
2043 ],
2044}
2045
2046cc_test {
2047 name: "xnnpack_channel_shuffle_nc_test",
2048 defaults: ["xnnpack_tests_default"],
2049 srcs: [
2050 "test/channel-shuffle-nc.cc",
2051 ],
2052 test_suites: [
2053 "general-tests",
2054 ],
2055}
2056
2057cc_test {
2058 name: "xnnpack_clamp_nc_test",
2059 defaults: ["xnnpack_tests_default"],
2060 srcs: [
2061 "test/clamp-nc.cc",
2062 ],
2063 test_suites: [
2064 "general-tests",
2065 ],
2066}
2067
2068cc_test {
2069 name: "xnnpack_convolution_nhwc_test",
2070 defaults: ["xnnpack_tests_default"],
2071 srcs: [
2072 "test/convolution-nhwc.cc",
2073 ],
2074 test_suites: [
2075 "general-tests",
2076 ],
2077}
2078
2079cc_test {
2080 name: "xnnpack_convolution_nchw_test",
2081 defaults: ["xnnpack_tests_default"],
2082 srcs: [
2083 "test/convolution-nchw.cc",
2084 ],
2085 test_suites: [
2086 "general-tests",
2087 ],
2088}
2089
2090cc_test {
2091 name: "xnnpack_deconvolution_nhwc_test",
2092 defaults: ["xnnpack_tests_default"],
2093 srcs: [
2094 "test/deconvolution-nhwc.cc",
2095 ],
2096 test_suites: [
2097 "general-tests",
2098 ],
2099}
2100
2101cc_test {
2102 name: "xnnpack_divide_nd_test",
2103 defaults: ["xnnpack_tests_default"],
2104 srcs: [
2105 "test/divide-nd.cc",
2106 ],
2107 test_suites: [
2108 "general-tests",
2109 ],
2110}
2111
2112cc_test {
2113 name: "xnnpack_fully_connected_nc_test",
2114 defaults: ["xnnpack_tests_default"],
2115 srcs: [
2116 "test/fully-connected-nc.cc",
2117 ],
2118 test_suites: [
2119 "general-tests",
2120 ],
2121}
2122
2123cc_test {
2124 name: "xnnpack_global_average_pooling_nwc_test",
2125 defaults: ["xnnpack_tests_default"],
2126 srcs: [
2127 "test/global-average-pooling-nwc.cc",
2128 ],
2129 test_suites: [
2130 "general-tests",
2131 ],
2132}
2133
2134cc_test {
2135 name: "xnnpack_global_average_pooling_ncw_test",
2136 defaults: ["xnnpack_tests_default"],
2137 srcs: [
2138 "test/global-average-pooling-ncw.cc",
2139 ],
2140 test_suites: [
2141 "general-tests",
2142 ],
2143}
2144
2145cc_test {
2146 name: "xnnpack_hardswish_nc_test",
2147 defaults: ["xnnpack_tests_default"],
2148 srcs: [
2149 "test/hardswish-nc.cc",
2150 ],
2151 test_suites: [
2152 "general-tests",
2153 ],
2154}
2155
2156cc_test {
2157 name: "xnnpack_leaky_relu_nc_test",
2158 defaults: ["xnnpack_tests_default"],
2159 srcs: [
2160 "test/leaky-relu-nc.cc",
2161 ],
2162 test_suites: [
2163 "general-tests",
2164 ],
2165}
2166
2167cc_test {
2168 name: "xnnpack_max_pooling_nhwc_test",
2169 defaults: ["xnnpack_tests_default"],
2170 srcs: [
2171 "test/max-pooling-nhwc.cc",
2172 ],
2173 test_suites: [
2174 "general-tests",
2175 ],
2176}
2177
2178cc_test {
2179 name: "xnnpack_maximum_nd_test",
2180 defaults: ["xnnpack_tests_default"],
2181 srcs: [
2182 "test/maximum-nd.cc",
2183 ],
2184 test_suites: [
2185 "general-tests",
2186 ],
2187}
2188
2189cc_test {
2190 name: "xnnpack_minimum_nd_test",
2191 defaults: ["xnnpack_tests_default"],
2192 srcs: [
2193 "test/minimum-nd.cc",
2194 ],
2195 test_suites: [
2196 "general-tests",
2197 ],
2198}
2199
2200cc_test {
2201 name: "xnnpack_multiply_nd_test",
2202 defaults: ["xnnpack_tests_default"],
2203 srcs: [
2204 "test/multiply-nd.cc",
2205 ],
2206 test_suites: [
2207 "general-tests",
2208 ],
2209}
2210
2211cc_test {
2212 name: "xnnpack_prelu_nc_test",
2213 defaults: ["xnnpack_tests_default"],
2214 srcs: [
2215 "test/prelu-nc.cc",
2216 ],
2217 test_suites: [
2218 "general-tests",
2219 ],
2220}
2221
2222cc_test {
2223 name: "xnnpack_resize_bilinear_nhwc_test",
2224 defaults: ["xnnpack_tests_default"],
2225 srcs: [
2226 "test/resize-bilinear-nhwc.cc",
2227 ],
2228 test_suites: [
2229 "general-tests",
2230 ],
2231}
2232
2233cc_test {
2234 name: "xnnpack_sigmoid_nc_test",
2235 defaults: ["xnnpack_tests_default"],
2236 srcs: [
2237 "test/sigmoid-nc.cc",
2238 ],
2239 test_suites: [
2240 "general-tests",
2241 ],
2242}
2243
2244cc_test {
2245 name: "xnnpack_softmax_nc_test",
2246 defaults: ["xnnpack_tests_default"],
2247 srcs: [
2248 "test/softmax-nc.cc",
2249 ],
2250 test_suites: [
2251 "general-tests",
2252 ],
2253}
2254
2255cc_test {
2256 name: "xnnpack_subtract_nd_test",
2257 defaults: ["xnnpack_tests_default"],
2258 srcs: [
2259 "test/subtract-nd.cc",
2260 ],
2261 test_suites: [
2262 "general-tests",
2263 ],
2264}
2265
2266cc_test {
2267 name: "xnnpack_unpooling_nhwc_test",
2268 defaults: ["xnnpack_tests_default"],
2269 srcs: [
2270 "test/unpooling-nhwc.cc",
2271 ],
2272 test_suites: [
2273 "general-tests",
2274 ],
2275}