blob: da35aed60b45e91e8e37799619d20fe50fe27da9 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 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#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan8d3c6932020-03-06 20:27:27 -08009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_copts", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan08c4a432019-10-03 09:29:21 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
22MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
46OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
52OPERATOR_SRCS = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
Marat Dukhanfd8e6892020-01-27 15:25:25 -080072 "src/softmax-nc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080073 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -080085 "src/f32-avgpool/9p8x-scalar-c1.c",
86 "src/f32-avgpool/9x-scalar-c1.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080087 "src/f32-bilinear/gen/scalar-c1.c",
88 "src/f32-bilinear/gen/scalar-c2.c",
89 "src/f32-bilinear/gen/scalar-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070090 "src/f32-clamp/scalar.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080091 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080092 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070093 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070094 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070095 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080096 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080097 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080098 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
99 "src/f32-dwconv/gen/up1x25-scalar.c",
100 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
101 "src/f32-dwconv/gen/up1x4-scalar.c",
102 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
103 "src/f32-dwconv/gen/up1x9-scalar.c",
104 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
105 "src/f32-dwconv/gen/up2x25-scalar.c",
106 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
107 "src/f32-dwconv/gen/up2x4-scalar.c",
108 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
109 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700110 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700111 "src/f32-gavgpool/mp7p7q-scalar.c",
112 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800113 "src/f32-gemm/gen-inc/1x4-scalar.c",
114 "src/f32-gemm/gen-inc/2x4-scalar.c",
115 "src/f32-gemm/gen-inc/4x4-scalar.c",
116 "src/f32-gemm/gen/1x4-scalar.c",
117 "src/f32-gemm/gen/2x4-scalar.c",
118 "src/f32-gemm/gen/4x2-scalar.c",
119 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800120 "src/f32-hswish/gen/scalar-x1.c",
121 "src/f32-hswish/gen/scalar-x2.c",
122 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800123 "src/f32-igemm/gen/1x4-scalar.c",
124 "src/f32-igemm/gen/2x4-scalar.c",
125 "src/f32-igemm/gen/4x2-scalar.c",
126 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800127 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800128 "src/f32-pavgpool/9p8x-scalar-c1.c",
129 "src/f32-pavgpool/9x-scalar-c1.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800130 "src/f32-ppmm/gen/2x4-scalar.c",
131 "src/f32-ppmm/gen/3x3-scalar.c",
132 "src/f32-ppmm/gen/4x2-scalar.c",
133 "src/f32-ppmm/gen/4x4-scalar.c",
134 "src/f32-prelu/gen/scalar-2x1.c",
135 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800136 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
137 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
138 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
139 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
140 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
141 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
142 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
143 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
144 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
145 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
146 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
147 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700148 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800149 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
150 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
151 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
152 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
153 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
154 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
155 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
156 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
157 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800158 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
159 "src/f32-spmm/gen/1x1-scalar.c",
160 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
161 "src/f32-spmm/gen/2x1-scalar.c",
162 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
163 "src/f32-spmm/gen/4x1-scalar.c",
164 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
165 "src/f32-spmm/gen/8x1-scalar.c",
166 "src/f32-spmm/gen/8x2-scalar.c",
167 "src/f32-spmm/gen/8x4-scalar.c",
168 "src/f32-vbinary/gen/vadd-scalar-x1.c",
169 "src/f32-vbinary/gen/vadd-scalar-x2.c",
170 "src/f32-vbinary/gen/vadd-scalar-x4.c",
171 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
172 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
173 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800174 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
175 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
176 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
177 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
178 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
179 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800180 "src/f32-vbinary/gen/vmax-scalar-x1.c",
181 "src/f32-vbinary/gen/vmax-scalar-x2.c",
182 "src/f32-vbinary/gen/vmax-scalar-x4.c",
183 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
184 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
185 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
186 "src/f32-vbinary/gen/vmin-scalar-x1.c",
187 "src/f32-vbinary/gen/vmin-scalar-x2.c",
188 "src/f32-vbinary/gen/vmin-scalar-x4.c",
189 "src/f32-vbinary/gen/vminc-scalar-x1.c",
190 "src/f32-vbinary/gen/vminc-scalar-x2.c",
191 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800192 "src/f32-vbinary/gen/vmul-scalar-x1.c",
193 "src/f32-vbinary/gen/vmul-scalar-x2.c",
194 "src/f32-vbinary/gen/vmul-scalar-x4.c",
195 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
196 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
197 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800198 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
199 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
200 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
202 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
203 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
204 "src/f32-vbinary/gen/vsub-scalar-x1.c",
205 "src/f32-vbinary/gen/vsub-scalar-x2.c",
206 "src/f32-vbinary/gen/vsub-scalar-x4.c",
207 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
208 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
209 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
210 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
211 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
212 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800213 "src/math/expminus-scalar-lut2048-p1.c",
214 "src/math/expminus-scalar-lut64-p2.c",
215 "src/math/expminus-scalar-p5.c",
216 "src/math/sigmoid-scalar-lut2048-p1-div.c",
217 "src/math/sigmoid-scalar-lut64-p2-div.c",
218 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800219 "src/q8-avgpool/9p8x-scalar-c1.c",
220 "src/q8-avgpool/9x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700221 "src/q8-dwconv/up1x9-scalar.c",
222 "src/q8-gavgpool/mp7p7q-scalar.c",
223 "src/q8-gavgpool/up7-scalar.c",
224 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800225 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700226 "src/q8-vadd/scalar.c",
227 "src/u8-clamp/scalar.c",
228 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800229 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700230 "src/u8-rmax/scalar.c",
231 "src/x32-packx/x2-scalar.c",
232 "src/x32-packx/x3-scalar.c",
233 "src/x32-packx/x4-scalar.c",
234 "src/x32-pad/x2-scalar.c",
235 "src/x32-unpool/scalar.c",
236 "src/x32-zip/x2-scalar.c",
237 "src/x32-zip/x3-scalar.c",
238 "src/x32-zip/x4-scalar.c",
239 "src/x32-zip/xm-scalar.c",
240 "src/x8-lut/scalar.c",
241 "src/x8-zip/x2-scalar.c",
242 "src/x8-zip/x3-scalar.c",
243 "src/x8-zip/x4-scalar.c",
244 "src/x8-zip/xm-scalar.c",
245]
246
Marat Dukhan436ebe62019-12-04 15:10:12 -0800247WASM_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800248 "src/f32-avgpool/9p8x-wasm-c1.c",
249 "src/f32-avgpool/9x-wasm-c1.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800250 "src/f32-clamp/wasm.c",
251 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
252 "src/f32-dwconv/gen/up1x25-wasm.c",
253 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
254 "src/f32-dwconv/gen/up1x4-wasm.c",
255 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
256 "src/f32-dwconv/gen/up1x9-wasm.c",
257 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
258 "src/f32-dwconv/gen/up2x25-wasm.c",
259 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
260 "src/f32-dwconv/gen/up2x4-wasm.c",
261 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
262 "src/f32-dwconv/gen/up2x9-wasm.c",
263 "src/f32-gavgpool/mp7p7q-wasm.c",
264 "src/f32-gavgpool/up7-wasm.c",
265 "src/f32-gemm/gen-inc/1x4-wasm.c",
266 "src/f32-gemm/gen-inc/2x4-wasm.c",
267 "src/f32-gemm/gen-inc/4x4-wasm.c",
268 "src/f32-gemm/gen/1x4-wasm.c",
269 "src/f32-gemm/gen/2x4-wasm.c",
270 "src/f32-gemm/gen/4x2-wasm.c",
271 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800272 "src/f32-hswish/gen/wasm-x1.c",
273 "src/f32-hswish/gen/wasm-x2.c",
274 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800275 "src/f32-igemm/gen/1x4-wasm.c",
276 "src/f32-igemm/gen/2x4-wasm.c",
277 "src/f32-igemm/gen/4x2-wasm.c",
278 "src/f32-igemm/gen/4x4-wasm.c",
279 "src/f32-maxpool/9p8x-wasm-c1.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800280 "src/f32-pavgpool/9p8x-wasm-c1.c",
281 "src/f32-pavgpool/9x-wasm-c1.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800282 "src/f32-vbinary/gen/vadd-wasm-x1.c",
283 "src/f32-vbinary/gen/vadd-wasm-x2.c",
284 "src/f32-vbinary/gen/vadd-wasm-x4.c",
285 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
286 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
287 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800288 "src/f32-vbinary/gen/vdiv-wasm-x1.c",
289 "src/f32-vbinary/gen/vdiv-wasm-x2.c",
290 "src/f32-vbinary/gen/vdiv-wasm-x4.c",
291 "src/f32-vbinary/gen/vdivc-wasm-x1.c",
292 "src/f32-vbinary/gen/vdivc-wasm-x2.c",
293 "src/f32-vbinary/gen/vdivc-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800294 "src/f32-vbinary/gen/vmax-wasm-x1.c",
295 "src/f32-vbinary/gen/vmax-wasm-x2.c",
296 "src/f32-vbinary/gen/vmax-wasm-x4.c",
297 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
298 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
299 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
300 "src/f32-vbinary/gen/vmin-wasm-x1.c",
301 "src/f32-vbinary/gen/vmin-wasm-x2.c",
302 "src/f32-vbinary/gen/vmin-wasm-x4.c",
303 "src/f32-vbinary/gen/vminc-wasm-x1.c",
304 "src/f32-vbinary/gen/vminc-wasm-x2.c",
305 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800306 "src/f32-vbinary/gen/vmul-wasm-x1.c",
307 "src/f32-vbinary/gen/vmul-wasm-x2.c",
308 "src/f32-vbinary/gen/vmul-wasm-x4.c",
309 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
310 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
311 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800312 "src/f32-vbinary/gen/vrdivc-wasm-x1.c",
313 "src/f32-vbinary/gen/vrdivc-wasm-x2.c",
314 "src/f32-vbinary/gen/vrdivc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800315 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
316 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
317 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
318 "src/f32-vbinary/gen/vsub-wasm-x1.c",
319 "src/f32-vbinary/gen/vsub-wasm-x2.c",
320 "src/f32-vbinary/gen/vsub-wasm-x4.c",
321 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
322 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
323 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
324 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
325 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
326 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
327]
328
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800329PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800330 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800331 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800332 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800333 "src/f32-avgpool/9p8x-psimd-c4.c",
334 "src/f32-avgpool/9x-psimd-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800335 "src/f32-bilinear/gen/psimd-c4.c",
336 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700337 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800338 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
339 "src/f32-dwconv/gen/up4x25-psimd.c",
340 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
341 "src/f32-dwconv/gen/up4x4-psimd.c",
342 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
343 "src/f32-dwconv/gen/up4x9-psimd.c",
344 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
345 "src/f32-dwconv/gen/up8x25-psimd.c",
346 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
347 "src/f32-dwconv/gen/up8x4-psimd.c",
348 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
349 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700350 "src/f32-gavgpool/mp7p7q-psimd.c",
351 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800352 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
353 "src/f32-gemm/gen/1x8-psimd-splat.c",
354 "src/f32-gemm/gen/1x8s4-psimd.c",
Marat Dukhanb00004d2020-02-13 09:28:35 -0800355 "src/f32-gemm/gen/4x2c4-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800356 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
357 "src/f32-gemm/gen/4x8-psimd-splat.c",
358 "src/f32-gemm/gen/4x8s4-psimd.c",
359 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
360 "src/f32-gemm/gen/6x8-psimd-splat.c",
361 "src/f32-gemm/gen/6x8s4-psimd.c",
362 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
363 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
364 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
365 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
366 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
367 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
368 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
369 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
370 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800371 "src/f32-hswish/gen/psimd-x4.c",
372 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800373 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
374 "src/f32-igemm/gen/1x8-psimd-splat.c",
375 "src/f32-igemm/gen/1x8s4-psimd.c",
376 "src/f32-igemm/gen/4x2c4-psimd.c",
377 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
378 "src/f32-igemm/gen/4x8-psimd-splat.c",
379 "src/f32-igemm/gen/4x8s4-psimd.c",
380 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
381 "src/f32-igemm/gen/6x8-psimd-splat.c",
382 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800383 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800384 "src/f32-pavgpool/9p8x-psimd-c4.c",
385 "src/f32-pavgpool/9x-psimd-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800386 "src/f32-ppmm/gen/4x8-psimd.c",
387 "src/f32-prelu/gen/psimd-2x4.c",
388 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800389 "src/f32-rmax/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800390 "src/f32-vbinary/gen/vadd-psimd-x4.c",
391 "src/f32-vbinary/gen/vadd-psimd-x8.c",
392 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
393 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800394 "src/f32-vbinary/gen/vdiv-psimd-x4.c",
395 "src/f32-vbinary/gen/vdiv-psimd-x8.c",
396 "src/f32-vbinary/gen/vdivc-psimd-x4.c",
397 "src/f32-vbinary/gen/vdivc-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800398 "src/f32-vbinary/gen/vmax-psimd-x4.c",
399 "src/f32-vbinary/gen/vmax-psimd-x8.c",
400 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
401 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
402 "src/f32-vbinary/gen/vmin-psimd-x4.c",
403 "src/f32-vbinary/gen/vmin-psimd-x8.c",
404 "src/f32-vbinary/gen/vminc-psimd-x4.c",
405 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800406 "src/f32-vbinary/gen/vmul-psimd-x4.c",
407 "src/f32-vbinary/gen/vmul-psimd-x8.c",
408 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
409 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800410 "src/f32-vbinary/gen/vrdivc-psimd-x4.c",
411 "src/f32-vbinary/gen/vrdivc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800412 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
413 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
414 "src/f32-vbinary/gen/vsub-psimd-x4.c",
415 "src/f32-vbinary/gen/vsub-psimd-x8.c",
416 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
417 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
418 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
419 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700420 "src/x32-packx/x4-psimd.c",
421 "src/x32-pad/x2-psimd.c",
422 "src/x32-unpool/psimd.c",
423 "src/x32-zip/x2-psimd.c",
424 "src/x32-zip/x3-psimd.c",
425 "src/x32-zip/x4-psimd.c",
426 "src/x32-zip/xm-psimd.c",
427]
428
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800429PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800430 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
431 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
432 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
433 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
434 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
435 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
436 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
437 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
438 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
439 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
440 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
441 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800442 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
443 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
444 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
445 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
446 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
447 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
448 "src/math/sigmoid-psimd-p5-div.c",
449]
450
Marat Dukhan08c4a432019-10-03 09:29:21 -0700451# ISA-specific micro-kernels
452NEON_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800453 "src/f32-avgpool/9p8x-neon-c4.c",
454 "src/f32-avgpool/9x-neon-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800455 "src/f32-bilinear/gen/neon-c4.c",
456 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700457 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800458 "src/f32-dwconv/gen/up4x9-neon.c",
459 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
460 "src/f32-dwconv/gen/up8x9-neon.c",
461 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700462 "src/f32-gavgpool-spchw/neon-x4.c",
463 "src/f32-gavgpool/mp7p7q-neon.c",
464 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800465 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
466 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
467 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
468 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
469 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
470 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800471 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800472 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
473 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
474 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
475 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800476 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800477 "src/f32-gemm/gen/1x8s4-neon.c",
478 "src/f32-gemm/gen/4x8s4-neon.c",
479 "src/f32-gemm/gen/6x8s4-neon.c",
480 "src/f32-gemm/gen/8x8s4-neon.c",
481 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
482 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
483 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
484 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
485 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800486 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800487 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
488 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
489 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
490 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800491 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800492 "src/f32-gemm/gen-inc/1x8s4-neon.c",
493 "src/f32-gemm/gen-inc/4x8s4-neon.c",
494 "src/f32-gemm/gen-inc/6x8s4-neon.c",
495 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800496 "src/f32-hswish/gen/neon-x4.c",
497 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800498 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
499 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
500 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
501 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
502 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
503 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800504 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800505 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
506 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
507 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
508 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800509 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800510 "src/f32-igemm/gen/1x8s4-neon.c",
511 "src/f32-igemm/gen/4x8s4-neon.c",
512 "src/f32-igemm/gen/6x8s4-neon.c",
513 "src/f32-igemm/gen/8x8s4-neon.c",
Frank Barchardf092a4a2020-03-03 14:22:46 -0800514 "src/f32-maxpool/9p8x-neon-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800515 "src/f32-pavgpool/9p8x-neon-c4.c",
516 "src/f32-pavgpool/9x-neon-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800517 "src/f32-ppmm/gen/4x8-neon.c",
518 "src/f32-ppmm/gen/8x8-neon.c",
519 "src/f32-prelu/gen/neon-2x4.c",
520 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800521 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
522 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
523 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
524 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
525 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
526 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
527 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
528 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
529 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
530 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
531 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
532 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
533 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
534 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
535 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
536 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
537 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
538 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
539 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
540 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
541 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
542 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
543 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
544 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700545 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800546 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800547 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
548 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
549 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
550 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
551 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
552 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
553 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
554 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
555 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
556 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
557 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
558 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
559 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
560 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
561 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
562 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
563 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
564 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800565 "src/f32-vbinary/gen/vadd-neon-x4.c",
566 "src/f32-vbinary/gen/vadd-neon-x8.c",
567 "src/f32-vbinary/gen/vaddc-neon-x4.c",
568 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800569 "src/f32-vbinary/gen/vmax-neon-x4.c",
570 "src/f32-vbinary/gen/vmax-neon-x8.c",
571 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
572 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
573 "src/f32-vbinary/gen/vmin-neon-x4.c",
574 "src/f32-vbinary/gen/vmin-neon-x8.c",
575 "src/f32-vbinary/gen/vminc-neon-x4.c",
576 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800577 "src/f32-vbinary/gen/vmul-neon-x4.c",
578 "src/f32-vbinary/gen/vmul-neon-x8.c",
579 "src/f32-vbinary/gen/vmulc-neon-x4.c",
580 "src/f32-vbinary/gen/vmulc-neon-x8.c",
581 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
582 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
583 "src/f32-vbinary/gen/vsub-neon-x4.c",
584 "src/f32-vbinary/gen/vsub-neon-x8.c",
585 "src/f32-vbinary/gen/vsubc-neon-x4.c",
586 "src/f32-vbinary/gen/vsubc-neon-x8.c",
587 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
588 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800589 "src/q8-avgpool/9p8x-neon-c8.c",
590 "src/q8-avgpool/9x-neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700591 "src/q8-dwconv/up8x9-neon.c",
592 "src/q8-gavgpool/mp7p7q-neon.c",
593 "src/q8-gavgpool/up7-neon.c",
594 "src/q8-gemm/4x8-neon.c",
595 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800596 "src/q8-igemm/4x8-neon.c",
597 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700598 "src/q8-vadd/neon.c",
599 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800600 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700601 "src/u8-rmax/neon.c",
602 "src/x32-packx/x4-neon-st4.c",
603 "src/x32-pad/x2-neon.c",
604 "src/x32-zip/x2-neon.c",
605 "src/x32-zip/x3-neon.c",
606 "src/x32-zip/x4-neon.c",
607 "src/x32-zip/xm-neon.c",
608 "src/x8-zip/x2-neon.c",
609 "src/x8-zip/x3-neon.c",
610 "src/x8-zip/x4-neon.c",
611 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800612 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800613 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
614 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
615 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
616 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
617 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
618 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700619]
620
621NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800622 "src/f32-bilinear/gen/neonfma-c4.c",
623 "src/f32-bilinear/gen/neonfma-c8.c",
624 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
625 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
626 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
627 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800628 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800629 "src/f32-igemm/gen/1x8s4-neonfma.c",
630 "src/f32-igemm/gen/4x8s4-neonfma.c",
631 "src/f32-igemm/gen/6x8s4-neonfma.c",
632 "src/f32-igemm/gen/8x8s4-neonfma.c",
633 "src/f32-dwconv/gen/up4x9-neonfma.c",
634 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
635 "src/f32-dwconv/gen/up8x9-neonfma.c",
636 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
637 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
638 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
639 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
640 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800641 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800642 "src/f32-gemm/gen/1x8s4-neonfma.c",
643 "src/f32-gemm/gen/4x8s4-neonfma.c",
644 "src/f32-gemm/gen/6x8s4-neonfma.c",
645 "src/f32-gemm/gen/8x8s4-neonfma.c",
646 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
647 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
648 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
649 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800650 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800651 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
652 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
653 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
654 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800655 "src/f32-hswish/gen/neonfma-x4.c",
656 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800657 "src/f32-ppmm/gen/4x8-neonfma.c",
658 "src/f32-ppmm/gen/8x8-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800659 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
660 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
661 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
662 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
663 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
664 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
665 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
666 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
667 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
668 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
669 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
670 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
671 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
672 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
673 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
674 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
675 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
676 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
677 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
678 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
679 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
680 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
681 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
682 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800683 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
684 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
685 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
686 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
687 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
688 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
689 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
690 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
691 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
692 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
693 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
694 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
695 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
696 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
697 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
698 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
699 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
700 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
701 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
702 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
703 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
704 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
705 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
706 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
707 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
708 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
709 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
710 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
711 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
712 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
713 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
714 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
715 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
716 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
717 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
718 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
719 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
720 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
721 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
722 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
723 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
724 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
725 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
726 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
727 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
728 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
729 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
730 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
731 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
732 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
733 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
734 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
735 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
736 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800737 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
738 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800739 "src/math/exp-neonfma-lut64-p2.c",
740 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800741 "src/math/expminus-neonfma-lut2048-p1.c",
742 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800743 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800744 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
745 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
746 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
747 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
748 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
749 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
750 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
751 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
752 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
753 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
754 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
755 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
756 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
757 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
758 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
759 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
760 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
761 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700762]
763
764AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800765 "src/f32-vbinary/gen/vdiv-neon-x4.c",
766 "src/f32-vbinary/gen/vdiv-neon-x8.c",
767 "src/f32-vbinary/gen/vdivc-neon-x4.c",
768 "src/f32-vbinary/gen/vdivc-neon-x8.c",
769 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
770 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800771 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
772 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
773 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
774 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
775 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
776 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800777 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800778 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
779 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
780 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
781 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
782 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800783 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800784 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
785 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
786 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
787 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
788 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
789 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800790 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700791 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
792 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
793 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
794 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
795 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
796 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
797 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800798 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
813 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
814 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800816 "src/f32-spmm/gen/12x1-neonfma.c",
817 "src/f32-spmm/gen/12x2-neonfma.c",
818 "src/f32-spmm/gen/12x4-neonfma.c",
819 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
820 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
821 "src/f32-spmm/gen/16x1-neonfma.c",
822 "src/f32-spmm/gen/16x2-neonfma.c",
823 "src/f32-spmm/gen/16x4-neonfma.c",
824 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
825 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
826 "src/f32-spmm/gen/4x1-neonfma.c",
827 "src/f32-spmm/gen/4x2-neonfma.c",
828 "src/f32-spmm/gen/4x4-neonfma.c",
829 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
830 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
831 "src/f32-spmm/gen/8x1-neonfma.c",
832 "src/f32-spmm/gen/8x2-neonfma.c",
833 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800834 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
835 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
836 "src/math/sigmoid-neonfma-rr1-p5-div.c",
837 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
838 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
839 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700840]
841
842AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800843 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
844 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
845 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhanbdb56f52020-02-05 21:42:49 -0800846 "src/f16-spmm/gen/8x1-neonfp16arith.c",
847 "src/f16-spmm/gen/8x1-neonfp16arith-unroll2.c",
848 "src/f16-spmm/gen/16x1-neonfp16arith.c",
849 "src/f16-spmm/gen/16x1-neonfp16arith-unroll2.c",
850 "src/f16-spmm/gen/24x1-neonfp16arith.c",
851 "src/f16-spmm/gen/24x1-neonfp16arith-unroll2.c",
852 "src/f16-spmm/gen/32x1-neonfp16arith.c",
853 "src/f16-spmm/gen/32x1-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700854]
855
856SSE_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800857 "src/f32-avgpool/9p8x-sse-c4.c",
858 "src/f32-avgpool/9x-sse-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800859 "src/f32-bilinear/gen/sse-c4.c",
860 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700861 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800862 "src/f32-dwconv-spchw/3x3p1-sse.c",
863 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800864 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
865 "src/f32-dwconv/gen/up4x25-sse.c",
866 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
867 "src/f32-dwconv/gen/up4x4-sse.c",
868 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
869 "src/f32-dwconv/gen/up4x9-sse.c",
870 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
871 "src/f32-dwconv/gen/up8x25-sse.c",
872 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
873 "src/f32-dwconv/gen/up8x4-sse.c",
874 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
875 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700876 "src/f32-gavgpool-spchw/sse-x4.c",
877 "src/f32-gavgpool/mp7p7q-sse.c",
878 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800879 "src/f32-gemm/gen/1x8-sse-dup.c",
880 "src/f32-gemm/gen/1x8-sse-load1.c",
881 "src/f32-gemm/gen/1x8s4-sse.c",
Marat Dukhanb00004d2020-02-13 09:28:35 -0800882 "src/f32-gemm/gen/4x2c4-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800883 "src/f32-gemm/gen/4x8-sse-dup.c",
884 "src/f32-gemm/gen/4x8-sse-load1.c",
885 "src/f32-gemm/gen/4x8s4-sse.c",
886 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
887 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
888 "src/f32-gemm/gen-inc/1x8s4-sse.c",
889 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
890 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
891 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800892 "src/f32-hswish/gen/sse-x4.c",
893 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800894 "src/f32-igemm/gen/1x8-sse-dup.c",
895 "src/f32-igemm/gen/1x8-sse-load1.c",
896 "src/f32-igemm/gen/1x8s4-sse.c",
897 "src/f32-igemm/gen/4x2c4-sse.c",
898 "src/f32-igemm/gen/4x8-sse-dup.c",
899 "src/f32-igemm/gen/4x8-sse-load1.c",
900 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800901 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800902 "src/f32-pavgpool/9p8x-sse-c4.c",
903 "src/f32-pavgpool/9x-sse-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800904 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700905 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800906 "src/f32-spmm/gen/4x1-sse.c",
907 "src/f32-spmm/gen/8x1-sse.c",
908 "src/f32-vbinary/gen/vadd-sse-x4.c",
909 "src/f32-vbinary/gen/vadd-sse-x8.c",
910 "src/f32-vbinary/gen/vaddc-sse-x4.c",
911 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800912 "src/f32-vbinary/gen/vdiv-sse-x4.c",
913 "src/f32-vbinary/gen/vdiv-sse-x8.c",
914 "src/f32-vbinary/gen/vdivc-sse-x4.c",
915 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800916 "src/f32-vbinary/gen/vmax-sse-x4.c",
917 "src/f32-vbinary/gen/vmax-sse-x8.c",
918 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
919 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
920 "src/f32-vbinary/gen/vmin-sse-x4.c",
921 "src/f32-vbinary/gen/vmin-sse-x8.c",
922 "src/f32-vbinary/gen/vminc-sse-x4.c",
923 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800924 "src/f32-vbinary/gen/vmul-sse-x4.c",
925 "src/f32-vbinary/gen/vmul-sse-x8.c",
926 "src/f32-vbinary/gen/vmulc-sse-x4.c",
927 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800928 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
929 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800930 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
931 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
932 "src/f32-vbinary/gen/vsub-sse-x4.c",
933 "src/f32-vbinary/gen/vsub-sse-x8.c",
934 "src/f32-vbinary/gen/vsubc-sse-x4.c",
935 "src/f32-vbinary/gen/vsubc-sse-x8.c",
936 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
937 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700938 "src/x32-packx/x4-sse.c",
939]
940
941SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800942 "src/f32-argmaxpool/9p8x-sse2-c4.c",
943 "src/f32-argmaxpool/4x-sse2-c4.c",
944 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800945 "src/f32-prelu/gen/sse2-2x4.c",
946 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800947 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
948 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
949 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
950 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
951 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
952 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
953 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
954 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
955 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
956 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
957 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
958 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800959 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800960 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800961 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800962 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800963 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
964 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800965 "src/q8-avgpool/9p8x-sse2-c8.c",
966 "src/q8-avgpool/9x-sse2-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700967 "src/q8-igemm/4x4c2-sse2.c",
968 "src/q8-dwconv/up8x9-sse2.c",
969 "src/q8-gavgpool/mp7p7q-sse2.c",
970 "src/q8-gavgpool/up7-sse2.c",
971 "src/q8-gemm/2x4c8-sse2.c",
972 "src/q8-gemm/4x4c2-sse2.c",
973 "src/q8-vadd/sse2.c",
974 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800975 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700976 "src/u8-rmax/sse2.c",
977 "src/x32-pad/x2-sse2.c",
978 "src/x32-zip/x2-sse2.c",
979 "src/x32-zip/x3-sse2.c",
980 "src/x32-zip/x4-sse2.c",
981 "src/x32-zip/xm-sse2.c",
982 "src/x8-zip/x2-sse2.c",
983 "src/x8-zip/x3-sse2.c",
984 "src/x8-zip/x4-sse2.c",
985 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800986 "src/math/exp-sse2-p5.c",
987 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800988 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700989]
990
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800991SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800992 "src/f32-prelu/gen/sse41-2x4.c",
993 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800994 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800995 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800996 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800997 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800998 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
999 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001000]
1001
Marat Dukhan08c4a432019-10-03 09:29:21 -07001002AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -08001003 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001004 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
1005 "src/f32-dwconv/gen/up16x4-avx.c",
1006 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
1007 "src/f32-dwconv/gen/up8x4-avx.c",
1008 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
1009 "src/f32-dwconv/gen/up16x9-avx.c",
1010 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
1011 "src/f32-dwconv/gen/up8x9-avx.c",
1012 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
1013 "src/f32-dwconv/gen/up16x25-avx.c",
1014 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
1015 "src/f32-dwconv/gen/up8x25-avx.c",
1016 "src/f32-gemm/gen/1x8-avx-broadcast.c",
1017 "src/f32-gemm/gen/4x8-avx-broadcast.c",
1018 "src/f32-gemm/gen/5x8-avx-broadcast.c",
1019 "src/f32-gemm/gen/6x8-avx-broadcast.c",
1020 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001021 "src/f32-gemm/gen/1x16-avx-broadcast.c",
1022 "src/f32-gemm/gen/3x16-avx-broadcast.c",
1023 "src/f32-gemm/gen/4x16-avx-broadcast.c",
1024 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001025 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
1026 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
1027 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
1028 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
1029 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001030 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
1031 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
1032 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
1033 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001034 "src/f32-hswish/gen/avx-x8.c",
1035 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001036 "src/f32-igemm/gen/1x8-avx-broadcast.c",
1037 "src/f32-igemm/gen/4x8-avx-broadcast.c",
1038 "src/f32-igemm/gen/5x8-avx-broadcast.c",
1039 "src/f32-igemm/gen/6x8-avx-broadcast.c",
1040 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001041 "src/f32-igemm/gen/1x16-avx-broadcast.c",
1042 "src/f32-igemm/gen/3x16-avx-broadcast.c",
1043 "src/f32-igemm/gen/4x16-avx-broadcast.c",
1044 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001045 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001046 "src/f32-vbinary/gen/vadd-avx-x8.c",
1047 "src/f32-vbinary/gen/vadd-avx-x16.c",
1048 "src/f32-vbinary/gen/vaddc-avx-x8.c",
1049 "src/f32-vbinary/gen/vaddc-avx-x16.c",
1050 "src/f32-vbinary/gen/vdiv-avx-x8.c",
1051 "src/f32-vbinary/gen/vdiv-avx-x16.c",
1052 "src/f32-vbinary/gen/vdivc-avx-x8.c",
1053 "src/f32-vbinary/gen/vdivc-avx-x16.c",
1054 "src/f32-vbinary/gen/vmax-avx-x8.c",
1055 "src/f32-vbinary/gen/vmax-avx-x16.c",
1056 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1057 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1058 "src/f32-vbinary/gen/vmin-avx-x8.c",
1059 "src/f32-vbinary/gen/vmin-avx-x16.c",
1060 "src/f32-vbinary/gen/vminc-avx-x8.c",
1061 "src/f32-vbinary/gen/vminc-avx-x16.c",
1062 "src/f32-vbinary/gen/vmul-avx-x8.c",
1063 "src/f32-vbinary/gen/vmul-avx-x16.c",
1064 "src/f32-vbinary/gen/vmulc-avx-x8.c",
1065 "src/f32-vbinary/gen/vmulc-avx-x16.c",
1066 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
1067 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
1068 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
1069 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
1070 "src/f32-vbinary/gen/vsub-avx-x8.c",
1071 "src/f32-vbinary/gen/vsub-avx-x16.c",
1072 "src/f32-vbinary/gen/vsubc-avx-x8.c",
1073 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001074 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001075]
1076
Marat Dukhanfda12b82019-11-21 12:27:59 -08001077FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001078 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
1079 "src/f32-dwconv/gen/up16x4-fma3.c",
1080 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
1081 "src/f32-dwconv/gen/up8x4-fma3.c",
1082 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
1083 "src/f32-dwconv/gen/up16x9-fma3.c",
1084 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
1085 "src/f32-dwconv/gen/up8x9-fma3.c",
1086 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
1087 "src/f32-dwconv/gen/up16x25-fma3.c",
1088 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
1089 "src/f32-dwconv/gen/up8x25-fma3.c",
1090 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
1091 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
1092 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
1093 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
1094 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
1095 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001096 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
1097 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
1098 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
1099 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001100 "src/f32-gemm/gen/1x16s4-fma3-broadcast.c",
1101 "src/f32-gemm/gen/3x16s4-fma3-broadcast.c",
1102 "src/f32-gemm/gen/4x16s4-fma3-broadcast.c",
1103 "src/f32-gemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001104 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
1105 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
1106 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
1107 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
1108 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
1109 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001110 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
1111 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
1112 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
1113 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001114 "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c",
1115 "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c",
1116 "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c",
1117 "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001118 "src/f32-hswish/gen/fma3-x8.c",
1119 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001120 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
1121 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
1122 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
1123 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
1124 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
1125 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001126 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
1127 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
1128 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
1129 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001130 "src/f32-igemm/gen/1x16s4-fma3-broadcast.c",
1131 "src/f32-igemm/gen/3x16s4-fma3-broadcast.c",
1132 "src/f32-igemm/gen/4x16s4-fma3-broadcast.c",
1133 "src/f32-igemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001134]
1135
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001136AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001137 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1138 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1139 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1140 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1141 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1142 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1143 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1144 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1145 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1146 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1147 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1148 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1149 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1150 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1151 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1152 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1153 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1154 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1155 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1156 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1157 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1158 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1159 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1160 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1161 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1162 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1163 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1164 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1165 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1166 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1167 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1168 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1169 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1170 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1171 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1172 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001173 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1174 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1175 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1176 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1177 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1178 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1179 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1180 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1181 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1182 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1183 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1184 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1185 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1186 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1187 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1188 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1189 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1190 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1191 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1192 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1193 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1194 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1195 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1196 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1197 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1198 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1199 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1200 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1201 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1202 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001203 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1204 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1205 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1206 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1207 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1208 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1209 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1210 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1211 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1212 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1213 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1214 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1215 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1216 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1217 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1218 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1219 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1220 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1221 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1222 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1223 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1224 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1225 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1226 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001227 "src/math/exp-avx2-p5.c",
1228 "src/math/exp-avx2-perm-p3.c",
1229 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001230 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001231 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001232 "src/math/sigmoid-avx2-rr2-p5-div.c",
1233 "src/math/sigmoid-avx2-rr1-p5-div.c",
1234 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1235 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1236 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1237 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001238]
1239
Marat Dukhan08c4a432019-10-03 09:29:21 -07001240AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -08001241 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -08001242 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
1243 "src/f32-dwconv/gen/up32x4-avx512f.c",
1244 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
1245 "src/f32-dwconv/gen/up16x4-avx512f.c",
1246 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
1247 "src/f32-dwconv/gen/up32x9-avx512f.c",
1248 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
1249 "src/f32-dwconv/gen/up16x9-avx512f.c",
1250 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
1251 "src/f32-dwconv/gen/up32x25-avx512f.c",
1252 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
1253 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001254 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
1255 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
1256 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
1257 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
1258 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
1259 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
1260 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
1261 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
1262 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
1263 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
1264 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
1265 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001266 "src/f32-hswish/gen/avx512f-x16.c",
1267 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001268 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
1269 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
1270 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
1271 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
1272 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
1273 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001274 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1275 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1276 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1277 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1278 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1279 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1280 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1281 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1282 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1283 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1284 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1285 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1286 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1287 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1288 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1289 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1290 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1291 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1292 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1293 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1294 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1295 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1296 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1297 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1298 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1299 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1300 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1301 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1302 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1303 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1304 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1305 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1306 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1307 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1308 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1309 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001310 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001311 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1312 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1313 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1314 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1315 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1316 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1317 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1318 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1319 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1320 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1321 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1322 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1323 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1324 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1325 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1326 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1327 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1328 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1329 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1330 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1331 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1332 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1333 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1334 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1335 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1336 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1337 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1338 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001339 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001340 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1341 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1342 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1343 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1344 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1345 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1346 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1347 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1348 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1349 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1350 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1351 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1352 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1353 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1354 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1355 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1356 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1357 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1358 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1359 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1360 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1361 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1362 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1363 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001364 "src/math/exp-avx512f-p5-scalef.c",
1365 "src/math/exp-avx512f-p5.c",
1366 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001367 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001368 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001369]
1370
1371AARCH32_ASM_UKERNELS = [
1372 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard13916042019-12-11 10:56:34 -08001373 "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S",
Frank Barchardabf81542019-12-13 16:18:30 -08001374 "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S",
1375 "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001376 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Frank Barcharddc38f072020-02-10 13:21:42 -08001377 "src/f32-igemm/4x8-aarch32-neon-ld64.S",
Frank Barchard90ce7892020-02-10 23:35:45 -08001378 "src/f32-igemm/gen/4x8-aarch32-neon-cortex-a75.S",
1379 "src/f32-igemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchardc87a8fd2020-02-12 13:02:52 -08001380 "src/f32-igemm/4x8-aarch32-neon-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001381]
1382
1383AARCH64_ASM_UKERNELS = [
1384 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1385 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001386 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1387 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1388 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1389 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1390 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1391 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1392 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1393 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1394 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1395 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001396 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001397 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1398 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001399 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001400 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001401 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1402 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1403 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1404 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1405 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1406 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1407 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1408 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1409 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1410 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1411 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1412 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1413 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001414 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001415 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1416 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001417 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001418 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001419 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1420 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1421 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001422 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001423 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001424 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1425 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001426 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001427 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001428 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1429 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1430 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
1431 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001432 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001433 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001434 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1435 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001436]
1437
1438INTERNAL_MICROKERNEL_HDRS = [
1439 "src/xnnpack/argmaxpool.h",
1440 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001441 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001442 "src/xnnpack/clamp.h",
1443 "src/xnnpack/common.h",
1444 "src/xnnpack/conv.h",
1445 "src/xnnpack/dwconv.h",
1446 "src/xnnpack/gavgpool.h",
1447 "src/xnnpack/gemm.h",
1448 "src/xnnpack/hswish.h",
1449 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001450 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001451 "src/xnnpack/lut.h",
1452 "src/xnnpack/math.h",
1453 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001454 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001455 "src/xnnpack/packx.h",
1456 "src/xnnpack/pad.h",
1457 "src/xnnpack/params.h",
1458 "src/xnnpack/pavgpool.h",
1459 "src/xnnpack/ppmm.h",
1460 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001461 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001462 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001463 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001464 "src/xnnpack/rmax.h",
1465 "src/xnnpack/scalar-utils.h",
1466 "src/xnnpack/spmm.h",
1467 "src/xnnpack/unpool.h",
1468 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001469 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001470 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001471 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001472 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001473 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001474 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001475 "src/xnnpack/zip.h",
1476]
1477
1478INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1479 "include/xnnpack.h",
1480 "src/xnnpack/allocator.h",
1481 "src/xnnpack/compute.h",
1482 "src/xnnpack/im2col.h",
1483 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001484 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001485 "src/xnnpack/operator.h",
1486 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001487 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001488 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001489 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001490 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001491]
1492
1493ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1494 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001495]
1496
1497MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001498 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001499 "include/xnnpack.h",
1500]
1501
1502MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1503 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001504 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001505 "src/xnnpack/requantization.h",
1506 "include/xnnpack.h",
1507]
1508
1509OPERATOR_TEST_PARAMS_HDRS = [
1510 "src/xnnpack/params.h",
1511 "src/xnnpack/common.h",
1512]
1513
1514WEIGHTS_PACK_HDRS = [
1515 "src/xnnpack/pack.h",
1516 "src/xnnpack/operator.h",
1517 "src/xnnpack/compute.h",
1518]
1519
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001520LOGGING_COPTS = select({
1521 # No logging in optimized mode
1522 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1523 # Full logging in debug mode
1524 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1525 # Error-only logging in default (fastbuild) mode
1526 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1527})
1528
1529LOGGING_HDRS = [
1530 "src/xnnpack/log.h",
1531]
1532
Marat Dukhan08c4a432019-10-03 09:29:21 -07001533xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001534 name = "tables",
1535 srcs = TABLE_SRCS,
1536 hdrs = INTERNAL_HDRS,
1537 copts = xnnpack_std_copts(),
1538)
1539
1540xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001541 name = "scalar_ukernels",
1542 srcs = SCALAR_UKERNELS,
1543 hdrs = INTERNAL_HDRS,
1544 aarch32_copts = ["-marm"],
1545 copts = xnnpack_std_copts(),
1546 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001547 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001548 "@FP16",
1549 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001550 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001551 ],
1552)
1553
1554xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001555 name = "wasm_ukernels",
1556 hdrs = INTERNAL_HDRS,
1557 copts = xnnpack_std_copts(),
1558 wasm_srcs = WASM_UKERNELS,
1559 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001560 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001561 "@FP16",
1562 "@FXdiv",
1563 "@pthreadpool",
1564 ],
1565)
1566
1567xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001568 name = "psimd_fastmath_ukernels",
1569 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001570 hdrs = INTERNAL_HDRS,
1571 aarch32_copts = [
1572 "-marm",
1573 "-mfpu=neon",
1574 ],
1575 copts = xnnpack_std_copts(),
1576 optimized_copts = [
1577 "-O3",
1578 "-ffast-math",
1579 ],
1580 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001581 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "@FP16",
1583 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001584 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001585 ],
1586)
1587
1588xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001589 name = "psimd_accmath_ukernels",
1590 srcs = PSIMD_ACCMATH_UKERNELS,
1591 hdrs = INTERNAL_HDRS,
1592 aarch32_copts = [
1593 "-marm",
1594 "-mfpu=neon",
1595 ],
1596 copts = xnnpack_std_copts(),
1597 optimized_copts = [
1598 "-O3",
1599 ],
1600 deps = [
1601 ":tables",
1602 "@FP16",
1603 "@psimd",
1604 "@pthreadpool",
1605 ],
1606)
1607
1608xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001609 name = "neon_ukernels",
1610 hdrs = INTERNAL_HDRS,
1611 aarch32_copts = [
1612 "-marm",
1613 "-mfpu=neon",
1614 ],
1615 aarch32_srcs = NEON_UKERNELS,
1616 aarch64_srcs = NEON_UKERNELS,
1617 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001618 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001619 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001620 "@FP16",
Frank Barchardf092a4a2020-03-03 14:22:46 -08001621 "@psimd", # remove me
Marat Dukhan04f03be2019-11-19 12:36:47 -08001622 "@pthreadpool",
1623 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001624)
1625
1626xnnpack_cc_library(
1627 name = "neonfma_ukernels",
1628 hdrs = INTERNAL_HDRS,
1629 aarch32_copts = [
1630 "-marm",
1631 "-mfpu=neon-vfpv4",
1632 ],
1633 aarch32_srcs = NEONFMA_UKERNELS,
1634 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1635 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001636 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001637 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001638 "@FP16",
1639 "@pthreadpool",
1640 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001641)
1642
1643xnnpack_cc_library(
1644 name = "neonfp16arith_ukernels",
1645 hdrs = INTERNAL_HDRS,
1646 aarch64_copts = ["-march=armv8.2-a+fp16"],
1647 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1648 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001649 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001650 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001651 "@FP16",
1652 "@pthreadpool",
1653 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001654)
1655
1656xnnpack_cc_library(
1657 name = "sse2_ukernels",
1658 hdrs = INTERNAL_HDRS,
1659 copts = xnnpack_std_copts(),
1660 x86_copts = ["-msse2"],
1661 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001662 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001663 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001664 "@FP16",
1665 "@pthreadpool",
1666 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001667)
1668
1669xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001670 name = "sse41_ukernels",
1671 hdrs = INTERNAL_HDRS,
1672 copts = xnnpack_std_copts(),
1673 x86_copts = ["-msse4.1"],
1674 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001675 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001676 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001677 "@FP16",
1678 "@pthreadpool",
1679 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001680)
1681
1682xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001683 name = "avx_ukernels",
1684 hdrs = INTERNAL_HDRS,
1685 copts = xnnpack_std_copts(),
1686 x86_copts = ["-mavx"],
1687 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001688 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001689 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001690 "@FP16",
1691 "@pthreadpool",
1692 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693)
1694
1695xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001696 name = "fma3_ukernels",
1697 hdrs = INTERNAL_HDRS,
1698 copts = xnnpack_std_copts(),
1699 x86_copts = [
1700 "-mfma",
1701 ],
1702 x86_srcs = FMA3_UKERNELS,
1703 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001704 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001705 "@FP16",
1706 "@pthreadpool",
1707 ],
1708)
1709
1710xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001711 name = "avx2_ukernels",
1712 hdrs = INTERNAL_HDRS,
1713 copts = xnnpack_std_copts(),
1714 x86_copts = [
1715 "-mfma",
1716 "-mavx2",
1717 ],
1718 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001719 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001720 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001721 "@FP16",
1722 "@pthreadpool",
1723 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001724)
1725
1726xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001727 name = "avx512f_ukernels",
1728 hdrs = INTERNAL_HDRS,
1729 copts = xnnpack_std_copts(),
1730 x86_copts = ["-mavx512f"],
1731 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001732 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001733 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001734 "@FP16",
1735 "@pthreadpool",
1736 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001737)
1738
1739xnnpack_cc_library(
1740 name = "asm_ukernels",
1741 hdrs = ["src/xnnpack/assembly.h"],
1742 aarch32_srcs = AARCH32_ASM_UKERNELS,
1743 aarch64_srcs = AARCH64_ASM_UKERNELS,
1744)
1745
1746xnnpack_aggregate_library(
1747 name = "ukernels",
1748 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001749 ":psimd_fastmath_ukernels",
1750 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001751 ":neon_ukernels",
1752 ":neonfma_ukernels",
1753 ":asm_ukernels",
1754 ],
1755 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001756 ":psimd_fastmath_ukernels",
1757 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001758 ":neon_ukernels",
1759 ":neonfma_ukernels",
1760 ":neonfp16arith_ukernels",
1761 ":asm_ukernels",
1762 ],
1763 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001764 wasm_deps = [
1765 ":wasm_ukernels",
1766 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001767 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001768 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001769 ":psimd_fastmath_ukernels",
1770 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001771 ],
1772 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001773 ":psimd_fastmath_ukernels",
1774 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001775 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001776 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001778 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001779 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001780 ":avx512f_ukernels",
1781 ],
1782)
1783
1784xnnpack_cc_library(
1785 name = "im2col",
1786 srcs = ["src/im2col.c"],
1787 hdrs = [
1788 "src/xnnpack/common.h",
1789 "src/xnnpack/im2col.h",
1790 ],
1791 copts = xnnpack_std_copts(),
1792)
1793
1794xnnpack_cc_library(
1795 name = "indirection",
1796 srcs = ["src/indirection.c"],
1797 hdrs = INTERNAL_HDRS,
1798 copts = xnnpack_std_copts(),
1799 deps = [
1800 "@FP16",
1801 "@FXdiv",
1802 "@pthreadpool",
1803 ],
1804)
1805
1806xnnpack_cc_library(
1807 name = "operator_run",
1808 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001809 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1810 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001811 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1812 "-Wno-vla",
1813 ],
1814 deps = [
1815 "@FP16",
1816 "@FXdiv",
1817 "@clog",
1818 "@pthreadpool",
1819 ],
1820)
1821
1822cc_library(
1823 name = "enable_assembly",
1824 defines = select({
1825 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1826 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001827 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001828 }),
1829)
1830
Marat Dukhancf056b22019-10-07 10:26:29 -07001831xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001832 name = "operators",
1833 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001834 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001836 ],
1837 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001838 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001839 "-Isrc",
1840 "-Iinclude",
1841 ] + select({
1842 ":debug_build": [],
1843 "//conditions:default": xnnpack_min_size_copts(),
1844 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001845 wasm_srcs = ["src/wasm-stubs.c"],
1846 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001847 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001848 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001849 "@FP16",
1850 "@FXdiv",
1851 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001852 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001853 ],
1854)
1855
1856cc_library(
1857 name = "XNNPACK",
1858 srcs = [
1859 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08001860 "src/runtime.c",
1861 "src/subgraph.c",
1862 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001863 ],
1864 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1865 "-Isrc",
1866 "-Iinclude",
1867 ] + select({
1868 ":debug_build": [],
1869 "//conditions:default": xnnpack_min_size_copts(),
1870 }),
1871 includes = ["include"],
1872 linkstatic = True,
1873 textual_hdrs = ["include/xnnpack.h"],
1874 visibility = xnnpack_visibility(),
1875 deps = [
1876 ":enable_assembly",
1877 ":ukernels",
1878 ":operator_run",
1879 ":operators",
1880 "@clog",
1881 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001882 ] + select({
1883 ":emscripten": [],
1884 "//conditions:default": ["@cpuinfo"],
1885 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001886)
1887
1888cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001889 name = "xnnpack_operators_nhwc_f32",
1890 srcs = [
1891 "src/init.c",
1892 ],
1893 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1894 "-Isrc",
1895 "-Iinclude",
1896 ] + select({
1897 ":debug_build": [],
1898 "//conditions:default": xnnpack_min_size_copts(),
1899 }),
1900 defines = [
1901 "XNN_NO_Q8_OPERATORS",
1902 "XNN_NO_U8_OPERATORS",
1903 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001904 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001905 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001906 includes = ["include"],
1907 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001908 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001909 visibility = xnnpack_visibility(),
1910 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001911 ":enable_assembly",
1912 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001913 ":operator_run",
1914 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001915 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001916 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001917 ] + select({
1918 ":emscripten": [],
1919 "//conditions:default": ["@cpuinfo"],
1920 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001921)
1922
Marat Dukhancf056b22019-10-07 10:26:29 -07001923xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001924 name = "bench_utils",
1925 srcs = ["bench/utils.cc"],
1926 hdrs = ["bench/utils.h"],
1927 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001928 deps = [
1929 "@com_google_benchmark//:benchmark",
1930 "@cpuinfo",
1931 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001932)
1933
Frank Barchard7e955972019-10-11 10:34:25 -07001934######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001935
1936xnnpack_benchmark(
1937 name = "q8_gemm_bench",
1938 srcs = [
1939 "bench/gemm.h",
1940 "bench/q8-gemm.cc",
1941 "src/xnnpack/AlignedAllocator.h",
1942 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1943 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1944 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1945)
1946
1947xnnpack_benchmark(
1948 name = "f16_gemm_bench",
1949 srcs = [
1950 "bench/f16-gemm.cc",
1951 "bench/gemm.h",
1952 "src/xnnpack/AlignedAllocator.h",
1953 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1954 copts = ["-Wno-unused-function"],
1955 deps = MICROKERNEL_BENCHMARK_DEPS,
1956)
1957
1958xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08001959 name = "f16_spmm_bench",
1960 srcs = [
1961 "bench/f16-spmm.cc",
1962 "bench/gemm.h",
1963 "bench/google/gemm.h",
1964 "src/xnnpack/AlignedAllocator.h",
1965 ] + MICROKERNEL_BENCHMARK_HDRS,
1966 copts = ["-Wno-unused-function"],
1967 deps = MICROKERNEL_BENCHMARK_DEPS,
1968)
1969
1970xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001971 name = "f32_igemm_bench",
1972 srcs = [
1973 "bench/f32-igemm.cc",
1974 "bench/conv.h",
1975 "src/xnnpack/AlignedAllocator.h",
1976 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001977 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001978)
1979
1980xnnpack_benchmark(
1981 name = "f32_conv_hwc_bench",
1982 srcs = [
1983 "bench/f32-conv-hwc.cc",
1984 "bench/dconv.h",
1985 "src/xnnpack/AlignedAllocator.h",
1986 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1987 copts = ["-Wno-unused-function"],
1988 deps = MICROKERNEL_BENCHMARK_DEPS,
1989)
1990
1991xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001992 name = "f32_conv_hwc2spchw_bench",
1993 srcs = [
1994 "bench/f32-conv-hwc2spchw.cc",
1995 "bench/dconv.h",
1996 "src/xnnpack/AlignedAllocator.h",
1997 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1998 copts = ["-Wno-unused-function"],
1999 deps = MICROKERNEL_BENCHMARK_DEPS,
2000)
2001
2002xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002003 name = "f32_dwconv_bench",
2004 srcs = [
2005 "bench/f32-dwconv.cc",
2006 "bench/dwconv.h",
2007 "src/xnnpack/AlignedAllocator.h",
2008 ] + MICROKERNEL_BENCHMARK_HDRS,
2009 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2010)
2011
2012xnnpack_benchmark(
2013 name = "f32_dwconv_spchw_bench",
2014 srcs = [
2015 "bench/f32-dwconv-spchw.cc",
2016 "bench/dwconv.h",
2017 "src/xnnpack/AlignedAllocator.h",
2018 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2019 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2020)
2021
2022xnnpack_benchmark(
2023 name = "f32_gemm_bench",
2024 srcs = [
2025 "bench/f32-gemm.cc",
2026 "bench/gemm.h",
2027 "src/xnnpack/AlignedAllocator.h",
2028 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2029 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002030 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002031)
2032
2033xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002034 name = "f32_raddexpminusmax_bench",
2035 srcs = [
2036 "bench/f32-raddexpminusmax.cc",
2037 "src/xnnpack/AlignedAllocator.h",
2038 ] + MICROKERNEL_BENCHMARK_HDRS,
2039 deps = MICROKERNEL_BENCHMARK_DEPS,
2040)
2041
2042xnnpack_benchmark(
2043 name = "f32_raddextexp_bench",
2044 srcs = [
2045 "bench/f32-raddextexp.cc",
2046 "src/xnnpack/AlignedAllocator.h",
2047 ] + MICROKERNEL_BENCHMARK_HDRS,
2048 deps = MICROKERNEL_BENCHMARK_DEPS,
2049)
2050
2051xnnpack_benchmark(
2052 name = "f32_raddstoreexpminusmax_bench",
2053 srcs = [
2054 "bench/f32-raddstoreexpminusmax.cc",
2055 "src/xnnpack/AlignedAllocator.h",
2056 ] + MICROKERNEL_BENCHMARK_HDRS,
2057 deps = MICROKERNEL_BENCHMARK_DEPS,
2058)
2059
2060xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002061 name = "f32_rmax_bench",
2062 srcs = [
2063 "bench/f32-rmax.cc",
2064 "src/xnnpack/AlignedAllocator.h",
2065 ] + MICROKERNEL_BENCHMARK_HDRS,
2066 deps = MICROKERNEL_BENCHMARK_DEPS,
2067)
2068
2069xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002070 name = "f32_sigmoid_bench",
2071 srcs = [
2072 "bench/f32-sigmoid.cc",
2073 "src/xnnpack/AlignedAllocator.h",
2074 ] + MICROKERNEL_BENCHMARK_HDRS,
2075 copts = ["-Wno-unused-function"],
2076 deps = MICROKERNEL_BENCHMARK_DEPS,
2077)
2078
2079xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002080 name = "f32_spmm_bench",
2081 srcs = [
2082 "bench/f32-spmm.cc",
2083 "bench/gemm.h",
2084 "src/xnnpack/AlignedAllocator.h",
2085 ] + MICROKERNEL_BENCHMARK_HDRS,
2086 copts = ["-Wno-unused-function"],
2087 deps = MICROKERNEL_BENCHMARK_DEPS,
2088)
2089
2090xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002091 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002092 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002093 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002094 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002095 copts = ["-Wno-unused-function"] + xnnpack_optional_dnnl_copts(),
2096 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002097)
2098
2099xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002100 name = "f32_vscaleexpminusmax_bench",
2101 srcs = [
2102 "bench/f32-vscaleexpminusmax.cc",
2103 "src/xnnpack/AlignedAllocator.h",
2104 ] + MICROKERNEL_BENCHMARK_HDRS,
2105 deps = MICROKERNEL_BENCHMARK_DEPS,
2106)
2107
2108xnnpack_benchmark(
2109 name = "f32_vscaleextexp_bench",
2110 srcs = [
2111 "bench/f32-vscaleextexp.cc",
2112 "src/xnnpack/AlignedAllocator.h",
2113 ] + MICROKERNEL_BENCHMARK_HDRS,
2114 deps = MICROKERNEL_BENCHMARK_DEPS,
2115)
2116
2117xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002118 name = "f32_im2col_gemm_bench",
2119 srcs = [
2120 "bench/f32-im2col-gemm.cc",
2121 "bench/conv.h",
2122 "src/xnnpack/AlignedAllocator.h",
2123 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2124 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2125)
2126
2127########################### Benchmarks for operators ###########################
2128
2129xnnpack_benchmark(
2130 name = "add_bench",
2131 srcs = ["bench/add.cc"],
2132 deps = OPERATOR_BENCHMARK_DEPS,
2133)
2134
2135xnnpack_benchmark(
2136 name = "average_pooling_bench",
2137 srcs = ["bench/average-pooling.cc"],
2138 deps = OPERATOR_BENCHMARK_DEPS,
2139)
2140
2141xnnpack_benchmark(
2142 name = "channel_shuffle_bench",
2143 srcs = ["bench/channel-shuffle.cc"],
2144 deps = OPERATOR_BENCHMARK_DEPS,
2145)
2146
2147xnnpack_benchmark(
2148 name = "convolution_bench",
2149 srcs = ["bench/convolution.cc"],
2150 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
2151 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
2152)
2153
2154xnnpack_benchmark(
2155 name = "deconvolution_bench",
2156 srcs = ["bench/deconvolution.cc"],
2157 copts = xnnpack_optional_tflite_copts(),
2158 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2159)
2160
2161xnnpack_benchmark(
2162 name = "global_average_pooling_bench",
2163 srcs = ["bench/global-average-pooling.cc"],
2164 deps = OPERATOR_BENCHMARK_DEPS,
2165)
2166
2167xnnpack_benchmark(
2168 name = "max_pooling_bench",
2169 srcs = ["bench/max-pooling.cc"],
2170 deps = OPERATOR_BENCHMARK_DEPS,
2171)
2172
2173xnnpack_benchmark(
2174 name = "sigmoid_bench",
2175 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002176 copts = xnnpack_optional_tflite_copts(),
2177 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002178)
2179
2180xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002181 name = "prelu_bench",
2182 srcs = ["bench/prelu.cc"],
2183 copts = xnnpack_optional_tflite_copts(),
2184 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2185)
2186
2187xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002188 name = "softmax_bench",
2189 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002190 copts = xnnpack_optional_tflite_copts(),
2191 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002192)
2193
Marat Dukhanc068bb62019-10-04 13:24:39 -07002194############################# End-to-end benchmarks ############################
2195
2196cc_library(
2197 name = "mobilenet_v1",
2198 srcs = ["models/mobilenet-v1.cc"],
2199 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002200 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002201 linkstatic = True,
2202 deps = [
2203 ":XNNPACK",
2204 "@pthreadpool",
2205 ],
2206)
2207
2208cc_library(
2209 name = "mobilenet_v2",
2210 srcs = ["models/mobilenet-v2.cc"],
2211 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002212 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002213 linkstatic = True,
2214 deps = [
2215 ":XNNPACK",
2216 "@pthreadpool",
2217 ],
2218)
2219
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002220cc_library(
2221 name = "mobilenet_v3_large",
2222 srcs = ["models/mobilenet-v3-large.cc"],
2223 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002224 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002225 linkstatic = True,
2226 deps = [
2227 ":XNNPACK",
2228 "@pthreadpool",
2229 ],
2230)
2231
2232cc_library(
2233 name = "mobilenet_v3_small",
2234 srcs = ["models/mobilenet-v3-small.cc"],
2235 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002236 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002237 linkstatic = True,
2238 deps = [
2239 ":XNNPACK",
2240 "@pthreadpool",
2241 ],
2242)
2243
Marat Dukhanc068bb62019-10-04 13:24:39 -07002244xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002245 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002246 srcs = [
2247 "bench/f32-dwconv-e2e.cc",
2248 "bench/end2end.h",
2249 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002250 copts = ["-Wno-unused-function"],
2251 deps = MICROKERNEL_BENCHMARK_DEPS + [
2252 ":XNNPACK",
2253 ":mobilenet_v1",
2254 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002255 ":mobilenet_v3_large",
2256 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002257 ],
2258)
2259
2260xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002261 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002262 srcs = [
2263 "bench/f32-gemm-e2e.cc",
2264 "bench/end2end.h",
2265 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002266 copts = ["-Wno-unused-function"],
2267 deps = MICROKERNEL_BENCHMARK_DEPS + [
2268 ":XNNPACK",
2269 ":mobilenet_v1",
2270 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002271 ":mobilenet_v3_large",
2272 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002273 ],
2274)
2275
2276xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002277 name = "end2end_bench",
2278 srcs = ["bench/end2end.cc"],
2279 deps = [
2280 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002281 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002282 ":mobilenet_v1",
2283 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002284 ":mobilenet_v3_large",
2285 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002286 "@pthreadpool",
2287 ],
2288)
2289
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002290#################### Accuracy evaluation for math functions ####################
2291
2292xnnpack_benchmark(
2293 name = "f32_exp_eval",
2294 srcs = [
2295 "eval/f32-exp.cc",
2296 "src/xnnpack/AlignedAllocator.h",
2297 ] + ACCURACY_EVAL_HDRS,
2298 deps = ACCURACY_EVAL_DEPS,
2299)
2300
Marat Dukhan515c9772019-10-17 18:07:57 -07002301xnnpack_benchmark(
2302 name = "f32_expminus_eval",
2303 srcs = [
2304 "eval/f32-expminus.cc",
2305 "src/xnnpack/AlignedAllocator.h",
2306 ] + ACCURACY_EVAL_HDRS,
2307 deps = ACCURACY_EVAL_DEPS,
2308)
2309
Marat Dukhan98ba4412019-10-23 02:14:28 -07002310xnnpack_benchmark(
2311 name = "f32_extexp_eval",
2312 srcs = [
2313 "eval/f32-extexp.cc",
2314 "src/xnnpack/AlignedAllocator.h",
2315 ] + ACCURACY_EVAL_HDRS,
2316 deps = ACCURACY_EVAL_DEPS,
2317)
2318
Marat Dukhan346a9e52019-11-15 09:06:30 -08002319xnnpack_benchmark(
2320 name = "f32_sigmoid_eval",
2321 srcs = [
2322 "eval/f32-sigmoid.cc",
2323 "src/xnnpack/AlignedAllocator.h",
2324 ] + ACCURACY_EVAL_HDRS,
2325 deps = ACCURACY_EVAL_DEPS,
2326)
2327
Marat Dukhan08c4a432019-10-03 09:29:21 -07002328######################### Unit tests for micro-kernels #########################
2329
2330xnnpack_unit_test(
2331 name = "f16_gemm_test",
2332 srcs = [
2333 "test/f16-gemm.cc",
2334 "test/gemm-microkernel-tester.h",
2335 "src/xnnpack/AlignedAllocator.h",
2336 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2337 deps = MICROKERNEL_TEST_DEPS,
2338)
2339
2340xnnpack_unit_test(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002341 name = "f16_spmm_test",
2342 srcs = [
2343 "test/f16-spmm.cc",
2344 "test/spmm-microkernel-tester.h",
2345 "src/xnnpack/AlignedAllocator.h",
2346 ] + MICROKERNEL_TEST_HDRS,
2347 deps = MICROKERNEL_TEST_DEPS,
2348)
2349
2350xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002351 name = "f32_argmaxpool_test",
2352 srcs = [
2353 "test/f32-argmaxpool.cc",
2354 "test/argmaxpool-microkernel-tester.h",
2355 "src/xnnpack/AlignedAllocator.h",
2356 ] + MICROKERNEL_TEST_HDRS,
2357 deps = MICROKERNEL_TEST_DEPS,
2358)
2359
2360xnnpack_unit_test(
2361 name = "f32_avgpool_test",
2362 srcs = [
2363 "test/f32-avgpool.cc",
2364 "test/avgpool-microkernel-tester.h",
2365 "src/xnnpack/AlignedAllocator.h",
2366 ] + MICROKERNEL_TEST_HDRS,
2367 deps = MICROKERNEL_TEST_DEPS,
2368)
2369
2370xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002371 name = "f32_bilinear_test",
2372 srcs = [
2373 "test/f32-bilinear.cc",
2374 "test/bilinear-microkernel-tester.h",
2375 "src/xnnpack/AlignedAllocator.h",
2376 ] + MICROKERNEL_TEST_HDRS,
2377 deps = MICROKERNEL_TEST_DEPS,
2378)
2379
2380xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002381 name = "f32_clamp_test",
2382 srcs = [
2383 "test/f32-clamp.cc",
2384 "test/clamp-microkernel-tester.h",
2385 ] + MICROKERNEL_TEST_HDRS,
2386 deps = MICROKERNEL_TEST_DEPS,
2387)
2388
2389xnnpack_unit_test(
2390 name = "f32_igemm_test",
2391 srcs = [
2392 "test/f32-igemm.cc",
2393 "test/gemm-microkernel-tester.h",
2394 "src/xnnpack/AlignedAllocator.h",
2395 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2396 deps = MICROKERNEL_TEST_DEPS,
2397)
2398
2399xnnpack_unit_test(
2400 name = "f32_conv_hwc_test",
2401 srcs = [
2402 "test/f32-conv-hwc.cc",
2403 "test/conv-hwc-microkernel-tester.h",
2404 "src/xnnpack/AlignedAllocator.h",
2405 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2406 deps = MICROKERNEL_TEST_DEPS,
2407)
2408
2409xnnpack_unit_test(
2410 name = "f32_conv_hwc2spchw_test",
2411 srcs = [
2412 "test/f32-conv-hwc2spchw.cc",
2413 "test/conv-hwc2spchw-microkernel-tester.h",
2414 "src/xnnpack/AlignedAllocator.h",
2415 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2416 deps = MICROKERNEL_TEST_DEPS,
2417)
2418
2419xnnpack_unit_test(
2420 name = "f32_dwconv_test",
2421 srcs = [
2422 "test/f32-dwconv.cc",
2423 "test/dwconv-microkernel-tester.h",
2424 "src/xnnpack/AlignedAllocator.h",
2425 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2426 deps = MICROKERNEL_TEST_DEPS,
2427)
2428
2429xnnpack_unit_test(
2430 name = "f32_dwconv_spchw_test",
2431 srcs = [
2432 "test/f32-dwconv-spchw.cc",
2433 "test/dwconv-spchw-microkernel-tester.h",
2434 "src/xnnpack/AlignedAllocator.h",
2435 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2436 deps = MICROKERNEL_TEST_DEPS,
2437)
2438
2439xnnpack_unit_test(
2440 name = "f32_gavgpool_test",
2441 srcs = [
2442 "test/f32-gavgpool.cc",
2443 "test/gavgpool-microkernel-tester.h",
2444 "src/xnnpack/AlignedAllocator.h",
2445 ] + MICROKERNEL_TEST_HDRS,
2446 deps = MICROKERNEL_TEST_DEPS,
2447)
2448
2449xnnpack_unit_test(
2450 name = "f32_gavgpool_spchw_test",
2451 srcs = [
2452 "test/f32-gavgpool-spchw.cc",
2453 "test/gavgpool-spchw-microkernel-tester.h",
2454 "src/xnnpack/AlignedAllocator.h",
2455 ] + MICROKERNEL_TEST_HDRS,
2456 deps = MICROKERNEL_TEST_DEPS,
2457)
2458
2459xnnpack_unit_test(
2460 name = "f32_gemm_test",
2461 srcs = [
2462 "test/f32-gemm.cc",
2463 "test/gemm-microkernel-tester.h",
2464 "src/xnnpack/AlignedAllocator.h",
2465 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2466 deps = MICROKERNEL_TEST_DEPS,
2467)
2468
2469xnnpack_unit_test(
2470 name = "f32_gemminc_test",
2471 srcs = [
2472 "test/f32-gemminc.cc",
2473 "test/gemm-microkernel-tester.h",
2474 "src/xnnpack/AlignedAllocator.h",
2475 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2476 deps = MICROKERNEL_TEST_DEPS,
2477)
2478
2479xnnpack_unit_test(
2480 name = "f32_hswish_test",
2481 srcs = [
2482 "test/f32-hswish.cc",
2483 "test/hswish-microkernel-tester.h",
2484 ] + MICROKERNEL_TEST_HDRS,
2485 deps = MICROKERNEL_TEST_DEPS,
2486)
2487
2488xnnpack_unit_test(
2489 name = "f32_maxpool_test",
2490 srcs = [
2491 "test/f32-maxpool.cc",
2492 "test/maxpool-microkernel-tester.h",
2493 ] + MICROKERNEL_TEST_HDRS,
2494 deps = MICROKERNEL_TEST_DEPS,
2495)
2496
2497xnnpack_unit_test(
2498 name = "f32_pavgpool_test",
2499 srcs = [
2500 "test/f32-pavgpool.cc",
2501 "test/avgpool-microkernel-tester.h",
2502 "src/xnnpack/AlignedAllocator.h",
2503 ] + MICROKERNEL_TEST_HDRS,
2504 deps = MICROKERNEL_TEST_DEPS,
2505)
2506
2507xnnpack_unit_test(
2508 name = "f32_ppmm_test",
2509 srcs = [
2510 "test/f32-ppmm.cc",
2511 "test/gemm-microkernel-tester.h",
2512 "src/xnnpack/AlignedAllocator.h",
2513 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2514 deps = MICROKERNEL_TEST_DEPS,
2515)
2516
2517xnnpack_unit_test(
2518 name = "f32_prelu_test",
2519 srcs = [
2520 "test/f32-prelu.cc",
2521 "test/prelu-microkernel-tester.h",
2522 "src/xnnpack/AlignedAllocator.h",
2523 ] + MICROKERNEL_TEST_HDRS,
2524 deps = MICROKERNEL_TEST_DEPS,
2525)
2526
2527xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002528 name = "f32_raddexpminusmax_test",
2529 srcs = [
2530 "test/f32-raddexpminusmax.cc",
2531 "test/raddexpminusmax-microkernel-tester.h",
2532 ] + MICROKERNEL_TEST_HDRS,
2533 deps = MICROKERNEL_TEST_DEPS,
2534)
2535
2536xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002537 name = "f32_raddextexp_test",
2538 srcs = [
2539 "test/f32-raddextexp.cc",
2540 "test/raddextexp-microkernel-tester.h",
2541 ] + MICROKERNEL_TEST_HDRS,
2542 deps = MICROKERNEL_TEST_DEPS,
2543)
2544
2545xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002546 name = "f32_raddstoreexpminusmax_test",
2547 srcs = [
2548 "test/f32-raddstoreexpminusmax.cc",
2549 "test/raddstoreexpminusmax-microkernel-tester.h",
2550 ] + MICROKERNEL_TEST_HDRS,
2551 deps = MICROKERNEL_TEST_DEPS,
2552)
2553
2554xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002555 name = "f32_rmax_test",
2556 srcs = [
2557 "test/f32-rmax.cc",
2558 "test/rmax-microkernel-tester.h",
2559 ] + MICROKERNEL_TEST_HDRS,
2560 deps = MICROKERNEL_TEST_DEPS,
2561)
2562
2563xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002564 name = "f32_sigmoid_test",
2565 srcs = [
2566 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002567 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002568 ] + MICROKERNEL_TEST_HDRS,
2569 deps = MICROKERNEL_TEST_DEPS,
2570)
2571
2572xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002573 name = "f32_spmm_test",
2574 srcs = [
2575 "test/f32-spmm.cc",
2576 "test/spmm-microkernel-tester.h",
2577 "src/xnnpack/AlignedAllocator.h",
2578 ] + MICROKERNEL_TEST_HDRS,
2579 deps = MICROKERNEL_TEST_DEPS,
2580)
2581
2582xnnpack_unit_test(
2583 name = "f32_vadd_test",
2584 srcs = [
2585 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002586 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002587 ] + MICROKERNEL_TEST_HDRS,
2588 deps = MICROKERNEL_TEST_DEPS,
2589)
2590
2591xnnpack_unit_test(
2592 name = "f32_vaddc_test",
2593 srcs = [
2594 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002595 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002596 ] + MICROKERNEL_TEST_HDRS,
2597 deps = MICROKERNEL_TEST_DEPS,
2598)
2599
2600xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002601 name = "f32_vdiv_test",
2602 srcs = [
2603 "test/f32-vdiv.cc",
2604 "test/vbinary-microkernel-tester.h",
2605 ] + MICROKERNEL_TEST_HDRS,
2606 deps = MICROKERNEL_TEST_DEPS,
2607)
2608
2609xnnpack_unit_test(
2610 name = "f32_vdivc_test",
2611 srcs = [
2612 "test/f32-vdivc.cc",
2613 "test/vbinaryc-microkernel-tester.h",
2614 ] + MICROKERNEL_TEST_HDRS,
2615 deps = MICROKERNEL_TEST_DEPS,
2616)
2617
2618xnnpack_unit_test(
2619 name = "f32_vrdivc_test",
2620 srcs = [
2621 "test/f32-vrdivc.cc",
2622 "test/vbinaryc-microkernel-tester.h",
2623 ] + MICROKERNEL_TEST_HDRS,
2624 deps = MICROKERNEL_TEST_DEPS,
2625)
2626
2627xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002628 name = "f32_vmax_test",
2629 srcs = [
2630 "test/f32-vmax.cc",
2631 "test/vbinary-microkernel-tester.h",
2632 ] + MICROKERNEL_TEST_HDRS,
2633 deps = MICROKERNEL_TEST_DEPS,
2634)
2635
2636xnnpack_unit_test(
2637 name = "f32_vmaxc_test",
2638 srcs = [
2639 "test/f32-vmaxc.cc",
2640 "test/vbinaryc-microkernel-tester.h",
2641 ] + MICROKERNEL_TEST_HDRS,
2642 deps = MICROKERNEL_TEST_DEPS,
2643)
2644
2645xnnpack_unit_test(
2646 name = "f32_vmin_test",
2647 srcs = [
2648 "test/f32-vmin.cc",
2649 "test/vbinary-microkernel-tester.h",
2650 ] + MICROKERNEL_TEST_HDRS,
2651 deps = MICROKERNEL_TEST_DEPS,
2652)
2653
2654xnnpack_unit_test(
2655 name = "f32_vminc_test",
2656 srcs = [
2657 "test/f32-vminc.cc",
2658 "test/vbinaryc-microkernel-tester.h",
2659 ] + MICROKERNEL_TEST_HDRS,
2660 deps = MICROKERNEL_TEST_DEPS,
2661)
2662
2663xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002664 name = "f32_vmul_test",
2665 srcs = [
2666 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002667 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002668 ] + MICROKERNEL_TEST_HDRS,
2669 deps = MICROKERNEL_TEST_DEPS,
2670)
2671
2672xnnpack_unit_test(
2673 name = "f32_vmulc_test",
2674 srcs = [
2675 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002676 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002677 ] + MICROKERNEL_TEST_HDRS,
2678 deps = MICROKERNEL_TEST_DEPS,
2679)
2680
2681xnnpack_unit_test(
2682 name = "f32_vmulcaddc_test",
2683 srcs = [
2684 "test/f32-vmulcaddc.cc",
2685 "test/vmulcaddc-microkernel-tester.h",
2686 "src/xnnpack/AlignedAllocator.h",
2687 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2688 deps = MICROKERNEL_TEST_DEPS,
2689)
2690
2691xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002692 name = "f32_vscale_test",
2693 srcs = [
2694 "test/f32-vscale.cc",
2695 "test/vscale-microkernel-tester.h",
2696 ] + MICROKERNEL_TEST_HDRS,
2697 deps = MICROKERNEL_TEST_DEPS,
2698)
2699
2700xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002701 name = "f32_vscaleexpminusmax_test",
2702 srcs = [
2703 "test/f32-vscaleexpminusmax.cc",
2704 "test/vscaleexpminusmax-microkernel-tester.h",
2705 ] + MICROKERNEL_TEST_HDRS,
2706 deps = MICROKERNEL_TEST_DEPS,
2707)
2708
2709xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002710 name = "f32_vscaleextexp_test",
2711 srcs = [
2712 "test/f32-vscaleextexp.cc",
2713 "test/vscaleextexp-microkernel-tester.h",
2714 ] + MICROKERNEL_TEST_HDRS,
2715 deps = MICROKERNEL_TEST_DEPS,
2716)
2717
2718xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002719 name = "f32_vsub_test",
2720 srcs = [
2721 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002722 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002723 ] + MICROKERNEL_TEST_HDRS,
2724 deps = MICROKERNEL_TEST_DEPS,
2725)
2726
2727xnnpack_unit_test(
2728 name = "f32_vsubc_test",
2729 srcs = [
2730 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002731 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002732 ] + MICROKERNEL_TEST_HDRS,
2733 deps = MICROKERNEL_TEST_DEPS,
2734)
2735
2736xnnpack_unit_test(
2737 name = "f32_vrsubc_test",
2738 srcs = [
2739 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002740 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002741 ] + MICROKERNEL_TEST_HDRS,
2742 deps = MICROKERNEL_TEST_DEPS,
2743)
2744
2745xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002746 name = "q8_avgpool_test",
2747 srcs = [
2748 "test/q8-avgpool.cc",
2749 "test/avgpool-microkernel-tester.h",
2750 "src/xnnpack/AlignedAllocator.h",
2751 ] + MICROKERNEL_TEST_HDRS,
2752 deps = MICROKERNEL_TEST_DEPS,
2753)
2754
2755xnnpack_unit_test(
2756 name = "q8_igemm_test",
2757 srcs = [
2758 "test/q8-igemm.cc",
2759 "test/gemm-microkernel-tester.h",
2760 "src/xnnpack/AlignedAllocator.h",
2761 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2762 deps = MICROKERNEL_TEST_DEPS,
2763)
2764
2765xnnpack_unit_test(
2766 name = "q8_dwconv_test",
2767 srcs = [
2768 "test/q8-dwconv.cc",
2769 "test/dwconv-microkernel-tester.h",
2770 "src/xnnpack/AlignedAllocator.h",
2771 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2772 deps = MICROKERNEL_TEST_DEPS,
2773)
2774
2775xnnpack_unit_test(
2776 name = "q8_gavgpool_test",
2777 srcs = [
2778 "test/q8-gavgpool.cc",
2779 "test/gavgpool-microkernel-tester.h",
2780 "src/xnnpack/AlignedAllocator.h",
2781 ] + MICROKERNEL_TEST_HDRS,
2782 deps = MICROKERNEL_TEST_DEPS,
2783)
2784
2785xnnpack_unit_test(
2786 name = "q8_gemm_test",
2787 srcs = [
2788 "test/q8-gemm.cc",
2789 "test/gemm-microkernel-tester.h",
2790 "src/xnnpack/AlignedAllocator.h",
2791 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2792 deps = MICROKERNEL_TEST_DEPS,
2793)
2794
2795xnnpack_unit_test(
2796 name = "q8_vadd_test",
2797 srcs = [
2798 "test/q8-vadd.cc",
2799 "test/vadd-microkernel-tester.h",
2800 ] + MICROKERNEL_TEST_HDRS,
2801 deps = MICROKERNEL_TEST_DEPS,
2802)
2803
2804xnnpack_unit_test(
2805 name = "u8_clamp_test",
2806 srcs = [
2807 "test/u8-clamp.cc",
2808 "test/clamp-microkernel-tester.h",
2809 ] + MICROKERNEL_TEST_HDRS,
2810 deps = MICROKERNEL_TEST_DEPS,
2811)
2812
2813xnnpack_unit_test(
2814 name = "u8_lut32norm_test",
2815 srcs = [
2816 "test/u8-lut32norm.cc",
2817 "test/lut-norm-microkernel-tester.h",
2818 ] + MICROKERNEL_TEST_HDRS,
2819 deps = MICROKERNEL_TEST_DEPS,
2820)
2821
2822xnnpack_unit_test(
2823 name = "u8_maxpool_test",
2824 srcs = [
2825 "test/u8-maxpool.cc",
2826 "test/maxpool-microkernel-tester.h",
2827 ] + MICROKERNEL_TEST_HDRS,
2828 deps = MICROKERNEL_TEST_DEPS,
2829)
2830
2831xnnpack_unit_test(
2832 name = "u8_rmax_test",
2833 srcs = [
2834 "test/u8-rmax.cc",
2835 "test/rmax-microkernel-tester.h",
2836 ] + MICROKERNEL_TEST_HDRS,
2837 deps = MICROKERNEL_TEST_DEPS,
2838)
2839
2840xnnpack_unit_test(
2841 name = "x32_packx_test",
2842 srcs = [
2843 "test/x32-packx.cc",
2844 "test/pack-microkernel-tester.h",
2845 "src/xnnpack/AlignedAllocator.h",
2846 ] + MICROKERNEL_TEST_HDRS,
2847 deps = MICROKERNEL_TEST_DEPS,
2848)
2849
2850xnnpack_unit_test(
2851 name = "x32_pad_test",
2852 srcs = [
2853 "test/x32-pad.cc",
2854 "test/pad-microkernel-tester.h",
2855 ] + MICROKERNEL_TEST_HDRS,
2856 deps = MICROKERNEL_TEST_DEPS,
2857)
2858
2859xnnpack_unit_test(
2860 name = "x32_unpool_test",
2861 srcs = [
2862 "test/x32-unpool.cc",
2863 "test/unpool-microkernel-tester.h",
2864 ] + MICROKERNEL_TEST_HDRS,
2865 deps = MICROKERNEL_TEST_DEPS,
2866)
2867
2868xnnpack_unit_test(
2869 name = "x32_zip_test",
2870 srcs = [
2871 "test/x32-zip.cc",
2872 "test/zip-microkernel-tester.h",
2873 ] + MICROKERNEL_TEST_HDRS,
2874 deps = MICROKERNEL_TEST_DEPS,
2875)
2876
2877xnnpack_unit_test(
2878 name = "x8_lut_test",
2879 srcs = [
2880 "test/x8-lut.cc",
2881 "test/lut-microkernel-tester.h",
2882 ] + MICROKERNEL_TEST_HDRS,
2883 deps = MICROKERNEL_TEST_DEPS,
2884)
2885
2886xnnpack_unit_test(
2887 name = "x8_zip_test",
2888 srcs = [
2889 "test/x8-zip.cc",
2890 "test/zip-microkernel-tester.h",
2891 ] + MICROKERNEL_TEST_HDRS,
2892 deps = MICROKERNEL_TEST_DEPS,
2893)
2894
2895########################### Size test for the library ##########################
2896
2897xnnpack_binary(
2898 name = "size_test",
2899 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002900 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901)
2902
2903########################### Unit tests for operators ###########################
2904
2905xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002906 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002907 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002908 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002909 "test/add-operator-tester.h",
2910 ],
2911 deps = OPERATOR_TEST_DEPS,
2912)
2913
2914xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002915 name = "add_nd_test",
2916 srcs = [
2917 "test/add-nd.cc",
2918 "test/binary-elementwise-operator-tester.h",
2919 ],
2920 deps = OPERATOR_TEST_DEPS,
2921)
2922
2923xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002924 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002925 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002926 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002927 "test/argmax-pooling-operator-tester.h",
2928 ] + OPERATOR_TEST_PARAMS_HDRS,
2929 deps = OPERATOR_TEST_DEPS,
2930)
2931
2932xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002933 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002935 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 "test/average-pooling-operator-tester.h",
2937 ] + OPERATOR_TEST_PARAMS_HDRS,
2938 deps = OPERATOR_TEST_DEPS,
2939)
2940
2941xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002942 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002943 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002944 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002945 "test/channel-pad-operator-tester.h",
2946 ] + OPERATOR_TEST_PARAMS_HDRS,
2947 deps = OPERATOR_TEST_DEPS,
2948)
2949
2950xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002951 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002952 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002953 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002954 "test/channel-shuffle-operator-tester.h",
2955 ],
2956 deps = OPERATOR_TEST_DEPS,
2957)
2958
2959xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002960 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002962 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002963 "test/clamp-operator-tester.h",
2964 ],
2965 deps = OPERATOR_TEST_DEPS,
2966)
2967
2968xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002969 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002971 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002972 "test/convolution-operator-tester.h",
2973 ],
2974 deps = OPERATOR_TEST_DEPS,
2975)
2976
2977xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002978 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002980 "test/convolution-nchw.cc",
2981 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 ],
2983 deps = OPERATOR_TEST_DEPS,
2984)
2985
2986xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002987 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002988 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002989 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002990 "test/deconvolution-operator-tester.h",
2991 ] + OPERATOR_TEST_PARAMS_HDRS,
2992 deps = OPERATOR_TEST_DEPS,
2993)
2994
2995xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002996 name = "divide_nd_test",
2997 srcs = [
2998 "test/binary-elementwise-operator-tester.h",
2999 "test/divide-nd.cc",
3000 ],
3001 deps = OPERATOR_TEST_DEPS,
3002)
3003
3004xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003005 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003006 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003007 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008 "test/fully-connected-operator-tester.h",
3009 ],
3010 deps = OPERATOR_TEST_DEPS,
3011)
3012
3013xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003014 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003015 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003016 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003017 "test/global-average-pooling-operator-tester.h",
3018 ] + OPERATOR_TEST_PARAMS_HDRS,
3019 deps = OPERATOR_TEST_DEPS,
3020)
3021
3022xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003023 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003025 "test/global-average-pooling-ncw.cc",
3026 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003027 ],
3028 deps = OPERATOR_TEST_DEPS,
3029)
3030
3031xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003032 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003033 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003034 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003035 "test/hardswish-operator-tester.h",
3036 ],
3037 deps = OPERATOR_TEST_DEPS,
3038)
3039
3040xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003041 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003042 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003043 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003044 "test/leaky-relu-operator-tester.h",
3045 ],
3046 deps = OPERATOR_TEST_DEPS,
3047)
3048
3049xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003050 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003051 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003052 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003053 "test/max-pooling-operator-tester.h",
3054 ] + OPERATOR_TEST_PARAMS_HDRS,
3055 deps = OPERATOR_TEST_DEPS,
3056)
3057
3058xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003059 name = "maximum_nd_test",
3060 srcs = [
3061 "test/binary-elementwise-operator-tester.h",
3062 "test/maximum-nd.cc",
3063 ],
3064 deps = OPERATOR_TEST_DEPS,
3065)
3066
3067xnnpack_unit_test(
3068 name = "minimum_nd_test",
3069 srcs = [
3070 "test/binary-elementwise-operator-tester.h",
3071 "test/minimum-nd.cc",
3072 ],
3073 deps = OPERATOR_TEST_DEPS,
3074)
3075
3076xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003077 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003078 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003079 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003080 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003081 ],
3082 deps = OPERATOR_TEST_DEPS,
3083)
3084
3085xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003086 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003087 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003088 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003089 "test/prelu-operator-tester.h",
3090 ] + OPERATOR_TEST_PARAMS_HDRS,
3091 deps = OPERATOR_TEST_DEPS,
3092)
3093
3094xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003095 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003096 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003097 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003098 "test/resize-bilinear-operator-tester.h",
3099 ] + OPERATOR_TEST_PARAMS_HDRS,
3100 deps = OPERATOR_TEST_DEPS,
3101)
3102
3103xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003104 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003105 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003106 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003107 "test/sigmoid-operator-tester.h",
3108 ],
3109 deps = OPERATOR_TEST_DEPS,
3110)
3111
3112xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003113 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003115 "test/softmax-nc.cc",
3116 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003117 ],
3118 deps = OPERATOR_TEST_DEPS,
3119)
3120
3121xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003122 name = "subtract_nd_test",
3123 srcs = [
3124 "test/binary-elementwise-operator-tester.h",
3125 "test/subtract-nd.cc",
3126 ],
3127 deps = OPERATOR_TEST_DEPS,
3128)
3129
3130xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003131 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003132 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003133 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003134 "test/unpooling-operator-tester.h",
3135 ],
3136 deps = OPERATOR_TEST_DEPS,
3137)
3138
3139############################# Build configurations #############################
3140
Marat Dukhanb8642352019-10-30 15:43:02 -07003141# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003143 name = "xnn_enable_assembly_explicit_true",
3144 define_values = {"xnn_enable_assembly": "true"},
3145)
3146
3147# Disables usage of assembly kernels.
3148config_setting(
3149 name = "xnn_enable_assembly_explicit_false",
3150 define_values = {"xnn_enable_assembly": "false"},
3151)
3152
3153# Builds with -c dbg
3154config_setting(
3155 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003157 "compilation_mode": "dbg",
3158 },
3159)
3160
3161# Builds with -c opt
3162config_setting(
3163 name = "optimized_build",
3164 values = {
3165 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166 },
3167)
3168
3169config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003170 name = "linux_k8",
3171 values = {"cpu": "k8"},
3172)
3173
3174config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003175 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003176 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003177)
3178
3179config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180 name = "android",
3181 values = {"crosstool_top": "//external:android/crosstool"},
3182)
3183
3184config_setting(
3185 name = "android_armv7",
3186 values = {
3187 "crosstool_top": "//external:android/crosstool",
3188 "cpu": "armeabi-v7a",
3189 },
3190)
3191
3192config_setting(
3193 name = "android_arm64",
3194 values = {
3195 "crosstool_top": "//external:android/crosstool",
3196 "cpu": "arm64-v8a",
3197 },
3198)
3199
3200config_setting(
3201 name = "android_x86",
3202 values = {
3203 "crosstool_top": "//external:android/crosstool",
3204 "cpu": "x86",
3205 },
3206)
3207
3208config_setting(
3209 name = "android_x86_64",
3210 values = {
3211 "crosstool_top": "//external:android/crosstool",
3212 "cpu": "x86_64",
3213 },
3214)
3215
3216config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003217 name = "macos_x86_64",
3218 values = {
3219 "apple_platform_type": "macos",
3220 "cpu": "darwin",
3221 },
3222)
3223
3224config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003225 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003226 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227)
3228
3229config_setting(
3230 name = "emscripten_wasm",
3231 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003232 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003233 "cpu": "wasm",
3234 },
3235)
3236
3237config_setting(
3238 name = "emscripten_wasmsimd",
3239 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003240 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003241 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003242 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003243 },
3244)
3245
3246config_setting(
3247 name = "emscripten_asmjs",
3248 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003249 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250 "cpu": "asmjs",
3251 },
3252)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003253
3254config_setting(
3255 name = "ios_armv7",
3256 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003257 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003258 "cpu": "ios_armv7",
3259 },
3260)
3261
3262config_setting(
3263 name = "ios_arm64",
3264 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003265 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003266 "cpu": "ios_arm64",
3267 },
3268)
3269
3270config_setting(
3271 name = "ios_arm64e",
3272 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003273 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003274 "cpu": "ios_arm64e",
3275 },
3276)
3277
3278config_setting(
3279 name = "ios_x86",
3280 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003281 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003282 "cpu": "ios_i386",
3283 },
3284)
3285
3286config_setting(
3287 name = "ios_x86_64",
3288 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003289 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003290 "cpu": "ios_x86_64",
3291 },
3292)
3293
3294config_setting(
3295 name = "watchos_armv7k",
3296 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003297 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003298 "cpu": "watchos_armv7k",
3299 },
3300)
3301
3302config_setting(
3303 name = "watchos_arm64_32",
3304 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003305 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003306 "cpu": "watchos_arm64_32",
3307 },
3308)
3309
3310config_setting(
3311 name = "watchos_x86",
3312 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003313 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003314 "cpu": "watchos_i386",
3315 },
3316)
3317
3318config_setting(
3319 name = "watchos_x86_64",
3320 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003321 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003322 "cpu": "watchos_x86_64",
3323 },
3324)
3325
3326config_setting(
3327 name = "tvos_arm64",
3328 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003329 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003330 "cpu": "tvos_arm64",
3331 },
3332)
3333
3334config_setting(
3335 name = "tvos_x86_64",
3336 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003337 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003338 "cpu": "tvos_x86_64",
3339 },
3340)