blob: 1421c6aba0663a6e284d12591e83a27c6469806f [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 Dukhan69c3f2c2019-11-06 12:30:01 -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_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_unit_test", "xnnpack_visibility")
10
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",
72 "src/softargmax-nc.c",
73 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
76SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080077 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080078 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080079 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070080 "src/f32-avgpool/mp9p8q-scalar.c",
81 "src/f32-avgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080082 "src/f32-bilinear/gen/scalar-c1.c",
83 "src/f32-bilinear/gen/scalar-c2.c",
84 "src/f32-bilinear/gen/scalar-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070085 "src/f32-clamp/scalar.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080086 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080087 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070088 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070089 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070090 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080091 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080092 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080093 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
94 "src/f32-dwconv/gen/up1x25-scalar.c",
95 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
96 "src/f32-dwconv/gen/up1x4-scalar.c",
97 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x9-scalar.c",
99 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
100 "src/f32-dwconv/gen/up2x25-scalar.c",
101 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
102 "src/f32-dwconv/gen/up2x4-scalar.c",
103 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700105 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700106 "src/f32-gavgpool/mp7p7q-scalar.c",
107 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800108 "src/f32-gemm/gen-inc/1x4-scalar.c",
109 "src/f32-gemm/gen-inc/2x4-scalar.c",
110 "src/f32-gemm/gen-inc/4x4-scalar.c",
111 "src/f32-gemm/gen/1x4-scalar.c",
112 "src/f32-gemm/gen/2x4-scalar.c",
113 "src/f32-gemm/gen/4x2-scalar.c",
114 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115 "src/f32-hswish/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800116 "src/f32-igemm/gen/1x4-scalar.c",
117 "src/f32-igemm/gen/2x4-scalar.c",
118 "src/f32-igemm/gen/4x2-scalar.c",
119 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800120 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700121 "src/f32-pavgpool/mp9p8q-scalar.c",
122 "src/f32-pavgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800123 "src/f32-ppmm/gen/2x4-scalar.c",
124 "src/f32-ppmm/gen/3x3-scalar.c",
125 "src/f32-ppmm/gen/4x2-scalar.c",
126 "src/f32-ppmm/gen/4x4-scalar.c",
127 "src/f32-prelu/gen/scalar-2x1.c",
128 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700129 "src/f32-rmax/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800130 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
131 "src/f32-spmm/gen/1x1-scalar.c",
132 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
133 "src/f32-spmm/gen/2x1-scalar.c",
134 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
135 "src/f32-spmm/gen/4x1-scalar.c",
136 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
137 "src/f32-spmm/gen/8x1-scalar.c",
138 "src/f32-spmm/gen/8x2-scalar.c",
139 "src/f32-spmm/gen/8x4-scalar.c",
140 "src/f32-vbinary/gen/vadd-scalar-x1.c",
141 "src/f32-vbinary/gen/vadd-scalar-x2.c",
142 "src/f32-vbinary/gen/vadd-scalar-x4.c",
143 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
144 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
145 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800146 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
147 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
148 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
149 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
150 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
151 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800152 "src/f32-vbinary/gen/vmax-scalar-x1.c",
153 "src/f32-vbinary/gen/vmax-scalar-x2.c",
154 "src/f32-vbinary/gen/vmax-scalar-x4.c",
155 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
156 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
157 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
158 "src/f32-vbinary/gen/vmin-scalar-x1.c",
159 "src/f32-vbinary/gen/vmin-scalar-x2.c",
160 "src/f32-vbinary/gen/vmin-scalar-x4.c",
161 "src/f32-vbinary/gen/vminc-scalar-x1.c",
162 "src/f32-vbinary/gen/vminc-scalar-x2.c",
163 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800164 "src/f32-vbinary/gen/vmul-scalar-x1.c",
165 "src/f32-vbinary/gen/vmul-scalar-x2.c",
166 "src/f32-vbinary/gen/vmul-scalar-x4.c",
167 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
168 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
169 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800170 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
171 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
172 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800173 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
174 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
175 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
176 "src/f32-vbinary/gen/vsub-scalar-x1.c",
177 "src/f32-vbinary/gen/vsub-scalar-x2.c",
178 "src/f32-vbinary/gen/vsub-scalar-x4.c",
179 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
180 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
181 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
182 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
183 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
184 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700185 "src/q8-avgpool/mp9p8q-scalar.c",
186 "src/q8-avgpool/up9-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700187 "src/q8-dwconv/up1x9-scalar.c",
188 "src/q8-gavgpool/mp7p7q-scalar.c",
189 "src/q8-gavgpool/up7-scalar.c",
190 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800191 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700192 "src/q8-vadd/scalar.c",
193 "src/u8-clamp/scalar.c",
194 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800195 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700196 "src/u8-rmax/scalar.c",
197 "src/x32-packx/x2-scalar.c",
198 "src/x32-packx/x3-scalar.c",
199 "src/x32-packx/x4-scalar.c",
200 "src/x32-pad/x2-scalar.c",
201 "src/x32-unpool/scalar.c",
202 "src/x32-zip/x2-scalar.c",
203 "src/x32-zip/x3-scalar.c",
204 "src/x32-zip/x4-scalar.c",
205 "src/x32-zip/xm-scalar.c",
206 "src/x8-lut/scalar.c",
207 "src/x8-zip/x2-scalar.c",
208 "src/x8-zip/x3-scalar.c",
209 "src/x8-zip/x4-scalar.c",
210 "src/x8-zip/xm-scalar.c",
211]
212
Marat Dukhan436ebe62019-12-04 15:10:12 -0800213WASM_UKERNELS = [
214 "src/f32-avgpool/mp9p8q-wasm.c",
215 "src/f32-avgpool/up9-wasm.c",
216 "src/f32-clamp/wasm.c",
217 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
218 "src/f32-dwconv/gen/up1x25-wasm.c",
219 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
220 "src/f32-dwconv/gen/up1x4-wasm.c",
221 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
222 "src/f32-dwconv/gen/up1x9-wasm.c",
223 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
224 "src/f32-dwconv/gen/up2x25-wasm.c",
225 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
226 "src/f32-dwconv/gen/up2x4-wasm.c",
227 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
228 "src/f32-dwconv/gen/up2x9-wasm.c",
229 "src/f32-gavgpool/mp7p7q-wasm.c",
230 "src/f32-gavgpool/up7-wasm.c",
231 "src/f32-gemm/gen-inc/1x4-wasm.c",
232 "src/f32-gemm/gen-inc/2x4-wasm.c",
233 "src/f32-gemm/gen-inc/4x4-wasm.c",
234 "src/f32-gemm/gen/1x4-wasm.c",
235 "src/f32-gemm/gen/2x4-wasm.c",
236 "src/f32-gemm/gen/4x2-wasm.c",
237 "src/f32-gemm/gen/4x4-wasm.c",
238 "src/f32-hswish/wasm.c",
239 "src/f32-igemm/gen/1x4-wasm.c",
240 "src/f32-igemm/gen/2x4-wasm.c",
241 "src/f32-igemm/gen/4x2-wasm.c",
242 "src/f32-igemm/gen/4x4-wasm.c",
243 "src/f32-maxpool/9p8x-wasm-c1.c",
244 "src/f32-pavgpool/mp9p8q-wasm.c",
245 "src/f32-pavgpool/up9-wasm.c",
246 "src/f32-prelu/gen/wasm-2x1.c",
247 "src/f32-prelu/gen/wasm-2x4.c",
248 "src/f32-vbinary/gen/vadd-wasm-x1.c",
249 "src/f32-vbinary/gen/vadd-wasm-x2.c",
250 "src/f32-vbinary/gen/vadd-wasm-x4.c",
251 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
252 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
253 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800254 "src/f32-vbinary/gen/vdiv-wasm-x1.c",
255 "src/f32-vbinary/gen/vdiv-wasm-x2.c",
256 "src/f32-vbinary/gen/vdiv-wasm-x4.c",
257 "src/f32-vbinary/gen/vdivc-wasm-x1.c",
258 "src/f32-vbinary/gen/vdivc-wasm-x2.c",
259 "src/f32-vbinary/gen/vdivc-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800260 "src/f32-vbinary/gen/vmax-wasm-x1.c",
261 "src/f32-vbinary/gen/vmax-wasm-x2.c",
262 "src/f32-vbinary/gen/vmax-wasm-x4.c",
263 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
264 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
265 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
266 "src/f32-vbinary/gen/vmin-wasm-x1.c",
267 "src/f32-vbinary/gen/vmin-wasm-x2.c",
268 "src/f32-vbinary/gen/vmin-wasm-x4.c",
269 "src/f32-vbinary/gen/vminc-wasm-x1.c",
270 "src/f32-vbinary/gen/vminc-wasm-x2.c",
271 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800272 "src/f32-vbinary/gen/vmul-wasm-x1.c",
273 "src/f32-vbinary/gen/vmul-wasm-x2.c",
274 "src/f32-vbinary/gen/vmul-wasm-x4.c",
275 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
276 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
277 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800278 "src/f32-vbinary/gen/vrdivc-wasm-x1.c",
279 "src/f32-vbinary/gen/vrdivc-wasm-x2.c",
280 "src/f32-vbinary/gen/vrdivc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800281 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
282 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
283 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
284 "src/f32-vbinary/gen/vsub-wasm-x1.c",
285 "src/f32-vbinary/gen/vsub-wasm-x2.c",
286 "src/f32-vbinary/gen/vsub-wasm-x4.c",
287 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
288 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
289 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
290 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
291 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
292 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
293]
294
Marat Dukhan08c4a432019-10-03 09:29:21 -0700295PSIMD_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800296 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800297 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800298 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700299 "src/f32-avgpool/mp9p8q-psimd.c",
300 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800301 "src/f32-bilinear/gen/psimd-c4.c",
302 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700303 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800304 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
305 "src/f32-dwconv/gen/up4x25-psimd.c",
306 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
307 "src/f32-dwconv/gen/up4x4-psimd.c",
308 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
309 "src/f32-dwconv/gen/up4x9-psimd.c",
310 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
311 "src/f32-dwconv/gen/up8x25-psimd.c",
312 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
313 "src/f32-dwconv/gen/up8x4-psimd.c",
314 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
315 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700316 "src/f32-gavgpool/mp7p7q-psimd.c",
317 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800318 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
319 "src/f32-gemm/gen/1x8-psimd-splat.c",
320 "src/f32-gemm/gen/1x8s4-psimd.c",
321 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
322 "src/f32-gemm/gen/4x8-psimd-splat.c",
323 "src/f32-gemm/gen/4x8s4-psimd.c",
324 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
325 "src/f32-gemm/gen/6x8-psimd-splat.c",
326 "src/f32-gemm/gen/6x8s4-psimd.c",
327 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
328 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
329 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
330 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
331 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
332 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
333 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
334 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
335 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700336 "src/f32-hswish/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800337 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
338 "src/f32-igemm/gen/1x8-psimd-splat.c",
339 "src/f32-igemm/gen/1x8s4-psimd.c",
340 "src/f32-igemm/gen/4x2c4-psimd.c",
341 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
342 "src/f32-igemm/gen/4x8-psimd-splat.c",
343 "src/f32-igemm/gen/4x8s4-psimd.c",
344 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
345 "src/f32-igemm/gen/6x8-psimd-splat.c",
346 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800347 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700348 "src/f32-pavgpool/mp9p8q-psimd.c",
349 "src/f32-pavgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800350 "src/f32-ppmm/gen/4x8-psimd.c",
351 "src/f32-prelu/gen/psimd-2x4.c",
352 "src/f32-prelu/gen/psimd-2x8.c",
353 "src/f32-vbinary/gen/vadd-psimd-x4.c",
354 "src/f32-vbinary/gen/vadd-psimd-x8.c",
355 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
356 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800357 "src/f32-vbinary/gen/vdiv-psimd-x4.c",
358 "src/f32-vbinary/gen/vdiv-psimd-x8.c",
359 "src/f32-vbinary/gen/vdivc-psimd-x4.c",
360 "src/f32-vbinary/gen/vdivc-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800361 "src/f32-vbinary/gen/vmax-psimd-x4.c",
362 "src/f32-vbinary/gen/vmax-psimd-x8.c",
363 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
364 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
365 "src/f32-vbinary/gen/vmin-psimd-x4.c",
366 "src/f32-vbinary/gen/vmin-psimd-x8.c",
367 "src/f32-vbinary/gen/vminc-psimd-x4.c",
368 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800369 "src/f32-vbinary/gen/vmul-psimd-x4.c",
370 "src/f32-vbinary/gen/vmul-psimd-x8.c",
371 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
372 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800373 "src/f32-vbinary/gen/vrdivc-psimd-x4.c",
374 "src/f32-vbinary/gen/vrdivc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800375 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
376 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
377 "src/f32-vbinary/gen/vsub-psimd-x4.c",
378 "src/f32-vbinary/gen/vsub-psimd-x8.c",
379 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
380 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
381 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
382 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700383 "src/x32-packx/x4-psimd.c",
384 "src/x32-pad/x2-psimd.c",
385 "src/x32-unpool/psimd.c",
386 "src/x32-zip/x2-psimd.c",
387 "src/x32-zip/x3-psimd.c",
388 "src/x32-zip/x4-psimd.c",
389 "src/x32-zip/xm-psimd.c",
390]
391
392# ISA-specific micro-kernels
393NEON_UKERNELS = [
394 "src/f32-avgpool/mp9p8q-neon.c",
395 "src/f32-avgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800396 "src/f32-bilinear/gen/neon-c4.c",
397 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700398 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800399 "src/f32-dwconv/gen/up4x9-neon.c",
400 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
401 "src/f32-dwconv/gen/up8x9-neon.c",
402 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700403 "src/f32-gavgpool-spchw/neon-x4.c",
404 "src/f32-gavgpool/mp7p7q-neon.c",
405 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800406 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
407 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
408 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
409 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
410 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
411 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800412 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800413 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
414 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
415 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
416 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800417 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800418 "src/f32-gemm/gen/1x8s4-neon.c",
419 "src/f32-gemm/gen/4x8s4-neon.c",
420 "src/f32-gemm/gen/6x8s4-neon.c",
421 "src/f32-gemm/gen/8x8s4-neon.c",
422 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
423 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
424 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
425 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
426 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800427 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800428 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
429 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
430 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
431 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800432 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800433 "src/f32-gemm/gen-inc/1x8s4-neon.c",
434 "src/f32-gemm/gen-inc/4x8s4-neon.c",
435 "src/f32-gemm/gen-inc/6x8s4-neon.c",
436 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700437 "src/f32-hswish/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800438 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
439 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
440 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
441 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
442 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
443 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800444 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800445 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
446 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
447 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
448 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800449 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800450 "src/f32-igemm/gen/1x8s4-neon.c",
451 "src/f32-igemm/gen/4x8s4-neon.c",
452 "src/f32-igemm/gen/6x8s4-neon.c",
453 "src/f32-igemm/gen/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700454 "src/f32-pavgpool/mp9p8q-neon.c",
455 "src/f32-pavgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800456 "src/f32-ppmm/gen/4x8-neon.c",
457 "src/f32-ppmm/gen/8x8-neon.c",
458 "src/f32-prelu/gen/neon-2x4.c",
459 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700460 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800461 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
462 "src/f32-vbinary/gen/vadd-neon-x4.c",
463 "src/f32-vbinary/gen/vadd-neon-x8.c",
464 "src/f32-vbinary/gen/vaddc-neon-x4.c",
465 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800466 "src/f32-vbinary/gen/vmax-neon-x4.c",
467 "src/f32-vbinary/gen/vmax-neon-x8.c",
468 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
469 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
470 "src/f32-vbinary/gen/vmin-neon-x4.c",
471 "src/f32-vbinary/gen/vmin-neon-x8.c",
472 "src/f32-vbinary/gen/vminc-neon-x4.c",
473 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800474 "src/f32-vbinary/gen/vmul-neon-x4.c",
475 "src/f32-vbinary/gen/vmul-neon-x8.c",
476 "src/f32-vbinary/gen/vmulc-neon-x4.c",
477 "src/f32-vbinary/gen/vmulc-neon-x8.c",
478 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
479 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
480 "src/f32-vbinary/gen/vsub-neon-x4.c",
481 "src/f32-vbinary/gen/vsub-neon-x8.c",
482 "src/f32-vbinary/gen/vsubc-neon-x4.c",
483 "src/f32-vbinary/gen/vsubc-neon-x8.c",
484 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
485 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700486 "src/q8-avgpool/mp9p8q-neon.c",
487 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700488 "src/q8-dwconv/up8x9-neon.c",
489 "src/q8-gavgpool/mp7p7q-neon.c",
490 "src/q8-gavgpool/up7-neon.c",
491 "src/q8-gemm/4x8-neon.c",
492 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800493 "src/q8-igemm/4x8-neon.c",
494 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700495 "src/q8-vadd/neon.c",
496 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800497 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700498 "src/u8-rmax/neon.c",
499 "src/x32-packx/x4-neon-st4.c",
500 "src/x32-pad/x2-neon.c",
501 "src/x32-zip/x2-neon.c",
502 "src/x32-zip/x3-neon.c",
503 "src/x32-zip/x4-neon.c",
504 "src/x32-zip/xm-neon.c",
505 "src/x8-zip/x2-neon.c",
506 "src/x8-zip/x3-neon.c",
507 "src/x8-zip/x4-neon.c",
508 "src/x8-zip/xm-neon.c",
509]
510
511NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800512 "src/f32-bilinear/gen/neonfma-c4.c",
513 "src/f32-bilinear/gen/neonfma-c8.c",
514 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
515 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
516 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
517 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800518 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800519 "src/f32-igemm/gen/1x8s4-neonfma.c",
520 "src/f32-igemm/gen/4x8s4-neonfma.c",
521 "src/f32-igemm/gen/6x8s4-neonfma.c",
522 "src/f32-igemm/gen/8x8s4-neonfma.c",
523 "src/f32-dwconv/gen/up4x9-neonfma.c",
524 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
525 "src/f32-dwconv/gen/up8x9-neonfma.c",
526 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
527 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
528 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
529 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
530 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800531 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800532 "src/f32-gemm/gen/1x8s4-neonfma.c",
533 "src/f32-gemm/gen/4x8s4-neonfma.c",
534 "src/f32-gemm/gen/6x8s4-neonfma.c",
535 "src/f32-gemm/gen/8x8s4-neonfma.c",
536 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
537 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
538 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
539 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800540 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800541 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
542 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
543 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
544 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700545 "src/f32-hswish/neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800546 "src/f32-ppmm/gen/4x8-neonfma.c",
547 "src/f32-ppmm/gen/8x8-neonfma.c",
548 "src/f32-sigmoid/gen/neonfma-p5-nr2fma-x16.c",
549 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
550 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800551 "src/math/exp-neonfma-lut64-p2.c",
552 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800553 "src/math/expminus-neonfma-lut2048-p1.c",
554 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800555 "src/math/expminus-neonfma-p5.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800556 "src/math/sigmoid-neonfma-lut2048-p1-nr1recps1fma.c",
557 "src/math/sigmoid-neonfma-lut2048-p1-nr2fma.c",
558 "src/math/sigmoid-neonfma-lut2048-p1-nr2recps.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800559 "src/math/sigmoid-neonfma-p5-nr1recps1fma.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800560 "src/math/sigmoid-neonfma-p5-nr2fma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800561 "src/math/sigmoid-neonfma-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700562]
563
564AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800565 "src/f32-vbinary/gen/vdiv-neon-x4.c",
566 "src/f32-vbinary/gen/vdiv-neon-x8.c",
567 "src/f32-vbinary/gen/vdivc-neon-x4.c",
568 "src/f32-vbinary/gen/vdivc-neon-x8.c",
569 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
570 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800571 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
572 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
573 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
574 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
575 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
576 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800577 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800578 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
579 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
580 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
581 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
582 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800583 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800584 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
585 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
586 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
587 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
588 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
589 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800590 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700591 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
592 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
593 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
594 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
595 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
596 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
597 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800598 "src/f32-spmm/gen/12x1-neonfma.c",
599 "src/f32-spmm/gen/12x2-neonfma.c",
600 "src/f32-spmm/gen/12x4-neonfma.c",
601 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
602 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
603 "src/f32-spmm/gen/16x1-neonfma.c",
604 "src/f32-spmm/gen/16x2-neonfma.c",
605 "src/f32-spmm/gen/16x4-neonfma.c",
606 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
607 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
608 "src/f32-spmm/gen/4x1-neonfma.c",
609 "src/f32-spmm/gen/4x2-neonfma.c",
610 "src/f32-spmm/gen/4x4-neonfma.c",
611 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
612 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
613 "src/f32-spmm/gen/8x1-neonfma.c",
614 "src/f32-spmm/gen/8x2-neonfma.c",
615 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800616 "src/math/sigmoid-neonfma-p5-div.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800617 "src/math/sigmoid-neonfma-lut2048-p1-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700618]
619
620AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800621 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
622 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
623 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700624]
625
626SSE_UKERNELS = [
627 "src/f32-avgpool/mp9p8q-sse.c",
628 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800629 "src/f32-bilinear/gen/sse-c4.c",
630 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700631 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800632 "src/f32-dwconv-spchw/3x3p1-sse.c",
633 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800634 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
635 "src/f32-dwconv/gen/up4x25-sse.c",
636 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
637 "src/f32-dwconv/gen/up4x4-sse.c",
638 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
639 "src/f32-dwconv/gen/up4x9-sse.c",
640 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
641 "src/f32-dwconv/gen/up8x25-sse.c",
642 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
643 "src/f32-dwconv/gen/up8x4-sse.c",
644 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
645 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700646 "src/f32-gavgpool-spchw/sse-x4.c",
647 "src/f32-gavgpool/mp7p7q-sse.c",
648 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800649 "src/f32-gemm/gen/1x8-sse-dup.c",
650 "src/f32-gemm/gen/1x8-sse-load1.c",
651 "src/f32-gemm/gen/1x8s4-sse.c",
652 "src/f32-gemm/gen/4x8-sse-dup.c",
653 "src/f32-gemm/gen/4x8-sse-load1.c",
654 "src/f32-gemm/gen/4x8s4-sse.c",
655 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
656 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
657 "src/f32-gemm/gen-inc/1x8s4-sse.c",
658 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
659 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
660 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700661 "src/f32-hswish/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800662 "src/f32-igemm/gen/1x8-sse-dup.c",
663 "src/f32-igemm/gen/1x8-sse-load1.c",
664 "src/f32-igemm/gen/1x8s4-sse.c",
665 "src/f32-igemm/gen/4x2c4-sse.c",
666 "src/f32-igemm/gen/4x8-sse-dup.c",
667 "src/f32-igemm/gen/4x8-sse-load1.c",
668 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800669 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700670 "src/f32-pavgpool/mp9p8q-sse.c",
671 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800672 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700673 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800674 "src/f32-spmm/gen/4x1-sse.c",
675 "src/f32-spmm/gen/8x1-sse.c",
676 "src/f32-vbinary/gen/vadd-sse-x4.c",
677 "src/f32-vbinary/gen/vadd-sse-x8.c",
678 "src/f32-vbinary/gen/vaddc-sse-x4.c",
679 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800680 "src/f32-vbinary/gen/vdiv-sse-x4.c",
681 "src/f32-vbinary/gen/vdiv-sse-x8.c",
682 "src/f32-vbinary/gen/vdivc-sse-x4.c",
683 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800684 "src/f32-vbinary/gen/vmax-sse-x4.c",
685 "src/f32-vbinary/gen/vmax-sse-x8.c",
686 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
687 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
688 "src/f32-vbinary/gen/vmin-sse-x4.c",
689 "src/f32-vbinary/gen/vmin-sse-x8.c",
690 "src/f32-vbinary/gen/vminc-sse-x4.c",
691 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800692 "src/f32-vbinary/gen/vmul-sse-x4.c",
693 "src/f32-vbinary/gen/vmul-sse-x8.c",
694 "src/f32-vbinary/gen/vmulc-sse-x4.c",
695 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800696 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
697 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800698 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
699 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
700 "src/f32-vbinary/gen/vsub-sse-x4.c",
701 "src/f32-vbinary/gen/vsub-sse-x8.c",
702 "src/f32-vbinary/gen/vsubc-sse-x4.c",
703 "src/f32-vbinary/gen/vsubc-sse-x8.c",
704 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
705 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700706 "src/x32-packx/x4-sse.c",
707]
708
709SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800710 "src/f32-argmaxpool/9p8x-sse2-c4.c",
711 "src/f32-argmaxpool/4x-sse2-c4.c",
712 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800713 "src/f32-prelu/gen/sse2-2x4.c",
714 "src/f32-prelu/gen/sse2-2x8.c",
715 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
716 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700717 "src/q8-avgpool/mp9p8q-sse2.c",
718 "src/q8-avgpool/up9-sse2.c",
719 "src/q8-igemm/4x4c2-sse2.c",
720 "src/q8-dwconv/up8x9-sse2.c",
721 "src/q8-gavgpool/mp7p7q-sse2.c",
722 "src/q8-gavgpool/up7-sse2.c",
723 "src/q8-gemm/2x4c8-sse2.c",
724 "src/q8-gemm/4x4c2-sse2.c",
725 "src/q8-vadd/sse2.c",
726 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800727 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700728 "src/u8-rmax/sse2.c",
729 "src/x32-pad/x2-sse2.c",
730 "src/x32-zip/x2-sse2.c",
731 "src/x32-zip/x3-sse2.c",
732 "src/x32-zip/x4-sse2.c",
733 "src/x32-zip/xm-sse2.c",
734 "src/x8-zip/x2-sse2.c",
735 "src/x8-zip/x3-sse2.c",
736 "src/x8-zip/x4-sse2.c",
737 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800738 "src/math/exp-sse2-p5.c",
739 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800740 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700741]
742
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800743SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800744 "src/f32-prelu/gen/sse41-2x4.c",
745 "src/f32-prelu/gen/sse41-2x8.c",
746 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
747 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800748]
749
Marat Dukhan08c4a432019-10-03 09:29:21 -0700750AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800751 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800752 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
753 "src/f32-dwconv/gen/up16x4-avx.c",
754 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
755 "src/f32-dwconv/gen/up8x4-avx.c",
756 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
757 "src/f32-dwconv/gen/up16x9-avx.c",
758 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
759 "src/f32-dwconv/gen/up8x9-avx.c",
760 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
761 "src/f32-dwconv/gen/up16x25-avx.c",
762 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
763 "src/f32-dwconv/gen/up8x25-avx.c",
764 "src/f32-gemm/gen/1x8-avx-broadcast.c",
765 "src/f32-gemm/gen/4x8-avx-broadcast.c",
766 "src/f32-gemm/gen/5x8-avx-broadcast.c",
767 "src/f32-gemm/gen/6x8-avx-broadcast.c",
768 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800769 "src/f32-gemm/gen/1x16-avx-broadcast.c",
770 "src/f32-gemm/gen/3x16-avx-broadcast.c",
771 "src/f32-gemm/gen/4x16-avx-broadcast.c",
772 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800773 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
774 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
775 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
776 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
777 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800778 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
779 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
780 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
781 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800782 "src/f32-igemm/gen/1x8-avx-broadcast.c",
783 "src/f32-igemm/gen/4x8-avx-broadcast.c",
784 "src/f32-igemm/gen/5x8-avx-broadcast.c",
785 "src/f32-igemm/gen/6x8-avx-broadcast.c",
786 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800787 "src/f32-igemm/gen/1x16-avx-broadcast.c",
788 "src/f32-igemm/gen/3x16-avx-broadcast.c",
789 "src/f32-igemm/gen/4x16-avx-broadcast.c",
790 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700791 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700792 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700793]
794
Marat Dukhanfda12b82019-11-21 12:27:59 -0800795FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800796 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
797 "src/f32-dwconv/gen/up16x4-fma3.c",
798 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
799 "src/f32-dwconv/gen/up8x4-fma3.c",
800 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
801 "src/f32-dwconv/gen/up16x9-fma3.c",
802 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
803 "src/f32-dwconv/gen/up8x9-fma3.c",
804 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
805 "src/f32-dwconv/gen/up16x25-fma3.c",
806 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
807 "src/f32-dwconv/gen/up8x25-fma3.c",
808 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
809 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
810 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
811 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
812 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
813 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800814 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
815 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
816 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
817 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800818 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
819 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
820 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
821 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
822 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
823 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800824 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
825 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
826 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
827 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800828 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
829 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
830 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
831 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
832 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
833 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800834 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
835 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
836 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
837 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -0800838]
839
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700840AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700841 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700842 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700843 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
844 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700845 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700846 "src/math/exp-avx2-p5.c",
847 "src/math/exp-avx2-perm-p3.c",
848 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700849 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700850 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700851]
852
Marat Dukhan08c4a432019-10-03 09:29:21 -0700853AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800854 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -0800855 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
856 "src/f32-dwconv/gen/up32x4-avx512f.c",
857 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
858 "src/f32-dwconv/gen/up16x4-avx512f.c",
859 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
860 "src/f32-dwconv/gen/up32x9-avx512f.c",
861 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
862 "src/f32-dwconv/gen/up16x9-avx512f.c",
863 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
864 "src/f32-dwconv/gen/up32x25-avx512f.c",
865 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
866 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -0800867 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
868 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
869 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
870 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
871 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
872 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
873 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
874 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
875 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
876 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
877 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
878 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
879 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
880 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
881 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
882 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
883 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
884 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700885 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700886 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700887 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700888 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700889 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700890 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700891 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700892 "src/math/exp-avx512f-p5-scalef.c",
893 "src/math/exp-avx512f-p5.c",
894 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700895 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700896 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700897]
898
899AARCH32_ASM_UKERNELS = [
900 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard3e237f22019-12-04 23:08:51 -0800901 "src/f32-gemm/4x8-aarch32-neon-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -0800902 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700903]
904
905AARCH64_ASM_UKERNELS = [
906 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
907 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800908 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
909 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
910 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
911 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
912 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
913 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
914 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
915 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
916 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
917 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
918 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
919 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
920 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
921 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
922 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
923 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
924 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
925 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
926 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
927 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
928 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
929 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
930 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
931 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
932 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
933 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
934 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
935 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
936 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
937 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
938 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
939 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
940 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
941 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700942 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700943 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700944 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
945 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
946 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700947 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700948 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
949 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700950 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700951 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
952 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
953 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
954]
955
956INTERNAL_MICROKERNEL_HDRS = [
957 "src/xnnpack/argmaxpool.h",
958 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800959 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700960 "src/xnnpack/clamp.h",
961 "src/xnnpack/common.h",
962 "src/xnnpack/conv.h",
963 "src/xnnpack/dwconv.h",
964 "src/xnnpack/gavgpool.h",
965 "src/xnnpack/gemm.h",
966 "src/xnnpack/hswish.h",
967 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -0800968 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700969 "src/xnnpack/lut.h",
970 "src/xnnpack/math.h",
971 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800972 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700973 "src/xnnpack/packx.h",
974 "src/xnnpack/pad.h",
975 "src/xnnpack/params.h",
976 "src/xnnpack/pavgpool.h",
977 "src/xnnpack/ppmm.h",
978 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700979 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700980 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700981 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700982 "src/xnnpack/rmax.h",
983 "src/xnnpack/scalar-utils.h",
984 "src/xnnpack/spmm.h",
985 "src/xnnpack/unpool.h",
986 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800987 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700988 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700989 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700990 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700991 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800992 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700993 "src/xnnpack/zip.h",
994]
995
996INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
997 "include/xnnpack.h",
998 "src/xnnpack/allocator.h",
999 "src/xnnpack/compute.h",
1000 "src/xnnpack/im2col.h",
1001 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001002 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001003 "src/xnnpack/operator.h",
1004 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001005 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001006 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001007 "src/xnnpack/requantization.h",
1008]
1009
1010ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1011 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001012]
1013
1014MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001015 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001016 "include/xnnpack.h",
1017]
1018
1019MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1020 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001021 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001022 "src/xnnpack/requantization.h",
1023 "include/xnnpack.h",
1024]
1025
1026OPERATOR_TEST_PARAMS_HDRS = [
1027 "src/xnnpack/params.h",
1028 "src/xnnpack/common.h",
1029]
1030
1031WEIGHTS_PACK_HDRS = [
1032 "src/xnnpack/pack.h",
1033 "src/xnnpack/operator.h",
1034 "src/xnnpack/compute.h",
1035]
1036
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001037LOGGING_COPTS = select({
1038 # No logging in optimized mode
1039 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1040 # Full logging in debug mode
1041 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1042 # Error-only logging in default (fastbuild) mode
1043 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1044})
1045
1046LOGGING_HDRS = [
1047 "src/xnnpack/log.h",
1048]
1049
Marat Dukhan08c4a432019-10-03 09:29:21 -07001050xnnpack_cc_library(
1051 name = "scalar_ukernels",
1052 srcs = SCALAR_UKERNELS,
1053 hdrs = INTERNAL_HDRS,
1054 aarch32_copts = ["-marm"],
1055 copts = xnnpack_std_copts(),
1056 deps = [
1057 "@FP16",
1058 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001059 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001060 ],
1061)
1062
1063xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001064 name = "wasm_ukernels",
1065 hdrs = INTERNAL_HDRS,
1066 copts = xnnpack_std_copts(),
1067 wasm_srcs = WASM_UKERNELS,
1068 deps = [
1069 "@FP16",
1070 "@FXdiv",
1071 "@pthreadpool",
1072 ],
1073)
1074
1075xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001076 name = "psimd_ukernels",
1077 srcs = PSIMD_UKERNELS,
1078 hdrs = INTERNAL_HDRS,
1079 aarch32_copts = [
1080 "-marm",
1081 "-mfpu=neon",
1082 ],
1083 copts = xnnpack_std_copts(),
1084 optimized_copts = [
1085 "-O3",
1086 "-ffast-math",
1087 ],
1088 deps = [
1089 "@FP16",
1090 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001091 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001092 ],
1093)
1094
1095xnnpack_cc_library(
1096 name = "neon_ukernels",
1097 hdrs = INTERNAL_HDRS,
1098 aarch32_copts = [
1099 "-marm",
1100 "-mfpu=neon",
1101 ],
1102 aarch32_srcs = NEON_UKERNELS,
1103 aarch64_srcs = NEON_UKERNELS,
1104 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001105 deps = [
1106 "@FP16",
1107 "@pthreadpool",
1108 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001109)
1110
1111xnnpack_cc_library(
1112 name = "neonfma_ukernels",
1113 hdrs = INTERNAL_HDRS,
1114 aarch32_copts = [
1115 "-marm",
1116 "-mfpu=neon-vfpv4",
1117 ],
1118 aarch32_srcs = NEONFMA_UKERNELS,
1119 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1120 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001121 deps = [
1122 "@FP16",
1123 "@pthreadpool",
1124 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001125)
1126
1127xnnpack_cc_library(
1128 name = "neonfp16arith_ukernels",
1129 hdrs = INTERNAL_HDRS,
1130 aarch64_copts = ["-march=armv8.2-a+fp16"],
1131 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1132 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001133 deps = [
1134 "@FP16",
1135 "@pthreadpool",
1136 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001137)
1138
1139xnnpack_cc_library(
1140 name = "sse2_ukernels",
1141 hdrs = INTERNAL_HDRS,
1142 copts = xnnpack_std_copts(),
1143 x86_copts = ["-msse2"],
1144 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001145 deps = [
1146 "@FP16",
1147 "@pthreadpool",
1148 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001149)
1150
1151xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001152 name = "sse41_ukernels",
1153 hdrs = INTERNAL_HDRS,
1154 copts = xnnpack_std_copts(),
1155 x86_copts = ["-msse4.1"],
1156 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001157 deps = [
1158 "@FP16",
1159 "@pthreadpool",
1160 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001161)
1162
1163xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001164 name = "avx_ukernels",
1165 hdrs = INTERNAL_HDRS,
1166 copts = xnnpack_std_copts(),
1167 x86_copts = ["-mavx"],
1168 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001169 deps = [
1170 "@FP16",
1171 "@pthreadpool",
1172 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001173)
1174
1175xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001176 name = "fma3_ukernels",
1177 hdrs = INTERNAL_HDRS,
1178 copts = xnnpack_std_copts(),
1179 x86_copts = [
1180 "-mfma",
1181 ],
1182 x86_srcs = FMA3_UKERNELS,
1183 deps = [
1184 "@FP16",
1185 "@pthreadpool",
1186 ],
1187)
1188
1189xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001190 name = "avx2_ukernels",
1191 hdrs = INTERNAL_HDRS,
1192 copts = xnnpack_std_copts(),
1193 x86_copts = [
1194 "-mfma",
1195 "-mavx2",
1196 ],
1197 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001198 deps = [
1199 "@FP16",
1200 "@pthreadpool",
1201 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001202)
1203
1204xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001205 name = "avx512f_ukernels",
1206 hdrs = INTERNAL_HDRS,
1207 copts = xnnpack_std_copts(),
1208 x86_copts = ["-mavx512f"],
1209 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001210 deps = [
1211 "@FP16",
1212 "@pthreadpool",
1213 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001214)
1215
1216xnnpack_cc_library(
1217 name = "asm_ukernels",
1218 hdrs = ["src/xnnpack/assembly.h"],
1219 aarch32_srcs = AARCH32_ASM_UKERNELS,
1220 aarch64_srcs = AARCH64_ASM_UKERNELS,
1221)
1222
1223xnnpack_aggregate_library(
1224 name = "ukernels",
1225 aarch32_deps = [
1226 ":psimd_ukernels",
1227 ":neon_ukernels",
1228 ":neonfma_ukernels",
1229 ":asm_ukernels",
1230 ],
1231 aarch64_deps = [
1232 ":psimd_ukernels",
1233 ":neon_ukernels",
1234 ":neonfma_ukernels",
1235 ":neonfp16arith_ukernels",
1236 ":asm_ukernels",
1237 ],
1238 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001239 wasm_deps = [
1240 ":wasm_ukernels",
1241 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001242 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001243 ":wasm_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001244 ":psimd_ukernels",
1245 ],
1246 x86_deps = [
1247 ":psimd_ukernels",
1248 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001249 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001250 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001251 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001252 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001253 ":avx512f_ukernels",
1254 ],
1255)
1256
1257xnnpack_cc_library(
1258 name = "im2col",
1259 srcs = ["src/im2col.c"],
1260 hdrs = [
1261 "src/xnnpack/common.h",
1262 "src/xnnpack/im2col.h",
1263 ],
1264 copts = xnnpack_std_copts(),
1265)
1266
1267xnnpack_cc_library(
1268 name = "indirection",
1269 srcs = ["src/indirection.c"],
1270 hdrs = INTERNAL_HDRS,
1271 copts = xnnpack_std_copts(),
1272 deps = [
1273 "@FP16",
1274 "@FXdiv",
1275 "@pthreadpool",
1276 ],
1277)
1278
1279xnnpack_cc_library(
1280 name = "operator_run",
1281 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001282 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1283 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001284 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1285 "-Wno-vla",
1286 ],
1287 deps = [
1288 "@FP16",
1289 "@FXdiv",
1290 "@clog",
1291 "@pthreadpool",
1292 ],
1293)
1294
1295cc_library(
1296 name = "enable_assembly",
1297 defines = select({
1298 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1299 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001300 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001301 }),
1302)
1303
Marat Dukhancf056b22019-10-07 10:26:29 -07001304xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001305 name = "operators",
1306 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001307 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001308 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001309 ],
1310 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001311 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001312 "-Isrc",
1313 "-Iinclude",
1314 ] + select({
1315 ":debug_build": [],
1316 "//conditions:default": xnnpack_min_size_copts(),
1317 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001318 wasm_srcs = ["src/wasm-stubs.c"],
1319 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001320 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001321 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001322 "@FP16",
1323 "@FXdiv",
1324 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001325 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001326 ],
1327)
1328
1329cc_library(
1330 name = "XNNPACK",
1331 srcs = [
1332 "src/init.c",
1333 ],
1334 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1335 "-Isrc",
1336 "-Iinclude",
1337 ] + select({
1338 ":debug_build": [],
1339 "//conditions:default": xnnpack_min_size_copts(),
1340 }),
1341 includes = ["include"],
1342 linkstatic = True,
1343 textual_hdrs = ["include/xnnpack.h"],
1344 visibility = xnnpack_visibility(),
1345 deps = [
1346 ":enable_assembly",
1347 ":ukernels",
1348 ":operator_run",
1349 ":operators",
1350 "@clog",
1351 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001352 ] + select({
1353 ":emscripten": [],
1354 "//conditions:default": ["@cpuinfo"],
1355 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001356)
1357
1358cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001359 name = "xnnpack_operators_nhwc_f32",
1360 srcs = [
1361 "src/init.c",
1362 ],
1363 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1364 "-Isrc",
1365 "-Iinclude",
1366 ] + select({
1367 ":debug_build": [],
1368 "//conditions:default": xnnpack_min_size_copts(),
1369 }),
1370 defines = [
1371 "XNN_NO_Q8_OPERATORS",
1372 "XNN_NO_U8_OPERATORS",
1373 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001374 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001375 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001376 includes = ["include"],
1377 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001378 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001379 visibility = xnnpack_visibility(),
1380 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001381 ":enable_assembly",
1382 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001383 ":operator_run",
1384 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001385 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001386 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001387 ] + select({
1388 ":emscripten": [],
1389 "//conditions:default": ["@cpuinfo"],
1390 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001391)
1392
Marat Dukhancf056b22019-10-07 10:26:29 -07001393xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001394 name = "bench_utils",
1395 srcs = ["bench/utils.cc"],
1396 hdrs = ["bench/utils.h"],
1397 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001398 deps = [
1399 "@com_google_benchmark//:benchmark",
1400 "@cpuinfo",
1401 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001402)
1403
Frank Barchard7e955972019-10-11 10:34:25 -07001404######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001405
1406xnnpack_benchmark(
1407 name = "q8_gemm_bench",
1408 srcs = [
1409 "bench/gemm.h",
1410 "bench/q8-gemm.cc",
1411 "src/xnnpack/AlignedAllocator.h",
1412 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1413 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1414 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1415)
1416
1417xnnpack_benchmark(
1418 name = "f16_gemm_bench",
1419 srcs = [
1420 "bench/f16-gemm.cc",
1421 "bench/gemm.h",
1422 "src/xnnpack/AlignedAllocator.h",
1423 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1424 copts = ["-Wno-unused-function"],
1425 deps = MICROKERNEL_BENCHMARK_DEPS,
1426)
1427
1428xnnpack_benchmark(
1429 name = "f32_igemm_bench",
1430 srcs = [
1431 "bench/f32-igemm.cc",
1432 "bench/conv.h",
1433 "src/xnnpack/AlignedAllocator.h",
1434 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001435 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001436)
1437
1438xnnpack_benchmark(
1439 name = "f32_conv_hwc_bench",
1440 srcs = [
1441 "bench/f32-conv-hwc.cc",
1442 "bench/dconv.h",
1443 "src/xnnpack/AlignedAllocator.h",
1444 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1445 copts = ["-Wno-unused-function"],
1446 deps = MICROKERNEL_BENCHMARK_DEPS,
1447)
1448
1449xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001450 name = "f32_conv_hwc2spchw_bench",
1451 srcs = [
1452 "bench/f32-conv-hwc2spchw.cc",
1453 "bench/dconv.h",
1454 "src/xnnpack/AlignedAllocator.h",
1455 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1456 copts = ["-Wno-unused-function"],
1457 deps = MICROKERNEL_BENCHMARK_DEPS,
1458)
1459
1460xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001461 name = "f32_dwconv_bench",
1462 srcs = [
1463 "bench/f32-dwconv.cc",
1464 "bench/dwconv.h",
1465 "src/xnnpack/AlignedAllocator.h",
1466 ] + MICROKERNEL_BENCHMARK_HDRS,
1467 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1468)
1469
1470xnnpack_benchmark(
1471 name = "f32_dwconv_spchw_bench",
1472 srcs = [
1473 "bench/f32-dwconv-spchw.cc",
1474 "bench/dwconv.h",
1475 "src/xnnpack/AlignedAllocator.h",
1476 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1477 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1478)
1479
1480xnnpack_benchmark(
1481 name = "f32_gemm_bench",
1482 srcs = [
1483 "bench/f32-gemm.cc",
1484 "bench/gemm.h",
1485 "src/xnnpack/AlignedAllocator.h",
1486 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1487 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001488 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001489)
1490
1491xnnpack_benchmark(
1492 name = "f32_rmax_bench",
1493 srcs = [
1494 "bench/f32-rmax.cc",
1495 "src/xnnpack/AlignedAllocator.h",
1496 ] + MICROKERNEL_BENCHMARK_HDRS,
1497 deps = MICROKERNEL_BENCHMARK_DEPS,
1498)
1499
1500xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001501 name = "f32_sigmoid_bench",
1502 srcs = [
1503 "bench/f32-sigmoid.cc",
1504 "src/xnnpack/AlignedAllocator.h",
1505 ] + MICROKERNEL_BENCHMARK_HDRS,
1506 copts = ["-Wno-unused-function"],
1507 deps = MICROKERNEL_BENCHMARK_DEPS,
1508)
1509
1510xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001511 name = "f32_spmm_bench",
1512 srcs = [
1513 "bench/f32-spmm.cc",
1514 "bench/gemm.h",
1515 "src/xnnpack/AlignedAllocator.h",
1516 ] + MICROKERNEL_BENCHMARK_HDRS,
1517 copts = ["-Wno-unused-function"],
1518 deps = MICROKERNEL_BENCHMARK_DEPS,
1519)
1520
1521xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001522 name = "f32_softargmax_bench",
1523 srcs = [
1524 "bench/f32-softargmax.cc",
1525 ] + MICROKERNEL_BENCHMARK_HDRS,
1526 copts = ["-Wno-unused-function"],
1527 deps = MICROKERNEL_BENCHMARK_DEPS,
1528)
1529
1530xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001531 name = "f32_im2col_gemm_bench",
1532 srcs = [
1533 "bench/f32-im2col-gemm.cc",
1534 "bench/conv.h",
1535 "src/xnnpack/AlignedAllocator.h",
1536 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1537 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1538)
1539
1540########################### Benchmarks for operators ###########################
1541
1542xnnpack_benchmark(
1543 name = "add_bench",
1544 srcs = ["bench/add.cc"],
1545 deps = OPERATOR_BENCHMARK_DEPS,
1546)
1547
1548xnnpack_benchmark(
1549 name = "average_pooling_bench",
1550 srcs = ["bench/average-pooling.cc"],
1551 deps = OPERATOR_BENCHMARK_DEPS,
1552)
1553
1554xnnpack_benchmark(
1555 name = "channel_shuffle_bench",
1556 srcs = ["bench/channel-shuffle.cc"],
1557 deps = OPERATOR_BENCHMARK_DEPS,
1558)
1559
1560xnnpack_benchmark(
1561 name = "convolution_bench",
1562 srcs = ["bench/convolution.cc"],
1563 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1564 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1565)
1566
1567xnnpack_benchmark(
1568 name = "deconvolution_bench",
1569 srcs = ["bench/deconvolution.cc"],
1570 copts = xnnpack_optional_tflite_copts(),
1571 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1572)
1573
1574xnnpack_benchmark(
1575 name = "global_average_pooling_bench",
1576 srcs = ["bench/global-average-pooling.cc"],
1577 deps = OPERATOR_BENCHMARK_DEPS,
1578)
1579
1580xnnpack_benchmark(
1581 name = "max_pooling_bench",
1582 srcs = ["bench/max-pooling.cc"],
1583 deps = OPERATOR_BENCHMARK_DEPS,
1584)
1585
1586xnnpack_benchmark(
1587 name = "sigmoid_bench",
1588 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001589 copts = xnnpack_optional_tflite_copts(),
1590 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001591)
1592
1593xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001594 name = "prelu_bench",
1595 srcs = ["bench/prelu.cc"],
1596 copts = xnnpack_optional_tflite_copts(),
1597 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1598)
1599
1600xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001601 name = "softargmax_bench",
1602 srcs = ["bench/softargmax.cc"],
1603 deps = OPERATOR_BENCHMARK_DEPS,
1604)
1605
Marat Dukhanc068bb62019-10-04 13:24:39 -07001606############################# End-to-end benchmarks ############################
1607
1608cc_library(
1609 name = "mobilenet_v1",
1610 srcs = ["models/mobilenet-v1.cc"],
1611 hdrs = ["models/models.h"],
1612 linkstatic = True,
1613 deps = [
1614 ":XNNPACK",
1615 "@pthreadpool",
1616 ],
1617)
1618
1619cc_library(
1620 name = "mobilenet_v2",
1621 srcs = ["models/mobilenet-v2.cc"],
1622 hdrs = ["models/models.h"],
1623 linkstatic = True,
1624 deps = [
1625 ":XNNPACK",
1626 "@pthreadpool",
1627 ],
1628)
1629
1630xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001631 name = "f32_dwconv_e2e_bench",
1632 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1633 copts = ["-Wno-unused-function"],
1634 deps = MICROKERNEL_BENCHMARK_DEPS + [
1635 ":XNNPACK",
1636 ":mobilenet_v1",
1637 ":mobilenet_v2",
1638 ],
1639)
1640
1641xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001642 name = "f32_gemm_e2e_bench",
1643 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1644 copts = ["-Wno-unused-function"],
1645 deps = MICROKERNEL_BENCHMARK_DEPS + [
1646 ":XNNPACK",
1647 ":mobilenet_v1",
1648 ":mobilenet_v2",
1649 ],
1650)
1651
1652xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001653 name = "end2end_bench",
1654 srcs = ["bench/end2end.cc"],
1655 deps = [
1656 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001657 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001658 ":mobilenet_v1",
1659 ":mobilenet_v2",
1660 "@pthreadpool",
1661 ],
1662)
1663
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001664#################### Accuracy evaluation for math functions ####################
1665
1666xnnpack_benchmark(
1667 name = "f32_exp_eval",
1668 srcs = [
1669 "eval/f32-exp.cc",
1670 "src/xnnpack/AlignedAllocator.h",
1671 ] + ACCURACY_EVAL_HDRS,
1672 deps = ACCURACY_EVAL_DEPS,
1673)
1674
Marat Dukhan515c9772019-10-17 18:07:57 -07001675xnnpack_benchmark(
1676 name = "f32_expminus_eval",
1677 srcs = [
1678 "eval/f32-expminus.cc",
1679 "src/xnnpack/AlignedAllocator.h",
1680 ] + ACCURACY_EVAL_HDRS,
1681 deps = ACCURACY_EVAL_DEPS,
1682)
1683
Marat Dukhan98ba4412019-10-23 02:14:28 -07001684xnnpack_benchmark(
1685 name = "f32_extexp_eval",
1686 srcs = [
1687 "eval/f32-extexp.cc",
1688 "src/xnnpack/AlignedAllocator.h",
1689 ] + ACCURACY_EVAL_HDRS,
1690 deps = ACCURACY_EVAL_DEPS,
1691)
1692
Marat Dukhan346a9e52019-11-15 09:06:30 -08001693xnnpack_benchmark(
1694 name = "f32_sigmoid_eval",
1695 srcs = [
1696 "eval/f32-sigmoid.cc",
1697 "src/xnnpack/AlignedAllocator.h",
1698 ] + ACCURACY_EVAL_HDRS,
1699 deps = ACCURACY_EVAL_DEPS,
1700)
1701
Marat Dukhan08c4a432019-10-03 09:29:21 -07001702######################### Unit tests for micro-kernels #########################
1703
1704xnnpack_unit_test(
1705 name = "f16_gemm_test",
1706 srcs = [
1707 "test/f16-gemm.cc",
1708 "test/gemm-microkernel-tester.h",
1709 "src/xnnpack/AlignedAllocator.h",
1710 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1711 deps = MICROKERNEL_TEST_DEPS,
1712)
1713
1714xnnpack_unit_test(
1715 name = "f32_argmaxpool_test",
1716 srcs = [
1717 "test/f32-argmaxpool.cc",
1718 "test/argmaxpool-microkernel-tester.h",
1719 "src/xnnpack/AlignedAllocator.h",
1720 ] + MICROKERNEL_TEST_HDRS,
1721 deps = MICROKERNEL_TEST_DEPS,
1722)
1723
1724xnnpack_unit_test(
1725 name = "f32_avgpool_test",
1726 srcs = [
1727 "test/f32-avgpool.cc",
1728 "test/avgpool-microkernel-tester.h",
1729 "src/xnnpack/AlignedAllocator.h",
1730 ] + MICROKERNEL_TEST_HDRS,
1731 deps = MICROKERNEL_TEST_DEPS,
1732)
1733
1734xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001735 name = "f32_bilinear_test",
1736 srcs = [
1737 "test/f32-bilinear.cc",
1738 "test/bilinear-microkernel-tester.h",
1739 "src/xnnpack/AlignedAllocator.h",
1740 ] + MICROKERNEL_TEST_HDRS,
1741 deps = MICROKERNEL_TEST_DEPS,
1742)
1743
1744xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001745 name = "f32_clamp_test",
1746 srcs = [
1747 "test/f32-clamp.cc",
1748 "test/clamp-microkernel-tester.h",
1749 ] + MICROKERNEL_TEST_HDRS,
1750 deps = MICROKERNEL_TEST_DEPS,
1751)
1752
1753xnnpack_unit_test(
1754 name = "f32_igemm_test",
1755 srcs = [
1756 "test/f32-igemm.cc",
1757 "test/gemm-microkernel-tester.h",
1758 "src/xnnpack/AlignedAllocator.h",
1759 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1760 deps = MICROKERNEL_TEST_DEPS,
1761)
1762
1763xnnpack_unit_test(
1764 name = "f32_conv_hwc_test",
1765 srcs = [
1766 "test/f32-conv-hwc.cc",
1767 "test/conv-hwc-microkernel-tester.h",
1768 "src/xnnpack/AlignedAllocator.h",
1769 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1770 deps = MICROKERNEL_TEST_DEPS,
1771)
1772
1773xnnpack_unit_test(
1774 name = "f32_conv_hwc2spchw_test",
1775 srcs = [
1776 "test/f32-conv-hwc2spchw.cc",
1777 "test/conv-hwc2spchw-microkernel-tester.h",
1778 "src/xnnpack/AlignedAllocator.h",
1779 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1780 deps = MICROKERNEL_TEST_DEPS,
1781)
1782
1783xnnpack_unit_test(
1784 name = "f32_dwconv_test",
1785 srcs = [
1786 "test/f32-dwconv.cc",
1787 "test/dwconv-microkernel-tester.h",
1788 "src/xnnpack/AlignedAllocator.h",
1789 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1790 deps = MICROKERNEL_TEST_DEPS,
1791)
1792
1793xnnpack_unit_test(
1794 name = "f32_dwconv_spchw_test",
1795 srcs = [
1796 "test/f32-dwconv-spchw.cc",
1797 "test/dwconv-spchw-microkernel-tester.h",
1798 "src/xnnpack/AlignedAllocator.h",
1799 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1800 deps = MICROKERNEL_TEST_DEPS,
1801)
1802
1803xnnpack_unit_test(
1804 name = "f32_gavgpool_test",
1805 srcs = [
1806 "test/f32-gavgpool.cc",
1807 "test/gavgpool-microkernel-tester.h",
1808 "src/xnnpack/AlignedAllocator.h",
1809 ] + MICROKERNEL_TEST_HDRS,
1810 deps = MICROKERNEL_TEST_DEPS,
1811)
1812
1813xnnpack_unit_test(
1814 name = "f32_gavgpool_spchw_test",
1815 srcs = [
1816 "test/f32-gavgpool-spchw.cc",
1817 "test/gavgpool-spchw-microkernel-tester.h",
1818 "src/xnnpack/AlignedAllocator.h",
1819 ] + MICROKERNEL_TEST_HDRS,
1820 deps = MICROKERNEL_TEST_DEPS,
1821)
1822
1823xnnpack_unit_test(
1824 name = "f32_gemm_test",
1825 srcs = [
1826 "test/f32-gemm.cc",
1827 "test/gemm-microkernel-tester.h",
1828 "src/xnnpack/AlignedAllocator.h",
1829 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1830 deps = MICROKERNEL_TEST_DEPS,
1831)
1832
1833xnnpack_unit_test(
1834 name = "f32_gemminc_test",
1835 srcs = [
1836 "test/f32-gemminc.cc",
1837 "test/gemm-microkernel-tester.h",
1838 "src/xnnpack/AlignedAllocator.h",
1839 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1840 deps = MICROKERNEL_TEST_DEPS,
1841)
1842
1843xnnpack_unit_test(
1844 name = "f32_hswish_test",
1845 srcs = [
1846 "test/f32-hswish.cc",
1847 "test/hswish-microkernel-tester.h",
1848 ] + MICROKERNEL_TEST_HDRS,
1849 deps = MICROKERNEL_TEST_DEPS,
1850)
1851
1852xnnpack_unit_test(
1853 name = "f32_maxpool_test",
1854 srcs = [
1855 "test/f32-maxpool.cc",
1856 "test/maxpool-microkernel-tester.h",
1857 ] + MICROKERNEL_TEST_HDRS,
1858 deps = MICROKERNEL_TEST_DEPS,
1859)
1860
1861xnnpack_unit_test(
1862 name = "f32_pavgpool_test",
1863 srcs = [
1864 "test/f32-pavgpool.cc",
1865 "test/avgpool-microkernel-tester.h",
1866 "src/xnnpack/AlignedAllocator.h",
1867 ] + MICROKERNEL_TEST_HDRS,
1868 deps = MICROKERNEL_TEST_DEPS,
1869)
1870
1871xnnpack_unit_test(
1872 name = "f32_ppmm_test",
1873 srcs = [
1874 "test/f32-ppmm.cc",
1875 "test/gemm-microkernel-tester.h",
1876 "src/xnnpack/AlignedAllocator.h",
1877 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1878 deps = MICROKERNEL_TEST_DEPS,
1879)
1880
1881xnnpack_unit_test(
1882 name = "f32_prelu_test",
1883 srcs = [
1884 "test/f32-prelu.cc",
1885 "test/prelu-microkernel-tester.h",
1886 "src/xnnpack/AlignedAllocator.h",
1887 ] + MICROKERNEL_TEST_HDRS,
1888 deps = MICROKERNEL_TEST_DEPS,
1889)
1890
1891xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001892 name = "f32_raddexpminusmax_test",
1893 srcs = [
1894 "test/f32-raddexpminusmax.cc",
1895 "test/raddexpminusmax-microkernel-tester.h",
1896 ] + MICROKERNEL_TEST_HDRS,
1897 deps = MICROKERNEL_TEST_DEPS,
1898)
1899
1900xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001901 name = "f32_raddextexp_test",
1902 srcs = [
1903 "test/f32-raddextexp.cc",
1904 "test/raddextexp-microkernel-tester.h",
1905 ] + MICROKERNEL_TEST_HDRS,
1906 deps = MICROKERNEL_TEST_DEPS,
1907)
1908
1909xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001910 name = "f32_raddstoreexpminusmax_test",
1911 srcs = [
1912 "test/f32-raddstoreexpminusmax.cc",
1913 "test/raddstoreexpminusmax-microkernel-tester.h",
1914 ] + MICROKERNEL_TEST_HDRS,
1915 deps = MICROKERNEL_TEST_DEPS,
1916)
1917
1918xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001919 name = "f32_rmax_test",
1920 srcs = [
1921 "test/f32-rmax.cc",
1922 "test/rmax-microkernel-tester.h",
1923 ] + MICROKERNEL_TEST_HDRS,
1924 deps = MICROKERNEL_TEST_DEPS,
1925)
1926
1927xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08001928 name = "f32_sigmoid_test",
1929 srcs = [
1930 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001931 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001932 ] + MICROKERNEL_TEST_HDRS,
1933 deps = MICROKERNEL_TEST_DEPS,
1934)
1935
1936xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001937 name = "f32_spmm_test",
1938 srcs = [
1939 "test/f32-spmm.cc",
1940 "test/spmm-microkernel-tester.h",
1941 "src/xnnpack/AlignedAllocator.h",
1942 ] + MICROKERNEL_TEST_HDRS,
1943 deps = MICROKERNEL_TEST_DEPS,
1944)
1945
1946xnnpack_unit_test(
1947 name = "f32_vadd_test",
1948 srcs = [
1949 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001950 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001951 ] + MICROKERNEL_TEST_HDRS,
1952 deps = MICROKERNEL_TEST_DEPS,
1953)
1954
1955xnnpack_unit_test(
1956 name = "f32_vaddc_test",
1957 srcs = [
1958 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001959 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001960 ] + MICROKERNEL_TEST_HDRS,
1961 deps = MICROKERNEL_TEST_DEPS,
1962)
1963
1964xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08001965 name = "f32_vdiv_test",
1966 srcs = [
1967 "test/f32-vdiv.cc",
1968 "test/vbinary-microkernel-tester.h",
1969 ] + MICROKERNEL_TEST_HDRS,
1970 deps = MICROKERNEL_TEST_DEPS,
1971)
1972
1973xnnpack_unit_test(
1974 name = "f32_vdivc_test",
1975 srcs = [
1976 "test/f32-vdivc.cc",
1977 "test/vbinaryc-microkernel-tester.h",
1978 ] + MICROKERNEL_TEST_HDRS,
1979 deps = MICROKERNEL_TEST_DEPS,
1980)
1981
1982xnnpack_unit_test(
1983 name = "f32_vrdivc_test",
1984 srcs = [
1985 "test/f32-vrdivc.cc",
1986 "test/vbinaryc-microkernel-tester.h",
1987 ] + MICROKERNEL_TEST_HDRS,
1988 deps = MICROKERNEL_TEST_DEPS,
1989)
1990
1991xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08001992 name = "f32_vmax_test",
1993 srcs = [
1994 "test/f32-vmax.cc",
1995 "test/vbinary-microkernel-tester.h",
1996 ] + MICROKERNEL_TEST_HDRS,
1997 deps = MICROKERNEL_TEST_DEPS,
1998)
1999
2000xnnpack_unit_test(
2001 name = "f32_vmaxc_test",
2002 srcs = [
2003 "test/f32-vmaxc.cc",
2004 "test/vbinaryc-microkernel-tester.h",
2005 ] + MICROKERNEL_TEST_HDRS,
2006 deps = MICROKERNEL_TEST_DEPS,
2007)
2008
2009xnnpack_unit_test(
2010 name = "f32_vmin_test",
2011 srcs = [
2012 "test/f32-vmin.cc",
2013 "test/vbinary-microkernel-tester.h",
2014 ] + MICROKERNEL_TEST_HDRS,
2015 deps = MICROKERNEL_TEST_DEPS,
2016)
2017
2018xnnpack_unit_test(
2019 name = "f32_vminc_test",
2020 srcs = [
2021 "test/f32-vminc.cc",
2022 "test/vbinaryc-microkernel-tester.h",
2023 ] + MICROKERNEL_TEST_HDRS,
2024 deps = MICROKERNEL_TEST_DEPS,
2025)
2026
2027xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002028 name = "f32_vmul_test",
2029 srcs = [
2030 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002031 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002032 ] + MICROKERNEL_TEST_HDRS,
2033 deps = MICROKERNEL_TEST_DEPS,
2034)
2035
2036xnnpack_unit_test(
2037 name = "f32_vmulc_test",
2038 srcs = [
2039 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002040 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002041 ] + MICROKERNEL_TEST_HDRS,
2042 deps = MICROKERNEL_TEST_DEPS,
2043)
2044
2045xnnpack_unit_test(
2046 name = "f32_vmulcaddc_test",
2047 srcs = [
2048 "test/f32-vmulcaddc.cc",
2049 "test/vmulcaddc-microkernel-tester.h",
2050 "src/xnnpack/AlignedAllocator.h",
2051 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2052 deps = MICROKERNEL_TEST_DEPS,
2053)
2054
2055xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002056 name = "f32_vscale_test",
2057 srcs = [
2058 "test/f32-vscale.cc",
2059 "test/vscale-microkernel-tester.h",
2060 ] + MICROKERNEL_TEST_HDRS,
2061 deps = MICROKERNEL_TEST_DEPS,
2062)
2063
2064xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002065 name = "f32_vscaleexpminusmax_test",
2066 srcs = [
2067 "test/f32-vscaleexpminusmax.cc",
2068 "test/vscaleexpminusmax-microkernel-tester.h",
2069 ] + MICROKERNEL_TEST_HDRS,
2070 deps = MICROKERNEL_TEST_DEPS,
2071)
2072
2073xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002074 name = "f32_vscaleextexp_test",
2075 srcs = [
2076 "test/f32-vscaleextexp.cc",
2077 "test/vscaleextexp-microkernel-tester.h",
2078 ] + MICROKERNEL_TEST_HDRS,
2079 deps = MICROKERNEL_TEST_DEPS,
2080)
2081
2082xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002083 name = "f32_vsub_test",
2084 srcs = [
2085 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002086 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002087 ] + MICROKERNEL_TEST_HDRS,
2088 deps = MICROKERNEL_TEST_DEPS,
2089)
2090
2091xnnpack_unit_test(
2092 name = "f32_vsubc_test",
2093 srcs = [
2094 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002095 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002096 ] + MICROKERNEL_TEST_HDRS,
2097 deps = MICROKERNEL_TEST_DEPS,
2098)
2099
2100xnnpack_unit_test(
2101 name = "f32_vrsubc_test",
2102 srcs = [
2103 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002104 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002105 ] + MICROKERNEL_TEST_HDRS,
2106 deps = MICROKERNEL_TEST_DEPS,
2107)
2108
2109xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002110 name = "q8_avgpool_test",
2111 srcs = [
2112 "test/q8-avgpool.cc",
2113 "test/avgpool-microkernel-tester.h",
2114 "src/xnnpack/AlignedAllocator.h",
2115 ] + MICROKERNEL_TEST_HDRS,
2116 deps = MICROKERNEL_TEST_DEPS,
2117)
2118
2119xnnpack_unit_test(
2120 name = "q8_igemm_test",
2121 srcs = [
2122 "test/q8-igemm.cc",
2123 "test/gemm-microkernel-tester.h",
2124 "src/xnnpack/AlignedAllocator.h",
2125 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2126 deps = MICROKERNEL_TEST_DEPS,
2127)
2128
2129xnnpack_unit_test(
2130 name = "q8_dwconv_test",
2131 srcs = [
2132 "test/q8-dwconv.cc",
2133 "test/dwconv-microkernel-tester.h",
2134 "src/xnnpack/AlignedAllocator.h",
2135 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2136 deps = MICROKERNEL_TEST_DEPS,
2137)
2138
2139xnnpack_unit_test(
2140 name = "q8_gavgpool_test",
2141 srcs = [
2142 "test/q8-gavgpool.cc",
2143 "test/gavgpool-microkernel-tester.h",
2144 "src/xnnpack/AlignedAllocator.h",
2145 ] + MICROKERNEL_TEST_HDRS,
2146 deps = MICROKERNEL_TEST_DEPS,
2147)
2148
2149xnnpack_unit_test(
2150 name = "q8_gemm_test",
2151 srcs = [
2152 "test/q8-gemm.cc",
2153 "test/gemm-microkernel-tester.h",
2154 "src/xnnpack/AlignedAllocator.h",
2155 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2156 deps = MICROKERNEL_TEST_DEPS,
2157)
2158
2159xnnpack_unit_test(
2160 name = "q8_vadd_test",
2161 srcs = [
2162 "test/q8-vadd.cc",
2163 "test/vadd-microkernel-tester.h",
2164 ] + MICROKERNEL_TEST_HDRS,
2165 deps = MICROKERNEL_TEST_DEPS,
2166)
2167
2168xnnpack_unit_test(
2169 name = "u8_clamp_test",
2170 srcs = [
2171 "test/u8-clamp.cc",
2172 "test/clamp-microkernel-tester.h",
2173 ] + MICROKERNEL_TEST_HDRS,
2174 deps = MICROKERNEL_TEST_DEPS,
2175)
2176
2177xnnpack_unit_test(
2178 name = "u8_lut32norm_test",
2179 srcs = [
2180 "test/u8-lut32norm.cc",
2181 "test/lut-norm-microkernel-tester.h",
2182 ] + MICROKERNEL_TEST_HDRS,
2183 deps = MICROKERNEL_TEST_DEPS,
2184)
2185
2186xnnpack_unit_test(
2187 name = "u8_maxpool_test",
2188 srcs = [
2189 "test/u8-maxpool.cc",
2190 "test/maxpool-microkernel-tester.h",
2191 ] + MICROKERNEL_TEST_HDRS,
2192 deps = MICROKERNEL_TEST_DEPS,
2193)
2194
2195xnnpack_unit_test(
2196 name = "u8_rmax_test",
2197 srcs = [
2198 "test/u8-rmax.cc",
2199 "test/rmax-microkernel-tester.h",
2200 ] + MICROKERNEL_TEST_HDRS,
2201 deps = MICROKERNEL_TEST_DEPS,
2202)
2203
2204xnnpack_unit_test(
2205 name = "x32_packx_test",
2206 srcs = [
2207 "test/x32-packx.cc",
2208 "test/pack-microkernel-tester.h",
2209 "src/xnnpack/AlignedAllocator.h",
2210 ] + MICROKERNEL_TEST_HDRS,
2211 deps = MICROKERNEL_TEST_DEPS,
2212)
2213
2214xnnpack_unit_test(
2215 name = "x32_pad_test",
2216 srcs = [
2217 "test/x32-pad.cc",
2218 "test/pad-microkernel-tester.h",
2219 ] + MICROKERNEL_TEST_HDRS,
2220 deps = MICROKERNEL_TEST_DEPS,
2221)
2222
2223xnnpack_unit_test(
2224 name = "x32_unpool_test",
2225 srcs = [
2226 "test/x32-unpool.cc",
2227 "test/unpool-microkernel-tester.h",
2228 ] + MICROKERNEL_TEST_HDRS,
2229 deps = MICROKERNEL_TEST_DEPS,
2230)
2231
2232xnnpack_unit_test(
2233 name = "x32_zip_test",
2234 srcs = [
2235 "test/x32-zip.cc",
2236 "test/zip-microkernel-tester.h",
2237 ] + MICROKERNEL_TEST_HDRS,
2238 deps = MICROKERNEL_TEST_DEPS,
2239)
2240
2241xnnpack_unit_test(
2242 name = "x8_lut_test",
2243 srcs = [
2244 "test/x8-lut.cc",
2245 "test/lut-microkernel-tester.h",
2246 ] + MICROKERNEL_TEST_HDRS,
2247 deps = MICROKERNEL_TEST_DEPS,
2248)
2249
2250xnnpack_unit_test(
2251 name = "x8_zip_test",
2252 srcs = [
2253 "test/x8-zip.cc",
2254 "test/zip-microkernel-tester.h",
2255 ] + MICROKERNEL_TEST_HDRS,
2256 deps = MICROKERNEL_TEST_DEPS,
2257)
2258
2259########################### Size test for the library ##########################
2260
2261xnnpack_binary(
2262 name = "size_test",
2263 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002264 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002265)
2266
2267########################### Unit tests for operators ###########################
2268
2269xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002270 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002271 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002272 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002273 "test/add-operator-tester.h",
2274 ],
2275 deps = OPERATOR_TEST_DEPS,
2276)
2277
2278xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002279 name = "add_nd_test",
2280 srcs = [
2281 "test/add-nd.cc",
2282 "test/binary-elementwise-operator-tester.h",
2283 ],
2284 deps = OPERATOR_TEST_DEPS,
2285)
2286
2287xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002288 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002290 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002291 "test/argmax-pooling-operator-tester.h",
2292 ] + OPERATOR_TEST_PARAMS_HDRS,
2293 deps = OPERATOR_TEST_DEPS,
2294)
2295
2296xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002297 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002298 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002299 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002300 "test/average-pooling-operator-tester.h",
2301 ] + OPERATOR_TEST_PARAMS_HDRS,
2302 deps = OPERATOR_TEST_DEPS,
2303)
2304
2305xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002306 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002307 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002308 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002309 "test/channel-pad-operator-tester.h",
2310 ] + OPERATOR_TEST_PARAMS_HDRS,
2311 deps = OPERATOR_TEST_DEPS,
2312)
2313
2314xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002315 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002316 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002317 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002318 "test/channel-shuffle-operator-tester.h",
2319 ],
2320 deps = OPERATOR_TEST_DEPS,
2321)
2322
2323xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002324 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002326 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002327 "test/clamp-operator-tester.h",
2328 ],
2329 deps = OPERATOR_TEST_DEPS,
2330)
2331
2332xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002333 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002334 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002335 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002336 "test/convolution-operator-tester.h",
2337 ],
2338 deps = OPERATOR_TEST_DEPS,
2339)
2340
2341xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002342 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002343 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002344 "test/convolution-nchw.cc",
2345 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002346 ],
2347 deps = OPERATOR_TEST_DEPS,
2348)
2349
2350xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002351 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002353 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002354 "test/deconvolution-operator-tester.h",
2355 ] + OPERATOR_TEST_PARAMS_HDRS,
2356 deps = OPERATOR_TEST_DEPS,
2357)
2358
2359xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002360 name = "divide_nd_test",
2361 srcs = [
2362 "test/binary-elementwise-operator-tester.h",
2363 "test/divide-nd.cc",
2364 ],
2365 deps = OPERATOR_TEST_DEPS,
2366)
2367
2368xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002369 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002370 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002371 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002372 "test/fully-connected-operator-tester.h",
2373 ],
2374 deps = OPERATOR_TEST_DEPS,
2375)
2376
2377xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002378 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002379 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002380 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002381 "test/global-average-pooling-operator-tester.h",
2382 ] + OPERATOR_TEST_PARAMS_HDRS,
2383 deps = OPERATOR_TEST_DEPS,
2384)
2385
2386xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002387 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002388 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002389 "test/global-average-pooling-ncw.cc",
2390 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391 ],
2392 deps = OPERATOR_TEST_DEPS,
2393)
2394
2395xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002396 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002397 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002398 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002399 "test/hardswish-operator-tester.h",
2400 ],
2401 deps = OPERATOR_TEST_DEPS,
2402)
2403
2404xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002405 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002406 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002407 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002408 "test/leaky-relu-operator-tester.h",
2409 ],
2410 deps = OPERATOR_TEST_DEPS,
2411)
2412
2413xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002414 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002415 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002416 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002417 "test/max-pooling-operator-tester.h",
2418 ] + OPERATOR_TEST_PARAMS_HDRS,
2419 deps = OPERATOR_TEST_DEPS,
2420)
2421
2422xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08002423 name = "maximum_nd_test",
2424 srcs = [
2425 "test/binary-elementwise-operator-tester.h",
2426 "test/maximum-nd.cc",
2427 ],
2428 deps = OPERATOR_TEST_DEPS,
2429)
2430
2431xnnpack_unit_test(
2432 name = "minimum_nd_test",
2433 srcs = [
2434 "test/binary-elementwise-operator-tester.h",
2435 "test/minimum-nd.cc",
2436 ],
2437 deps = OPERATOR_TEST_DEPS,
2438)
2439
2440xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002441 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002442 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002443 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002444 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002445 ],
2446 deps = OPERATOR_TEST_DEPS,
2447)
2448
2449xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002450 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002451 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002452 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002453 "test/prelu-operator-tester.h",
2454 ] + OPERATOR_TEST_PARAMS_HDRS,
2455 deps = OPERATOR_TEST_DEPS,
2456)
2457
2458xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002459 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002460 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002461 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002462 "test/resize-bilinear-operator-tester.h",
2463 ] + OPERATOR_TEST_PARAMS_HDRS,
2464 deps = OPERATOR_TEST_DEPS,
2465)
2466
2467xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002468 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002469 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002470 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002471 "test/sigmoid-operator-tester.h",
2472 ],
2473 deps = OPERATOR_TEST_DEPS,
2474)
2475
2476xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002477 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002478 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002479 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002480 "test/softargmax-operator-tester.h",
2481 ],
2482 deps = OPERATOR_TEST_DEPS,
2483)
2484
2485xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08002486 name = "subtract_nd_test",
2487 srcs = [
2488 "test/binary-elementwise-operator-tester.h",
2489 "test/subtract-nd.cc",
2490 ],
2491 deps = OPERATOR_TEST_DEPS,
2492)
2493
2494xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002495 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002496 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002497 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002498 "test/unpooling-operator-tester.h",
2499 ],
2500 deps = OPERATOR_TEST_DEPS,
2501)
2502
2503############################# Build configurations #############################
2504
Marat Dukhanb8642352019-10-30 15:43:02 -07002505# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002506config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002507 name = "xnn_enable_assembly_explicit_true",
2508 define_values = {"xnn_enable_assembly": "true"},
2509)
2510
2511# Disables usage of assembly kernels.
2512config_setting(
2513 name = "xnn_enable_assembly_explicit_false",
2514 define_values = {"xnn_enable_assembly": "false"},
2515)
2516
2517# Builds with -c dbg
2518config_setting(
2519 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002520 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002521 "compilation_mode": "dbg",
2522 },
2523)
2524
2525# Builds with -c opt
2526config_setting(
2527 name = "optimized_build",
2528 values = {
2529 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002530 },
2531)
2532
2533config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002534 name = "linux_k8",
2535 values = {"cpu": "k8"},
2536)
2537
2538config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002539 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002540 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002541)
2542
2543config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002544 name = "android",
2545 values = {"crosstool_top": "//external:android/crosstool"},
2546)
2547
2548config_setting(
2549 name = "android_armv7",
2550 values = {
2551 "crosstool_top": "//external:android/crosstool",
2552 "cpu": "armeabi-v7a",
2553 },
2554)
2555
2556config_setting(
2557 name = "android_arm64",
2558 values = {
2559 "crosstool_top": "//external:android/crosstool",
2560 "cpu": "arm64-v8a",
2561 },
2562)
2563
2564config_setting(
2565 name = "android_x86",
2566 values = {
2567 "crosstool_top": "//external:android/crosstool",
2568 "cpu": "x86",
2569 },
2570)
2571
2572config_setting(
2573 name = "android_x86_64",
2574 values = {
2575 "crosstool_top": "//external:android/crosstool",
2576 "cpu": "x86_64",
2577 },
2578)
2579
2580config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002581 name = "macos_x86_64",
2582 values = {
2583 "apple_platform_type": "macos",
2584 "cpu": "darwin",
2585 },
2586)
2587
2588config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002589 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002590 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002591)
2592
2593config_setting(
2594 name = "emscripten_wasm",
2595 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002596 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002597 "cpu": "wasm",
2598 },
2599)
2600
2601config_setting(
2602 name = "emscripten_wasmsimd",
2603 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002604 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002605 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002606 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002607 },
2608)
2609
2610config_setting(
2611 name = "emscripten_asmjs",
2612 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002613 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002614 "cpu": "asmjs",
2615 },
2616)