blob: 8b5d6dfbfb8c697cb3f4041a77cc6471562e0a66 [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 Dukhana84e40b2019-12-11 15:38:03 -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_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan08c4a432019-10-03 09:29:21 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
22MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
46OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
52OPERATOR_SRCS = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
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 Dukhan662faa02019-12-09 22:48:16 -0800115 "src/f32-hswish/gen/scalar-x1.c",
116 "src/f32-hswish/gen/scalar-x2.c",
117 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800118 "src/f32-igemm/gen/1x4-scalar.c",
119 "src/f32-igemm/gen/2x4-scalar.c",
120 "src/f32-igemm/gen/4x2-scalar.c",
121 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800122 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700123 "src/f32-pavgpool/mp9p8q-scalar.c",
124 "src/f32-pavgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800125 "src/f32-ppmm/gen/2x4-scalar.c",
126 "src/f32-ppmm/gen/3x3-scalar.c",
127 "src/f32-ppmm/gen/4x2-scalar.c",
128 "src/f32-ppmm/gen/4x4-scalar.c",
129 "src/f32-prelu/gen/scalar-2x1.c",
130 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700131 "src/f32-rmax/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800132 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
133 "src/f32-spmm/gen/1x1-scalar.c",
134 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
135 "src/f32-spmm/gen/2x1-scalar.c",
136 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
137 "src/f32-spmm/gen/4x1-scalar.c",
138 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
139 "src/f32-spmm/gen/8x1-scalar.c",
140 "src/f32-spmm/gen/8x2-scalar.c",
141 "src/f32-spmm/gen/8x4-scalar.c",
142 "src/f32-vbinary/gen/vadd-scalar-x1.c",
143 "src/f32-vbinary/gen/vadd-scalar-x2.c",
144 "src/f32-vbinary/gen/vadd-scalar-x4.c",
145 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
146 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
147 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800148 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
149 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
150 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
151 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
152 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
153 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800154 "src/f32-vbinary/gen/vmax-scalar-x1.c",
155 "src/f32-vbinary/gen/vmax-scalar-x2.c",
156 "src/f32-vbinary/gen/vmax-scalar-x4.c",
157 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
158 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
159 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
160 "src/f32-vbinary/gen/vmin-scalar-x1.c",
161 "src/f32-vbinary/gen/vmin-scalar-x2.c",
162 "src/f32-vbinary/gen/vmin-scalar-x4.c",
163 "src/f32-vbinary/gen/vminc-scalar-x1.c",
164 "src/f32-vbinary/gen/vminc-scalar-x2.c",
165 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800166 "src/f32-vbinary/gen/vmul-scalar-x1.c",
167 "src/f32-vbinary/gen/vmul-scalar-x2.c",
168 "src/f32-vbinary/gen/vmul-scalar-x4.c",
169 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
170 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
171 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800172 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
173 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
174 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800175 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
176 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
177 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
178 "src/f32-vbinary/gen/vsub-scalar-x1.c",
179 "src/f32-vbinary/gen/vsub-scalar-x2.c",
180 "src/f32-vbinary/gen/vsub-scalar-x4.c",
181 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
182 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
183 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
184 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
185 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
186 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700187 "src/q8-avgpool/mp9p8q-scalar.c",
188 "src/q8-avgpool/up9-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700189 "src/q8-dwconv/up1x9-scalar.c",
190 "src/q8-gavgpool/mp7p7q-scalar.c",
191 "src/q8-gavgpool/up7-scalar.c",
192 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800193 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700194 "src/q8-vadd/scalar.c",
195 "src/u8-clamp/scalar.c",
196 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800197 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700198 "src/u8-rmax/scalar.c",
199 "src/x32-packx/x2-scalar.c",
200 "src/x32-packx/x3-scalar.c",
201 "src/x32-packx/x4-scalar.c",
202 "src/x32-pad/x2-scalar.c",
203 "src/x32-unpool/scalar.c",
204 "src/x32-zip/x2-scalar.c",
205 "src/x32-zip/x3-scalar.c",
206 "src/x32-zip/x4-scalar.c",
207 "src/x32-zip/xm-scalar.c",
208 "src/x8-lut/scalar.c",
209 "src/x8-zip/x2-scalar.c",
210 "src/x8-zip/x3-scalar.c",
211 "src/x8-zip/x4-scalar.c",
212 "src/x8-zip/xm-scalar.c",
213]
214
Marat Dukhan436ebe62019-12-04 15:10:12 -0800215WASM_UKERNELS = [
216 "src/f32-avgpool/mp9p8q-wasm.c",
217 "src/f32-avgpool/up9-wasm.c",
218 "src/f32-clamp/wasm.c",
219 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
220 "src/f32-dwconv/gen/up1x25-wasm.c",
221 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
222 "src/f32-dwconv/gen/up1x4-wasm.c",
223 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
224 "src/f32-dwconv/gen/up1x9-wasm.c",
225 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
226 "src/f32-dwconv/gen/up2x25-wasm.c",
227 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
228 "src/f32-dwconv/gen/up2x4-wasm.c",
229 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
230 "src/f32-dwconv/gen/up2x9-wasm.c",
231 "src/f32-gavgpool/mp7p7q-wasm.c",
232 "src/f32-gavgpool/up7-wasm.c",
233 "src/f32-gemm/gen-inc/1x4-wasm.c",
234 "src/f32-gemm/gen-inc/2x4-wasm.c",
235 "src/f32-gemm/gen-inc/4x4-wasm.c",
236 "src/f32-gemm/gen/1x4-wasm.c",
237 "src/f32-gemm/gen/2x4-wasm.c",
238 "src/f32-gemm/gen/4x2-wasm.c",
239 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800240 "src/f32-hswish/gen/wasm-x1.c",
241 "src/f32-hswish/gen/wasm-x2.c",
242 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800243 "src/f32-igemm/gen/1x4-wasm.c",
244 "src/f32-igemm/gen/2x4-wasm.c",
245 "src/f32-igemm/gen/4x2-wasm.c",
246 "src/f32-igemm/gen/4x4-wasm.c",
247 "src/f32-maxpool/9p8x-wasm-c1.c",
248 "src/f32-pavgpool/mp9p8q-wasm.c",
249 "src/f32-pavgpool/up9-wasm.c",
250 "src/f32-prelu/gen/wasm-2x1.c",
251 "src/f32-prelu/gen/wasm-2x4.c",
252 "src/f32-vbinary/gen/vadd-wasm-x1.c",
253 "src/f32-vbinary/gen/vadd-wasm-x2.c",
254 "src/f32-vbinary/gen/vadd-wasm-x4.c",
255 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
256 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
257 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800258 "src/f32-vbinary/gen/vdiv-wasm-x1.c",
259 "src/f32-vbinary/gen/vdiv-wasm-x2.c",
260 "src/f32-vbinary/gen/vdiv-wasm-x4.c",
261 "src/f32-vbinary/gen/vdivc-wasm-x1.c",
262 "src/f32-vbinary/gen/vdivc-wasm-x2.c",
263 "src/f32-vbinary/gen/vdivc-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800264 "src/f32-vbinary/gen/vmax-wasm-x1.c",
265 "src/f32-vbinary/gen/vmax-wasm-x2.c",
266 "src/f32-vbinary/gen/vmax-wasm-x4.c",
267 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
268 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
269 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
270 "src/f32-vbinary/gen/vmin-wasm-x1.c",
271 "src/f32-vbinary/gen/vmin-wasm-x2.c",
272 "src/f32-vbinary/gen/vmin-wasm-x4.c",
273 "src/f32-vbinary/gen/vminc-wasm-x1.c",
274 "src/f32-vbinary/gen/vminc-wasm-x2.c",
275 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800276 "src/f32-vbinary/gen/vmul-wasm-x1.c",
277 "src/f32-vbinary/gen/vmul-wasm-x2.c",
278 "src/f32-vbinary/gen/vmul-wasm-x4.c",
279 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
280 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
281 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800282 "src/f32-vbinary/gen/vrdivc-wasm-x1.c",
283 "src/f32-vbinary/gen/vrdivc-wasm-x2.c",
284 "src/f32-vbinary/gen/vrdivc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800285 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
286 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
287 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
288 "src/f32-vbinary/gen/vsub-wasm-x1.c",
289 "src/f32-vbinary/gen/vsub-wasm-x2.c",
290 "src/f32-vbinary/gen/vsub-wasm-x4.c",
291 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
292 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
293 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
294 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
295 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
296 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
297]
298
Marat Dukhan08c4a432019-10-03 09:29:21 -0700299PSIMD_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800300 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800301 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800302 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700303 "src/f32-avgpool/mp9p8q-psimd.c",
304 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800305 "src/f32-bilinear/gen/psimd-c4.c",
306 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700307 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800308 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
309 "src/f32-dwconv/gen/up4x25-psimd.c",
310 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
311 "src/f32-dwconv/gen/up4x4-psimd.c",
312 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
313 "src/f32-dwconv/gen/up4x9-psimd.c",
314 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
315 "src/f32-dwconv/gen/up8x25-psimd.c",
316 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
317 "src/f32-dwconv/gen/up8x4-psimd.c",
318 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
319 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700320 "src/f32-gavgpool/mp7p7q-psimd.c",
321 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800322 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
323 "src/f32-gemm/gen/1x8-psimd-splat.c",
324 "src/f32-gemm/gen/1x8s4-psimd.c",
325 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
326 "src/f32-gemm/gen/4x8-psimd-splat.c",
327 "src/f32-gemm/gen/4x8s4-psimd.c",
328 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
329 "src/f32-gemm/gen/6x8-psimd-splat.c",
330 "src/f32-gemm/gen/6x8s4-psimd.c",
331 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
332 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
333 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
334 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
335 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
336 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
337 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
338 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
339 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800340 "src/f32-hswish/gen/psimd-x4.c",
341 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800342 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
343 "src/f32-igemm/gen/1x8-psimd-splat.c",
344 "src/f32-igemm/gen/1x8s4-psimd.c",
345 "src/f32-igemm/gen/4x2c4-psimd.c",
346 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
347 "src/f32-igemm/gen/4x8-psimd-splat.c",
348 "src/f32-igemm/gen/4x8s4-psimd.c",
349 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
350 "src/f32-igemm/gen/6x8-psimd-splat.c",
351 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800352 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700353 "src/f32-pavgpool/mp9p8q-psimd.c",
354 "src/f32-pavgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800355 "src/f32-ppmm/gen/4x8-psimd.c",
356 "src/f32-prelu/gen/psimd-2x4.c",
357 "src/f32-prelu/gen/psimd-2x8.c",
358 "src/f32-vbinary/gen/vadd-psimd-x4.c",
359 "src/f32-vbinary/gen/vadd-psimd-x8.c",
360 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
361 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800362 "src/f32-vbinary/gen/vdiv-psimd-x4.c",
363 "src/f32-vbinary/gen/vdiv-psimd-x8.c",
364 "src/f32-vbinary/gen/vdivc-psimd-x4.c",
365 "src/f32-vbinary/gen/vdivc-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800366 "src/f32-vbinary/gen/vmax-psimd-x4.c",
367 "src/f32-vbinary/gen/vmax-psimd-x8.c",
368 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
369 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
370 "src/f32-vbinary/gen/vmin-psimd-x4.c",
371 "src/f32-vbinary/gen/vmin-psimd-x8.c",
372 "src/f32-vbinary/gen/vminc-psimd-x4.c",
373 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800374 "src/f32-vbinary/gen/vmul-psimd-x4.c",
375 "src/f32-vbinary/gen/vmul-psimd-x8.c",
376 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
377 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800378 "src/f32-vbinary/gen/vrdivc-psimd-x4.c",
379 "src/f32-vbinary/gen/vrdivc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800380 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
381 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
382 "src/f32-vbinary/gen/vsub-psimd-x4.c",
383 "src/f32-vbinary/gen/vsub-psimd-x8.c",
384 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
385 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
386 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
387 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700388 "src/x32-packx/x4-psimd.c",
389 "src/x32-pad/x2-psimd.c",
390 "src/x32-unpool/psimd.c",
391 "src/x32-zip/x2-psimd.c",
392 "src/x32-zip/x3-psimd.c",
393 "src/x32-zip/x4-psimd.c",
394 "src/x32-zip/xm-psimd.c",
395]
396
397# ISA-specific micro-kernels
398NEON_UKERNELS = [
399 "src/f32-avgpool/mp9p8q-neon.c",
400 "src/f32-avgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800401 "src/f32-bilinear/gen/neon-c4.c",
402 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700403 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800404 "src/f32-dwconv/gen/up4x9-neon.c",
405 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
406 "src/f32-dwconv/gen/up8x9-neon.c",
407 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700408 "src/f32-gavgpool-spchw/neon-x4.c",
409 "src/f32-gavgpool/mp7p7q-neon.c",
410 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800411 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
412 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
413 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
414 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
415 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
416 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800417 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800418 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
419 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
420 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
421 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800422 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800423 "src/f32-gemm/gen/1x8s4-neon.c",
424 "src/f32-gemm/gen/4x8s4-neon.c",
425 "src/f32-gemm/gen/6x8s4-neon.c",
426 "src/f32-gemm/gen/8x8s4-neon.c",
427 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
428 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
429 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
430 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
431 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800432 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800433 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
434 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
435 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
436 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800437 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800438 "src/f32-gemm/gen-inc/1x8s4-neon.c",
439 "src/f32-gemm/gen-inc/4x8s4-neon.c",
440 "src/f32-gemm/gen-inc/6x8s4-neon.c",
441 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800442 "src/f32-hswish/gen/neon-x4.c",
443 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800444 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
445 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
446 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
447 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
448 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
449 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800450 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800451 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
452 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
453 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
454 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800455 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800456 "src/f32-igemm/gen/1x8s4-neon.c",
457 "src/f32-igemm/gen/4x8s4-neon.c",
458 "src/f32-igemm/gen/6x8s4-neon.c",
459 "src/f32-igemm/gen/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700460 "src/f32-pavgpool/mp9p8q-neon.c",
461 "src/f32-pavgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800462 "src/f32-ppmm/gen/4x8-neon.c",
463 "src/f32-ppmm/gen/8x8-neon.c",
464 "src/f32-prelu/gen/neon-2x4.c",
465 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700466 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800467 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
468 "src/f32-vbinary/gen/vadd-neon-x4.c",
469 "src/f32-vbinary/gen/vadd-neon-x8.c",
470 "src/f32-vbinary/gen/vaddc-neon-x4.c",
471 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800472 "src/f32-vbinary/gen/vmax-neon-x4.c",
473 "src/f32-vbinary/gen/vmax-neon-x8.c",
474 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
475 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
476 "src/f32-vbinary/gen/vmin-neon-x4.c",
477 "src/f32-vbinary/gen/vmin-neon-x8.c",
478 "src/f32-vbinary/gen/vminc-neon-x4.c",
479 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800480 "src/f32-vbinary/gen/vmul-neon-x4.c",
481 "src/f32-vbinary/gen/vmul-neon-x8.c",
482 "src/f32-vbinary/gen/vmulc-neon-x4.c",
483 "src/f32-vbinary/gen/vmulc-neon-x8.c",
484 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
485 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
486 "src/f32-vbinary/gen/vsub-neon-x4.c",
487 "src/f32-vbinary/gen/vsub-neon-x8.c",
488 "src/f32-vbinary/gen/vsubc-neon-x4.c",
489 "src/f32-vbinary/gen/vsubc-neon-x8.c",
490 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
491 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700492 "src/q8-avgpool/mp9p8q-neon.c",
493 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700494 "src/q8-dwconv/up8x9-neon.c",
495 "src/q8-gavgpool/mp7p7q-neon.c",
496 "src/q8-gavgpool/up7-neon.c",
497 "src/q8-gemm/4x8-neon.c",
498 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800499 "src/q8-igemm/4x8-neon.c",
500 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700501 "src/q8-vadd/neon.c",
502 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800503 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700504 "src/u8-rmax/neon.c",
505 "src/x32-packx/x4-neon-st4.c",
506 "src/x32-pad/x2-neon.c",
507 "src/x32-zip/x2-neon.c",
508 "src/x32-zip/x3-neon.c",
509 "src/x32-zip/x4-neon.c",
510 "src/x32-zip/xm-neon.c",
511 "src/x8-zip/x2-neon.c",
512 "src/x8-zip/x3-neon.c",
513 "src/x8-zip/x4-neon.c",
514 "src/x8-zip/xm-neon.c",
515]
516
517NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800518 "src/f32-bilinear/gen/neonfma-c4.c",
519 "src/f32-bilinear/gen/neonfma-c8.c",
520 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
521 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
522 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
523 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800524 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800525 "src/f32-igemm/gen/1x8s4-neonfma.c",
526 "src/f32-igemm/gen/4x8s4-neonfma.c",
527 "src/f32-igemm/gen/6x8s4-neonfma.c",
528 "src/f32-igemm/gen/8x8s4-neonfma.c",
529 "src/f32-dwconv/gen/up4x9-neonfma.c",
530 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
531 "src/f32-dwconv/gen/up8x9-neonfma.c",
532 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
533 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
534 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
535 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
536 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800537 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800538 "src/f32-gemm/gen/1x8s4-neonfma.c",
539 "src/f32-gemm/gen/4x8s4-neonfma.c",
540 "src/f32-gemm/gen/6x8s4-neonfma.c",
541 "src/f32-gemm/gen/8x8s4-neonfma.c",
542 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
543 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
544 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
545 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800546 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800547 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
548 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
549 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
550 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800551 "src/f32-hswish/gen/neonfma-x4.c",
552 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800553 "src/f32-ppmm/gen/4x8-neonfma.c",
554 "src/f32-ppmm/gen/8x8-neonfma.c",
555 "src/f32-sigmoid/gen/neonfma-p5-nr2fma-x16.c",
556 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
557 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800558 "src/math/exp-neonfma-lut64-p2.c",
559 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800560 "src/math/expminus-neonfma-lut2048-p1.c",
561 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800562 "src/math/expminus-neonfma-p5.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800563 "src/math/sigmoid-neonfma-lut2048-p1-nr1recps1fma.c",
564 "src/math/sigmoid-neonfma-lut2048-p1-nr2fma.c",
565 "src/math/sigmoid-neonfma-lut2048-p1-nr2recps.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800566 "src/math/sigmoid-neonfma-p5-nr1recps1fma.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800567 "src/math/sigmoid-neonfma-p5-nr2fma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800568 "src/math/sigmoid-neonfma-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700569]
570
571AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800572 "src/f32-vbinary/gen/vdiv-neon-x4.c",
573 "src/f32-vbinary/gen/vdiv-neon-x8.c",
574 "src/f32-vbinary/gen/vdivc-neon-x4.c",
575 "src/f32-vbinary/gen/vdivc-neon-x8.c",
576 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
577 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800578 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
579 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
580 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
581 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
582 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
583 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800584 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800585 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
586 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
587 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
588 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
589 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800590 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800591 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
592 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
593 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
594 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
595 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
596 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800597 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700598 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
599 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
600 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
601 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
602 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
603 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
604 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800605 "src/f32-spmm/gen/12x1-neonfma.c",
606 "src/f32-spmm/gen/12x2-neonfma.c",
607 "src/f32-spmm/gen/12x4-neonfma.c",
608 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
609 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
610 "src/f32-spmm/gen/16x1-neonfma.c",
611 "src/f32-spmm/gen/16x2-neonfma.c",
612 "src/f32-spmm/gen/16x4-neonfma.c",
613 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
614 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
615 "src/f32-spmm/gen/4x1-neonfma.c",
616 "src/f32-spmm/gen/4x2-neonfma.c",
617 "src/f32-spmm/gen/4x4-neonfma.c",
618 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
619 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
620 "src/f32-spmm/gen/8x1-neonfma.c",
621 "src/f32-spmm/gen/8x2-neonfma.c",
622 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800623 "src/math/sigmoid-neonfma-p5-div.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800624 "src/math/sigmoid-neonfma-lut2048-p1-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700625]
626
627AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800628 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
629 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
630 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700631]
632
633SSE_UKERNELS = [
634 "src/f32-avgpool/mp9p8q-sse.c",
635 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800636 "src/f32-bilinear/gen/sse-c4.c",
637 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700638 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800639 "src/f32-dwconv-spchw/3x3p1-sse.c",
640 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800641 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
642 "src/f32-dwconv/gen/up4x25-sse.c",
643 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
644 "src/f32-dwconv/gen/up4x4-sse.c",
645 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
646 "src/f32-dwconv/gen/up4x9-sse.c",
647 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
648 "src/f32-dwconv/gen/up8x25-sse.c",
649 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
650 "src/f32-dwconv/gen/up8x4-sse.c",
651 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
652 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700653 "src/f32-gavgpool-spchw/sse-x4.c",
654 "src/f32-gavgpool/mp7p7q-sse.c",
655 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800656 "src/f32-gemm/gen/1x8-sse-dup.c",
657 "src/f32-gemm/gen/1x8-sse-load1.c",
658 "src/f32-gemm/gen/1x8s4-sse.c",
659 "src/f32-gemm/gen/4x8-sse-dup.c",
660 "src/f32-gemm/gen/4x8-sse-load1.c",
661 "src/f32-gemm/gen/4x8s4-sse.c",
662 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
663 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
664 "src/f32-gemm/gen-inc/1x8s4-sse.c",
665 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
666 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
667 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800668 "src/f32-hswish/gen/sse-x4.c",
669 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800670 "src/f32-igemm/gen/1x8-sse-dup.c",
671 "src/f32-igemm/gen/1x8-sse-load1.c",
672 "src/f32-igemm/gen/1x8s4-sse.c",
673 "src/f32-igemm/gen/4x2c4-sse.c",
674 "src/f32-igemm/gen/4x8-sse-dup.c",
675 "src/f32-igemm/gen/4x8-sse-load1.c",
676 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800677 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700678 "src/f32-pavgpool/mp9p8q-sse.c",
679 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800680 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700681 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800682 "src/f32-spmm/gen/4x1-sse.c",
683 "src/f32-spmm/gen/8x1-sse.c",
684 "src/f32-vbinary/gen/vadd-sse-x4.c",
685 "src/f32-vbinary/gen/vadd-sse-x8.c",
686 "src/f32-vbinary/gen/vaddc-sse-x4.c",
687 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800688 "src/f32-vbinary/gen/vdiv-sse-x4.c",
689 "src/f32-vbinary/gen/vdiv-sse-x8.c",
690 "src/f32-vbinary/gen/vdivc-sse-x4.c",
691 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800692 "src/f32-vbinary/gen/vmax-sse-x4.c",
693 "src/f32-vbinary/gen/vmax-sse-x8.c",
694 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
695 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
696 "src/f32-vbinary/gen/vmin-sse-x4.c",
697 "src/f32-vbinary/gen/vmin-sse-x8.c",
698 "src/f32-vbinary/gen/vminc-sse-x4.c",
699 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800700 "src/f32-vbinary/gen/vmul-sse-x4.c",
701 "src/f32-vbinary/gen/vmul-sse-x8.c",
702 "src/f32-vbinary/gen/vmulc-sse-x4.c",
703 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800704 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
705 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800706 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
707 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
708 "src/f32-vbinary/gen/vsub-sse-x4.c",
709 "src/f32-vbinary/gen/vsub-sse-x8.c",
710 "src/f32-vbinary/gen/vsubc-sse-x4.c",
711 "src/f32-vbinary/gen/vsubc-sse-x8.c",
712 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
713 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700714 "src/x32-packx/x4-sse.c",
715]
716
717SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800718 "src/f32-argmaxpool/9p8x-sse2-c4.c",
719 "src/f32-argmaxpool/4x-sse2-c4.c",
720 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800721 "src/f32-prelu/gen/sse2-2x4.c",
722 "src/f32-prelu/gen/sse2-2x8.c",
723 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
724 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700725 "src/q8-avgpool/mp9p8q-sse2.c",
726 "src/q8-avgpool/up9-sse2.c",
727 "src/q8-igemm/4x4c2-sse2.c",
728 "src/q8-dwconv/up8x9-sse2.c",
729 "src/q8-gavgpool/mp7p7q-sse2.c",
730 "src/q8-gavgpool/up7-sse2.c",
731 "src/q8-gemm/2x4c8-sse2.c",
732 "src/q8-gemm/4x4c2-sse2.c",
733 "src/q8-vadd/sse2.c",
734 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800735 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700736 "src/u8-rmax/sse2.c",
737 "src/x32-pad/x2-sse2.c",
738 "src/x32-zip/x2-sse2.c",
739 "src/x32-zip/x3-sse2.c",
740 "src/x32-zip/x4-sse2.c",
741 "src/x32-zip/xm-sse2.c",
742 "src/x8-zip/x2-sse2.c",
743 "src/x8-zip/x3-sse2.c",
744 "src/x8-zip/x4-sse2.c",
745 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800746 "src/math/exp-sse2-p5.c",
747 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800748 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700749]
750
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800751SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800752 "src/f32-prelu/gen/sse41-2x4.c",
753 "src/f32-prelu/gen/sse41-2x8.c",
754 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
755 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800756]
757
Marat Dukhan08c4a432019-10-03 09:29:21 -0700758AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800759 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800760 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
761 "src/f32-dwconv/gen/up16x4-avx.c",
762 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
763 "src/f32-dwconv/gen/up8x4-avx.c",
764 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
765 "src/f32-dwconv/gen/up16x9-avx.c",
766 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
767 "src/f32-dwconv/gen/up8x9-avx.c",
768 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
769 "src/f32-dwconv/gen/up16x25-avx.c",
770 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
771 "src/f32-dwconv/gen/up8x25-avx.c",
772 "src/f32-gemm/gen/1x8-avx-broadcast.c",
773 "src/f32-gemm/gen/4x8-avx-broadcast.c",
774 "src/f32-gemm/gen/5x8-avx-broadcast.c",
775 "src/f32-gemm/gen/6x8-avx-broadcast.c",
776 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800777 "src/f32-gemm/gen/1x16-avx-broadcast.c",
778 "src/f32-gemm/gen/3x16-avx-broadcast.c",
779 "src/f32-gemm/gen/4x16-avx-broadcast.c",
780 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800781 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
782 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
783 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
784 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
785 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800786 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
787 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
788 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
789 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800790 "src/f32-hswish/gen/avx-x8.c",
791 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800792 "src/f32-igemm/gen/1x8-avx-broadcast.c",
793 "src/f32-igemm/gen/4x8-avx-broadcast.c",
794 "src/f32-igemm/gen/5x8-avx-broadcast.c",
795 "src/f32-igemm/gen/6x8-avx-broadcast.c",
796 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800797 "src/f32-igemm/gen/1x16-avx-broadcast.c",
798 "src/f32-igemm/gen/3x16-avx-broadcast.c",
799 "src/f32-igemm/gen/4x16-avx-broadcast.c",
800 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700801 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -0800802 "src/f32-vbinary/gen/vadd-avx-x8.c",
803 "src/f32-vbinary/gen/vadd-avx-x16.c",
804 "src/f32-vbinary/gen/vaddc-avx-x8.c",
805 "src/f32-vbinary/gen/vaddc-avx-x16.c",
806 "src/f32-vbinary/gen/vdiv-avx-x8.c",
807 "src/f32-vbinary/gen/vdiv-avx-x16.c",
808 "src/f32-vbinary/gen/vdivc-avx-x8.c",
809 "src/f32-vbinary/gen/vdivc-avx-x16.c",
810 "src/f32-vbinary/gen/vmax-avx-x8.c",
811 "src/f32-vbinary/gen/vmax-avx-x16.c",
812 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
813 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
814 "src/f32-vbinary/gen/vmin-avx-x8.c",
815 "src/f32-vbinary/gen/vmin-avx-x16.c",
816 "src/f32-vbinary/gen/vminc-avx-x8.c",
817 "src/f32-vbinary/gen/vminc-avx-x16.c",
818 "src/f32-vbinary/gen/vmul-avx-x8.c",
819 "src/f32-vbinary/gen/vmul-avx-x16.c",
820 "src/f32-vbinary/gen/vmulc-avx-x8.c",
821 "src/f32-vbinary/gen/vmulc-avx-x16.c",
822 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
823 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
824 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
825 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
826 "src/f32-vbinary/gen/vsub-avx-x8.c",
827 "src/f32-vbinary/gen/vsub-avx-x16.c",
828 "src/f32-vbinary/gen/vsubc-avx-x8.c",
829 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700830 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700831]
832
Marat Dukhanfda12b82019-11-21 12:27:59 -0800833FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800834 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
835 "src/f32-dwconv/gen/up16x4-fma3.c",
836 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
837 "src/f32-dwconv/gen/up8x4-fma3.c",
838 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
839 "src/f32-dwconv/gen/up16x9-fma3.c",
840 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
841 "src/f32-dwconv/gen/up8x9-fma3.c",
842 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
843 "src/f32-dwconv/gen/up16x25-fma3.c",
844 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
845 "src/f32-dwconv/gen/up8x25-fma3.c",
846 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
847 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
848 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
849 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
850 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
851 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800852 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
853 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
854 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
855 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -0800856 "src/f32-gemm/gen/1x16s4-fma3-broadcast.c",
857 "src/f32-gemm/gen/3x16s4-fma3-broadcast.c",
858 "src/f32-gemm/gen/4x16s4-fma3-broadcast.c",
859 "src/f32-gemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800860 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
861 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
862 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
863 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
864 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
865 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800866 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
867 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
868 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
869 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -0800870 "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c",
871 "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c",
872 "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c",
873 "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800874 "src/f32-hswish/gen/fma3-x8.c",
875 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800876 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
877 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
878 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
879 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
880 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
881 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800882 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
883 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
884 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
885 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -0800886 "src/f32-igemm/gen/1x16s4-fma3-broadcast.c",
887 "src/f32-igemm/gen/3x16s4-fma3-broadcast.c",
888 "src/f32-igemm/gen/4x16s4-fma3-broadcast.c",
889 "src/f32-igemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -0800890]
891
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700892AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -0800893 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
894 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
895 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
896 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
897 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
898 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
899 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
900 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
901 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
902 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
903 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
904 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
905 "src/f32-raddextexp/gen/avx2-p5-x64.c",
906 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
907 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
908 "src/f32-raddextexp/gen/avx2-p5-x72.c",
909 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
910 "src/f32-raddextexp/gen/avx2-p5-x80.c",
911 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
912 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
913 "src/f32-raddextexp/gen/avx2-p5-x96.c",
914 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
915 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
916 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
917 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
918 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
919 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
920 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
921 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
922 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
923 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
924 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
925 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
926 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
927 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
928 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
929 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
930 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
931 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
932 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
933 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
934 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
935 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
936 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
937 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
938 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
939 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
940 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
941 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
942 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
943 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
944 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
945 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
946 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
947 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
948 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
949 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
950 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
951 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
952 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700953 "src/math/exp-avx2-p5.c",
954 "src/math/exp-avx2-perm-p3.c",
955 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700956 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700957 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700958]
959
Marat Dukhan08c4a432019-10-03 09:29:21 -0700960AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800961 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -0800962 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
963 "src/f32-dwconv/gen/up32x4-avx512f.c",
964 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
965 "src/f32-dwconv/gen/up16x4-avx512f.c",
966 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
967 "src/f32-dwconv/gen/up32x9-avx512f.c",
968 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
969 "src/f32-dwconv/gen/up16x9-avx512f.c",
970 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
971 "src/f32-dwconv/gen/up32x25-avx512f.c",
972 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
973 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -0800974 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
975 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
976 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
977 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
978 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
979 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
980 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
981 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
982 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
983 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
984 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
985 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800986 "src/f32-hswish/gen/avx512f-x16.c",
987 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -0800988 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
989 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
990 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
991 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
992 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
993 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -0800994 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
995 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
996 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
997 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
998 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
999 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1000 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1001 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1002 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1003 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1004 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1005 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1006 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1007 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1008 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1009 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1010 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1011 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1012 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1013 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1014 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1015 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1016 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1017 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1018 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1019 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1020 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1021 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1022 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1023 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1024 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1025 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1026 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1027 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1028 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1029 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001030 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001031 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1032 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1033 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1034 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1035 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1036 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1037 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1038 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1039 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1040 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1041 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1042 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1043 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1044 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1045 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1046 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1047 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1048 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1049 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1050 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1051 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1052 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1053 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1054 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1055 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1056 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1057 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1058 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001059 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001060 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1061 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1062 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1063 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1064 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1065 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1066 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1067 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1068 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1069 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1070 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1071 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1072 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1073 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1074 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1075 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1076 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1077 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1078 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1079 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1080 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1081 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1082 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1083 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001084 "src/math/exp-avx512f-p5-scalef.c",
1085 "src/math/exp-avx512f-p5.c",
1086 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001087 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001088 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001089]
1090
1091AARCH32_ASM_UKERNELS = [
1092 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard13916042019-12-11 10:56:34 -08001093 "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S",
Frank Barchard3e237f22019-12-04 23:08:51 -08001094 "src/f32-gemm/4x8-aarch32-neon-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001095 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001096]
1097
1098AARCH64_ASM_UKERNELS = [
1099 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1100 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001101 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1102 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1103 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1104 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1105 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1106 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1107 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1108 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1109 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1110 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
1111 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1112 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
1113 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1114 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
1115 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1116 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1117 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1118 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1119 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1120 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1121 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1122 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1123 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1124 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1125 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1126 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1127 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
1128 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1129 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
1130 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
1131 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
1132 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1133 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1134 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001135 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001136 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001137 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
1138 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
1139 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001140 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001141 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
1142 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001143 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001144 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
1145 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
1146 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
1147]
1148
1149INTERNAL_MICROKERNEL_HDRS = [
1150 "src/xnnpack/argmaxpool.h",
1151 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001152 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001153 "src/xnnpack/clamp.h",
1154 "src/xnnpack/common.h",
1155 "src/xnnpack/conv.h",
1156 "src/xnnpack/dwconv.h",
1157 "src/xnnpack/gavgpool.h",
1158 "src/xnnpack/gemm.h",
1159 "src/xnnpack/hswish.h",
1160 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001161 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001162 "src/xnnpack/lut.h",
1163 "src/xnnpack/math.h",
1164 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001165 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001166 "src/xnnpack/packx.h",
1167 "src/xnnpack/pad.h",
1168 "src/xnnpack/params.h",
1169 "src/xnnpack/pavgpool.h",
1170 "src/xnnpack/ppmm.h",
1171 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001172 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001173 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001174 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001175 "src/xnnpack/rmax.h",
1176 "src/xnnpack/scalar-utils.h",
1177 "src/xnnpack/spmm.h",
1178 "src/xnnpack/unpool.h",
1179 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001180 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001181 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001182 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001183 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001184 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001185 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001186 "src/xnnpack/zip.h",
1187]
1188
1189INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1190 "include/xnnpack.h",
1191 "src/xnnpack/allocator.h",
1192 "src/xnnpack/compute.h",
1193 "src/xnnpack/im2col.h",
1194 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001195 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001196 "src/xnnpack/operator.h",
1197 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001198 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001199 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001200 "src/xnnpack/requantization.h",
1201]
1202
1203ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1204 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001205]
1206
1207MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001208 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001209 "include/xnnpack.h",
1210]
1211
1212MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1213 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001214 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001215 "src/xnnpack/requantization.h",
1216 "include/xnnpack.h",
1217]
1218
1219OPERATOR_TEST_PARAMS_HDRS = [
1220 "src/xnnpack/params.h",
1221 "src/xnnpack/common.h",
1222]
1223
1224WEIGHTS_PACK_HDRS = [
1225 "src/xnnpack/pack.h",
1226 "src/xnnpack/operator.h",
1227 "src/xnnpack/compute.h",
1228]
1229
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001230LOGGING_COPTS = select({
1231 # No logging in optimized mode
1232 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1233 # Full logging in debug mode
1234 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1235 # Error-only logging in default (fastbuild) mode
1236 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1237})
1238
1239LOGGING_HDRS = [
1240 "src/xnnpack/log.h",
1241]
1242
Marat Dukhan08c4a432019-10-03 09:29:21 -07001243xnnpack_cc_library(
1244 name = "scalar_ukernels",
1245 srcs = SCALAR_UKERNELS,
1246 hdrs = INTERNAL_HDRS,
1247 aarch32_copts = ["-marm"],
1248 copts = xnnpack_std_copts(),
1249 deps = [
1250 "@FP16",
1251 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001252 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001253 ],
1254)
1255
1256xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001257 name = "wasm_ukernels",
1258 hdrs = INTERNAL_HDRS,
1259 copts = xnnpack_std_copts(),
1260 wasm_srcs = WASM_UKERNELS,
1261 deps = [
1262 "@FP16",
1263 "@FXdiv",
1264 "@pthreadpool",
1265 ],
1266)
1267
1268xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001269 name = "psimd_ukernels",
1270 srcs = PSIMD_UKERNELS,
1271 hdrs = INTERNAL_HDRS,
1272 aarch32_copts = [
1273 "-marm",
1274 "-mfpu=neon",
1275 ],
1276 copts = xnnpack_std_copts(),
1277 optimized_copts = [
1278 "-O3",
1279 "-ffast-math",
1280 ],
1281 deps = [
1282 "@FP16",
1283 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001284 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001285 ],
1286)
1287
1288xnnpack_cc_library(
1289 name = "neon_ukernels",
1290 hdrs = INTERNAL_HDRS,
1291 aarch32_copts = [
1292 "-marm",
1293 "-mfpu=neon",
1294 ],
1295 aarch32_srcs = NEON_UKERNELS,
1296 aarch64_srcs = NEON_UKERNELS,
1297 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001298 deps = [
1299 "@FP16",
1300 "@pthreadpool",
1301 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001302)
1303
1304xnnpack_cc_library(
1305 name = "neonfma_ukernels",
1306 hdrs = INTERNAL_HDRS,
1307 aarch32_copts = [
1308 "-marm",
1309 "-mfpu=neon-vfpv4",
1310 ],
1311 aarch32_srcs = NEONFMA_UKERNELS,
1312 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1313 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001314 deps = [
1315 "@FP16",
1316 "@pthreadpool",
1317 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001318)
1319
1320xnnpack_cc_library(
1321 name = "neonfp16arith_ukernels",
1322 hdrs = INTERNAL_HDRS,
1323 aarch64_copts = ["-march=armv8.2-a+fp16"],
1324 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1325 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001326 deps = [
1327 "@FP16",
1328 "@pthreadpool",
1329 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001330)
1331
1332xnnpack_cc_library(
1333 name = "sse2_ukernels",
1334 hdrs = INTERNAL_HDRS,
1335 copts = xnnpack_std_copts(),
1336 x86_copts = ["-msse2"],
1337 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001338 deps = [
1339 "@FP16",
1340 "@pthreadpool",
1341 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001342)
1343
1344xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001345 name = "sse41_ukernels",
1346 hdrs = INTERNAL_HDRS,
1347 copts = xnnpack_std_copts(),
1348 x86_copts = ["-msse4.1"],
1349 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001350 deps = [
1351 "@FP16",
1352 "@pthreadpool",
1353 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001354)
1355
1356xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001357 name = "avx_ukernels",
1358 hdrs = INTERNAL_HDRS,
1359 copts = xnnpack_std_copts(),
1360 x86_copts = ["-mavx"],
1361 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001362 deps = [
1363 "@FP16",
1364 "@pthreadpool",
1365 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001366)
1367
1368xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001369 name = "fma3_ukernels",
1370 hdrs = INTERNAL_HDRS,
1371 copts = xnnpack_std_copts(),
1372 x86_copts = [
1373 "-mfma",
1374 ],
1375 x86_srcs = FMA3_UKERNELS,
1376 deps = [
1377 "@FP16",
1378 "@pthreadpool",
1379 ],
1380)
1381
1382xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001383 name = "avx2_ukernels",
1384 hdrs = INTERNAL_HDRS,
1385 copts = xnnpack_std_copts(),
1386 x86_copts = [
1387 "-mfma",
1388 "-mavx2",
1389 ],
1390 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001391 deps = [
1392 "@FP16",
1393 "@pthreadpool",
1394 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001395)
1396
1397xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001398 name = "avx512f_ukernels",
1399 hdrs = INTERNAL_HDRS,
1400 copts = xnnpack_std_copts(),
1401 x86_copts = ["-mavx512f"],
1402 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001403 deps = [
1404 "@FP16",
1405 "@pthreadpool",
1406 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001407)
1408
1409xnnpack_cc_library(
1410 name = "asm_ukernels",
1411 hdrs = ["src/xnnpack/assembly.h"],
1412 aarch32_srcs = AARCH32_ASM_UKERNELS,
1413 aarch64_srcs = AARCH64_ASM_UKERNELS,
1414)
1415
1416xnnpack_aggregate_library(
1417 name = "ukernels",
1418 aarch32_deps = [
1419 ":psimd_ukernels",
1420 ":neon_ukernels",
1421 ":neonfma_ukernels",
1422 ":asm_ukernels",
1423 ],
1424 aarch64_deps = [
1425 ":psimd_ukernels",
1426 ":neon_ukernels",
1427 ":neonfma_ukernels",
1428 ":neonfp16arith_ukernels",
1429 ":asm_ukernels",
1430 ],
1431 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001432 wasm_deps = [
1433 ":wasm_ukernels",
1434 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001435 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001436 ":wasm_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001437 ":psimd_ukernels",
1438 ],
1439 x86_deps = [
1440 ":psimd_ukernels",
1441 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001442 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001443 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001444 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001445 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001446 ":avx512f_ukernels",
1447 ],
1448)
1449
1450xnnpack_cc_library(
1451 name = "im2col",
1452 srcs = ["src/im2col.c"],
1453 hdrs = [
1454 "src/xnnpack/common.h",
1455 "src/xnnpack/im2col.h",
1456 ],
1457 copts = xnnpack_std_copts(),
1458)
1459
1460xnnpack_cc_library(
1461 name = "indirection",
1462 srcs = ["src/indirection.c"],
1463 hdrs = INTERNAL_HDRS,
1464 copts = xnnpack_std_copts(),
1465 deps = [
1466 "@FP16",
1467 "@FXdiv",
1468 "@pthreadpool",
1469 ],
1470)
1471
1472xnnpack_cc_library(
1473 name = "operator_run",
1474 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001475 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1476 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001477 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1478 "-Wno-vla",
1479 ],
1480 deps = [
1481 "@FP16",
1482 "@FXdiv",
1483 "@clog",
1484 "@pthreadpool",
1485 ],
1486)
1487
1488cc_library(
1489 name = "enable_assembly",
1490 defines = select({
1491 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1492 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001493 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001494 }),
1495)
1496
Marat Dukhancf056b22019-10-07 10:26:29 -07001497xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001498 name = "operators",
1499 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001500 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001501 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001502 ],
1503 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001504 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001505 "-Isrc",
1506 "-Iinclude",
1507 ] + select({
1508 ":debug_build": [],
1509 "//conditions:default": xnnpack_min_size_copts(),
1510 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001511 wasm_srcs = ["src/wasm-stubs.c"],
1512 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001513 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001514 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001515 "@FP16",
1516 "@FXdiv",
1517 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001518 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001519 ],
1520)
1521
1522cc_library(
1523 name = "XNNPACK",
1524 srcs = [
1525 "src/init.c",
1526 ],
1527 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1528 "-Isrc",
1529 "-Iinclude",
1530 ] + select({
1531 ":debug_build": [],
1532 "//conditions:default": xnnpack_min_size_copts(),
1533 }),
1534 includes = ["include"],
1535 linkstatic = True,
1536 textual_hdrs = ["include/xnnpack.h"],
1537 visibility = xnnpack_visibility(),
1538 deps = [
1539 ":enable_assembly",
1540 ":ukernels",
1541 ":operator_run",
1542 ":operators",
1543 "@clog",
1544 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001545 ] + select({
1546 ":emscripten": [],
1547 "//conditions:default": ["@cpuinfo"],
1548 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001549)
1550
1551cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001552 name = "xnnpack_operators_nhwc_f32",
1553 srcs = [
1554 "src/init.c",
1555 ],
1556 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1557 "-Isrc",
1558 "-Iinclude",
1559 ] + select({
1560 ":debug_build": [],
1561 "//conditions:default": xnnpack_min_size_copts(),
1562 }),
1563 defines = [
1564 "XNN_NO_Q8_OPERATORS",
1565 "XNN_NO_U8_OPERATORS",
1566 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001567 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001568 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001569 includes = ["include"],
1570 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001571 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001572 visibility = xnnpack_visibility(),
1573 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001574 ":enable_assembly",
1575 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001576 ":operator_run",
1577 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001578 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001579 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001580 ] + select({
1581 ":emscripten": [],
1582 "//conditions:default": ["@cpuinfo"],
1583 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001584)
1585
Marat Dukhancf056b22019-10-07 10:26:29 -07001586xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001587 name = "bench_utils",
1588 srcs = ["bench/utils.cc"],
1589 hdrs = ["bench/utils.h"],
1590 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001591 deps = [
1592 "@com_google_benchmark//:benchmark",
1593 "@cpuinfo",
1594 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595)
1596
Frank Barchard7e955972019-10-11 10:34:25 -07001597######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001598
1599xnnpack_benchmark(
1600 name = "q8_gemm_bench",
1601 srcs = [
1602 "bench/gemm.h",
1603 "bench/q8-gemm.cc",
1604 "src/xnnpack/AlignedAllocator.h",
1605 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1606 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1607 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1608)
1609
1610xnnpack_benchmark(
1611 name = "f16_gemm_bench",
1612 srcs = [
1613 "bench/f16-gemm.cc",
1614 "bench/gemm.h",
1615 "src/xnnpack/AlignedAllocator.h",
1616 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1617 copts = ["-Wno-unused-function"],
1618 deps = MICROKERNEL_BENCHMARK_DEPS,
1619)
1620
1621xnnpack_benchmark(
1622 name = "f32_igemm_bench",
1623 srcs = [
1624 "bench/f32-igemm.cc",
1625 "bench/conv.h",
1626 "src/xnnpack/AlignedAllocator.h",
1627 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001628 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001629)
1630
1631xnnpack_benchmark(
1632 name = "f32_conv_hwc_bench",
1633 srcs = [
1634 "bench/f32-conv-hwc.cc",
1635 "bench/dconv.h",
1636 "src/xnnpack/AlignedAllocator.h",
1637 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1638 copts = ["-Wno-unused-function"],
1639 deps = MICROKERNEL_BENCHMARK_DEPS,
1640)
1641
1642xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001643 name = "f32_conv_hwc2spchw_bench",
1644 srcs = [
1645 "bench/f32-conv-hwc2spchw.cc",
1646 "bench/dconv.h",
1647 "src/xnnpack/AlignedAllocator.h",
1648 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1649 copts = ["-Wno-unused-function"],
1650 deps = MICROKERNEL_BENCHMARK_DEPS,
1651)
1652
1653xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001654 name = "f32_dwconv_bench",
1655 srcs = [
1656 "bench/f32-dwconv.cc",
1657 "bench/dwconv.h",
1658 "src/xnnpack/AlignedAllocator.h",
1659 ] + MICROKERNEL_BENCHMARK_HDRS,
1660 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1661)
1662
1663xnnpack_benchmark(
1664 name = "f32_dwconv_spchw_bench",
1665 srcs = [
1666 "bench/f32-dwconv-spchw.cc",
1667 "bench/dwconv.h",
1668 "src/xnnpack/AlignedAllocator.h",
1669 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1670 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1671)
1672
1673xnnpack_benchmark(
1674 name = "f32_gemm_bench",
1675 srcs = [
1676 "bench/f32-gemm.cc",
1677 "bench/gemm.h",
1678 "src/xnnpack/AlignedAllocator.h",
1679 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1680 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001681 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001682)
1683
1684xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001685 name = "f32_raddexpminusmax_bench",
1686 srcs = [
1687 "bench/f32-raddexpminusmax.cc",
1688 "src/xnnpack/AlignedAllocator.h",
1689 ] + MICROKERNEL_BENCHMARK_HDRS,
1690 deps = MICROKERNEL_BENCHMARK_DEPS,
1691)
1692
1693xnnpack_benchmark(
1694 name = "f32_raddextexp_bench",
1695 srcs = [
1696 "bench/f32-raddextexp.cc",
1697 "src/xnnpack/AlignedAllocator.h",
1698 ] + MICROKERNEL_BENCHMARK_HDRS,
1699 deps = MICROKERNEL_BENCHMARK_DEPS,
1700)
1701
1702xnnpack_benchmark(
1703 name = "f32_raddstoreexpminusmax_bench",
1704 srcs = [
1705 "bench/f32-raddstoreexpminusmax.cc",
1706 "src/xnnpack/AlignedAllocator.h",
1707 ] + MICROKERNEL_BENCHMARK_HDRS,
1708 deps = MICROKERNEL_BENCHMARK_DEPS,
1709)
1710
1711xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001712 name = "f32_rmax_bench",
1713 srcs = [
1714 "bench/f32-rmax.cc",
1715 "src/xnnpack/AlignedAllocator.h",
1716 ] + MICROKERNEL_BENCHMARK_HDRS,
1717 deps = MICROKERNEL_BENCHMARK_DEPS,
1718)
1719
1720xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001721 name = "f32_sigmoid_bench",
1722 srcs = [
1723 "bench/f32-sigmoid.cc",
1724 "src/xnnpack/AlignedAllocator.h",
1725 ] + MICROKERNEL_BENCHMARK_HDRS,
1726 copts = ["-Wno-unused-function"],
1727 deps = MICROKERNEL_BENCHMARK_DEPS,
1728)
1729
1730xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001731 name = "f32_spmm_bench",
1732 srcs = [
1733 "bench/f32-spmm.cc",
1734 "bench/gemm.h",
1735 "src/xnnpack/AlignedAllocator.h",
1736 ] + MICROKERNEL_BENCHMARK_HDRS,
1737 copts = ["-Wno-unused-function"],
1738 deps = MICROKERNEL_BENCHMARK_DEPS,
1739)
1740
1741xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001742 name = "f32_softargmax_bench",
1743 srcs = [
1744 "bench/f32-softargmax.cc",
1745 ] + MICROKERNEL_BENCHMARK_HDRS,
1746 copts = ["-Wno-unused-function"],
1747 deps = MICROKERNEL_BENCHMARK_DEPS,
1748)
1749
1750xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001751 name = "f32_vscaleexpminusmax_bench",
1752 srcs = [
1753 "bench/f32-vscaleexpminusmax.cc",
1754 "src/xnnpack/AlignedAllocator.h",
1755 ] + MICROKERNEL_BENCHMARK_HDRS,
1756 deps = MICROKERNEL_BENCHMARK_DEPS,
1757)
1758
1759xnnpack_benchmark(
1760 name = "f32_vscaleextexp_bench",
1761 srcs = [
1762 "bench/f32-vscaleextexp.cc",
1763 "src/xnnpack/AlignedAllocator.h",
1764 ] + MICROKERNEL_BENCHMARK_HDRS,
1765 deps = MICROKERNEL_BENCHMARK_DEPS,
1766)
1767
1768xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769 name = "f32_im2col_gemm_bench",
1770 srcs = [
1771 "bench/f32-im2col-gemm.cc",
1772 "bench/conv.h",
1773 "src/xnnpack/AlignedAllocator.h",
1774 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1775 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1776)
1777
1778########################### Benchmarks for operators ###########################
1779
1780xnnpack_benchmark(
1781 name = "add_bench",
1782 srcs = ["bench/add.cc"],
1783 deps = OPERATOR_BENCHMARK_DEPS,
1784)
1785
1786xnnpack_benchmark(
1787 name = "average_pooling_bench",
1788 srcs = ["bench/average-pooling.cc"],
1789 deps = OPERATOR_BENCHMARK_DEPS,
1790)
1791
1792xnnpack_benchmark(
1793 name = "channel_shuffle_bench",
1794 srcs = ["bench/channel-shuffle.cc"],
1795 deps = OPERATOR_BENCHMARK_DEPS,
1796)
1797
1798xnnpack_benchmark(
1799 name = "convolution_bench",
1800 srcs = ["bench/convolution.cc"],
1801 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1802 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1803)
1804
1805xnnpack_benchmark(
1806 name = "deconvolution_bench",
1807 srcs = ["bench/deconvolution.cc"],
1808 copts = xnnpack_optional_tflite_copts(),
1809 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1810)
1811
1812xnnpack_benchmark(
1813 name = "global_average_pooling_bench",
1814 srcs = ["bench/global-average-pooling.cc"],
1815 deps = OPERATOR_BENCHMARK_DEPS,
1816)
1817
1818xnnpack_benchmark(
1819 name = "max_pooling_bench",
1820 srcs = ["bench/max-pooling.cc"],
1821 deps = OPERATOR_BENCHMARK_DEPS,
1822)
1823
1824xnnpack_benchmark(
1825 name = "sigmoid_bench",
1826 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001827 copts = xnnpack_optional_tflite_copts(),
1828 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001829)
1830
1831xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001832 name = "prelu_bench",
1833 srcs = ["bench/prelu.cc"],
1834 copts = xnnpack_optional_tflite_copts(),
1835 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1836)
1837
1838xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001839 name = "softargmax_bench",
1840 srcs = ["bench/softargmax.cc"],
1841 deps = OPERATOR_BENCHMARK_DEPS,
1842)
1843
Marat Dukhanc068bb62019-10-04 13:24:39 -07001844############################# End-to-end benchmarks ############################
1845
1846cc_library(
1847 name = "mobilenet_v1",
1848 srcs = ["models/mobilenet-v1.cc"],
1849 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08001850 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07001851 linkstatic = True,
1852 deps = [
1853 ":XNNPACK",
1854 "@pthreadpool",
1855 ],
1856)
1857
1858cc_library(
1859 name = "mobilenet_v2",
1860 srcs = ["models/mobilenet-v2.cc"],
1861 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08001862 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07001863 linkstatic = True,
1864 deps = [
1865 ":XNNPACK",
1866 "@pthreadpool",
1867 ],
1868)
1869
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001870cc_library(
1871 name = "mobilenet_v3_large",
1872 srcs = ["models/mobilenet-v3-large.cc"],
1873 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08001874 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001875 linkstatic = True,
1876 deps = [
1877 ":XNNPACK",
1878 "@pthreadpool",
1879 ],
1880)
1881
1882cc_library(
1883 name = "mobilenet_v3_small",
1884 srcs = ["models/mobilenet-v3-small.cc"],
1885 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08001886 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001887 linkstatic = True,
1888 deps = [
1889 ":XNNPACK",
1890 "@pthreadpool",
1891 ],
1892)
1893
Marat Dukhanc068bb62019-10-04 13:24:39 -07001894xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001895 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001896 srcs = [
1897 "bench/f32-dwconv-e2e.cc",
1898 "bench/end2end.h",
1899 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07001900 copts = ["-Wno-unused-function"],
1901 deps = MICROKERNEL_BENCHMARK_DEPS + [
1902 ":XNNPACK",
1903 ":mobilenet_v1",
1904 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001905 ":mobilenet_v3_large",
1906 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07001907 ],
1908)
1909
1910xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001911 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001912 srcs = [
1913 "bench/f32-gemm-e2e.cc",
1914 "bench/end2end.h",
1915 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07001916 copts = ["-Wno-unused-function"],
1917 deps = MICROKERNEL_BENCHMARK_DEPS + [
1918 ":XNNPACK",
1919 ":mobilenet_v1",
1920 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001921 ":mobilenet_v3_large",
1922 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07001923 ],
1924)
1925
1926xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001927 name = "end2end_bench",
1928 srcs = ["bench/end2end.cc"],
1929 deps = [
1930 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001931 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001932 ":mobilenet_v1",
1933 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001934 ":mobilenet_v3_large",
1935 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001936 "@pthreadpool",
1937 ],
1938)
1939
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001940#################### Accuracy evaluation for math functions ####################
1941
1942xnnpack_benchmark(
1943 name = "f32_exp_eval",
1944 srcs = [
1945 "eval/f32-exp.cc",
1946 "src/xnnpack/AlignedAllocator.h",
1947 ] + ACCURACY_EVAL_HDRS,
1948 deps = ACCURACY_EVAL_DEPS,
1949)
1950
Marat Dukhan515c9772019-10-17 18:07:57 -07001951xnnpack_benchmark(
1952 name = "f32_expminus_eval",
1953 srcs = [
1954 "eval/f32-expminus.cc",
1955 "src/xnnpack/AlignedAllocator.h",
1956 ] + ACCURACY_EVAL_HDRS,
1957 deps = ACCURACY_EVAL_DEPS,
1958)
1959
Marat Dukhan98ba4412019-10-23 02:14:28 -07001960xnnpack_benchmark(
1961 name = "f32_extexp_eval",
1962 srcs = [
1963 "eval/f32-extexp.cc",
1964 "src/xnnpack/AlignedAllocator.h",
1965 ] + ACCURACY_EVAL_HDRS,
1966 deps = ACCURACY_EVAL_DEPS,
1967)
1968
Marat Dukhan346a9e52019-11-15 09:06:30 -08001969xnnpack_benchmark(
1970 name = "f32_sigmoid_eval",
1971 srcs = [
1972 "eval/f32-sigmoid.cc",
1973 "src/xnnpack/AlignedAllocator.h",
1974 ] + ACCURACY_EVAL_HDRS,
1975 deps = ACCURACY_EVAL_DEPS,
1976)
1977
Marat Dukhan08c4a432019-10-03 09:29:21 -07001978######################### Unit tests for micro-kernels #########################
1979
1980xnnpack_unit_test(
1981 name = "f16_gemm_test",
1982 srcs = [
1983 "test/f16-gemm.cc",
1984 "test/gemm-microkernel-tester.h",
1985 "src/xnnpack/AlignedAllocator.h",
1986 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1987 deps = MICROKERNEL_TEST_DEPS,
1988)
1989
1990xnnpack_unit_test(
1991 name = "f32_argmaxpool_test",
1992 srcs = [
1993 "test/f32-argmaxpool.cc",
1994 "test/argmaxpool-microkernel-tester.h",
1995 "src/xnnpack/AlignedAllocator.h",
1996 ] + MICROKERNEL_TEST_HDRS,
1997 deps = MICROKERNEL_TEST_DEPS,
1998)
1999
2000xnnpack_unit_test(
2001 name = "f32_avgpool_test",
2002 srcs = [
2003 "test/f32-avgpool.cc",
2004 "test/avgpool-microkernel-tester.h",
2005 "src/xnnpack/AlignedAllocator.h",
2006 ] + MICROKERNEL_TEST_HDRS,
2007 deps = MICROKERNEL_TEST_DEPS,
2008)
2009
2010xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002011 name = "f32_bilinear_test",
2012 srcs = [
2013 "test/f32-bilinear.cc",
2014 "test/bilinear-microkernel-tester.h",
2015 "src/xnnpack/AlignedAllocator.h",
2016 ] + MICROKERNEL_TEST_HDRS,
2017 deps = MICROKERNEL_TEST_DEPS,
2018)
2019
2020xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002021 name = "f32_clamp_test",
2022 srcs = [
2023 "test/f32-clamp.cc",
2024 "test/clamp-microkernel-tester.h",
2025 ] + MICROKERNEL_TEST_HDRS,
2026 deps = MICROKERNEL_TEST_DEPS,
2027)
2028
2029xnnpack_unit_test(
2030 name = "f32_igemm_test",
2031 srcs = [
2032 "test/f32-igemm.cc",
2033 "test/gemm-microkernel-tester.h",
2034 "src/xnnpack/AlignedAllocator.h",
2035 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2036 deps = MICROKERNEL_TEST_DEPS,
2037)
2038
2039xnnpack_unit_test(
2040 name = "f32_conv_hwc_test",
2041 srcs = [
2042 "test/f32-conv-hwc.cc",
2043 "test/conv-hwc-microkernel-tester.h",
2044 "src/xnnpack/AlignedAllocator.h",
2045 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2046 deps = MICROKERNEL_TEST_DEPS,
2047)
2048
2049xnnpack_unit_test(
2050 name = "f32_conv_hwc2spchw_test",
2051 srcs = [
2052 "test/f32-conv-hwc2spchw.cc",
2053 "test/conv-hwc2spchw-microkernel-tester.h",
2054 "src/xnnpack/AlignedAllocator.h",
2055 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2056 deps = MICROKERNEL_TEST_DEPS,
2057)
2058
2059xnnpack_unit_test(
2060 name = "f32_dwconv_test",
2061 srcs = [
2062 "test/f32-dwconv.cc",
2063 "test/dwconv-microkernel-tester.h",
2064 "src/xnnpack/AlignedAllocator.h",
2065 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2066 deps = MICROKERNEL_TEST_DEPS,
2067)
2068
2069xnnpack_unit_test(
2070 name = "f32_dwconv_spchw_test",
2071 srcs = [
2072 "test/f32-dwconv-spchw.cc",
2073 "test/dwconv-spchw-microkernel-tester.h",
2074 "src/xnnpack/AlignedAllocator.h",
2075 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2076 deps = MICROKERNEL_TEST_DEPS,
2077)
2078
2079xnnpack_unit_test(
2080 name = "f32_gavgpool_test",
2081 srcs = [
2082 "test/f32-gavgpool.cc",
2083 "test/gavgpool-microkernel-tester.h",
2084 "src/xnnpack/AlignedAllocator.h",
2085 ] + MICROKERNEL_TEST_HDRS,
2086 deps = MICROKERNEL_TEST_DEPS,
2087)
2088
2089xnnpack_unit_test(
2090 name = "f32_gavgpool_spchw_test",
2091 srcs = [
2092 "test/f32-gavgpool-spchw.cc",
2093 "test/gavgpool-spchw-microkernel-tester.h",
2094 "src/xnnpack/AlignedAllocator.h",
2095 ] + MICROKERNEL_TEST_HDRS,
2096 deps = MICROKERNEL_TEST_DEPS,
2097)
2098
2099xnnpack_unit_test(
2100 name = "f32_gemm_test",
2101 srcs = [
2102 "test/f32-gemm.cc",
2103 "test/gemm-microkernel-tester.h",
2104 "src/xnnpack/AlignedAllocator.h",
2105 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2106 deps = MICROKERNEL_TEST_DEPS,
2107)
2108
2109xnnpack_unit_test(
2110 name = "f32_gemminc_test",
2111 srcs = [
2112 "test/f32-gemminc.cc",
2113 "test/gemm-microkernel-tester.h",
2114 "src/xnnpack/AlignedAllocator.h",
2115 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2116 deps = MICROKERNEL_TEST_DEPS,
2117)
2118
2119xnnpack_unit_test(
2120 name = "f32_hswish_test",
2121 srcs = [
2122 "test/f32-hswish.cc",
2123 "test/hswish-microkernel-tester.h",
2124 ] + MICROKERNEL_TEST_HDRS,
2125 deps = MICROKERNEL_TEST_DEPS,
2126)
2127
2128xnnpack_unit_test(
2129 name = "f32_maxpool_test",
2130 srcs = [
2131 "test/f32-maxpool.cc",
2132 "test/maxpool-microkernel-tester.h",
2133 ] + MICROKERNEL_TEST_HDRS,
2134 deps = MICROKERNEL_TEST_DEPS,
2135)
2136
2137xnnpack_unit_test(
2138 name = "f32_pavgpool_test",
2139 srcs = [
2140 "test/f32-pavgpool.cc",
2141 "test/avgpool-microkernel-tester.h",
2142 "src/xnnpack/AlignedAllocator.h",
2143 ] + MICROKERNEL_TEST_HDRS,
2144 deps = MICROKERNEL_TEST_DEPS,
2145)
2146
2147xnnpack_unit_test(
2148 name = "f32_ppmm_test",
2149 srcs = [
2150 "test/f32-ppmm.cc",
2151 "test/gemm-microkernel-tester.h",
2152 "src/xnnpack/AlignedAllocator.h",
2153 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2154 deps = MICROKERNEL_TEST_DEPS,
2155)
2156
2157xnnpack_unit_test(
2158 name = "f32_prelu_test",
2159 srcs = [
2160 "test/f32-prelu.cc",
2161 "test/prelu-microkernel-tester.h",
2162 "src/xnnpack/AlignedAllocator.h",
2163 ] + MICROKERNEL_TEST_HDRS,
2164 deps = MICROKERNEL_TEST_DEPS,
2165)
2166
2167xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002168 name = "f32_raddexpminusmax_test",
2169 srcs = [
2170 "test/f32-raddexpminusmax.cc",
2171 "test/raddexpminusmax-microkernel-tester.h",
2172 ] + MICROKERNEL_TEST_HDRS,
2173 deps = MICROKERNEL_TEST_DEPS,
2174)
2175
2176xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002177 name = "f32_raddextexp_test",
2178 srcs = [
2179 "test/f32-raddextexp.cc",
2180 "test/raddextexp-microkernel-tester.h",
2181 ] + MICROKERNEL_TEST_HDRS,
2182 deps = MICROKERNEL_TEST_DEPS,
2183)
2184
2185xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002186 name = "f32_raddstoreexpminusmax_test",
2187 srcs = [
2188 "test/f32-raddstoreexpminusmax.cc",
2189 "test/raddstoreexpminusmax-microkernel-tester.h",
2190 ] + MICROKERNEL_TEST_HDRS,
2191 deps = MICROKERNEL_TEST_DEPS,
2192)
2193
2194xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002195 name = "f32_rmax_test",
2196 srcs = [
2197 "test/f32-rmax.cc",
2198 "test/rmax-microkernel-tester.h",
2199 ] + MICROKERNEL_TEST_HDRS,
2200 deps = MICROKERNEL_TEST_DEPS,
2201)
2202
2203xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002204 name = "f32_sigmoid_test",
2205 srcs = [
2206 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002207 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002208 ] + MICROKERNEL_TEST_HDRS,
2209 deps = MICROKERNEL_TEST_DEPS,
2210)
2211
2212xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002213 name = "f32_spmm_test",
2214 srcs = [
2215 "test/f32-spmm.cc",
2216 "test/spmm-microkernel-tester.h",
2217 "src/xnnpack/AlignedAllocator.h",
2218 ] + MICROKERNEL_TEST_HDRS,
2219 deps = MICROKERNEL_TEST_DEPS,
2220)
2221
2222xnnpack_unit_test(
2223 name = "f32_vadd_test",
2224 srcs = [
2225 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002226 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002227 ] + MICROKERNEL_TEST_HDRS,
2228 deps = MICROKERNEL_TEST_DEPS,
2229)
2230
2231xnnpack_unit_test(
2232 name = "f32_vaddc_test",
2233 srcs = [
2234 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002235 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002236 ] + MICROKERNEL_TEST_HDRS,
2237 deps = MICROKERNEL_TEST_DEPS,
2238)
2239
2240xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002241 name = "f32_vdiv_test",
2242 srcs = [
2243 "test/f32-vdiv.cc",
2244 "test/vbinary-microkernel-tester.h",
2245 ] + MICROKERNEL_TEST_HDRS,
2246 deps = MICROKERNEL_TEST_DEPS,
2247)
2248
2249xnnpack_unit_test(
2250 name = "f32_vdivc_test",
2251 srcs = [
2252 "test/f32-vdivc.cc",
2253 "test/vbinaryc-microkernel-tester.h",
2254 ] + MICROKERNEL_TEST_HDRS,
2255 deps = MICROKERNEL_TEST_DEPS,
2256)
2257
2258xnnpack_unit_test(
2259 name = "f32_vrdivc_test",
2260 srcs = [
2261 "test/f32-vrdivc.cc",
2262 "test/vbinaryc-microkernel-tester.h",
2263 ] + MICROKERNEL_TEST_HDRS,
2264 deps = MICROKERNEL_TEST_DEPS,
2265)
2266
2267xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002268 name = "f32_vmax_test",
2269 srcs = [
2270 "test/f32-vmax.cc",
2271 "test/vbinary-microkernel-tester.h",
2272 ] + MICROKERNEL_TEST_HDRS,
2273 deps = MICROKERNEL_TEST_DEPS,
2274)
2275
2276xnnpack_unit_test(
2277 name = "f32_vmaxc_test",
2278 srcs = [
2279 "test/f32-vmaxc.cc",
2280 "test/vbinaryc-microkernel-tester.h",
2281 ] + MICROKERNEL_TEST_HDRS,
2282 deps = MICROKERNEL_TEST_DEPS,
2283)
2284
2285xnnpack_unit_test(
2286 name = "f32_vmin_test",
2287 srcs = [
2288 "test/f32-vmin.cc",
2289 "test/vbinary-microkernel-tester.h",
2290 ] + MICROKERNEL_TEST_HDRS,
2291 deps = MICROKERNEL_TEST_DEPS,
2292)
2293
2294xnnpack_unit_test(
2295 name = "f32_vminc_test",
2296 srcs = [
2297 "test/f32-vminc.cc",
2298 "test/vbinaryc-microkernel-tester.h",
2299 ] + MICROKERNEL_TEST_HDRS,
2300 deps = MICROKERNEL_TEST_DEPS,
2301)
2302
2303xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002304 name = "f32_vmul_test",
2305 srcs = [
2306 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002307 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002308 ] + MICROKERNEL_TEST_HDRS,
2309 deps = MICROKERNEL_TEST_DEPS,
2310)
2311
2312xnnpack_unit_test(
2313 name = "f32_vmulc_test",
2314 srcs = [
2315 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002316 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002317 ] + MICROKERNEL_TEST_HDRS,
2318 deps = MICROKERNEL_TEST_DEPS,
2319)
2320
2321xnnpack_unit_test(
2322 name = "f32_vmulcaddc_test",
2323 srcs = [
2324 "test/f32-vmulcaddc.cc",
2325 "test/vmulcaddc-microkernel-tester.h",
2326 "src/xnnpack/AlignedAllocator.h",
2327 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2328 deps = MICROKERNEL_TEST_DEPS,
2329)
2330
2331xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002332 name = "f32_vscale_test",
2333 srcs = [
2334 "test/f32-vscale.cc",
2335 "test/vscale-microkernel-tester.h",
2336 ] + MICROKERNEL_TEST_HDRS,
2337 deps = MICROKERNEL_TEST_DEPS,
2338)
2339
2340xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002341 name = "f32_vscaleexpminusmax_test",
2342 srcs = [
2343 "test/f32-vscaleexpminusmax.cc",
2344 "test/vscaleexpminusmax-microkernel-tester.h",
2345 ] + MICROKERNEL_TEST_HDRS,
2346 deps = MICROKERNEL_TEST_DEPS,
2347)
2348
2349xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002350 name = "f32_vscaleextexp_test",
2351 srcs = [
2352 "test/f32-vscaleextexp.cc",
2353 "test/vscaleextexp-microkernel-tester.h",
2354 ] + MICROKERNEL_TEST_HDRS,
2355 deps = MICROKERNEL_TEST_DEPS,
2356)
2357
2358xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002359 name = "f32_vsub_test",
2360 srcs = [
2361 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002362 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002363 ] + MICROKERNEL_TEST_HDRS,
2364 deps = MICROKERNEL_TEST_DEPS,
2365)
2366
2367xnnpack_unit_test(
2368 name = "f32_vsubc_test",
2369 srcs = [
2370 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002371 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002372 ] + MICROKERNEL_TEST_HDRS,
2373 deps = MICROKERNEL_TEST_DEPS,
2374)
2375
2376xnnpack_unit_test(
2377 name = "f32_vrsubc_test",
2378 srcs = [
2379 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002380 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002381 ] + MICROKERNEL_TEST_HDRS,
2382 deps = MICROKERNEL_TEST_DEPS,
2383)
2384
2385xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002386 name = "q8_avgpool_test",
2387 srcs = [
2388 "test/q8-avgpool.cc",
2389 "test/avgpool-microkernel-tester.h",
2390 "src/xnnpack/AlignedAllocator.h",
2391 ] + MICROKERNEL_TEST_HDRS,
2392 deps = MICROKERNEL_TEST_DEPS,
2393)
2394
2395xnnpack_unit_test(
2396 name = "q8_igemm_test",
2397 srcs = [
2398 "test/q8-igemm.cc",
2399 "test/gemm-microkernel-tester.h",
2400 "src/xnnpack/AlignedAllocator.h",
2401 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2402 deps = MICROKERNEL_TEST_DEPS,
2403)
2404
2405xnnpack_unit_test(
2406 name = "q8_dwconv_test",
2407 srcs = [
2408 "test/q8-dwconv.cc",
2409 "test/dwconv-microkernel-tester.h",
2410 "src/xnnpack/AlignedAllocator.h",
2411 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2412 deps = MICROKERNEL_TEST_DEPS,
2413)
2414
2415xnnpack_unit_test(
2416 name = "q8_gavgpool_test",
2417 srcs = [
2418 "test/q8-gavgpool.cc",
2419 "test/gavgpool-microkernel-tester.h",
2420 "src/xnnpack/AlignedAllocator.h",
2421 ] + MICROKERNEL_TEST_HDRS,
2422 deps = MICROKERNEL_TEST_DEPS,
2423)
2424
2425xnnpack_unit_test(
2426 name = "q8_gemm_test",
2427 srcs = [
2428 "test/q8-gemm.cc",
2429 "test/gemm-microkernel-tester.h",
2430 "src/xnnpack/AlignedAllocator.h",
2431 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2432 deps = MICROKERNEL_TEST_DEPS,
2433)
2434
2435xnnpack_unit_test(
2436 name = "q8_vadd_test",
2437 srcs = [
2438 "test/q8-vadd.cc",
2439 "test/vadd-microkernel-tester.h",
2440 ] + MICROKERNEL_TEST_HDRS,
2441 deps = MICROKERNEL_TEST_DEPS,
2442)
2443
2444xnnpack_unit_test(
2445 name = "u8_clamp_test",
2446 srcs = [
2447 "test/u8-clamp.cc",
2448 "test/clamp-microkernel-tester.h",
2449 ] + MICROKERNEL_TEST_HDRS,
2450 deps = MICROKERNEL_TEST_DEPS,
2451)
2452
2453xnnpack_unit_test(
2454 name = "u8_lut32norm_test",
2455 srcs = [
2456 "test/u8-lut32norm.cc",
2457 "test/lut-norm-microkernel-tester.h",
2458 ] + MICROKERNEL_TEST_HDRS,
2459 deps = MICROKERNEL_TEST_DEPS,
2460)
2461
2462xnnpack_unit_test(
2463 name = "u8_maxpool_test",
2464 srcs = [
2465 "test/u8-maxpool.cc",
2466 "test/maxpool-microkernel-tester.h",
2467 ] + MICROKERNEL_TEST_HDRS,
2468 deps = MICROKERNEL_TEST_DEPS,
2469)
2470
2471xnnpack_unit_test(
2472 name = "u8_rmax_test",
2473 srcs = [
2474 "test/u8-rmax.cc",
2475 "test/rmax-microkernel-tester.h",
2476 ] + MICROKERNEL_TEST_HDRS,
2477 deps = MICROKERNEL_TEST_DEPS,
2478)
2479
2480xnnpack_unit_test(
2481 name = "x32_packx_test",
2482 srcs = [
2483 "test/x32-packx.cc",
2484 "test/pack-microkernel-tester.h",
2485 "src/xnnpack/AlignedAllocator.h",
2486 ] + MICROKERNEL_TEST_HDRS,
2487 deps = MICROKERNEL_TEST_DEPS,
2488)
2489
2490xnnpack_unit_test(
2491 name = "x32_pad_test",
2492 srcs = [
2493 "test/x32-pad.cc",
2494 "test/pad-microkernel-tester.h",
2495 ] + MICROKERNEL_TEST_HDRS,
2496 deps = MICROKERNEL_TEST_DEPS,
2497)
2498
2499xnnpack_unit_test(
2500 name = "x32_unpool_test",
2501 srcs = [
2502 "test/x32-unpool.cc",
2503 "test/unpool-microkernel-tester.h",
2504 ] + MICROKERNEL_TEST_HDRS,
2505 deps = MICROKERNEL_TEST_DEPS,
2506)
2507
2508xnnpack_unit_test(
2509 name = "x32_zip_test",
2510 srcs = [
2511 "test/x32-zip.cc",
2512 "test/zip-microkernel-tester.h",
2513 ] + MICROKERNEL_TEST_HDRS,
2514 deps = MICROKERNEL_TEST_DEPS,
2515)
2516
2517xnnpack_unit_test(
2518 name = "x8_lut_test",
2519 srcs = [
2520 "test/x8-lut.cc",
2521 "test/lut-microkernel-tester.h",
2522 ] + MICROKERNEL_TEST_HDRS,
2523 deps = MICROKERNEL_TEST_DEPS,
2524)
2525
2526xnnpack_unit_test(
2527 name = "x8_zip_test",
2528 srcs = [
2529 "test/x8-zip.cc",
2530 "test/zip-microkernel-tester.h",
2531 ] + MICROKERNEL_TEST_HDRS,
2532 deps = MICROKERNEL_TEST_DEPS,
2533)
2534
2535########################### Size test for the library ##########################
2536
2537xnnpack_binary(
2538 name = "size_test",
2539 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002540 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002541)
2542
2543########################### Unit tests for operators ###########################
2544
2545xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002546 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002547 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002548 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002549 "test/add-operator-tester.h",
2550 ],
2551 deps = OPERATOR_TEST_DEPS,
2552)
2553
2554xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002555 name = "add_nd_test",
2556 srcs = [
2557 "test/add-nd.cc",
2558 "test/binary-elementwise-operator-tester.h",
2559 ],
2560 deps = OPERATOR_TEST_DEPS,
2561)
2562
2563xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002564 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002565 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002566 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002567 "test/argmax-pooling-operator-tester.h",
2568 ] + OPERATOR_TEST_PARAMS_HDRS,
2569 deps = OPERATOR_TEST_DEPS,
2570)
2571
2572xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002573 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002574 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002575 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002576 "test/average-pooling-operator-tester.h",
2577 ] + OPERATOR_TEST_PARAMS_HDRS,
2578 deps = OPERATOR_TEST_DEPS,
2579)
2580
2581xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002582 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002583 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002584 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002585 "test/channel-pad-operator-tester.h",
2586 ] + OPERATOR_TEST_PARAMS_HDRS,
2587 deps = OPERATOR_TEST_DEPS,
2588)
2589
2590xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002591 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002592 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002593 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002594 "test/channel-shuffle-operator-tester.h",
2595 ],
2596 deps = OPERATOR_TEST_DEPS,
2597)
2598
2599xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002600 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002601 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002602 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002603 "test/clamp-operator-tester.h",
2604 ],
2605 deps = OPERATOR_TEST_DEPS,
2606)
2607
2608xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002609 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002610 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002611 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002612 "test/convolution-operator-tester.h",
2613 ],
2614 deps = OPERATOR_TEST_DEPS,
2615)
2616
2617xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002618 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002619 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002620 "test/convolution-nchw.cc",
2621 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002622 ],
2623 deps = OPERATOR_TEST_DEPS,
2624)
2625
2626xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002627 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002628 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002629 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002630 "test/deconvolution-operator-tester.h",
2631 ] + OPERATOR_TEST_PARAMS_HDRS,
2632 deps = OPERATOR_TEST_DEPS,
2633)
2634
2635xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002636 name = "divide_nd_test",
2637 srcs = [
2638 "test/binary-elementwise-operator-tester.h",
2639 "test/divide-nd.cc",
2640 ],
2641 deps = OPERATOR_TEST_DEPS,
2642)
2643
2644xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002645 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002646 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002647 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002648 "test/fully-connected-operator-tester.h",
2649 ],
2650 deps = OPERATOR_TEST_DEPS,
2651)
2652
2653xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002654 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002655 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002656 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002657 "test/global-average-pooling-operator-tester.h",
2658 ] + OPERATOR_TEST_PARAMS_HDRS,
2659 deps = OPERATOR_TEST_DEPS,
2660)
2661
2662xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002663 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002664 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002665 "test/global-average-pooling-ncw.cc",
2666 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002667 ],
2668 deps = OPERATOR_TEST_DEPS,
2669)
2670
2671xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002672 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002673 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002674 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002675 "test/hardswish-operator-tester.h",
2676 ],
2677 deps = OPERATOR_TEST_DEPS,
2678)
2679
2680xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002681 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002682 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002683 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002684 "test/leaky-relu-operator-tester.h",
2685 ],
2686 deps = OPERATOR_TEST_DEPS,
2687)
2688
2689xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002690 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002691 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002692 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002693 "test/max-pooling-operator-tester.h",
2694 ] + OPERATOR_TEST_PARAMS_HDRS,
2695 deps = OPERATOR_TEST_DEPS,
2696)
2697
2698xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08002699 name = "maximum_nd_test",
2700 srcs = [
2701 "test/binary-elementwise-operator-tester.h",
2702 "test/maximum-nd.cc",
2703 ],
2704 deps = OPERATOR_TEST_DEPS,
2705)
2706
2707xnnpack_unit_test(
2708 name = "minimum_nd_test",
2709 srcs = [
2710 "test/binary-elementwise-operator-tester.h",
2711 "test/minimum-nd.cc",
2712 ],
2713 deps = OPERATOR_TEST_DEPS,
2714)
2715
2716xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002717 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002718 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002719 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002720 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002721 ],
2722 deps = OPERATOR_TEST_DEPS,
2723)
2724
2725xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002726 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002727 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002728 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002729 "test/prelu-operator-tester.h",
2730 ] + OPERATOR_TEST_PARAMS_HDRS,
2731 deps = OPERATOR_TEST_DEPS,
2732)
2733
2734xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002735 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002736 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002737 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002738 "test/resize-bilinear-operator-tester.h",
2739 ] + OPERATOR_TEST_PARAMS_HDRS,
2740 deps = OPERATOR_TEST_DEPS,
2741)
2742
2743xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002744 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002745 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002746 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747 "test/sigmoid-operator-tester.h",
2748 ],
2749 deps = OPERATOR_TEST_DEPS,
2750)
2751
2752xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002753 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002754 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002755 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002756 "test/softargmax-operator-tester.h",
2757 ],
2758 deps = OPERATOR_TEST_DEPS,
2759)
2760
2761xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08002762 name = "subtract_nd_test",
2763 srcs = [
2764 "test/binary-elementwise-operator-tester.h",
2765 "test/subtract-nd.cc",
2766 ],
2767 deps = OPERATOR_TEST_DEPS,
2768)
2769
2770xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002771 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002772 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002773 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002774 "test/unpooling-operator-tester.h",
2775 ],
2776 deps = OPERATOR_TEST_DEPS,
2777)
2778
2779############################# Build configurations #############################
2780
Marat Dukhanb8642352019-10-30 15:43:02 -07002781# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002782config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002783 name = "xnn_enable_assembly_explicit_true",
2784 define_values = {"xnn_enable_assembly": "true"},
2785)
2786
2787# Disables usage of assembly kernels.
2788config_setting(
2789 name = "xnn_enable_assembly_explicit_false",
2790 define_values = {"xnn_enable_assembly": "false"},
2791)
2792
2793# Builds with -c dbg
2794config_setting(
2795 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002796 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002797 "compilation_mode": "dbg",
2798 },
2799)
2800
2801# Builds with -c opt
2802config_setting(
2803 name = "optimized_build",
2804 values = {
2805 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002806 },
2807)
2808
2809config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002810 name = "linux_k8",
2811 values = {"cpu": "k8"},
2812)
2813
2814config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002815 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002816 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002817)
2818
2819config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820 name = "android",
2821 values = {"crosstool_top": "//external:android/crosstool"},
2822)
2823
2824config_setting(
2825 name = "android_armv7",
2826 values = {
2827 "crosstool_top": "//external:android/crosstool",
2828 "cpu": "armeabi-v7a",
2829 },
2830)
2831
2832config_setting(
2833 name = "android_arm64",
2834 values = {
2835 "crosstool_top": "//external:android/crosstool",
2836 "cpu": "arm64-v8a",
2837 },
2838)
2839
2840config_setting(
2841 name = "android_x86",
2842 values = {
2843 "crosstool_top": "//external:android/crosstool",
2844 "cpu": "x86",
2845 },
2846)
2847
2848config_setting(
2849 name = "android_x86_64",
2850 values = {
2851 "crosstool_top": "//external:android/crosstool",
2852 "cpu": "x86_64",
2853 },
2854)
2855
2856config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002857 name = "macos_x86_64",
2858 values = {
2859 "apple_platform_type": "macos",
2860 "cpu": "darwin",
2861 },
2862)
2863
2864config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002865 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002866 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867)
2868
2869config_setting(
2870 name = "emscripten_wasm",
2871 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002872 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002873 "cpu": "wasm",
2874 },
2875)
2876
2877config_setting(
2878 name = "emscripten_wasmsimd",
2879 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002880 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002881 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002882 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002883 },
2884)
2885
2886config_setting(
2887 name = "emscripten_asmjs",
2888 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002889 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 "cpu": "asmjs",
2891 },
2892)