blob: 9bdae62d2e9c3fc91aa6ed43e3e2ddbcb71d08e5 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_copts", "xnnpack_unit_test", "xnnpack_visibility")
10
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan08c4a432019-10-03 09:29:21 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
22MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
46OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
52OPERATOR_SRCS = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
72 "src/softargmax-nc.c",
73 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
76SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080077 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080078 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080079 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070080 "src/f32-avgpool/mp9p8q-scalar.c",
81 "src/f32-avgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080082 "src/f32-bilinear/gen/scalar-c1.c",
83 "src/f32-bilinear/gen/scalar-c2.c",
84 "src/f32-bilinear/gen/scalar-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070085 "src/f32-clamp/scalar.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080086 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080087 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070088 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070089 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070090 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080091 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080092 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080093 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
94 "src/f32-dwconv/gen/up1x25-scalar.c",
95 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
96 "src/f32-dwconv/gen/up1x4-scalar.c",
97 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x9-scalar.c",
99 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
100 "src/f32-dwconv/gen/up2x25-scalar.c",
101 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
102 "src/f32-dwconv/gen/up2x4-scalar.c",
103 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700105 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700106 "src/f32-gavgpool/mp7p7q-scalar.c",
107 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800108 "src/f32-gemm/gen-inc/1x4-scalar.c",
109 "src/f32-gemm/gen-inc/2x4-scalar.c",
110 "src/f32-gemm/gen-inc/4x4-scalar.c",
111 "src/f32-gemm/gen/1x4-scalar.c",
112 "src/f32-gemm/gen/2x4-scalar.c",
113 "src/f32-gemm/gen/4x2-scalar.c",
114 "src/f32-gemm/gen/4x4-scalar.c",
Marat 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",
Marat Dukhan27121322019-12-09 14:57:40 -0800856 "src/f32-gemm/gen/1x16s4-fma3.c",
857 "src/f32-gemm/gen/3x16s4-fma3.c",
858 "src/f32-gemm/gen/4x16s4-fma3.c",
859 "src/f32-gemm/gen/5x16s4-fma3.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",
Marat Dukhan27121322019-12-09 14:57:40 -0800870 "src/f32-gemm/gen-inc/1x16s4-fma3.c",
871 "src/f32-gemm/gen-inc/3x16s4-fma3.c",
872 "src/f32-gemm/gen-inc/4x16s4-fma3.c",
873 "src/f32-gemm/gen-inc/5x16s4-fma3.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",
Marat Dukhan27121322019-12-09 14:57:40 -0800886 "src/f32-igemm/gen/1x16s4-fma3.c",
887 "src/f32-igemm/gen/3x16s4-fma3.c",
888 "src/f32-igemm/gen/4x16s4-fma3.c",
889 "src/f32-igemm/gen/5x16s4-fma3.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 Barchard3e237f22019-12-04 23:08:51 -08001093 "src/f32-gemm/4x8-aarch32-neon-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001094 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001095]
1096
1097AARCH64_ASM_UKERNELS = [
1098 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1099 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001100 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1101 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1102 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1103 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1104 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1105 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1106 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1107 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1108 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1109 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
1110 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1111 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
1112 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1113 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
1114 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1115 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1116 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1117 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1118 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1119 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1120 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1121 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1122 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1123 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1124 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1125 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1126 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
1127 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1128 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
1129 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
1130 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
1131 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1132 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1133 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001134 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001135 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001136 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
1137 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
1138 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001139 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001140 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
1141 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001142 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
1144 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
1145 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
1146]
1147
1148INTERNAL_MICROKERNEL_HDRS = [
1149 "src/xnnpack/argmaxpool.h",
1150 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001151 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001152 "src/xnnpack/clamp.h",
1153 "src/xnnpack/common.h",
1154 "src/xnnpack/conv.h",
1155 "src/xnnpack/dwconv.h",
1156 "src/xnnpack/gavgpool.h",
1157 "src/xnnpack/gemm.h",
1158 "src/xnnpack/hswish.h",
1159 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001160 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001161 "src/xnnpack/lut.h",
1162 "src/xnnpack/math.h",
1163 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001164 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001165 "src/xnnpack/packx.h",
1166 "src/xnnpack/pad.h",
1167 "src/xnnpack/params.h",
1168 "src/xnnpack/pavgpool.h",
1169 "src/xnnpack/ppmm.h",
1170 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001171 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001172 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001173 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001174 "src/xnnpack/rmax.h",
1175 "src/xnnpack/scalar-utils.h",
1176 "src/xnnpack/spmm.h",
1177 "src/xnnpack/unpool.h",
1178 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001179 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001180 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001181 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001182 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001183 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001184 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001185 "src/xnnpack/zip.h",
1186]
1187
1188INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1189 "include/xnnpack.h",
1190 "src/xnnpack/allocator.h",
1191 "src/xnnpack/compute.h",
1192 "src/xnnpack/im2col.h",
1193 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001194 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001195 "src/xnnpack/operator.h",
1196 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001197 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001198 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001199 "src/xnnpack/requantization.h",
1200]
1201
1202ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1203 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001204]
1205
1206MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001207 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001208 "include/xnnpack.h",
1209]
1210
1211MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1212 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001213 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001214 "src/xnnpack/requantization.h",
1215 "include/xnnpack.h",
1216]
1217
1218OPERATOR_TEST_PARAMS_HDRS = [
1219 "src/xnnpack/params.h",
1220 "src/xnnpack/common.h",
1221]
1222
1223WEIGHTS_PACK_HDRS = [
1224 "src/xnnpack/pack.h",
1225 "src/xnnpack/operator.h",
1226 "src/xnnpack/compute.h",
1227]
1228
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001229LOGGING_COPTS = select({
1230 # No logging in optimized mode
1231 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1232 # Full logging in debug mode
1233 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1234 # Error-only logging in default (fastbuild) mode
1235 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1236})
1237
1238LOGGING_HDRS = [
1239 "src/xnnpack/log.h",
1240]
1241
Marat Dukhan08c4a432019-10-03 09:29:21 -07001242xnnpack_cc_library(
1243 name = "scalar_ukernels",
1244 srcs = SCALAR_UKERNELS,
1245 hdrs = INTERNAL_HDRS,
1246 aarch32_copts = ["-marm"],
1247 copts = xnnpack_std_copts(),
1248 deps = [
1249 "@FP16",
1250 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001251 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001252 ],
1253)
1254
1255xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001256 name = "wasm_ukernels",
1257 hdrs = INTERNAL_HDRS,
1258 copts = xnnpack_std_copts(),
1259 wasm_srcs = WASM_UKERNELS,
1260 deps = [
1261 "@FP16",
1262 "@FXdiv",
1263 "@pthreadpool",
1264 ],
1265)
1266
1267xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001268 name = "psimd_ukernels",
1269 srcs = PSIMD_UKERNELS,
1270 hdrs = INTERNAL_HDRS,
1271 aarch32_copts = [
1272 "-marm",
1273 "-mfpu=neon",
1274 ],
1275 copts = xnnpack_std_copts(),
1276 optimized_copts = [
1277 "-O3",
1278 "-ffast-math",
1279 ],
1280 deps = [
1281 "@FP16",
1282 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001283 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001284 ],
1285)
1286
1287xnnpack_cc_library(
1288 name = "neon_ukernels",
1289 hdrs = INTERNAL_HDRS,
1290 aarch32_copts = [
1291 "-marm",
1292 "-mfpu=neon",
1293 ],
1294 aarch32_srcs = NEON_UKERNELS,
1295 aarch64_srcs = NEON_UKERNELS,
1296 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001297 deps = [
1298 "@FP16",
1299 "@pthreadpool",
1300 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001301)
1302
1303xnnpack_cc_library(
1304 name = "neonfma_ukernels",
1305 hdrs = INTERNAL_HDRS,
1306 aarch32_copts = [
1307 "-marm",
1308 "-mfpu=neon-vfpv4",
1309 ],
1310 aarch32_srcs = NEONFMA_UKERNELS,
1311 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1312 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001313 deps = [
1314 "@FP16",
1315 "@pthreadpool",
1316 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001317)
1318
1319xnnpack_cc_library(
1320 name = "neonfp16arith_ukernels",
1321 hdrs = INTERNAL_HDRS,
1322 aarch64_copts = ["-march=armv8.2-a+fp16"],
1323 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1324 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001325 deps = [
1326 "@FP16",
1327 "@pthreadpool",
1328 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001329)
1330
1331xnnpack_cc_library(
1332 name = "sse2_ukernels",
1333 hdrs = INTERNAL_HDRS,
1334 copts = xnnpack_std_copts(),
1335 x86_copts = ["-msse2"],
1336 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001337 deps = [
1338 "@FP16",
1339 "@pthreadpool",
1340 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001341)
1342
1343xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001344 name = "sse41_ukernels",
1345 hdrs = INTERNAL_HDRS,
1346 copts = xnnpack_std_copts(),
1347 x86_copts = ["-msse4.1"],
1348 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001349 deps = [
1350 "@FP16",
1351 "@pthreadpool",
1352 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001353)
1354
1355xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001356 name = "avx_ukernels",
1357 hdrs = INTERNAL_HDRS,
1358 copts = xnnpack_std_copts(),
1359 x86_copts = ["-mavx"],
1360 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001361 deps = [
1362 "@FP16",
1363 "@pthreadpool",
1364 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001365)
1366
1367xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001368 name = "fma3_ukernels",
1369 hdrs = INTERNAL_HDRS,
1370 copts = xnnpack_std_copts(),
1371 x86_copts = [
1372 "-mfma",
1373 ],
1374 x86_srcs = FMA3_UKERNELS,
1375 deps = [
1376 "@FP16",
1377 "@pthreadpool",
1378 ],
1379)
1380
1381xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001382 name = "avx2_ukernels",
1383 hdrs = INTERNAL_HDRS,
1384 copts = xnnpack_std_copts(),
1385 x86_copts = [
1386 "-mfma",
1387 "-mavx2",
1388 ],
1389 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001390 deps = [
1391 "@FP16",
1392 "@pthreadpool",
1393 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001394)
1395
1396xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001397 name = "avx512f_ukernels",
1398 hdrs = INTERNAL_HDRS,
1399 copts = xnnpack_std_copts(),
1400 x86_copts = ["-mavx512f"],
1401 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001402 deps = [
1403 "@FP16",
1404 "@pthreadpool",
1405 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001406)
1407
1408xnnpack_cc_library(
1409 name = "asm_ukernels",
1410 hdrs = ["src/xnnpack/assembly.h"],
1411 aarch32_srcs = AARCH32_ASM_UKERNELS,
1412 aarch64_srcs = AARCH64_ASM_UKERNELS,
1413)
1414
1415xnnpack_aggregate_library(
1416 name = "ukernels",
1417 aarch32_deps = [
1418 ":psimd_ukernels",
1419 ":neon_ukernels",
1420 ":neonfma_ukernels",
1421 ":asm_ukernels",
1422 ],
1423 aarch64_deps = [
1424 ":psimd_ukernels",
1425 ":neon_ukernels",
1426 ":neonfma_ukernels",
1427 ":neonfp16arith_ukernels",
1428 ":asm_ukernels",
1429 ],
1430 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001431 wasm_deps = [
1432 ":wasm_ukernels",
1433 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001434 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001435 ":wasm_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001436 ":psimd_ukernels",
1437 ],
1438 x86_deps = [
1439 ":psimd_ukernels",
1440 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001441 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001442 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001443 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001444 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001445 ":avx512f_ukernels",
1446 ],
1447)
1448
1449xnnpack_cc_library(
1450 name = "im2col",
1451 srcs = ["src/im2col.c"],
1452 hdrs = [
1453 "src/xnnpack/common.h",
1454 "src/xnnpack/im2col.h",
1455 ],
1456 copts = xnnpack_std_copts(),
1457)
1458
1459xnnpack_cc_library(
1460 name = "indirection",
1461 srcs = ["src/indirection.c"],
1462 hdrs = INTERNAL_HDRS,
1463 copts = xnnpack_std_copts(),
1464 deps = [
1465 "@FP16",
1466 "@FXdiv",
1467 "@pthreadpool",
1468 ],
1469)
1470
1471xnnpack_cc_library(
1472 name = "operator_run",
1473 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001474 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1475 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001476 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1477 "-Wno-vla",
1478 ],
1479 deps = [
1480 "@FP16",
1481 "@FXdiv",
1482 "@clog",
1483 "@pthreadpool",
1484 ],
1485)
1486
1487cc_library(
1488 name = "enable_assembly",
1489 defines = select({
1490 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1491 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001492 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001493 }),
1494)
1495
Marat Dukhancf056b22019-10-07 10:26:29 -07001496xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001497 name = "operators",
1498 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001499 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001500 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001501 ],
1502 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001503 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001504 "-Isrc",
1505 "-Iinclude",
1506 ] + select({
1507 ":debug_build": [],
1508 "//conditions:default": xnnpack_min_size_copts(),
1509 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001510 wasm_srcs = ["src/wasm-stubs.c"],
1511 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001512 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001513 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001514 "@FP16",
1515 "@FXdiv",
1516 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001517 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001518 ],
1519)
1520
1521cc_library(
1522 name = "XNNPACK",
1523 srcs = [
1524 "src/init.c",
1525 ],
1526 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1527 "-Isrc",
1528 "-Iinclude",
1529 ] + select({
1530 ":debug_build": [],
1531 "//conditions:default": xnnpack_min_size_copts(),
1532 }),
1533 includes = ["include"],
1534 linkstatic = True,
1535 textual_hdrs = ["include/xnnpack.h"],
1536 visibility = xnnpack_visibility(),
1537 deps = [
1538 ":enable_assembly",
1539 ":ukernels",
1540 ":operator_run",
1541 ":operators",
1542 "@clog",
1543 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001544 ] + select({
1545 ":emscripten": [],
1546 "//conditions:default": ["@cpuinfo"],
1547 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001548)
1549
1550cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001551 name = "xnnpack_operators_nhwc_f32",
1552 srcs = [
1553 "src/init.c",
1554 ],
1555 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1556 "-Isrc",
1557 "-Iinclude",
1558 ] + select({
1559 ":debug_build": [],
1560 "//conditions:default": xnnpack_min_size_copts(),
1561 }),
1562 defines = [
1563 "XNN_NO_Q8_OPERATORS",
1564 "XNN_NO_U8_OPERATORS",
1565 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001566 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001567 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001568 includes = ["include"],
1569 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001570 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001571 visibility = xnnpack_visibility(),
1572 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001573 ":enable_assembly",
1574 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001575 ":operator_run",
1576 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001577 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001578 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001579 ] + select({
1580 ":emscripten": [],
1581 "//conditions:default": ["@cpuinfo"],
1582 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001583)
1584
Marat Dukhancf056b22019-10-07 10:26:29 -07001585xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001586 name = "bench_utils",
1587 srcs = ["bench/utils.cc"],
1588 hdrs = ["bench/utils.h"],
1589 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001590 deps = [
1591 "@com_google_benchmark//:benchmark",
1592 "@cpuinfo",
1593 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001594)
1595
Frank Barchard7e955972019-10-11 10:34:25 -07001596######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001597
1598xnnpack_benchmark(
1599 name = "q8_gemm_bench",
1600 srcs = [
1601 "bench/gemm.h",
1602 "bench/q8-gemm.cc",
1603 "src/xnnpack/AlignedAllocator.h",
1604 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1605 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1606 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1607)
1608
1609xnnpack_benchmark(
1610 name = "f16_gemm_bench",
1611 srcs = [
1612 "bench/f16-gemm.cc",
1613 "bench/gemm.h",
1614 "src/xnnpack/AlignedAllocator.h",
1615 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1616 copts = ["-Wno-unused-function"],
1617 deps = MICROKERNEL_BENCHMARK_DEPS,
1618)
1619
1620xnnpack_benchmark(
1621 name = "f32_igemm_bench",
1622 srcs = [
1623 "bench/f32-igemm.cc",
1624 "bench/conv.h",
1625 "src/xnnpack/AlignedAllocator.h",
1626 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001627 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001628)
1629
1630xnnpack_benchmark(
1631 name = "f32_conv_hwc_bench",
1632 srcs = [
1633 "bench/f32-conv-hwc.cc",
1634 "bench/dconv.h",
1635 "src/xnnpack/AlignedAllocator.h",
1636 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1637 copts = ["-Wno-unused-function"],
1638 deps = MICROKERNEL_BENCHMARK_DEPS,
1639)
1640
1641xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001642 name = "f32_conv_hwc2spchw_bench",
1643 srcs = [
1644 "bench/f32-conv-hwc2spchw.cc",
1645 "bench/dconv.h",
1646 "src/xnnpack/AlignedAllocator.h",
1647 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1648 copts = ["-Wno-unused-function"],
1649 deps = MICROKERNEL_BENCHMARK_DEPS,
1650)
1651
1652xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001653 name = "f32_dwconv_bench",
1654 srcs = [
1655 "bench/f32-dwconv.cc",
1656 "bench/dwconv.h",
1657 "src/xnnpack/AlignedAllocator.h",
1658 ] + MICROKERNEL_BENCHMARK_HDRS,
1659 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1660)
1661
1662xnnpack_benchmark(
1663 name = "f32_dwconv_spchw_bench",
1664 srcs = [
1665 "bench/f32-dwconv-spchw.cc",
1666 "bench/dwconv.h",
1667 "src/xnnpack/AlignedAllocator.h",
1668 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1669 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1670)
1671
1672xnnpack_benchmark(
1673 name = "f32_gemm_bench",
1674 srcs = [
1675 "bench/f32-gemm.cc",
1676 "bench/gemm.h",
1677 "src/xnnpack/AlignedAllocator.h",
1678 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1679 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001680 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001681)
1682
1683xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001684 name = "f32_raddexpminusmax_bench",
1685 srcs = [
1686 "bench/f32-raddexpminusmax.cc",
1687 "src/xnnpack/AlignedAllocator.h",
1688 ] + MICROKERNEL_BENCHMARK_HDRS,
1689 deps = MICROKERNEL_BENCHMARK_DEPS,
1690)
1691
1692xnnpack_benchmark(
1693 name = "f32_raddextexp_bench",
1694 srcs = [
1695 "bench/f32-raddextexp.cc",
1696 "src/xnnpack/AlignedAllocator.h",
1697 ] + MICROKERNEL_BENCHMARK_HDRS,
1698 deps = MICROKERNEL_BENCHMARK_DEPS,
1699)
1700
1701xnnpack_benchmark(
1702 name = "f32_raddstoreexpminusmax_bench",
1703 srcs = [
1704 "bench/f32-raddstoreexpminusmax.cc",
1705 "src/xnnpack/AlignedAllocator.h",
1706 ] + MICROKERNEL_BENCHMARK_HDRS,
1707 deps = MICROKERNEL_BENCHMARK_DEPS,
1708)
1709
1710xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001711 name = "f32_rmax_bench",
1712 srcs = [
1713 "bench/f32-rmax.cc",
1714 "src/xnnpack/AlignedAllocator.h",
1715 ] + MICROKERNEL_BENCHMARK_HDRS,
1716 deps = MICROKERNEL_BENCHMARK_DEPS,
1717)
1718
1719xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001720 name = "f32_sigmoid_bench",
1721 srcs = [
1722 "bench/f32-sigmoid.cc",
1723 "src/xnnpack/AlignedAllocator.h",
1724 ] + MICROKERNEL_BENCHMARK_HDRS,
1725 copts = ["-Wno-unused-function"],
1726 deps = MICROKERNEL_BENCHMARK_DEPS,
1727)
1728
1729xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730 name = "f32_spmm_bench",
1731 srcs = [
1732 "bench/f32-spmm.cc",
1733 "bench/gemm.h",
1734 "src/xnnpack/AlignedAllocator.h",
1735 ] + MICROKERNEL_BENCHMARK_HDRS,
1736 copts = ["-Wno-unused-function"],
1737 deps = MICROKERNEL_BENCHMARK_DEPS,
1738)
1739
1740xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001741 name = "f32_softargmax_bench",
1742 srcs = [
1743 "bench/f32-softargmax.cc",
1744 ] + MICROKERNEL_BENCHMARK_HDRS,
1745 copts = ["-Wno-unused-function"],
1746 deps = MICROKERNEL_BENCHMARK_DEPS,
1747)
1748
1749xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001750 name = "f32_vscaleexpminusmax_bench",
1751 srcs = [
1752 "bench/f32-vscaleexpminusmax.cc",
1753 "src/xnnpack/AlignedAllocator.h",
1754 ] + MICROKERNEL_BENCHMARK_HDRS,
1755 deps = MICROKERNEL_BENCHMARK_DEPS,
1756)
1757
1758xnnpack_benchmark(
1759 name = "f32_vscaleextexp_bench",
1760 srcs = [
1761 "bench/f32-vscaleextexp.cc",
1762 "src/xnnpack/AlignedAllocator.h",
1763 ] + MICROKERNEL_BENCHMARK_HDRS,
1764 deps = MICROKERNEL_BENCHMARK_DEPS,
1765)
1766
1767xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001768 name = "f32_im2col_gemm_bench",
1769 srcs = [
1770 "bench/f32-im2col-gemm.cc",
1771 "bench/conv.h",
1772 "src/xnnpack/AlignedAllocator.h",
1773 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1774 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1775)
1776
1777########################### Benchmarks for operators ###########################
1778
1779xnnpack_benchmark(
1780 name = "add_bench",
1781 srcs = ["bench/add.cc"],
1782 deps = OPERATOR_BENCHMARK_DEPS,
1783)
1784
1785xnnpack_benchmark(
1786 name = "average_pooling_bench",
1787 srcs = ["bench/average-pooling.cc"],
1788 deps = OPERATOR_BENCHMARK_DEPS,
1789)
1790
1791xnnpack_benchmark(
1792 name = "channel_shuffle_bench",
1793 srcs = ["bench/channel-shuffle.cc"],
1794 deps = OPERATOR_BENCHMARK_DEPS,
1795)
1796
1797xnnpack_benchmark(
1798 name = "convolution_bench",
1799 srcs = ["bench/convolution.cc"],
1800 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1801 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1802)
1803
1804xnnpack_benchmark(
1805 name = "deconvolution_bench",
1806 srcs = ["bench/deconvolution.cc"],
1807 copts = xnnpack_optional_tflite_copts(),
1808 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1809)
1810
1811xnnpack_benchmark(
1812 name = "global_average_pooling_bench",
1813 srcs = ["bench/global-average-pooling.cc"],
1814 deps = OPERATOR_BENCHMARK_DEPS,
1815)
1816
1817xnnpack_benchmark(
1818 name = "max_pooling_bench",
1819 srcs = ["bench/max-pooling.cc"],
1820 deps = OPERATOR_BENCHMARK_DEPS,
1821)
1822
1823xnnpack_benchmark(
1824 name = "sigmoid_bench",
1825 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001826 copts = xnnpack_optional_tflite_copts(),
1827 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001828)
1829
1830xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001831 name = "prelu_bench",
1832 srcs = ["bench/prelu.cc"],
1833 copts = xnnpack_optional_tflite_copts(),
1834 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1835)
1836
1837xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838 name = "softargmax_bench",
1839 srcs = ["bench/softargmax.cc"],
1840 deps = OPERATOR_BENCHMARK_DEPS,
1841)
1842
Marat Dukhanc068bb62019-10-04 13:24:39 -07001843############################# End-to-end benchmarks ############################
1844
1845cc_library(
1846 name = "mobilenet_v1",
1847 srcs = ["models/mobilenet-v1.cc"],
1848 hdrs = ["models/models.h"],
1849 linkstatic = True,
1850 deps = [
1851 ":XNNPACK",
1852 "@pthreadpool",
1853 ],
1854)
1855
1856cc_library(
1857 name = "mobilenet_v2",
1858 srcs = ["models/mobilenet-v2.cc"],
1859 hdrs = ["models/models.h"],
1860 linkstatic = True,
1861 deps = [
1862 ":XNNPACK",
1863 "@pthreadpool",
1864 ],
1865)
1866
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001867cc_library(
1868 name = "mobilenet_v3_large",
1869 srcs = ["models/mobilenet-v3-large.cc"],
1870 hdrs = ["models/models.h"],
1871 linkstatic = True,
1872 deps = [
1873 ":XNNPACK",
1874 "@pthreadpool",
1875 ],
1876)
1877
1878cc_library(
1879 name = "mobilenet_v3_small",
1880 srcs = ["models/mobilenet-v3-small.cc"],
1881 hdrs = ["models/models.h"],
1882 linkstatic = True,
1883 deps = [
1884 ":XNNPACK",
1885 "@pthreadpool",
1886 ],
1887)
1888
Marat Dukhanc068bb62019-10-04 13:24:39 -07001889xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001890 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001891 srcs = [
1892 "bench/f32-dwconv-e2e.cc",
1893 "bench/end2end.h",
1894 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07001895 copts = ["-Wno-unused-function"],
1896 deps = MICROKERNEL_BENCHMARK_DEPS + [
1897 ":XNNPACK",
1898 ":mobilenet_v1",
1899 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001900 ":mobilenet_v3_large",
1901 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07001902 ],
1903)
1904
1905xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001906 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001907 srcs = [
1908 "bench/f32-gemm-e2e.cc",
1909 "bench/end2end.h",
1910 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07001911 copts = ["-Wno-unused-function"],
1912 deps = MICROKERNEL_BENCHMARK_DEPS + [
1913 ":XNNPACK",
1914 ":mobilenet_v1",
1915 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001916 ":mobilenet_v3_large",
1917 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07001918 ],
1919)
1920
1921xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001922 name = "end2end_bench",
1923 srcs = ["bench/end2end.cc"],
1924 deps = [
1925 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001926 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001927 ":mobilenet_v1",
1928 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001929 ":mobilenet_v3_large",
1930 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001931 "@pthreadpool",
1932 ],
1933)
1934
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001935#################### Accuracy evaluation for math functions ####################
1936
1937xnnpack_benchmark(
1938 name = "f32_exp_eval",
1939 srcs = [
1940 "eval/f32-exp.cc",
1941 "src/xnnpack/AlignedAllocator.h",
1942 ] + ACCURACY_EVAL_HDRS,
1943 deps = ACCURACY_EVAL_DEPS,
1944)
1945
Marat Dukhan515c9772019-10-17 18:07:57 -07001946xnnpack_benchmark(
1947 name = "f32_expminus_eval",
1948 srcs = [
1949 "eval/f32-expminus.cc",
1950 "src/xnnpack/AlignedAllocator.h",
1951 ] + ACCURACY_EVAL_HDRS,
1952 deps = ACCURACY_EVAL_DEPS,
1953)
1954
Marat Dukhan98ba4412019-10-23 02:14:28 -07001955xnnpack_benchmark(
1956 name = "f32_extexp_eval",
1957 srcs = [
1958 "eval/f32-extexp.cc",
1959 "src/xnnpack/AlignedAllocator.h",
1960 ] + ACCURACY_EVAL_HDRS,
1961 deps = ACCURACY_EVAL_DEPS,
1962)
1963
Marat Dukhan346a9e52019-11-15 09:06:30 -08001964xnnpack_benchmark(
1965 name = "f32_sigmoid_eval",
1966 srcs = [
1967 "eval/f32-sigmoid.cc",
1968 "src/xnnpack/AlignedAllocator.h",
1969 ] + ACCURACY_EVAL_HDRS,
1970 deps = ACCURACY_EVAL_DEPS,
1971)
1972
Marat Dukhan08c4a432019-10-03 09:29:21 -07001973######################### Unit tests for micro-kernels #########################
1974
1975xnnpack_unit_test(
1976 name = "f16_gemm_test",
1977 srcs = [
1978 "test/f16-gemm.cc",
1979 "test/gemm-microkernel-tester.h",
1980 "src/xnnpack/AlignedAllocator.h",
1981 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1982 deps = MICROKERNEL_TEST_DEPS,
1983)
1984
1985xnnpack_unit_test(
1986 name = "f32_argmaxpool_test",
1987 srcs = [
1988 "test/f32-argmaxpool.cc",
1989 "test/argmaxpool-microkernel-tester.h",
1990 "src/xnnpack/AlignedAllocator.h",
1991 ] + MICROKERNEL_TEST_HDRS,
1992 deps = MICROKERNEL_TEST_DEPS,
1993)
1994
1995xnnpack_unit_test(
1996 name = "f32_avgpool_test",
1997 srcs = [
1998 "test/f32-avgpool.cc",
1999 "test/avgpool-microkernel-tester.h",
2000 "src/xnnpack/AlignedAllocator.h",
2001 ] + MICROKERNEL_TEST_HDRS,
2002 deps = MICROKERNEL_TEST_DEPS,
2003)
2004
2005xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002006 name = "f32_bilinear_test",
2007 srcs = [
2008 "test/f32-bilinear.cc",
2009 "test/bilinear-microkernel-tester.h",
2010 "src/xnnpack/AlignedAllocator.h",
2011 ] + MICROKERNEL_TEST_HDRS,
2012 deps = MICROKERNEL_TEST_DEPS,
2013)
2014
2015xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002016 name = "f32_clamp_test",
2017 srcs = [
2018 "test/f32-clamp.cc",
2019 "test/clamp-microkernel-tester.h",
2020 ] + MICROKERNEL_TEST_HDRS,
2021 deps = MICROKERNEL_TEST_DEPS,
2022)
2023
2024xnnpack_unit_test(
2025 name = "f32_igemm_test",
2026 srcs = [
2027 "test/f32-igemm.cc",
2028 "test/gemm-microkernel-tester.h",
2029 "src/xnnpack/AlignedAllocator.h",
2030 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2031 deps = MICROKERNEL_TEST_DEPS,
2032)
2033
2034xnnpack_unit_test(
2035 name = "f32_conv_hwc_test",
2036 srcs = [
2037 "test/f32-conv-hwc.cc",
2038 "test/conv-hwc-microkernel-tester.h",
2039 "src/xnnpack/AlignedAllocator.h",
2040 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2041 deps = MICROKERNEL_TEST_DEPS,
2042)
2043
2044xnnpack_unit_test(
2045 name = "f32_conv_hwc2spchw_test",
2046 srcs = [
2047 "test/f32-conv-hwc2spchw.cc",
2048 "test/conv-hwc2spchw-microkernel-tester.h",
2049 "src/xnnpack/AlignedAllocator.h",
2050 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2051 deps = MICROKERNEL_TEST_DEPS,
2052)
2053
2054xnnpack_unit_test(
2055 name = "f32_dwconv_test",
2056 srcs = [
2057 "test/f32-dwconv.cc",
2058 "test/dwconv-microkernel-tester.h",
2059 "src/xnnpack/AlignedAllocator.h",
2060 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2061 deps = MICROKERNEL_TEST_DEPS,
2062)
2063
2064xnnpack_unit_test(
2065 name = "f32_dwconv_spchw_test",
2066 srcs = [
2067 "test/f32-dwconv-spchw.cc",
2068 "test/dwconv-spchw-microkernel-tester.h",
2069 "src/xnnpack/AlignedAllocator.h",
2070 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2071 deps = MICROKERNEL_TEST_DEPS,
2072)
2073
2074xnnpack_unit_test(
2075 name = "f32_gavgpool_test",
2076 srcs = [
2077 "test/f32-gavgpool.cc",
2078 "test/gavgpool-microkernel-tester.h",
2079 "src/xnnpack/AlignedAllocator.h",
2080 ] + MICROKERNEL_TEST_HDRS,
2081 deps = MICROKERNEL_TEST_DEPS,
2082)
2083
2084xnnpack_unit_test(
2085 name = "f32_gavgpool_spchw_test",
2086 srcs = [
2087 "test/f32-gavgpool-spchw.cc",
2088 "test/gavgpool-spchw-microkernel-tester.h",
2089 "src/xnnpack/AlignedAllocator.h",
2090 ] + MICROKERNEL_TEST_HDRS,
2091 deps = MICROKERNEL_TEST_DEPS,
2092)
2093
2094xnnpack_unit_test(
2095 name = "f32_gemm_test",
2096 srcs = [
2097 "test/f32-gemm.cc",
2098 "test/gemm-microkernel-tester.h",
2099 "src/xnnpack/AlignedAllocator.h",
2100 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2101 deps = MICROKERNEL_TEST_DEPS,
2102)
2103
2104xnnpack_unit_test(
2105 name = "f32_gemminc_test",
2106 srcs = [
2107 "test/f32-gemminc.cc",
2108 "test/gemm-microkernel-tester.h",
2109 "src/xnnpack/AlignedAllocator.h",
2110 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2111 deps = MICROKERNEL_TEST_DEPS,
2112)
2113
2114xnnpack_unit_test(
2115 name = "f32_hswish_test",
2116 srcs = [
2117 "test/f32-hswish.cc",
2118 "test/hswish-microkernel-tester.h",
2119 ] + MICROKERNEL_TEST_HDRS,
2120 deps = MICROKERNEL_TEST_DEPS,
2121)
2122
2123xnnpack_unit_test(
2124 name = "f32_maxpool_test",
2125 srcs = [
2126 "test/f32-maxpool.cc",
2127 "test/maxpool-microkernel-tester.h",
2128 ] + MICROKERNEL_TEST_HDRS,
2129 deps = MICROKERNEL_TEST_DEPS,
2130)
2131
2132xnnpack_unit_test(
2133 name = "f32_pavgpool_test",
2134 srcs = [
2135 "test/f32-pavgpool.cc",
2136 "test/avgpool-microkernel-tester.h",
2137 "src/xnnpack/AlignedAllocator.h",
2138 ] + MICROKERNEL_TEST_HDRS,
2139 deps = MICROKERNEL_TEST_DEPS,
2140)
2141
2142xnnpack_unit_test(
2143 name = "f32_ppmm_test",
2144 srcs = [
2145 "test/f32-ppmm.cc",
2146 "test/gemm-microkernel-tester.h",
2147 "src/xnnpack/AlignedAllocator.h",
2148 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2149 deps = MICROKERNEL_TEST_DEPS,
2150)
2151
2152xnnpack_unit_test(
2153 name = "f32_prelu_test",
2154 srcs = [
2155 "test/f32-prelu.cc",
2156 "test/prelu-microkernel-tester.h",
2157 "src/xnnpack/AlignedAllocator.h",
2158 ] + MICROKERNEL_TEST_HDRS,
2159 deps = MICROKERNEL_TEST_DEPS,
2160)
2161
2162xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002163 name = "f32_raddexpminusmax_test",
2164 srcs = [
2165 "test/f32-raddexpminusmax.cc",
2166 "test/raddexpminusmax-microkernel-tester.h",
2167 ] + MICROKERNEL_TEST_HDRS,
2168 deps = MICROKERNEL_TEST_DEPS,
2169)
2170
2171xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002172 name = "f32_raddextexp_test",
2173 srcs = [
2174 "test/f32-raddextexp.cc",
2175 "test/raddextexp-microkernel-tester.h",
2176 ] + MICROKERNEL_TEST_HDRS,
2177 deps = MICROKERNEL_TEST_DEPS,
2178)
2179
2180xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002181 name = "f32_raddstoreexpminusmax_test",
2182 srcs = [
2183 "test/f32-raddstoreexpminusmax.cc",
2184 "test/raddstoreexpminusmax-microkernel-tester.h",
2185 ] + MICROKERNEL_TEST_HDRS,
2186 deps = MICROKERNEL_TEST_DEPS,
2187)
2188
2189xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002190 name = "f32_rmax_test",
2191 srcs = [
2192 "test/f32-rmax.cc",
2193 "test/rmax-microkernel-tester.h",
2194 ] + MICROKERNEL_TEST_HDRS,
2195 deps = MICROKERNEL_TEST_DEPS,
2196)
2197
2198xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002199 name = "f32_sigmoid_test",
2200 srcs = [
2201 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002202 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002203 ] + MICROKERNEL_TEST_HDRS,
2204 deps = MICROKERNEL_TEST_DEPS,
2205)
2206
2207xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002208 name = "f32_spmm_test",
2209 srcs = [
2210 "test/f32-spmm.cc",
2211 "test/spmm-microkernel-tester.h",
2212 "src/xnnpack/AlignedAllocator.h",
2213 ] + MICROKERNEL_TEST_HDRS,
2214 deps = MICROKERNEL_TEST_DEPS,
2215)
2216
2217xnnpack_unit_test(
2218 name = "f32_vadd_test",
2219 srcs = [
2220 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002221 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002222 ] + MICROKERNEL_TEST_HDRS,
2223 deps = MICROKERNEL_TEST_DEPS,
2224)
2225
2226xnnpack_unit_test(
2227 name = "f32_vaddc_test",
2228 srcs = [
2229 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002230 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002231 ] + MICROKERNEL_TEST_HDRS,
2232 deps = MICROKERNEL_TEST_DEPS,
2233)
2234
2235xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002236 name = "f32_vdiv_test",
2237 srcs = [
2238 "test/f32-vdiv.cc",
2239 "test/vbinary-microkernel-tester.h",
2240 ] + MICROKERNEL_TEST_HDRS,
2241 deps = MICROKERNEL_TEST_DEPS,
2242)
2243
2244xnnpack_unit_test(
2245 name = "f32_vdivc_test",
2246 srcs = [
2247 "test/f32-vdivc.cc",
2248 "test/vbinaryc-microkernel-tester.h",
2249 ] + MICROKERNEL_TEST_HDRS,
2250 deps = MICROKERNEL_TEST_DEPS,
2251)
2252
2253xnnpack_unit_test(
2254 name = "f32_vrdivc_test",
2255 srcs = [
2256 "test/f32-vrdivc.cc",
2257 "test/vbinaryc-microkernel-tester.h",
2258 ] + MICROKERNEL_TEST_HDRS,
2259 deps = MICROKERNEL_TEST_DEPS,
2260)
2261
2262xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002263 name = "f32_vmax_test",
2264 srcs = [
2265 "test/f32-vmax.cc",
2266 "test/vbinary-microkernel-tester.h",
2267 ] + MICROKERNEL_TEST_HDRS,
2268 deps = MICROKERNEL_TEST_DEPS,
2269)
2270
2271xnnpack_unit_test(
2272 name = "f32_vmaxc_test",
2273 srcs = [
2274 "test/f32-vmaxc.cc",
2275 "test/vbinaryc-microkernel-tester.h",
2276 ] + MICROKERNEL_TEST_HDRS,
2277 deps = MICROKERNEL_TEST_DEPS,
2278)
2279
2280xnnpack_unit_test(
2281 name = "f32_vmin_test",
2282 srcs = [
2283 "test/f32-vmin.cc",
2284 "test/vbinary-microkernel-tester.h",
2285 ] + MICROKERNEL_TEST_HDRS,
2286 deps = MICROKERNEL_TEST_DEPS,
2287)
2288
2289xnnpack_unit_test(
2290 name = "f32_vminc_test",
2291 srcs = [
2292 "test/f32-vminc.cc",
2293 "test/vbinaryc-microkernel-tester.h",
2294 ] + MICROKERNEL_TEST_HDRS,
2295 deps = MICROKERNEL_TEST_DEPS,
2296)
2297
2298xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002299 name = "f32_vmul_test",
2300 srcs = [
2301 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002302 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002303 ] + MICROKERNEL_TEST_HDRS,
2304 deps = MICROKERNEL_TEST_DEPS,
2305)
2306
2307xnnpack_unit_test(
2308 name = "f32_vmulc_test",
2309 srcs = [
2310 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002311 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002312 ] + MICROKERNEL_TEST_HDRS,
2313 deps = MICROKERNEL_TEST_DEPS,
2314)
2315
2316xnnpack_unit_test(
2317 name = "f32_vmulcaddc_test",
2318 srcs = [
2319 "test/f32-vmulcaddc.cc",
2320 "test/vmulcaddc-microkernel-tester.h",
2321 "src/xnnpack/AlignedAllocator.h",
2322 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2323 deps = MICROKERNEL_TEST_DEPS,
2324)
2325
2326xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002327 name = "f32_vscale_test",
2328 srcs = [
2329 "test/f32-vscale.cc",
2330 "test/vscale-microkernel-tester.h",
2331 ] + MICROKERNEL_TEST_HDRS,
2332 deps = MICROKERNEL_TEST_DEPS,
2333)
2334
2335xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002336 name = "f32_vscaleexpminusmax_test",
2337 srcs = [
2338 "test/f32-vscaleexpminusmax.cc",
2339 "test/vscaleexpminusmax-microkernel-tester.h",
2340 ] + MICROKERNEL_TEST_HDRS,
2341 deps = MICROKERNEL_TEST_DEPS,
2342)
2343
2344xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002345 name = "f32_vscaleextexp_test",
2346 srcs = [
2347 "test/f32-vscaleextexp.cc",
2348 "test/vscaleextexp-microkernel-tester.h",
2349 ] + MICROKERNEL_TEST_HDRS,
2350 deps = MICROKERNEL_TEST_DEPS,
2351)
2352
2353xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002354 name = "f32_vsub_test",
2355 srcs = [
2356 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002357 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002358 ] + MICROKERNEL_TEST_HDRS,
2359 deps = MICROKERNEL_TEST_DEPS,
2360)
2361
2362xnnpack_unit_test(
2363 name = "f32_vsubc_test",
2364 srcs = [
2365 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002366 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002367 ] + MICROKERNEL_TEST_HDRS,
2368 deps = MICROKERNEL_TEST_DEPS,
2369)
2370
2371xnnpack_unit_test(
2372 name = "f32_vrsubc_test",
2373 srcs = [
2374 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002375 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002376 ] + MICROKERNEL_TEST_HDRS,
2377 deps = MICROKERNEL_TEST_DEPS,
2378)
2379
2380xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002381 name = "q8_avgpool_test",
2382 srcs = [
2383 "test/q8-avgpool.cc",
2384 "test/avgpool-microkernel-tester.h",
2385 "src/xnnpack/AlignedAllocator.h",
2386 ] + MICROKERNEL_TEST_HDRS,
2387 deps = MICROKERNEL_TEST_DEPS,
2388)
2389
2390xnnpack_unit_test(
2391 name = "q8_igemm_test",
2392 srcs = [
2393 "test/q8-igemm.cc",
2394 "test/gemm-microkernel-tester.h",
2395 "src/xnnpack/AlignedAllocator.h",
2396 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2397 deps = MICROKERNEL_TEST_DEPS,
2398)
2399
2400xnnpack_unit_test(
2401 name = "q8_dwconv_test",
2402 srcs = [
2403 "test/q8-dwconv.cc",
2404 "test/dwconv-microkernel-tester.h",
2405 "src/xnnpack/AlignedAllocator.h",
2406 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2407 deps = MICROKERNEL_TEST_DEPS,
2408)
2409
2410xnnpack_unit_test(
2411 name = "q8_gavgpool_test",
2412 srcs = [
2413 "test/q8-gavgpool.cc",
2414 "test/gavgpool-microkernel-tester.h",
2415 "src/xnnpack/AlignedAllocator.h",
2416 ] + MICROKERNEL_TEST_HDRS,
2417 deps = MICROKERNEL_TEST_DEPS,
2418)
2419
2420xnnpack_unit_test(
2421 name = "q8_gemm_test",
2422 srcs = [
2423 "test/q8-gemm.cc",
2424 "test/gemm-microkernel-tester.h",
2425 "src/xnnpack/AlignedAllocator.h",
2426 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2427 deps = MICROKERNEL_TEST_DEPS,
2428)
2429
2430xnnpack_unit_test(
2431 name = "q8_vadd_test",
2432 srcs = [
2433 "test/q8-vadd.cc",
2434 "test/vadd-microkernel-tester.h",
2435 ] + MICROKERNEL_TEST_HDRS,
2436 deps = MICROKERNEL_TEST_DEPS,
2437)
2438
2439xnnpack_unit_test(
2440 name = "u8_clamp_test",
2441 srcs = [
2442 "test/u8-clamp.cc",
2443 "test/clamp-microkernel-tester.h",
2444 ] + MICROKERNEL_TEST_HDRS,
2445 deps = MICROKERNEL_TEST_DEPS,
2446)
2447
2448xnnpack_unit_test(
2449 name = "u8_lut32norm_test",
2450 srcs = [
2451 "test/u8-lut32norm.cc",
2452 "test/lut-norm-microkernel-tester.h",
2453 ] + MICROKERNEL_TEST_HDRS,
2454 deps = MICROKERNEL_TEST_DEPS,
2455)
2456
2457xnnpack_unit_test(
2458 name = "u8_maxpool_test",
2459 srcs = [
2460 "test/u8-maxpool.cc",
2461 "test/maxpool-microkernel-tester.h",
2462 ] + MICROKERNEL_TEST_HDRS,
2463 deps = MICROKERNEL_TEST_DEPS,
2464)
2465
2466xnnpack_unit_test(
2467 name = "u8_rmax_test",
2468 srcs = [
2469 "test/u8-rmax.cc",
2470 "test/rmax-microkernel-tester.h",
2471 ] + MICROKERNEL_TEST_HDRS,
2472 deps = MICROKERNEL_TEST_DEPS,
2473)
2474
2475xnnpack_unit_test(
2476 name = "x32_packx_test",
2477 srcs = [
2478 "test/x32-packx.cc",
2479 "test/pack-microkernel-tester.h",
2480 "src/xnnpack/AlignedAllocator.h",
2481 ] + MICROKERNEL_TEST_HDRS,
2482 deps = MICROKERNEL_TEST_DEPS,
2483)
2484
2485xnnpack_unit_test(
2486 name = "x32_pad_test",
2487 srcs = [
2488 "test/x32-pad.cc",
2489 "test/pad-microkernel-tester.h",
2490 ] + MICROKERNEL_TEST_HDRS,
2491 deps = MICROKERNEL_TEST_DEPS,
2492)
2493
2494xnnpack_unit_test(
2495 name = "x32_unpool_test",
2496 srcs = [
2497 "test/x32-unpool.cc",
2498 "test/unpool-microkernel-tester.h",
2499 ] + MICROKERNEL_TEST_HDRS,
2500 deps = MICROKERNEL_TEST_DEPS,
2501)
2502
2503xnnpack_unit_test(
2504 name = "x32_zip_test",
2505 srcs = [
2506 "test/x32-zip.cc",
2507 "test/zip-microkernel-tester.h",
2508 ] + MICROKERNEL_TEST_HDRS,
2509 deps = MICROKERNEL_TEST_DEPS,
2510)
2511
2512xnnpack_unit_test(
2513 name = "x8_lut_test",
2514 srcs = [
2515 "test/x8-lut.cc",
2516 "test/lut-microkernel-tester.h",
2517 ] + MICROKERNEL_TEST_HDRS,
2518 deps = MICROKERNEL_TEST_DEPS,
2519)
2520
2521xnnpack_unit_test(
2522 name = "x8_zip_test",
2523 srcs = [
2524 "test/x8-zip.cc",
2525 "test/zip-microkernel-tester.h",
2526 ] + MICROKERNEL_TEST_HDRS,
2527 deps = MICROKERNEL_TEST_DEPS,
2528)
2529
2530########################### Size test for the library ##########################
2531
2532xnnpack_binary(
2533 name = "size_test",
2534 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002535 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002536)
2537
2538########################### Unit tests for operators ###########################
2539
2540xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002541 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002543 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002544 "test/add-operator-tester.h",
2545 ],
2546 deps = OPERATOR_TEST_DEPS,
2547)
2548
2549xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002550 name = "add_nd_test",
2551 srcs = [
2552 "test/add-nd.cc",
2553 "test/binary-elementwise-operator-tester.h",
2554 ],
2555 deps = OPERATOR_TEST_DEPS,
2556)
2557
2558xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002559 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002560 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002561 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562 "test/argmax-pooling-operator-tester.h",
2563 ] + OPERATOR_TEST_PARAMS_HDRS,
2564 deps = OPERATOR_TEST_DEPS,
2565)
2566
2567xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002568 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002569 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002570 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002571 "test/average-pooling-operator-tester.h",
2572 ] + OPERATOR_TEST_PARAMS_HDRS,
2573 deps = OPERATOR_TEST_DEPS,
2574)
2575
2576xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002577 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002578 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002579 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002580 "test/channel-pad-operator-tester.h",
2581 ] + OPERATOR_TEST_PARAMS_HDRS,
2582 deps = OPERATOR_TEST_DEPS,
2583)
2584
2585xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002586 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002587 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002588 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002589 "test/channel-shuffle-operator-tester.h",
2590 ],
2591 deps = OPERATOR_TEST_DEPS,
2592)
2593
2594xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002595 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002596 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002597 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002598 "test/clamp-operator-tester.h",
2599 ],
2600 deps = OPERATOR_TEST_DEPS,
2601)
2602
2603xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002604 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002605 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002606 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002607 "test/convolution-operator-tester.h",
2608 ],
2609 deps = OPERATOR_TEST_DEPS,
2610)
2611
2612xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002613 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002614 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002615 "test/convolution-nchw.cc",
2616 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002617 ],
2618 deps = OPERATOR_TEST_DEPS,
2619)
2620
2621xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002622 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002623 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002624 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002625 "test/deconvolution-operator-tester.h",
2626 ] + OPERATOR_TEST_PARAMS_HDRS,
2627 deps = OPERATOR_TEST_DEPS,
2628)
2629
2630xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002631 name = "divide_nd_test",
2632 srcs = [
2633 "test/binary-elementwise-operator-tester.h",
2634 "test/divide-nd.cc",
2635 ],
2636 deps = OPERATOR_TEST_DEPS,
2637)
2638
2639xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002640 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002641 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002642 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002643 "test/fully-connected-operator-tester.h",
2644 ],
2645 deps = OPERATOR_TEST_DEPS,
2646)
2647
2648xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002649 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002650 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002651 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002652 "test/global-average-pooling-operator-tester.h",
2653 ] + OPERATOR_TEST_PARAMS_HDRS,
2654 deps = OPERATOR_TEST_DEPS,
2655)
2656
2657xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002658 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002659 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002660 "test/global-average-pooling-ncw.cc",
2661 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002662 ],
2663 deps = OPERATOR_TEST_DEPS,
2664)
2665
2666xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002667 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002668 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002669 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002670 "test/hardswish-operator-tester.h",
2671 ],
2672 deps = OPERATOR_TEST_DEPS,
2673)
2674
2675xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002676 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002677 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002678 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002679 "test/leaky-relu-operator-tester.h",
2680 ],
2681 deps = OPERATOR_TEST_DEPS,
2682)
2683
2684xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002685 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002687 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002688 "test/max-pooling-operator-tester.h",
2689 ] + OPERATOR_TEST_PARAMS_HDRS,
2690 deps = OPERATOR_TEST_DEPS,
2691)
2692
2693xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08002694 name = "maximum_nd_test",
2695 srcs = [
2696 "test/binary-elementwise-operator-tester.h",
2697 "test/maximum-nd.cc",
2698 ],
2699 deps = OPERATOR_TEST_DEPS,
2700)
2701
2702xnnpack_unit_test(
2703 name = "minimum_nd_test",
2704 srcs = [
2705 "test/binary-elementwise-operator-tester.h",
2706 "test/minimum-nd.cc",
2707 ],
2708 deps = OPERATOR_TEST_DEPS,
2709)
2710
2711xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002712 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002713 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002714 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002715 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002716 ],
2717 deps = OPERATOR_TEST_DEPS,
2718)
2719
2720xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002721 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002722 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002723 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002724 "test/prelu-operator-tester.h",
2725 ] + OPERATOR_TEST_PARAMS_HDRS,
2726 deps = OPERATOR_TEST_DEPS,
2727)
2728
2729xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002730 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002731 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002732 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002733 "test/resize-bilinear-operator-tester.h",
2734 ] + OPERATOR_TEST_PARAMS_HDRS,
2735 deps = OPERATOR_TEST_DEPS,
2736)
2737
2738xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002739 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002741 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002742 "test/sigmoid-operator-tester.h",
2743 ],
2744 deps = OPERATOR_TEST_DEPS,
2745)
2746
2747xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002748 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002749 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002750 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002751 "test/softargmax-operator-tester.h",
2752 ],
2753 deps = OPERATOR_TEST_DEPS,
2754)
2755
2756xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08002757 name = "subtract_nd_test",
2758 srcs = [
2759 "test/binary-elementwise-operator-tester.h",
2760 "test/subtract-nd.cc",
2761 ],
2762 deps = OPERATOR_TEST_DEPS,
2763)
2764
2765xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002766 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002768 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002769 "test/unpooling-operator-tester.h",
2770 ],
2771 deps = OPERATOR_TEST_DEPS,
2772)
2773
2774############################# Build configurations #############################
2775
Marat Dukhanb8642352019-10-30 15:43:02 -07002776# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002777config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002778 name = "xnn_enable_assembly_explicit_true",
2779 define_values = {"xnn_enable_assembly": "true"},
2780)
2781
2782# Disables usage of assembly kernels.
2783config_setting(
2784 name = "xnn_enable_assembly_explicit_false",
2785 define_values = {"xnn_enable_assembly": "false"},
2786)
2787
2788# Builds with -c dbg
2789config_setting(
2790 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002791 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002792 "compilation_mode": "dbg",
2793 },
2794)
2795
2796# Builds with -c opt
2797config_setting(
2798 name = "optimized_build",
2799 values = {
2800 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002801 },
2802)
2803
2804config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002805 name = "linux_k8",
2806 values = {"cpu": "k8"},
2807)
2808
2809config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002810 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002811 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002812)
2813
2814config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002815 name = "android",
2816 values = {"crosstool_top": "//external:android/crosstool"},
2817)
2818
2819config_setting(
2820 name = "android_armv7",
2821 values = {
2822 "crosstool_top": "//external:android/crosstool",
2823 "cpu": "armeabi-v7a",
2824 },
2825)
2826
2827config_setting(
2828 name = "android_arm64",
2829 values = {
2830 "crosstool_top": "//external:android/crosstool",
2831 "cpu": "arm64-v8a",
2832 },
2833)
2834
2835config_setting(
2836 name = "android_x86",
2837 values = {
2838 "crosstool_top": "//external:android/crosstool",
2839 "cpu": "x86",
2840 },
2841)
2842
2843config_setting(
2844 name = "android_x86_64",
2845 values = {
2846 "crosstool_top": "//external:android/crosstool",
2847 "cpu": "x86_64",
2848 },
2849)
2850
2851config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002852 name = "macos_x86_64",
2853 values = {
2854 "apple_platform_type": "macos",
2855 "cpu": "darwin",
2856 },
2857)
2858
2859config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002860 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002861 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002862)
2863
2864config_setting(
2865 name = "emscripten_wasm",
2866 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002867 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002868 "cpu": "wasm",
2869 },
2870)
2871
2872config_setting(
2873 name = "emscripten_wasmsimd",
2874 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002875 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002876 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002877 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002878 },
2879)
2880
2881config_setting(
2882 name = "emscripten_asmjs",
2883 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002884 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 "cpu": "asmjs",
2886 },
2887)