blob: 492754f04337f73f412ef0a654df904ef5efe854 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan8d3c6932020-03-06 20:27:27 -08009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_copts", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
Marat Dukhan08c4a432019-10-03 09:29:21 -070022MICROKERNEL_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
Marat Dukhan1b354632020-03-23 12:50:22 -070046OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
Marat Dukhanfd8e6892020-01-27 15:25:25 -080072 "src/softmax-nc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080073 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -070085 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
86 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -070087 "src/f32-clamp/gen/scalar-x1.c",
88 "src/f32-clamp/gen/scalar-x2.c",
89 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080090 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080091 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070092 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070093 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070094 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080095 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080096 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -070097 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x4-scalar.c",
99 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
100 "src/f32-dwconv/gen/up1x9-scalar.c",
101 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
102 "src/f32-dwconv/gen/up1x25-scalar.c",
103 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x4-scalar.c",
105 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
106 "src/f32-dwconv/gen/up2x9-scalar.c",
107 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
108 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700109 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
110 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
111 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
112 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700113 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
114 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700115 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
116 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
117 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
118 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700119 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
120 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700121 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700122 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
123 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700124 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
125 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
126 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700127 "src/f32-gemm/gen/1x4-scalar.c",
128 "src/f32-gemm/gen/2x4-scalar.c",
129 "src/f32-gemm/gen/4x2-scalar.c",
130 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700131 "src/f32-gemm/gen/1x4-minmax-scalar.c",
132 "src/f32-gemm/gen/2x4-minmax-scalar.c",
133 "src/f32-gemm/gen/4x2-minmax-scalar.c",
134 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800135 "src/f32-hswish/gen/scalar-x1.c",
136 "src/f32-hswish/gen/scalar-x2.c",
137 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700138 "src/f32-ibilinear/gen/scalar-c1.c",
139 "src/f32-ibilinear/gen/scalar-c2.c",
140 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-igemm/gen/1x4-scalar.c",
142 "src/f32-igemm/gen/2x4-scalar.c",
143 "src/f32-igemm/gen/4x2-scalar.c",
144 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-igemm/gen/1x4-minmax-scalar.c",
146 "src/f32-igemm/gen/2x4-minmax-scalar.c",
147 "src/f32-igemm/gen/4x2-minmax-scalar.c",
148 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700149 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
150 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
151 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700152 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
153 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
154 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
155 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800156 "src/f32-prelu/gen/scalar-2x1.c",
157 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800158 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
159 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
160 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
161 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
162 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
163 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
164 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
165 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
166 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
167 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
168 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
169 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700170 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800171 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
172 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
173 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
174 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
175 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
176 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
177 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
178 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
179 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700180 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
181 "src/f32-spmm/gen/1x1-minmax-scalar.c",
182 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
183 "src/f32-spmm/gen/2x1-minmax-scalar.c",
184 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
185 "src/f32-spmm/gen/4x1-minmax-scalar.c",
186 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
187 "src/f32-spmm/gen/8x1-minmax-scalar.c",
188 "src/f32-spmm/gen/8x2-minmax-scalar.c",
189 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700190 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
191 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
192 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
193 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
194 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
195 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
196 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
197 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
198 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
199 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
200 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
201 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800202 "src/f32-vbinary/gen/vmax-scalar-x1.c",
203 "src/f32-vbinary/gen/vmax-scalar-x2.c",
204 "src/f32-vbinary/gen/vmax-scalar-x4.c",
205 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
206 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
207 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
208 "src/f32-vbinary/gen/vmin-scalar-x1.c",
209 "src/f32-vbinary/gen/vmin-scalar-x2.c",
210 "src/f32-vbinary/gen/vmin-scalar-x4.c",
211 "src/f32-vbinary/gen/vminc-scalar-x1.c",
212 "src/f32-vbinary/gen/vminc-scalar-x2.c",
213 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700214 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
215 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
216 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
217 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
218 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
219 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
220 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
221 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
222 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
223 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
224 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
225 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
226 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
227 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
228 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
229 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700232 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
233 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
234 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800235 "src/math/expminus-scalar-lut2048-p1.c",
236 "src/math/expminus-scalar-lut64-p2.c",
237 "src/math/expminus-scalar-p5.c",
238 "src/math/sigmoid-scalar-lut2048-p1-div.c",
239 "src/math/sigmoid-scalar-lut64-p2-div.c",
240 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700241 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
242 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700243 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700244 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
245 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700246 "src/q8-gemm/2x2-minmax-scalar.c",
247 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700248 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700249 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700250 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700251 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700252 "src/u8-rmax/scalar.c",
253 "src/x32-packx/x2-scalar.c",
254 "src/x32-packx/x3-scalar.c",
255 "src/x32-packx/x4-scalar.c",
256 "src/x32-pad/x2-scalar.c",
257 "src/x32-unpool/scalar.c",
258 "src/x32-zip/x2-scalar.c",
259 "src/x32-zip/x3-scalar.c",
260 "src/x32-zip/x4-scalar.c",
261 "src/x32-zip/xm-scalar.c",
262 "src/x8-lut/scalar.c",
263 "src/x8-zip/x2-scalar.c",
264 "src/x8-zip/x3-scalar.c",
265 "src/x8-zip/x4-scalar.c",
266 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700267 "src/requantization/precise-scalar.c",
268 "src/requantization/fp32-scalar.c",
269 "src/requantization/q31-scalar.c",
270 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700271]
272
Marat Dukhan436ebe62019-12-04 15:10:12 -0800273WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700274 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
275 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700276 "src/f32-clamp/gen/wasm-x1.c",
277 "src/f32-clamp/gen/wasm-x2.c",
278 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700279 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
280 "src/f32-dwconv/gen/up1x4-wasm.c",
281 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
282 "src/f32-dwconv/gen/up1x9-wasm.c",
283 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
284 "src/f32-dwconv/gen/up1x25-wasm.c",
285 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
286 "src/f32-dwconv/gen/up2x4-wasm.c",
287 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
288 "src/f32-dwconv/gen/up2x9-wasm.c",
289 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
290 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700291 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
292 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
293 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
294 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700295 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
296 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700297 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
298 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
299 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
300 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700301 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
302 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700303 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
304 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700305 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
306 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
307 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700308 "src/f32-gemm/gen/1x4-wasm.c",
309 "src/f32-gemm/gen/2x4-wasm.c",
310 "src/f32-gemm/gen/4x2-wasm.c",
311 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700312 "src/f32-gemm/gen/1x4-minmax-wasm.c",
313 "src/f32-gemm/gen/2x4-minmax-wasm.c",
314 "src/f32-gemm/gen/4x2-minmax-wasm.c",
315 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800316 "src/f32-hswish/gen/wasm-x1.c",
317 "src/f32-hswish/gen/wasm-x2.c",
318 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700319 "src/f32-igemm/gen/1x4-wasm.c",
320 "src/f32-igemm/gen/2x4-wasm.c",
321 "src/f32-igemm/gen/4x2-wasm.c",
322 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700323 "src/f32-igemm/gen/1x4-minmax-wasm.c",
324 "src/f32-igemm/gen/2x4-minmax-wasm.c",
325 "src/f32-igemm/gen/4x2-minmax-wasm.c",
326 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700327 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
328 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
329 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700330 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
331 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
332 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
333 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
334 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
335 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
336 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
337 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
338 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
339 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
340 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
341 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800342 "src/f32-vbinary/gen/vmax-wasm-x1.c",
343 "src/f32-vbinary/gen/vmax-wasm-x2.c",
344 "src/f32-vbinary/gen/vmax-wasm-x4.c",
345 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
346 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
347 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
348 "src/f32-vbinary/gen/vmin-wasm-x1.c",
349 "src/f32-vbinary/gen/vmin-wasm-x2.c",
350 "src/f32-vbinary/gen/vmin-wasm-x4.c",
351 "src/f32-vbinary/gen/vminc-wasm-x1.c",
352 "src/f32-vbinary/gen/vminc-wasm-x2.c",
353 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700354 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
355 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
356 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
357 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
358 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
359 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
360 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
361 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
362 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
363 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
364 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
365 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
366 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
367 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
368 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
369 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
370 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
371 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700372 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
373 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
374 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800375]
376
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800377PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800378 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800379 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800380 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700381 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
382 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700383 "src/f32-clamp/gen/psimd-x4.c",
384 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700385 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
386 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
387 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
388 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
389 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
390 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
391 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
392 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
393 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
394 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
395 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
396 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700397 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
398 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700399 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
400 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
401 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
402 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
403 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
404 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
405 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
406 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
407 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
408 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
409 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
410 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
411 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
412 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
413 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
414 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
415 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
416 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
417 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800418 "src/f32-hswish/gen/psimd-x4.c",
419 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700420 "src/f32-ibilinear/gen/psimd-c4.c",
421 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
423 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
424 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
425 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
426 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
427 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
428 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
429 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
430 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
431 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700432 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
433 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
434 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700435 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800436 "src/f32-prelu/gen/psimd-2x4.c",
437 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800438 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700439 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
440 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
441 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
442 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
443 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
444 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
445 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
446 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800447 "src/f32-vbinary/gen/vmax-psimd-x4.c",
448 "src/f32-vbinary/gen/vmax-psimd-x8.c",
449 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
450 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
451 "src/f32-vbinary/gen/vmin-psimd-x4.c",
452 "src/f32-vbinary/gen/vmin-psimd-x8.c",
453 "src/f32-vbinary/gen/vminc-psimd-x4.c",
454 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700455 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
456 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
457 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
458 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
459 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
460 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
461 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
462 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
463 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
464 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
465 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
466 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700467 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
468 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700469 "src/x32-packx/x4-psimd.c",
470 "src/x32-pad/x2-psimd.c",
471 "src/x32-unpool/psimd.c",
472 "src/x32-zip/x2-psimd.c",
473 "src/x32-zip/x3-psimd.c",
474 "src/x32-zip/x4-psimd.c",
475 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700476 "src/requantization/precise-psimd.c",
477 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700478]
479
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800480PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800481 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
482 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
483 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
484 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
485 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
486 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
487 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
488 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
489 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
490 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
491 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
492 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800493 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
494 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
495 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
496 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
497 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
498 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
499 "src/math/sigmoid-psimd-p5-div.c",
500]
501
Marat Dukhan08c4a432019-10-03 09:29:21 -0700502# ISA-specific micro-kernels
503NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700504 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
505 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700506 "src/f32-clamp/gen/neon-x4.c",
507 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700508 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
509 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
510 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
511 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700512 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700513 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
514 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700515 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
516 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
517 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
518 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
519 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
520 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
521 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
522 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
523 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
524 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
525 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
526 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
527 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
528 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
529 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
530 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
531 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
532 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
533 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
534 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
535 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
536 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
537 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
538 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
539 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
540 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
541 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
542 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
543 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
544 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
545 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800546 "src/f32-hswish/gen/neon-x4.c",
547 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700548 "src/f32-ibilinear/gen/neon-c4.c",
549 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700550 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
551 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
552 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
553 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
554 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
555 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
556 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
557 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
558 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
559 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
560 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
561 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
562 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
563 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
564 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
565 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700566 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
567 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
568 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700569 "src/f32-ppmm/gen/4x8-minmax-neon.c",
570 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800571 "src/f32-prelu/gen/neon-2x4.c",
572 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
580 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
581 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
582 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
583 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
584 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
585 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
586 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
587 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
588 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
589 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
590 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
591 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
592 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
593 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
594 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
595 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
596 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700597 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800598 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
601 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
602 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
603 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
604 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
605 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
606 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
607 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
608 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
609 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
610 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
611 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
612 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
613 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
614 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
615 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
616 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700617 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
618 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
619 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
620 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800621 "src/f32-vbinary/gen/vmax-neon-x4.c",
622 "src/f32-vbinary/gen/vmax-neon-x8.c",
623 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
624 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
625 "src/f32-vbinary/gen/vmin-neon-x4.c",
626 "src/f32-vbinary/gen/vmin-neon-x8.c",
627 "src/f32-vbinary/gen/vminc-neon-x4.c",
628 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700629 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
630 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
631 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
632 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
633 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
634 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
635 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
636 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
637 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
638 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700639 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
640 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
641 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
642 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700643 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700644 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
645 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700646 "src/q8-gemm/4x8-minmax-neon.c",
647 "src/q8-gemm/8x8-minmax-neon.c",
648 "src/q8-igemm/4x8-minmax-neon.c",
649 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700650 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700651 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700652 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700653 "src/u8-rmax/neon.c",
654 "src/x32-packx/x4-neon-st4.c",
655 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700656 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700657 "src/x32-zip/x2-neon.c",
658 "src/x32-zip/x3-neon.c",
659 "src/x32-zip/x4-neon.c",
660 "src/x32-zip/xm-neon.c",
661 "src/x8-zip/x2-neon.c",
662 "src/x8-zip/x3-neon.c",
663 "src/x8-zip/x4-neon.c",
664 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800665 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800666 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
667 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
668 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
669 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
670 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
671 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700672 "src/requantization/precise-neon.c",
673 "src/requantization/fp32-neon.c",
674 "src/requantization/q31-neon.c",
675 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700676]
677
678NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700679 "src/f32-ibilinear/gen/neonfma-c4.c",
680 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700681 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
682 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
683 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
684 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
685 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
686 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
687 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
688 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
689 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
690 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
691 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
692 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
693 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
694 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
695 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
696 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
697 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
698 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
699 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
700 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
701 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
702 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
703 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
704 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
705 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
706 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
707 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
708 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
709 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
710 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
711 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800712 "src/f32-hswish/gen/neonfma-x4.c",
713 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700714 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
715 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800716 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
717 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
718 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
719 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
720 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
722 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
723 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
725 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
726 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
727 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
728 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
729 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
730 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
731 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
732 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
733 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
734 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
735 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
736 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
737 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
738 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
739 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800740 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
741 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
742 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
743 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
744 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
745 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
746 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
747 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
748 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
749 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
750 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
751 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
752 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
753 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
754 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
755 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
756 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
757 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
758 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
759 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
768 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
769 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
770 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
771 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
772 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
773 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
774 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
775 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
776 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
777 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
778 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
779 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700794 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
795 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800796 "src/math/exp-neonfma-lut64-p2.c",
797 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800798 "src/math/expminus-neonfma-lut2048-p1.c",
799 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800800 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800801 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
802 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
803 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
804 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
805 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
806 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
807 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
808 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
809 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
810 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
811 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
812 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
813 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
814 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
815 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
816 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
817 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
818 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700819]
820
821AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700822 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
823 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
824 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
825 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
826 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
827 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700828 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
829 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
830 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
831 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
832 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
833 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
834 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
835 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
836 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
837 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
838 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
839 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
840 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
841 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
842 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
843 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
844 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
845 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
846 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
847 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700848 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
849 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
850 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
851 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
852 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
853 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
854 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800855 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
856 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
857 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
858 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
859 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
860 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
861 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
862 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
863 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
864 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
865 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
866 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
867 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
868 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
869 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
870 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
871 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
872 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700873 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
874 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
875 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
876 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
877 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
878 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
879 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
880 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
881 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
882 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
883 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
884 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
885 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
886 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
887 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
888 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
889 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
890 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800891 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
892 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
893 "src/math/sigmoid-neonfma-rr1-p5-div.c",
894 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
895 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
896 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700897]
898
899AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard1f4e4612020-04-13 18:24:54 -0700900 "src/f16-gemm/gen/1x8-neonfp16arith-ld64.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800901 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
902 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
903 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700904 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
905 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
906 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
907 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
908 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
909 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
910 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
911 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700912]
913
914SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700915 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
916 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700917 "src/f32-clamp/gen/sse-x4.c",
918 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800919 "src/f32-dwconv-spchw/3x3p1-sse.c",
920 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700921 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
922 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
923 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
924 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
925 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
926 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
927 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
928 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
929 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
930 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
931 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
932 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700933 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700934 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
935 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700936 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
937 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
938 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
939 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
940 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
941 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
942 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
943 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
944 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
945 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
946 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
947 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
948 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800949 "src/f32-hswish/gen/sse-x4.c",
950 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700951 "src/f32-ibilinear/gen/sse-c4.c",
952 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700953 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
954 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
955 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
956 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
957 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
958 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
959 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700960 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
961 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
962 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700963 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700964 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700965 "src/f32-spmm/gen/4x1-minmax-sse.c",
966 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700967 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
968 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
969 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
970 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
971 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
972 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
973 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
974 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800975 "src/f32-vbinary/gen/vmax-sse-x4.c",
976 "src/f32-vbinary/gen/vmax-sse-x8.c",
977 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
978 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
979 "src/f32-vbinary/gen/vmin-sse-x4.c",
980 "src/f32-vbinary/gen/vmin-sse-x8.c",
981 "src/f32-vbinary/gen/vminc-sse-x4.c",
982 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700983 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
984 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
985 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
986 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
987 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
988 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
989 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
990 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
991 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
992 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
993 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
994 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700995 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
996 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700997 "src/x32-packx/x4-sse.c",
998]
999
1000SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001001 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1002 "src/f32-argmaxpool/4x-sse2-c4.c",
1003 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001004 "src/f32-prelu/gen/sse2-2x4.c",
1005 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001006 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1007 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1008 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1009 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1010 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1012 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1013 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1014 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1015 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1016 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1017 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001018 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001019 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001020 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001021 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001022 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1023 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001024 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1025 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001026 "src/q8-igemm/4x4c2-minmax-sse2.c",
1027 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001028 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1029 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001030 "src/q8-gemm/2x4c8-minmax-sse2.c",
1031 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001032 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001033 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001034 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001035 "src/u8-rmax/sse2.c",
1036 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001037 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001038 "src/x32-zip/x2-sse2.c",
1039 "src/x32-zip/x3-sse2.c",
1040 "src/x32-zip/x4-sse2.c",
1041 "src/x32-zip/xm-sse2.c",
1042 "src/x8-zip/x2-sse2.c",
1043 "src/x8-zip/x3-sse2.c",
1044 "src/x8-zip/x4-sse2.c",
1045 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001046 "src/math/exp-sse2-p5.c",
1047 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001048 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001049 "src/requantization/precise-sse2.c",
1050 "src/requantization/fp32-sse2.c",
1051 "src/requantization/q31-sse2.c",
1052 "src/requantization/gemmlowp-sse2.c",
1053]
1054
1055SSSE3_UKERNELS = [
1056 "src/requantization/precise-ssse3.c",
1057 "src/requantization/q31-ssse3.c",
1058 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001059]
1060
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001061SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001062 "src/f32-prelu/gen/sse41-2x4.c",
1063 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001064 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001065 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001066 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001067 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001068 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1069 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001070 "src/requantization/precise-sse4.c",
1071 "src/requantization/q31-sse4.c",
1072 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001073]
1074
Marat Dukhan08c4a432019-10-03 09:29:21 -07001075AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001076 "src/f32-clamp/gen/avx-x8.c",
1077 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001078 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1079 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1080 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1081 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1082 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1083 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1084 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1085 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1086 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1087 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1088 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1089 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1090 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1091 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1092 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1093 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1094 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1095 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1096 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1097 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1098 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1099 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1100 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1101 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1102 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1103 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1104 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1105 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1106 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1107 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001108 "src/f32-hswish/gen/avx-x8.c",
1109 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001110 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1111 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1112 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1113 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1114 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1115 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1116 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1117 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1118 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001119 "src/f32-prelu/gen/avx-2x8.c",
1120 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001122 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1123 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1124 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1125 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1126 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1127 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1128 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1129 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001130 "src/f32-vbinary/gen/vmax-avx-x8.c",
1131 "src/f32-vbinary/gen/vmax-avx-x16.c",
1132 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1133 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1134 "src/f32-vbinary/gen/vmin-avx-x8.c",
1135 "src/f32-vbinary/gen/vmin-avx-x16.c",
1136 "src/f32-vbinary/gen/vminc-avx-x8.c",
1137 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001138 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1139 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1140 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1141 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1142 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1143 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1144 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1145 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1146 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1147 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1148 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1149 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001150 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001151]
1152
Marat Dukhanfda12b82019-11-21 12:27:59 -08001153FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001154 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1155 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1156 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1157 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1158 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1159 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1160 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1161 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1162 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1163 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1164 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1165 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1166 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1167 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1168 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1169 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1170 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1171 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1172 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1173 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1174 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1175 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1176 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1177 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1178 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1179 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1180 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1181 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1182 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1183 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1184 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1185 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1186 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1187 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1188 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1189 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1190 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1191 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1192 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1193 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001194 "src/f32-hswish/gen/fma3-x8.c",
1195 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001196 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1197 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1198 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1199 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1200 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1201 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1202 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1203 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1204 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1205 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1206 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1207 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1208 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1209 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001210]
1211
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001212AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001213 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1214 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1215 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1216 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1217 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1218 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1219 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1220 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1221 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1222 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1223 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1224 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1225 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1226 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1227 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1228 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1229 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1230 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1231 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1232 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1233 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1234 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1235 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1236 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1237 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1238 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1239 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1240 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1241 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1242 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1243 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1244 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1245 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1246 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1247 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1248 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001249 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1250 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1251 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1252 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1253 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1254 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1255 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1256 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1257 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1258 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1259 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1260 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1261 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1262 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1263 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1264 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1265 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1266 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1267 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1268 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1269 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1270 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1271 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1272 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1273 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1274 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1275 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1276 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1277 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1278 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001279 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1280 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1281 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1282 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1283 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1284 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1285 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1286 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1287 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1288 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1289 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1290 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1291 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1292 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1293 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1294 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1295 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1296 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1297 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1298 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1299 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1300 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1301 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1302 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001303 "src/math/exp-avx2-p5.c",
1304 "src/math/exp-avx2-perm-p3.c",
1305 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001306 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001307 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001308 "src/math/sigmoid-avx2-rr2-p5-div.c",
1309 "src/math/sigmoid-avx2-rr1-p5-div.c",
1310 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1311 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1312 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1313 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001314]
1315
Marat Dukhan08c4a432019-10-03 09:29:21 -07001316AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001317 "src/f32-clamp/gen/avx512f-x16.c",
1318 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001319 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1320 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1321 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1322 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1323 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1324 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1325 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1326 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1327 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1328 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1329 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1330 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1331 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1332 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1333 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1334 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1335 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1336 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1337 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1338 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1339 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1340 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1341 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1342 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001343 "src/f32-hswish/gen/avx512f-x16.c",
1344 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001345 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1346 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1347 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1348 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1349 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1350 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001351 "src/f32-prelu/gen/avx512f-2x16.c",
1352 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001353 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1354 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1355 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1356 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1357 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1358 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1359 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1360 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1361 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1362 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1363 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1364 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1365 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1366 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1367 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1368 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1369 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1370 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1371 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1372 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1373 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1374 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1375 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1376 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1377 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1378 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1379 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1380 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1381 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1382 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1383 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1384 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1385 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1386 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1387 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1388 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001389 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001390 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1391 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1392 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1393 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1394 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1395 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1396 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1397 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001398 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1399 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1400 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1401 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1402 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1403 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1404 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1405 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001406 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1407 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1408 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1409 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1410 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1411 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1412 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1413 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1414 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1415 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1416 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1417 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001418 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001419 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1420 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1421 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1422 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1423 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1424 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1425 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1426 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1427 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1428 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1429 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1430 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1431 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1432 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1433 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1434 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1435 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1436 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1437 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1438 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1439 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1440 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1441 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1442 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001443 "src/math/exp-avx512f-p5-scalef.c",
1444 "src/math/exp-avx512f-p5.c",
1445 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001446 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001447 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001448]
1449
1450AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001451 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001452 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1453 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1454 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1455 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1456 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1457 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1458 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1459 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1460 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1461 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001462]
1463
1464AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001465 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001466 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1467 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001468 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001469 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1470 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001471 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1472 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1473 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
1474 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1475 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1476 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001477 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1478 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001479 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001480 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1481 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1482 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1483 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1484 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1485 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1486 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1487 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1488 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1489 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1490 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1491 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1492 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1493 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1494 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1495 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1496 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1497 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1498 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1499 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1500 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001501 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001502 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1503 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1504 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1505 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1506 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1507 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1508 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1509 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1510 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1511 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1512 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1513 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1514 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1515 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1516 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1517 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1518 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1519 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1520 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1521 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1522 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1523 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1524 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1525 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1526 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1527 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1528 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1529 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1530 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1531 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1532 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1533 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1534 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1535 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1536 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1537 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1538 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1539 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001540]
1541
Marat Dukhan1b354632020-03-23 12:50:22 -07001542INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001543 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001544 "src/xnnpack/argmaxpool.h",
1545 "src/xnnpack/avgpool.h",
1546 "src/xnnpack/clamp.h",
1547 "src/xnnpack/common.h",
1548 "src/xnnpack/conv.h",
1549 "src/xnnpack/dwconv.h",
1550 "src/xnnpack/gavgpool.h",
1551 "src/xnnpack/gemm.h",
1552 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001553 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001554 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001555 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001556 "src/xnnpack/lut.h",
1557 "src/xnnpack/math.h",
1558 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001559 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001560 "src/xnnpack/packx.h",
1561 "src/xnnpack/pad.h",
1562 "src/xnnpack/params.h",
1563 "src/xnnpack/pavgpool.h",
1564 "src/xnnpack/ppmm.h",
1565 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001566 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001567 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001568 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001569 "src/xnnpack/rmax.h",
1570 "src/xnnpack/scalar-utils.h",
1571 "src/xnnpack/spmm.h",
1572 "src/xnnpack/unpool.h",
1573 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001574 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001575 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001576 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001577 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001578 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001579 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001580 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001581]
1582
1583INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001584 "include/xnnpack.h",
1585 "src/xnnpack/allocator.h",
1586 "src/xnnpack/compute.h",
1587 "src/xnnpack/im2col.h",
1588 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001589 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001590 "src/xnnpack/operator.h",
1591 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001592 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001593 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001594 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001595 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001596]
1597
Marat Dukhan1b354632020-03-23 12:50:22 -07001598ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001599 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001600]
1601
Marat Dukhan1b354632020-03-23 12:50:22 -07001602MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001603 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604 "include/xnnpack.h",
1605]
1606
Marat Dukhan1b354632020-03-23 12:50:22 -07001607MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001608 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001609 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001610 "src/xnnpack/requantization.h",
1611 "include/xnnpack.h",
1612]
1613
1614OPERATOR_TEST_PARAMS_HDRS = [
1615 "src/xnnpack/params.h",
1616 "src/xnnpack/common.h",
1617]
1618
1619WEIGHTS_PACK_HDRS = [
1620 "src/xnnpack/pack.h",
1621 "src/xnnpack/operator.h",
1622 "src/xnnpack/compute.h",
1623]
1624
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001625LOGGING_COPTS = select({
1626 # No logging in optimized mode
1627 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1628 # Full logging in debug mode
1629 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1630 # Error-only logging in default (fastbuild) mode
1631 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1632})
1633
1634LOGGING_HDRS = [
1635 "src/xnnpack/log.h",
1636]
1637
Marat Dukhan08c4a432019-10-03 09:29:21 -07001638xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001639 name = "tables",
1640 srcs = TABLE_SRCS,
1641 hdrs = INTERNAL_HDRS,
1642 copts = xnnpack_std_copts(),
1643)
1644
1645xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001646 name = "scalar_ukernels",
1647 srcs = SCALAR_UKERNELS,
1648 hdrs = INTERNAL_HDRS,
1649 aarch32_copts = ["-marm"],
1650 copts = xnnpack_std_copts(),
1651 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001652 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001653 "@FP16",
1654 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001655 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001656 ],
1657)
1658
1659xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001660 name = "wasm_ukernels",
1661 hdrs = INTERNAL_HDRS,
1662 copts = xnnpack_std_copts(),
1663 wasm_srcs = WASM_UKERNELS,
1664 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001665 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001666 "@FP16",
1667 "@FXdiv",
1668 "@pthreadpool",
1669 ],
1670)
1671
1672xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001673 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001674 hdrs = INTERNAL_HDRS,
1675 aarch32_copts = [
1676 "-marm",
1677 "-mfpu=neon",
1678 ],
1679 copts = xnnpack_std_copts(),
1680 optimized_copts = [
1681 "-O3",
1682 "-ffast-math",
1683 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001684 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
1685 x86_copts = [
1686 "-msse2",
1687 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001688 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001689 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001690 "@FP16",
1691 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001692 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693 ],
1694)
1695
1696xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001697 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001698 hdrs = INTERNAL_HDRS,
1699 aarch32_copts = [
1700 "-marm",
1701 "-mfpu=neon",
1702 ],
1703 copts = xnnpack_std_copts(),
1704 optimized_copts = [
1705 "-O3",
1706 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001707 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
1708 x86_copts = [
1709 "-msse2",
1710 ],
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001711 deps = [
1712 ":tables",
1713 "@FP16",
1714 "@psimd",
1715 "@pthreadpool",
1716 ],
1717)
1718
1719xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001720 name = "neon_ukernels",
1721 hdrs = INTERNAL_HDRS,
1722 aarch32_copts = [
1723 "-marm",
1724 "-mfpu=neon",
1725 ],
1726 aarch32_srcs = NEON_UKERNELS,
1727 aarch64_srcs = NEON_UKERNELS,
1728 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001729 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001730 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001731 "@FP16",
1732 "@pthreadpool",
1733 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001734)
1735
1736xnnpack_cc_library(
1737 name = "neonfma_ukernels",
1738 hdrs = INTERNAL_HDRS,
1739 aarch32_copts = [
1740 "-marm",
1741 "-mfpu=neon-vfpv4",
1742 ],
1743 aarch32_srcs = NEONFMA_UKERNELS,
1744 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1745 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001746 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001747 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001748 "@FP16",
1749 "@pthreadpool",
1750 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001751)
1752
1753xnnpack_cc_library(
1754 name = "neonfp16arith_ukernels",
1755 hdrs = INTERNAL_HDRS,
1756 aarch64_copts = ["-march=armv8.2-a+fp16"],
1757 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1758 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001759 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001760 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001761 "@FP16",
1762 "@pthreadpool",
1763 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001764)
1765
1766xnnpack_cc_library(
1767 name = "sse2_ukernels",
1768 hdrs = INTERNAL_HDRS,
1769 copts = xnnpack_std_copts(),
1770 x86_copts = ["-msse2"],
1771 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001772 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001773 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001774 "@FP16",
1775 "@pthreadpool",
1776 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777)
1778
1779xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001780 name = "ssse3_ukernels",
1781 hdrs = INTERNAL_HDRS,
1782 copts = xnnpack_std_copts(),
1783 x86_copts = ["-mssse3"],
1784 x86_srcs = SSSE3_UKERNELS,
1785 deps = [
1786 ":tables",
1787 "@FP16",
1788 "@pthreadpool",
1789 ],
1790)
1791
1792xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001793 name = "sse41_ukernels",
1794 hdrs = INTERNAL_HDRS,
1795 copts = xnnpack_std_copts(),
1796 x86_copts = ["-msse4.1"],
1797 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001798 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001799 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001800 "@FP16",
1801 "@pthreadpool",
1802 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001803)
1804
1805xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001806 name = "avx_ukernels",
1807 hdrs = INTERNAL_HDRS,
1808 copts = xnnpack_std_copts(),
1809 x86_copts = ["-mavx"],
1810 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001811 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001812 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001813 "@FP16",
1814 "@pthreadpool",
1815 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001816)
1817
1818xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001819 name = "fma3_ukernels",
1820 hdrs = INTERNAL_HDRS,
1821 copts = xnnpack_std_copts(),
1822 x86_copts = [
1823 "-mfma",
1824 ],
1825 x86_srcs = FMA3_UKERNELS,
1826 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001827 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001828 "@FP16",
1829 "@pthreadpool",
1830 ],
1831)
1832
1833xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001834 name = "avx2_ukernels",
1835 hdrs = INTERNAL_HDRS,
1836 copts = xnnpack_std_copts(),
1837 x86_copts = [
1838 "-mfma",
1839 "-mavx2",
1840 ],
1841 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001842 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001843 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001844 "@FP16",
1845 "@pthreadpool",
1846 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001847)
1848
1849xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001850 name = "avx512f_ukernels",
1851 hdrs = INTERNAL_HDRS,
1852 copts = xnnpack_std_copts(),
1853 x86_copts = ["-mavx512f"],
1854 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001855 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001856 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001857 "@FP16",
1858 "@pthreadpool",
1859 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001860)
1861
1862xnnpack_cc_library(
1863 name = "asm_ukernels",
1864 hdrs = ["src/xnnpack/assembly.h"],
1865 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001866 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867 aarch64_srcs = AARCH64_ASM_UKERNELS,
1868)
1869
1870xnnpack_aggregate_library(
1871 name = "ukernels",
1872 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001873 ":neon_ukernels",
1874 ":neonfma_ukernels",
1875 ":asm_ukernels",
1876 ],
1877 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001878 ":neon_ukernels",
1879 ":neonfma_ukernels",
1880 ":neonfp16arith_ukernels",
1881 ":asm_ukernels",
1882 ],
1883 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07001884 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001885 ":psimd_fastmath_ukernels",
1886 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001887 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001888 wasm_deps = [":wasm_ukernels"],
1889 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001890 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001891 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001892 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001893 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001894 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001895 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001896 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001897 ":avx512f_ukernels",
1898 ],
1899)
1900
1901xnnpack_cc_library(
1902 name = "im2col",
1903 srcs = ["src/im2col.c"],
1904 hdrs = [
1905 "src/xnnpack/common.h",
1906 "src/xnnpack/im2col.h",
1907 ],
1908 copts = xnnpack_std_copts(),
1909)
1910
1911xnnpack_cc_library(
1912 name = "indirection",
1913 srcs = ["src/indirection.c"],
1914 hdrs = INTERNAL_HDRS,
1915 copts = xnnpack_std_copts(),
1916 deps = [
1917 "@FP16",
1918 "@FXdiv",
1919 "@pthreadpool",
1920 ],
1921)
1922
1923xnnpack_cc_library(
1924 name = "operator_run",
1925 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001926 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhand0cf9bd2020-04-13 20:19:19 -07001927 copts = xnnpack_std_copts() + LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07001928 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1929 "//conditions:default": [],
1930 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001931 deps = [
1932 "@FP16",
1933 "@FXdiv",
1934 "@clog",
1935 "@pthreadpool",
1936 ],
1937)
1938
1939cc_library(
1940 name = "enable_assembly",
1941 defines = select({
1942 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1943 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001944 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001945 }),
1946)
1947
Marat Dukhancf056b22019-10-07 10:26:29 -07001948xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001949 name = "operators",
1950 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001951 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001952 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001953 ],
1954 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001955 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001956 "-Isrc",
1957 "-Iinclude",
1958 ] + select({
1959 ":debug_build": [],
1960 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001961 }) + select({
1962 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1963 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001964 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001965 wasm_srcs = ["src/wasm-stubs.c"],
1966 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001967 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001968 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001969 "@FP16",
1970 "@FXdiv",
1971 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001972 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001973 ],
1974)
1975
1976cc_library(
1977 name = "XNNPACK",
1978 srcs = [
1979 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08001980 "src/runtime.c",
1981 "src/subgraph.c",
1982 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001983 ],
1984 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1985 "-Isrc",
1986 "-Iinclude",
1987 ] + select({
1988 ":debug_build": [],
1989 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001990 }) + select({
1991 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1992 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001993 }),
1994 includes = ["include"],
1995 linkstatic = True,
1996 textual_hdrs = ["include/xnnpack.h"],
1997 visibility = xnnpack_visibility(),
1998 deps = [
1999 ":enable_assembly",
2000 ":ukernels",
2001 ":operator_run",
2002 ":operators",
2003 "@clog",
2004 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002005 ] + select({
2006 ":emscripten": [],
2007 "//conditions:default": ["@cpuinfo"],
2008 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002009)
2010
2011cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002012 name = "xnnpack_operators_nhwc_f32",
2013 srcs = [
2014 "src/init.c",
2015 ],
2016 copts = xnnpack_std_copts() + LOGGING_COPTS + [
2017 "-Isrc",
2018 "-Iinclude",
2019 ] + select({
2020 ":debug_build": [],
2021 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002022 }) + select({
2023 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2024 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002025 }),
2026 defines = [
2027 "XNN_NO_Q8_OPERATORS",
2028 "XNN_NO_U8_OPERATORS",
2029 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002030 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002031 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002032 includes = ["include"],
2033 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002034 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002035 visibility = xnnpack_visibility(),
2036 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002037 ":enable_assembly",
2038 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002039 ":operator_run",
2040 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002041 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002042 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002043 ] + select({
2044 ":emscripten": [],
2045 "//conditions:default": ["@cpuinfo"],
2046 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002047)
2048
Marat Dukhancf056b22019-10-07 10:26:29 -07002049xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002050 name = "bench_utils",
2051 srcs = ["bench/utils.cc"],
2052 hdrs = ["bench/utils.h"],
2053 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002054 deps = [
2055 "@com_google_benchmark//:benchmark",
2056 "@cpuinfo",
2057 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002058)
2059
Frank Barchard7e955972019-10-11 10:34:25 -07002060######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002061
2062xnnpack_benchmark(
2063 name = "q8_gemm_bench",
2064 srcs = [
2065 "bench/gemm.h",
2066 "bench/q8-gemm.cc",
2067 "src/xnnpack/AlignedAllocator.h",
2068 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2069 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
2070 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2071)
2072
2073xnnpack_benchmark(
2074 name = "f16_gemm_bench",
2075 srcs = [
2076 "bench/f16-gemm.cc",
2077 "bench/gemm.h",
2078 "src/xnnpack/AlignedAllocator.h",
2079 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2080 copts = ["-Wno-unused-function"],
2081 deps = MICROKERNEL_BENCHMARK_DEPS,
2082)
2083
2084xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002085 name = "f16_spmm_bench",
2086 srcs = [
2087 "bench/f16-spmm.cc",
2088 "bench/gemm.h",
2089 "bench/google/gemm.h",
2090 "src/xnnpack/AlignedAllocator.h",
2091 ] + MICROKERNEL_BENCHMARK_HDRS,
2092 copts = ["-Wno-unused-function"],
2093 deps = MICROKERNEL_BENCHMARK_DEPS,
2094)
2095
2096xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002097 name = "f32_igemm_bench",
2098 srcs = [
2099 "bench/f32-igemm.cc",
2100 "bench/conv.h",
2101 "src/xnnpack/AlignedAllocator.h",
2102 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002103 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002104)
2105
2106xnnpack_benchmark(
2107 name = "f32_conv_hwc_bench",
2108 srcs = [
2109 "bench/f32-conv-hwc.cc",
2110 "bench/dconv.h",
2111 "src/xnnpack/AlignedAllocator.h",
2112 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2113 copts = ["-Wno-unused-function"],
2114 deps = MICROKERNEL_BENCHMARK_DEPS,
2115)
2116
2117xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002118 name = "f32_conv_hwc2spchw_bench",
2119 srcs = [
2120 "bench/f32-conv-hwc2spchw.cc",
2121 "bench/dconv.h",
2122 "src/xnnpack/AlignedAllocator.h",
2123 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2124 copts = ["-Wno-unused-function"],
2125 deps = MICROKERNEL_BENCHMARK_DEPS,
2126)
2127
2128xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002129 name = "f32_dwconv_bench",
2130 srcs = [
2131 "bench/f32-dwconv.cc",
2132 "bench/dwconv.h",
2133 "src/xnnpack/AlignedAllocator.h",
2134 ] + MICROKERNEL_BENCHMARK_HDRS,
2135 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2136)
2137
2138xnnpack_benchmark(
2139 name = "f32_dwconv_spchw_bench",
2140 srcs = [
2141 "bench/f32-dwconv-spchw.cc",
2142 "bench/dwconv.h",
2143 "src/xnnpack/AlignedAllocator.h",
2144 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2145 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2146)
2147
2148xnnpack_benchmark(
2149 name = "f32_gemm_bench",
2150 srcs = [
2151 "bench/f32-gemm.cc",
2152 "bench/gemm.h",
2153 "src/xnnpack/AlignedAllocator.h",
2154 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2155 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002156 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002157)
2158
2159xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002160 name = "f32_raddexpminusmax_bench",
2161 srcs = [
2162 "bench/f32-raddexpminusmax.cc",
2163 "src/xnnpack/AlignedAllocator.h",
2164 ] + MICROKERNEL_BENCHMARK_HDRS,
2165 deps = MICROKERNEL_BENCHMARK_DEPS,
2166)
2167
2168xnnpack_benchmark(
2169 name = "f32_raddextexp_bench",
2170 srcs = [
2171 "bench/f32-raddextexp.cc",
2172 "src/xnnpack/AlignedAllocator.h",
2173 ] + MICROKERNEL_BENCHMARK_HDRS,
2174 deps = MICROKERNEL_BENCHMARK_DEPS,
2175)
2176
2177xnnpack_benchmark(
2178 name = "f32_raddstoreexpminusmax_bench",
2179 srcs = [
2180 "bench/f32-raddstoreexpminusmax.cc",
2181 "src/xnnpack/AlignedAllocator.h",
2182 ] + MICROKERNEL_BENCHMARK_HDRS,
2183 deps = MICROKERNEL_BENCHMARK_DEPS,
2184)
2185
2186xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002187 name = "f32_rmax_bench",
2188 srcs = [
2189 "bench/f32-rmax.cc",
2190 "src/xnnpack/AlignedAllocator.h",
2191 ] + MICROKERNEL_BENCHMARK_HDRS,
2192 deps = MICROKERNEL_BENCHMARK_DEPS,
2193)
2194
2195xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002196 name = "f32_sigmoid_bench",
2197 srcs = [
2198 "bench/f32-sigmoid.cc",
2199 "src/xnnpack/AlignedAllocator.h",
2200 ] + MICROKERNEL_BENCHMARK_HDRS,
2201 copts = ["-Wno-unused-function"],
2202 deps = MICROKERNEL_BENCHMARK_DEPS,
2203)
2204
2205xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002206 name = "f32_spmm_bench",
2207 srcs = [
2208 "bench/f32-spmm.cc",
2209 "bench/gemm.h",
2210 "src/xnnpack/AlignedAllocator.h",
2211 ] + MICROKERNEL_BENCHMARK_HDRS,
2212 copts = ["-Wno-unused-function"],
2213 deps = MICROKERNEL_BENCHMARK_DEPS,
2214)
2215
2216xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002217 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002218 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002219 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002220 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002221 copts = ["-Wno-unused-function"] + xnnpack_optional_dnnl_copts(),
2222 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002223)
2224
2225xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002226 name = "f32_vscaleexpminusmax_bench",
2227 srcs = [
2228 "bench/f32-vscaleexpminusmax.cc",
2229 "src/xnnpack/AlignedAllocator.h",
2230 ] + MICROKERNEL_BENCHMARK_HDRS,
2231 deps = MICROKERNEL_BENCHMARK_DEPS,
2232)
2233
2234xnnpack_benchmark(
2235 name = "f32_vscaleextexp_bench",
2236 srcs = [
2237 "bench/f32-vscaleextexp.cc",
2238 "src/xnnpack/AlignedAllocator.h",
2239 ] + MICROKERNEL_BENCHMARK_HDRS,
2240 deps = MICROKERNEL_BENCHMARK_DEPS,
2241)
2242
2243xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244 name = "f32_im2col_gemm_bench",
2245 srcs = [
2246 "bench/f32-im2col-gemm.cc",
2247 "bench/conv.h",
2248 "src/xnnpack/AlignedAllocator.h",
2249 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2250 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2251)
2252
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002253xnnpack_benchmark(
2254 name = "requantization_bench",
2255 srcs = [
2256 "bench/requantization.cc",
2257 "src/xnnpack/requantization-stubs.h",
2258 "src/xnnpack/AlignedAllocator.h",
2259 ] + MICROKERNEL_BENCHMARK_HDRS,
2260 deps = MICROKERNEL_BENCHMARK_DEPS,
2261)
2262
Marat Dukhan08c4a432019-10-03 09:29:21 -07002263########################### Benchmarks for operators ###########################
2264
2265xnnpack_benchmark(
2266 name = "add_bench",
2267 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002268 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002269)
2270
2271xnnpack_benchmark(
2272 name = "average_pooling_bench",
2273 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002274 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002275 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002276)
2277
2278xnnpack_benchmark(
2279 name = "channel_shuffle_bench",
2280 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002281 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002282)
2283
2284xnnpack_benchmark(
2285 name = "convolution_bench",
2286 srcs = ["bench/convolution.cc"],
2287 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002288 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289)
2290
2291xnnpack_benchmark(
2292 name = "deconvolution_bench",
2293 srcs = ["bench/deconvolution.cc"],
2294 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002295 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002296)
2297
2298xnnpack_benchmark(
2299 name = "global_average_pooling_bench",
2300 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002301 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002302)
2303
2304xnnpack_benchmark(
2305 name = "max_pooling_bench",
2306 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002307 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002308)
2309
2310xnnpack_benchmark(
2311 name = "sigmoid_bench",
2312 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002313 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002314 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002315)
2316
2317xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002318 name = "prelu_bench",
2319 srcs = ["bench/prelu.cc"],
2320 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002321 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002322)
2323
2324xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002325 name = "softmax_bench",
2326 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002327 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002328 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002329)
2330
Marat Dukhanc068bb62019-10-04 13:24:39 -07002331############################# End-to-end benchmarks ############################
2332
2333cc_library(
2334 name = "mobilenet_v1",
2335 srcs = ["models/mobilenet-v1.cc"],
2336 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002337 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002338 linkstatic = True,
2339 deps = [
2340 ":XNNPACK",
2341 "@pthreadpool",
2342 ],
2343)
2344
2345cc_library(
2346 name = "mobilenet_v2",
2347 srcs = ["models/mobilenet-v2.cc"],
2348 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002349 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002350 linkstatic = True,
2351 deps = [
2352 ":XNNPACK",
2353 "@pthreadpool",
2354 ],
2355)
2356
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002357cc_library(
2358 name = "mobilenet_v3_large",
2359 srcs = ["models/mobilenet-v3-large.cc"],
2360 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002361 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002362 linkstatic = True,
2363 deps = [
2364 ":XNNPACK",
2365 "@pthreadpool",
2366 ],
2367)
2368
2369cc_library(
2370 name = "mobilenet_v3_small",
2371 srcs = ["models/mobilenet-v3-small.cc"],
2372 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002373 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002374 linkstatic = True,
2375 deps = [
2376 ":XNNPACK",
2377 "@pthreadpool",
2378 ],
2379)
2380
Marat Dukhanc068bb62019-10-04 13:24:39 -07002381xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002382 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002383 srcs = [
2384 "bench/f32-dwconv-e2e.cc",
2385 "bench/end2end.h",
2386 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002387 copts = ["-Wno-unused-function"],
2388 deps = MICROKERNEL_BENCHMARK_DEPS + [
2389 ":XNNPACK",
2390 ":mobilenet_v1",
2391 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002392 ":mobilenet_v3_large",
2393 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002394 ],
2395)
2396
2397xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002398 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002399 srcs = [
2400 "bench/f32-gemm-e2e.cc",
2401 "bench/end2end.h",
2402 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002403 copts = ["-Wno-unused-function"],
2404 deps = MICROKERNEL_BENCHMARK_DEPS + [
2405 ":XNNPACK",
2406 ":mobilenet_v1",
2407 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002408 ":mobilenet_v3_large",
2409 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002410 ],
2411)
2412
2413xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002414 name = "end2end_bench",
2415 srcs = ["bench/end2end.cc"],
2416 deps = [
2417 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002418 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002419 ":mobilenet_v1",
2420 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002421 ":mobilenet_v3_large",
2422 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002423 "@pthreadpool",
2424 ],
2425)
2426
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002427#################### Accuracy evaluation for math functions ####################
2428
2429xnnpack_benchmark(
2430 name = "f32_exp_eval",
2431 srcs = [
2432 "eval/f32-exp.cc",
2433 "src/xnnpack/AlignedAllocator.h",
2434 ] + ACCURACY_EVAL_HDRS,
2435 deps = ACCURACY_EVAL_DEPS,
2436)
2437
Marat Dukhan515c9772019-10-17 18:07:57 -07002438xnnpack_benchmark(
2439 name = "f32_expminus_eval",
2440 srcs = [
2441 "eval/f32-expminus.cc",
2442 "src/xnnpack/AlignedAllocator.h",
2443 ] + ACCURACY_EVAL_HDRS,
2444 deps = ACCURACY_EVAL_DEPS,
2445)
2446
Marat Dukhan98ba4412019-10-23 02:14:28 -07002447xnnpack_benchmark(
2448 name = "f32_extexp_eval",
2449 srcs = [
2450 "eval/f32-extexp.cc",
2451 "src/xnnpack/AlignedAllocator.h",
2452 ] + ACCURACY_EVAL_HDRS,
2453 deps = ACCURACY_EVAL_DEPS,
2454)
2455
Marat Dukhan346a9e52019-11-15 09:06:30 -08002456xnnpack_benchmark(
2457 name = "f32_sigmoid_eval",
2458 srcs = [
2459 "eval/f32-sigmoid.cc",
2460 "src/xnnpack/AlignedAllocator.h",
2461 ] + ACCURACY_EVAL_HDRS,
2462 deps = ACCURACY_EVAL_DEPS,
2463)
2464
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465######################### Unit tests for micro-kernels #########################
2466
2467xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002468 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002469 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002470 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002471 "test/gemm-microkernel-tester.h",
2472 "src/xnnpack/AlignedAllocator.h",
2473 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2474 deps = MICROKERNEL_TEST_DEPS,
2475)
2476
2477xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002478 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002479 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002480 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002481 "test/spmm-microkernel-tester.h",
2482 "src/xnnpack/AlignedAllocator.h",
2483 ] + MICROKERNEL_TEST_HDRS,
2484 deps = MICROKERNEL_TEST_DEPS,
2485)
2486
2487xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002488 name = "f32_argmaxpool_test",
2489 srcs = [
2490 "test/f32-argmaxpool.cc",
2491 "test/argmaxpool-microkernel-tester.h",
2492 "src/xnnpack/AlignedAllocator.h",
2493 ] + MICROKERNEL_TEST_HDRS,
2494 deps = MICROKERNEL_TEST_DEPS,
2495)
2496
2497xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002498 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002499 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002500 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002501 "test/avgpool-microkernel-tester.h",
2502 "src/xnnpack/AlignedAllocator.h",
2503 ] + MICROKERNEL_TEST_HDRS,
2504 deps = MICROKERNEL_TEST_DEPS,
2505)
2506
2507xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002508 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002509 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002510 "test/f32-ibilinear.cc",
2511 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002512 "src/xnnpack/AlignedAllocator.h",
2513 ] + MICROKERNEL_TEST_HDRS,
2514 deps = MICROKERNEL_TEST_DEPS,
2515)
2516
2517xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002518 name = "f32_clamp_test",
2519 srcs = [
2520 "test/f32-clamp.cc",
2521 "test/clamp-microkernel-tester.h",
2522 ] + MICROKERNEL_TEST_HDRS,
2523 deps = MICROKERNEL_TEST_DEPS,
2524)
2525
2526xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002527 name = "f32_igemm_test",
2528 srcs = [
2529 "test/f32-igemm.cc",
2530 "test/gemm-microkernel-tester.h",
2531 "src/xnnpack/AlignedAllocator.h",
2532 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2533 deps = MICROKERNEL_TEST_DEPS,
2534)
2535
2536xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002537 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002538 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002539 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002540 "test/gemm-microkernel-tester.h",
2541 "src/xnnpack/AlignedAllocator.h",
2542 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2543 deps = MICROKERNEL_TEST_DEPS,
2544)
2545
2546xnnpack_unit_test(
2547 name = "f32_conv_hwc_test",
2548 srcs = [
2549 "test/f32-conv-hwc.cc",
2550 "test/conv-hwc-microkernel-tester.h",
2551 "src/xnnpack/AlignedAllocator.h",
2552 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2553 deps = MICROKERNEL_TEST_DEPS,
2554)
2555
2556xnnpack_unit_test(
2557 name = "f32_conv_hwc2spchw_test",
2558 srcs = [
2559 "test/f32-conv-hwc2spchw.cc",
2560 "test/conv-hwc2spchw-microkernel-tester.h",
2561 "src/xnnpack/AlignedAllocator.h",
2562 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2563 deps = MICROKERNEL_TEST_DEPS,
2564)
2565
2566xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002567 name = "f32_dwconv_test",
2568 srcs = [
2569 "test/f32-dwconv.cc",
2570 "test/dwconv-microkernel-tester.h",
2571 "src/xnnpack/AlignedAllocator.h",
2572 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2573 deps = MICROKERNEL_TEST_DEPS,
2574)
2575
2576xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002577 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002578 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002579 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002580 "test/dwconv-microkernel-tester.h",
2581 "src/xnnpack/AlignedAllocator.h",
2582 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2583 deps = MICROKERNEL_TEST_DEPS,
2584)
2585
2586xnnpack_unit_test(
2587 name = "f32_dwconv_spchw_test",
2588 srcs = [
2589 "test/f32-dwconv-spchw.cc",
2590 "test/dwconv-spchw-microkernel-tester.h",
2591 "src/xnnpack/AlignedAllocator.h",
2592 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2593 deps = MICROKERNEL_TEST_DEPS,
2594)
2595
2596xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002597 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002598 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002599 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002600 "test/gavgpool-microkernel-tester.h",
2601 "src/xnnpack/AlignedAllocator.h",
2602 ] + MICROKERNEL_TEST_HDRS,
2603 deps = MICROKERNEL_TEST_DEPS,
2604)
2605
2606xnnpack_unit_test(
2607 name = "f32_gavgpool_spchw_test",
2608 srcs = [
2609 "test/f32-gavgpool-spchw.cc",
2610 "test/gavgpool-spchw-microkernel-tester.h",
2611 "src/xnnpack/AlignedAllocator.h",
2612 ] + MICROKERNEL_TEST_HDRS,
2613 deps = MICROKERNEL_TEST_DEPS,
2614)
2615
2616xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002617 name = "f32_gemm_test",
2618 srcs = [
2619 "test/f32-gemm.cc",
2620 "test/gemm-microkernel-tester.h",
2621 "src/xnnpack/AlignedAllocator.h",
2622 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2623 deps = MICROKERNEL_TEST_DEPS,
2624)
2625
2626xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002627 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002628 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002629 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002630 "test/gemm-microkernel-tester.h",
2631 "src/xnnpack/AlignedAllocator.h",
2632 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2633 deps = MICROKERNEL_TEST_DEPS,
2634)
2635
2636xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002637 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002638 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002639 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002640 "test/gemm-microkernel-tester.h",
2641 "src/xnnpack/AlignedAllocator.h",
2642 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2643 deps = MICROKERNEL_TEST_DEPS,
2644)
2645
2646xnnpack_unit_test(
2647 name = "f32_hswish_test",
2648 srcs = [
2649 "test/f32-hswish.cc",
2650 "test/hswish-microkernel-tester.h",
2651 ] + MICROKERNEL_TEST_HDRS,
2652 deps = MICROKERNEL_TEST_DEPS,
2653)
2654
2655xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002656 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002657 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002658 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002659 "test/maxpool-microkernel-tester.h",
2660 ] + MICROKERNEL_TEST_HDRS,
2661 deps = MICROKERNEL_TEST_DEPS,
2662)
2663
2664xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002665 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002666 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002667 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002668 "test/avgpool-microkernel-tester.h",
2669 "src/xnnpack/AlignedAllocator.h",
2670 ] + MICROKERNEL_TEST_HDRS,
2671 deps = MICROKERNEL_TEST_DEPS,
2672)
2673
2674xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002675 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002676 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002677 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002678 "test/gemm-microkernel-tester.h",
2679 "src/xnnpack/AlignedAllocator.h",
2680 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2681 deps = MICROKERNEL_TEST_DEPS,
2682)
2683
2684xnnpack_unit_test(
2685 name = "f32_prelu_test",
2686 srcs = [
2687 "test/f32-prelu.cc",
2688 "test/prelu-microkernel-tester.h",
2689 "src/xnnpack/AlignedAllocator.h",
2690 ] + MICROKERNEL_TEST_HDRS,
2691 deps = MICROKERNEL_TEST_DEPS,
2692)
2693
2694xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002695 name = "f32_raddexpminusmax_test",
2696 srcs = [
2697 "test/f32-raddexpminusmax.cc",
2698 "test/raddexpminusmax-microkernel-tester.h",
2699 ] + MICROKERNEL_TEST_HDRS,
2700 deps = MICROKERNEL_TEST_DEPS,
2701)
2702
2703xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002704 name = "f32_raddextexp_test",
2705 srcs = [
2706 "test/f32-raddextexp.cc",
2707 "test/raddextexp-microkernel-tester.h",
2708 ] + MICROKERNEL_TEST_HDRS,
2709 deps = MICROKERNEL_TEST_DEPS,
2710)
2711
2712xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002713 name = "f32_raddstoreexpminusmax_test",
2714 srcs = [
2715 "test/f32-raddstoreexpminusmax.cc",
2716 "test/raddstoreexpminusmax-microkernel-tester.h",
2717 ] + MICROKERNEL_TEST_HDRS,
2718 deps = MICROKERNEL_TEST_DEPS,
2719)
2720
2721xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002722 name = "f32_rmax_test",
2723 srcs = [
2724 "test/f32-rmax.cc",
2725 "test/rmax-microkernel-tester.h",
2726 ] + MICROKERNEL_TEST_HDRS,
2727 deps = MICROKERNEL_TEST_DEPS,
2728)
2729
2730xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002731 name = "f32_sigmoid_test",
2732 srcs = [
2733 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002734 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002735 ] + MICROKERNEL_TEST_HDRS,
2736 deps = MICROKERNEL_TEST_DEPS,
2737)
2738
2739xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002740 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002741 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002742 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002743 "test/spmm-microkernel-tester.h",
2744 "src/xnnpack/AlignedAllocator.h",
2745 ] + MICROKERNEL_TEST_HDRS,
2746 deps = MICROKERNEL_TEST_DEPS,
2747)
2748
2749xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002750 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002751 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002752 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002753 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002754 ] + MICROKERNEL_TEST_HDRS,
2755 deps = MICROKERNEL_TEST_DEPS,
2756)
2757
2758xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002759 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002760 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002761 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002762 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002763 ] + MICROKERNEL_TEST_HDRS,
2764 deps = MICROKERNEL_TEST_DEPS,
2765)
2766
2767xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002768 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002769 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002770 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002771 "test/vbinary-microkernel-tester.h",
2772 ] + MICROKERNEL_TEST_HDRS,
2773 deps = MICROKERNEL_TEST_DEPS,
2774)
2775
2776xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002777 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002778 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002779 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002780 "test/vbinaryc-microkernel-tester.h",
2781 ] + MICROKERNEL_TEST_HDRS,
2782 deps = MICROKERNEL_TEST_DEPS,
2783)
2784
2785xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002786 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002787 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002788 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002789 "test/vbinaryc-microkernel-tester.h",
2790 ] + MICROKERNEL_TEST_HDRS,
2791 deps = MICROKERNEL_TEST_DEPS,
2792)
2793
2794xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002795 name = "f32_vmax_test",
2796 srcs = [
2797 "test/f32-vmax.cc",
2798 "test/vbinary-microkernel-tester.h",
2799 ] + MICROKERNEL_TEST_HDRS,
2800 deps = MICROKERNEL_TEST_DEPS,
2801)
2802
2803xnnpack_unit_test(
2804 name = "f32_vmaxc_test",
2805 srcs = [
2806 "test/f32-vmaxc.cc",
2807 "test/vbinaryc-microkernel-tester.h",
2808 ] + MICROKERNEL_TEST_HDRS,
2809 deps = MICROKERNEL_TEST_DEPS,
2810)
2811
2812xnnpack_unit_test(
2813 name = "f32_vmin_test",
2814 srcs = [
2815 "test/f32-vmin.cc",
2816 "test/vbinary-microkernel-tester.h",
2817 ] + MICROKERNEL_TEST_HDRS,
2818 deps = MICROKERNEL_TEST_DEPS,
2819)
2820
2821xnnpack_unit_test(
2822 name = "f32_vminc_test",
2823 srcs = [
2824 "test/f32-vminc.cc",
2825 "test/vbinaryc-microkernel-tester.h",
2826 ] + MICROKERNEL_TEST_HDRS,
2827 deps = MICROKERNEL_TEST_DEPS,
2828)
2829
2830xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002831 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002832 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002833 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002834 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002835 ] + MICROKERNEL_TEST_HDRS,
2836 deps = MICROKERNEL_TEST_DEPS,
2837)
2838
2839xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002840 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002841 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002842 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002843 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002844 ] + MICROKERNEL_TEST_HDRS,
2845 deps = MICROKERNEL_TEST_DEPS,
2846)
2847
2848xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002849 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002850 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002851 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002852 "test/vmulcaddc-microkernel-tester.h",
2853 "src/xnnpack/AlignedAllocator.h",
2854 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2855 deps = MICROKERNEL_TEST_DEPS,
2856)
2857
2858xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002859 name = "f32_vscale_test",
2860 srcs = [
2861 "test/f32-vscale.cc",
2862 "test/vscale-microkernel-tester.h",
2863 ] + MICROKERNEL_TEST_HDRS,
2864 deps = MICROKERNEL_TEST_DEPS,
2865)
2866
2867xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002868 name = "f32_vscaleexpminusmax_test",
2869 srcs = [
2870 "test/f32-vscaleexpminusmax.cc",
2871 "test/vscaleexpminusmax-microkernel-tester.h",
2872 ] + MICROKERNEL_TEST_HDRS,
2873 deps = MICROKERNEL_TEST_DEPS,
2874)
2875
2876xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002877 name = "f32_vscaleextexp_test",
2878 srcs = [
2879 "test/f32-vscaleextexp.cc",
2880 "test/vscaleextexp-microkernel-tester.h",
2881 ] + MICROKERNEL_TEST_HDRS,
2882 deps = MICROKERNEL_TEST_DEPS,
2883)
2884
2885xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002886 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07002887 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002888 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002889 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002890 ] + MICROKERNEL_TEST_HDRS,
2891 deps = MICROKERNEL_TEST_DEPS,
2892)
2893
2894xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002895 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002896 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002897 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002898 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002899 ] + MICROKERNEL_TEST_HDRS,
2900 deps = MICROKERNEL_TEST_DEPS,
2901)
2902
2903xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002904 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002905 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002906 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002907 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002908 ] + MICROKERNEL_TEST_HDRS,
2909 deps = MICROKERNEL_TEST_DEPS,
2910)
2911
2912xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002913 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002914 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002915 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002916 "test/avgpool-microkernel-tester.h",
2917 "src/xnnpack/AlignedAllocator.h",
2918 ] + MICROKERNEL_TEST_HDRS,
2919 deps = MICROKERNEL_TEST_DEPS,
2920)
2921
2922xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002923 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002924 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002925 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002926 "test/gemm-microkernel-tester.h",
2927 "src/xnnpack/AlignedAllocator.h",
2928 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2929 deps = MICROKERNEL_TEST_DEPS,
2930)
2931
2932xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002933 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002935 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 "test/dwconv-microkernel-tester.h",
2937 "src/xnnpack/AlignedAllocator.h",
2938 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2939 deps = MICROKERNEL_TEST_DEPS,
2940)
2941
2942xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002943 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002944 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002945 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002946 "test/gavgpool-microkernel-tester.h",
2947 "src/xnnpack/AlignedAllocator.h",
2948 ] + MICROKERNEL_TEST_HDRS,
2949 deps = MICROKERNEL_TEST_DEPS,
2950)
2951
2952xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002953 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002954 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002955 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002956 "test/gemm-microkernel-tester.h",
2957 "src/xnnpack/AlignedAllocator.h",
2958 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2959 deps = MICROKERNEL_TEST_DEPS,
2960)
2961
2962xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002963 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002965 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 "test/vadd-microkernel-tester.h",
2967 ] + MICROKERNEL_TEST_HDRS,
2968 deps = MICROKERNEL_TEST_DEPS,
2969)
2970
2971xnnpack_unit_test(
2972 name = "u8_clamp_test",
2973 srcs = [
2974 "test/u8-clamp.cc",
2975 "test/clamp-microkernel-tester.h",
2976 ] + MICROKERNEL_TEST_HDRS,
2977 deps = MICROKERNEL_TEST_DEPS,
2978)
2979
2980xnnpack_unit_test(
2981 name = "u8_lut32norm_test",
2982 srcs = [
2983 "test/u8-lut32norm.cc",
2984 "test/lut-norm-microkernel-tester.h",
2985 ] + MICROKERNEL_TEST_HDRS,
2986 deps = MICROKERNEL_TEST_DEPS,
2987)
2988
2989xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002990 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002992 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 "test/maxpool-microkernel-tester.h",
2994 ] + MICROKERNEL_TEST_HDRS,
2995 deps = MICROKERNEL_TEST_DEPS,
2996)
2997
2998xnnpack_unit_test(
2999 name = "u8_rmax_test",
3000 srcs = [
3001 "test/u8-rmax.cc",
3002 "test/rmax-microkernel-tester.h",
3003 ] + MICROKERNEL_TEST_HDRS,
3004 deps = MICROKERNEL_TEST_DEPS,
3005)
3006
3007xnnpack_unit_test(
3008 name = "x32_packx_test",
3009 srcs = [
3010 "test/x32-packx.cc",
3011 "test/pack-microkernel-tester.h",
3012 "src/xnnpack/AlignedAllocator.h",
3013 ] + MICROKERNEL_TEST_HDRS,
3014 deps = MICROKERNEL_TEST_DEPS,
3015)
3016
3017xnnpack_unit_test(
3018 name = "x32_pad_test",
3019 srcs = [
3020 "test/x32-pad.cc",
3021 "test/pad-microkernel-tester.h",
3022 ] + MICROKERNEL_TEST_HDRS,
3023 deps = MICROKERNEL_TEST_DEPS,
3024)
3025
3026xnnpack_unit_test(
3027 name = "x32_unpool_test",
3028 srcs = [
3029 "test/x32-unpool.cc",
3030 "test/unpool-microkernel-tester.h",
3031 ] + MICROKERNEL_TEST_HDRS,
3032 deps = MICROKERNEL_TEST_DEPS,
3033)
3034
3035xnnpack_unit_test(
3036 name = "x32_zip_test",
3037 srcs = [
3038 "test/x32-zip.cc",
3039 "test/zip-microkernel-tester.h",
3040 ] + MICROKERNEL_TEST_HDRS,
3041 deps = MICROKERNEL_TEST_DEPS,
3042)
3043
3044xnnpack_unit_test(
3045 name = "x8_lut_test",
3046 srcs = [
3047 "test/x8-lut.cc",
3048 "test/lut-microkernel-tester.h",
3049 ] + MICROKERNEL_TEST_HDRS,
3050 deps = MICROKERNEL_TEST_DEPS,
3051)
3052
3053xnnpack_unit_test(
3054 name = "x8_zip_test",
3055 srcs = [
3056 "test/x8-zip.cc",
3057 "test/zip-microkernel-tester.h",
3058 ] + MICROKERNEL_TEST_HDRS,
3059 deps = MICROKERNEL_TEST_DEPS,
3060)
3061
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003062xnnpack_unit_test(
3063 name = "requantization_test",
3064 srcs = [
3065 "src/xnnpack/requantization-stubs.h",
3066 "test/requantization.cc",
3067 "test/requantization-tester.h",
3068 ] + MICROKERNEL_TEST_HDRS,
3069 deps = MICROKERNEL_TEST_DEPS,
3070)
3071
Marat Dukhan20c3b922020-03-10 03:45:06 -07003072########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003073
3074xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003075 name = "operator_size_test",
3076 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003077 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003078)
3079
Marat Dukhan20c3b922020-03-10 03:45:06 -07003080xnnpack_binary(
3081 name = "subgraph_size_test",
3082 srcs = ["test/subgraph-size.c"],
3083 deps = [":XNNPACK"],
3084)
3085
3086########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003087
3088xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003089 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003090 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003091 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003092 "test/add-operator-tester.h",
3093 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003094 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095)
3096
3097xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003098 name = "add_nd_test",
3099 srcs = [
3100 "test/add-nd.cc",
3101 "test/binary-elementwise-operator-tester.h",
3102 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003103 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003104)
3105
3106xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003107 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003108 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003109 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003110 "test/argmax-pooling-operator-tester.h",
3111 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003112 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003113)
3114
3115xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003116 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003117 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003118 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119 "test/average-pooling-operator-tester.h",
3120 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003121 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122)
3123
3124xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003125 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003126 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003127 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003128 "test/channel-pad-operator-tester.h",
3129 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003130 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003131)
3132
3133xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003134 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003135 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003136 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 "test/channel-shuffle-operator-tester.h",
3138 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003139 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140)
3141
3142xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003143 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003144 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003145 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003146 "test/clamp-operator-tester.h",
3147 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003148 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003149)
3150
3151xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003152 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003153 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003154 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003155 "test/convolution-operator-tester.h",
3156 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003157 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003158)
3159
3160xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003161 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003162 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003163 "test/convolution-nchw.cc",
3164 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003166 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167)
3168
3169xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003170 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003171 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003172 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173 "test/deconvolution-operator-tester.h",
3174 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003175 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003176)
3177
3178xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003179 name = "divide_nd_test",
3180 srcs = [
3181 "test/binary-elementwise-operator-tester.h",
3182 "test/divide-nd.cc",
3183 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003184 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003185)
3186
3187xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003188 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003190 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003191 "test/fully-connected-operator-tester.h",
3192 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003193 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003194)
3195
3196xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003197 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003198 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003199 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003200 "test/global-average-pooling-operator-tester.h",
3201 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003202 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203)
3204
3205xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003206 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003208 "test/global-average-pooling-ncw.cc",
3209 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003211 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212)
3213
3214xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003215 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003216 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003217 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003218 "test/hardswish-operator-tester.h",
3219 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003220 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003221)
3222
3223xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003224 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003225 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003226 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227 "test/leaky-relu-operator-tester.h",
3228 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003229 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003230)
3231
3232xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003233 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003234 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003235 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003236 "test/max-pooling-operator-tester.h",
3237 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003238 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003239)
3240
3241xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003242 name = "maximum_nd_test",
3243 srcs = [
3244 "test/binary-elementwise-operator-tester.h",
3245 "test/maximum-nd.cc",
3246 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003247 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003248)
3249
3250xnnpack_unit_test(
3251 name = "minimum_nd_test",
3252 srcs = [
3253 "test/binary-elementwise-operator-tester.h",
3254 "test/minimum-nd.cc",
3255 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003256 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003257)
3258
3259xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003260 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003261 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003262 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003263 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003264 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003265 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003266)
3267
3268xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003269 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003270 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003271 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272 "test/prelu-operator-tester.h",
3273 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003274 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003275)
3276
3277xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003278 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003279 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003280 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003281 "test/resize-bilinear-operator-tester.h",
3282 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003283 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003284)
3285
3286xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003287 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003288 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003289 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003290 "test/sigmoid-operator-tester.h",
3291 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003292 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003293)
3294
3295xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003296 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003297 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003298 "test/softmax-nc.cc",
3299 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003300 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003301 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003302)
3303
3304xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003305 name = "subtract_nd_test",
3306 srcs = [
3307 "test/binary-elementwise-operator-tester.h",
3308 "test/subtract-nd.cc",
3309 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003310 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003311)
3312
3313xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003314 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003315 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003316 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003317 "test/unpooling-operator-tester.h",
3318 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003319 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003320)
3321
3322############################# Build configurations #############################
3323
Marat Dukhanb8642352019-10-30 15:43:02 -07003324# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003325config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003326 name = "xnn_enable_assembly_explicit_true",
3327 define_values = {"xnn_enable_assembly": "true"},
3328)
3329
3330# Disables usage of assembly kernels.
3331config_setting(
3332 name = "xnn_enable_assembly_explicit_false",
3333 define_values = {"xnn_enable_assembly": "false"},
3334)
3335
Marat Dukhan05702cf2020-03-26 15:41:33 -07003336# Disables usage of HMP-aware optimizations.
3337config_setting(
3338 name = "xnn_enable_hmp_explicit_false",
3339 define_values = {"xnn_enable_hmp": "false"},
3340)
3341
Marat Dukhanb8642352019-10-30 15:43:02 -07003342# Builds with -c dbg
3343config_setting(
3344 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003345 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003346 "compilation_mode": "dbg",
3347 },
3348)
3349
3350# Builds with -c opt
3351config_setting(
3352 name = "optimized_build",
3353 values = {
3354 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003355 },
3356)
3357
3358config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003359 name = "linux_k8",
3360 values = {"cpu": "k8"},
3361)
3362
3363config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003364 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003365 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003366)
3367
3368config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003369 name = "linux_armhf",
3370 values = {"cpu": "armhf"},
3371)
3372
3373config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003374 name = "android",
3375 values = {"crosstool_top": "//external:android/crosstool"},
3376)
3377
3378config_setting(
3379 name = "android_armv7",
3380 values = {
3381 "crosstool_top": "//external:android/crosstool",
3382 "cpu": "armeabi-v7a",
3383 },
3384)
3385
3386config_setting(
3387 name = "android_arm64",
3388 values = {
3389 "crosstool_top": "//external:android/crosstool",
3390 "cpu": "arm64-v8a",
3391 },
3392)
3393
3394config_setting(
3395 name = "android_x86",
3396 values = {
3397 "crosstool_top": "//external:android/crosstool",
3398 "cpu": "x86",
3399 },
3400)
3401
3402config_setting(
3403 name = "android_x86_64",
3404 values = {
3405 "crosstool_top": "//external:android/crosstool",
3406 "cpu": "x86_64",
3407 },
3408)
3409
3410config_setting(
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003411 name = "windows_x86",
3412 values = {"cpu": "win_x86"},
3413)
3414
3415config_setting(
3416 name = "windows_x86_64",
3417 values = {"cpu": "win_x64"},
3418)
3419
3420config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003421 name = "macos_x86_64",
3422 values = {
3423 "apple_platform_type": "macos",
3424 "cpu": "darwin",
3425 },
3426)
3427
3428config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003429 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003430 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003431)
3432
3433config_setting(
3434 name = "emscripten_wasm",
3435 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003436 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003437 "cpu": "wasm",
3438 },
3439)
3440
3441config_setting(
3442 name = "emscripten_wasmsimd",
3443 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003444 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003445 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003446 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003447 },
3448)
3449
3450config_setting(
3451 name = "emscripten_asmjs",
3452 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003453 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003454 "cpu": "asmjs",
3455 },
3456)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003457
3458config_setting(
3459 name = "ios_armv7",
3460 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003461 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003462 "cpu": "ios_armv7",
3463 },
3464)
3465
3466config_setting(
3467 name = "ios_arm64",
3468 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003469 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003470 "cpu": "ios_arm64",
3471 },
3472)
3473
3474config_setting(
3475 name = "ios_arm64e",
3476 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003477 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003478 "cpu": "ios_arm64e",
3479 },
3480)
3481
3482config_setting(
3483 name = "ios_x86",
3484 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003485 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003486 "cpu": "ios_i386",
3487 },
3488)
3489
3490config_setting(
3491 name = "ios_x86_64",
3492 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003493 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003494 "cpu": "ios_x86_64",
3495 },
3496)
3497
3498config_setting(
3499 name = "watchos_armv7k",
3500 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003501 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003502 "cpu": "watchos_armv7k",
3503 },
3504)
3505
3506config_setting(
3507 name = "watchos_arm64_32",
3508 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003509 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003510 "cpu": "watchos_arm64_32",
3511 },
3512)
3513
3514config_setting(
3515 name = "watchos_x86",
3516 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003517 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003518 "cpu": "watchos_i386",
3519 },
3520)
3521
3522config_setting(
3523 name = "watchos_x86_64",
3524 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003525 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003526 "cpu": "watchos_x86_64",
3527 },
3528)
3529
3530config_setting(
3531 name = "tvos_arm64",
3532 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003533 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003534 "cpu": "tvos_arm64",
3535 },
3536)
3537
3538config_setting(
3539 name = "tvos_x86_64",
3540 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003541 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003542 "cpu": "tvos_x86_64",
3543 },
3544)