blob: bff5197c0b6c9c81691eee79470e365a8ddcba3a [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",
656 "src/x32-zip/x2-neon.c",
657 "src/x32-zip/x3-neon.c",
658 "src/x32-zip/x4-neon.c",
659 "src/x32-zip/xm-neon.c",
660 "src/x8-zip/x2-neon.c",
661 "src/x8-zip/x3-neon.c",
662 "src/x8-zip/x4-neon.c",
663 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800664 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800665 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
666 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
667 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
668 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
669 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
670 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700671 "src/requantization/precise-neon.c",
672 "src/requantization/fp32-neon.c",
673 "src/requantization/q31-neon.c",
674 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700675]
676
677NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700678 "src/f32-ibilinear/gen/neonfma-c4.c",
679 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700680 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
681 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
682 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
683 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
684 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
685 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
686 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
687 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
688 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
689 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
690 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
691 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
692 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
693 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
694 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
695 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
696 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
697 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
698 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
699 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
700 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
701 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
702 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
703 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
704 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
705 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
706 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
707 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
708 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
709 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
710 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800711 "src/f32-hswish/gen/neonfma-x4.c",
712 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700713 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
714 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800715 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
716 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
717 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
718 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
719 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
720 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
721 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
722 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
723 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
724 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
725 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
726 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
727 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
728 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
729 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
730 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
731 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
732 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
733 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
734 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
735 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
736 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
737 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
738 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800739 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
740 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
741 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
742 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
743 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
744 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
745 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
746 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
747 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
748 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
749 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
750 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
751 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
752 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
753 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
754 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
755 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
756 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
757 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
758 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
759 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
768 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
769 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
770 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
771 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
772 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
773 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
774 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
775 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
776 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
777 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
778 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
779 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700793 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
794 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800795 "src/math/exp-neonfma-lut64-p2.c",
796 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800797 "src/math/expminus-neonfma-lut2048-p1.c",
798 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800799 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800800 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
801 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
802 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
803 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
804 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
805 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
806 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
807 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
808 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
809 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
810 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
811 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
812 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
813 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
814 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
815 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
816 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
817 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700818]
819
820AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700821 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
822 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
823 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
824 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
825 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
826 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700827 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
828 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
829 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
830 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
831 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
832 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
833 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
834 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
835 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
836 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
837 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
838 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
839 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
840 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
841 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
842 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
843 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
844 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
845 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
846 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700847 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
848 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
849 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
850 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
851 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
852 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
853 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800854 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
855 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
856 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
857 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
858 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
859 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
860 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
861 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
862 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
863 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
864 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
865 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
866 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
867 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
868 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
869 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
870 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
871 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700872 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
873 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
874 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
875 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
876 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
877 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
878 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
879 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
880 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
881 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
882 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
883 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
884 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
885 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
886 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
887 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
888 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
889 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800890 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
891 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
892 "src/math/sigmoid-neonfma-rr1-p5-div.c",
893 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
894 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
895 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700896]
897
898AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800899 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
900 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
901 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700902 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
903 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
904 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
905 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
906 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
907 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
908 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
909 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700910]
911
912SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700913 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
914 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700915 "src/f32-clamp/gen/sse-x4.c",
916 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800917 "src/f32-dwconv-spchw/3x3p1-sse.c",
918 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700919 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
920 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
921 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
922 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
923 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
924 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
925 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
926 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
927 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
928 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
929 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
930 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700931 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700932 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
933 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700934 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
935 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
936 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
937 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
938 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
939 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
940 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
941 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
942 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
943 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
944 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
945 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
946 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800947 "src/f32-hswish/gen/sse-x4.c",
948 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700949 "src/f32-ibilinear/gen/sse-c4.c",
950 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700951 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
952 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
953 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
954 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
955 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
956 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
957 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700958 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
959 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
960 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700961 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700962 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700963 "src/f32-spmm/gen/4x1-minmax-sse.c",
964 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700965 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
966 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
967 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
968 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
969 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
970 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
971 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
972 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800973 "src/f32-vbinary/gen/vmax-sse-x4.c",
974 "src/f32-vbinary/gen/vmax-sse-x8.c",
975 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
976 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
977 "src/f32-vbinary/gen/vmin-sse-x4.c",
978 "src/f32-vbinary/gen/vmin-sse-x8.c",
979 "src/f32-vbinary/gen/vminc-sse-x4.c",
980 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700981 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
982 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
983 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
984 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
985 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
986 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
987 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
988 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
989 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
990 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
991 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
992 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700993 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
994 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700995 "src/x32-packx/x4-sse.c",
996]
997
998SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800999 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1000 "src/f32-argmaxpool/4x-sse2-c4.c",
1001 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001002 "src/f32-prelu/gen/sse2-2x4.c",
1003 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001004 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1005 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1006 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1007 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1008 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1009 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1010 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1011 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1012 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1013 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1014 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1015 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001016 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001017 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001018 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001019 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001020 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1021 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001022 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1023 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001024 "src/q8-igemm/4x4c2-minmax-sse2.c",
1025 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001026 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1027 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001028 "src/q8-gemm/2x4c8-minmax-sse2.c",
1029 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001030 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001031 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001032 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001033 "src/u8-rmax/sse2.c",
1034 "src/x32-pad/x2-sse2.c",
1035 "src/x32-zip/x2-sse2.c",
1036 "src/x32-zip/x3-sse2.c",
1037 "src/x32-zip/x4-sse2.c",
1038 "src/x32-zip/xm-sse2.c",
1039 "src/x8-zip/x2-sse2.c",
1040 "src/x8-zip/x3-sse2.c",
1041 "src/x8-zip/x4-sse2.c",
1042 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001043 "src/math/exp-sse2-p5.c",
1044 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001045 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001046 "src/requantization/precise-sse2.c",
1047 "src/requantization/fp32-sse2.c",
1048 "src/requantization/q31-sse2.c",
1049 "src/requantization/gemmlowp-sse2.c",
1050]
1051
1052SSSE3_UKERNELS = [
1053 "src/requantization/precise-ssse3.c",
1054 "src/requantization/q31-ssse3.c",
1055 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001056]
1057
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001058SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001059 "src/f32-prelu/gen/sse41-2x4.c",
1060 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001061 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001062 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001063 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001064 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001065 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1066 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001067 "src/requantization/precise-sse4.c",
1068 "src/requantization/q31-sse4.c",
1069 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001070]
1071
Marat Dukhan08c4a432019-10-03 09:29:21 -07001072AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001073 "src/f32-clamp/gen/avx-x8.c",
1074 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001075 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1076 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1077 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1078 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1079 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1080 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1081 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1082 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1083 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1084 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1085 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1086 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1087 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1088 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1089 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1090 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1091 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1092 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1093 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1094 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1095 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1096 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1097 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1098 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1099 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1100 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1101 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1102 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1103 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1104 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001105 "src/f32-hswish/gen/avx-x8.c",
1106 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001107 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1108 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1109 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1110 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1111 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1112 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1113 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1114 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1115 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001116 "src/f32-prelu/gen/avx-2x8.c",
1117 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001118 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001119 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1120 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1121 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1122 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1123 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1124 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1125 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1126 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001127 "src/f32-vbinary/gen/vmax-avx-x8.c",
1128 "src/f32-vbinary/gen/vmax-avx-x16.c",
1129 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1130 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1131 "src/f32-vbinary/gen/vmin-avx-x8.c",
1132 "src/f32-vbinary/gen/vmin-avx-x16.c",
1133 "src/f32-vbinary/gen/vminc-avx-x8.c",
1134 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001135 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1136 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1137 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1138 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1139 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1140 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1141 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1142 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1143 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1144 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1145 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1146 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001147 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001148]
1149
Marat Dukhanfda12b82019-11-21 12:27:59 -08001150FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001151 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1152 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1153 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1154 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1155 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1156 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1157 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1158 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1159 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1160 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1161 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1162 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1163 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1164 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1165 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1166 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1167 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1168 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1169 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1170 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1171 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1172 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1173 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1174 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1175 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1176 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1177 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1178 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1179 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1180 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1181 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1182 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1183 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1184 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1185 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1186 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1187 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1188 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1189 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1190 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001191 "src/f32-hswish/gen/fma3-x8.c",
1192 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001193 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1194 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1195 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1196 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1197 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1198 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1199 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1200 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1201 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1202 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1203 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1204 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1205 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1206 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001207]
1208
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001209AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001210 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1211 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1212 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1213 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1214 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1215 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1216 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1217 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1218 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1219 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1220 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1221 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1222 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1223 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1224 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1225 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1226 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1227 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1228 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1229 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1230 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1231 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1232 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1233 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1234 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1235 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1236 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1237 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1238 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1239 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1240 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1241 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1242 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1243 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1244 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1245 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001246 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1247 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1248 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1249 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1250 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1251 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1252 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1253 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1254 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1255 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1256 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1257 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1258 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1259 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1260 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1261 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1262 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1263 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1264 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1265 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1266 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1267 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1268 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1269 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1270 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1271 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1272 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1273 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1274 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1275 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001276 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1277 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1278 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1279 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1280 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1281 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1282 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1283 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1284 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1285 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1286 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1287 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1288 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1289 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1290 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1291 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1292 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1293 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1294 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1295 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1296 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1297 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1298 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1299 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001300 "src/math/exp-avx2-p5.c",
1301 "src/math/exp-avx2-perm-p3.c",
1302 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001303 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001304 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001305 "src/math/sigmoid-avx2-rr2-p5-div.c",
1306 "src/math/sigmoid-avx2-rr1-p5-div.c",
1307 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1308 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1309 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1310 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001311]
1312
Marat Dukhan08c4a432019-10-03 09:29:21 -07001313AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001314 "src/f32-clamp/gen/avx512f-x16.c",
1315 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001316 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1317 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1318 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1319 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1320 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1321 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1322 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1323 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1324 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1325 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1326 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1327 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1328 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1329 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1330 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1331 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1332 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1333 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1334 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1335 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1336 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1337 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1338 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1339 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001340 "src/f32-hswish/gen/avx512f-x16.c",
1341 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001342 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1343 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1344 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1345 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1346 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1347 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001348 "src/f32-prelu/gen/avx512f-2x16.c",
1349 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001350 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1351 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1352 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1353 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1354 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1355 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1356 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1357 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1358 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1359 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1360 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1361 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1362 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1363 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1364 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1365 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1366 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1367 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1368 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1369 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1370 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1371 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1372 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1373 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1374 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1375 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1376 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1377 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1378 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1379 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1380 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1381 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1382 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1383 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1384 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1385 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001386 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001387 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1388 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1389 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1390 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1391 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1392 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1393 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1394 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001395 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1396 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1397 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1398 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1399 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1400 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1401 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1402 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001403 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1404 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1405 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1406 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1407 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1408 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1409 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1410 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1411 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1412 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1413 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1414 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001415 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001416 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1417 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1418 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1419 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1420 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1421 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1422 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1423 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1424 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1425 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1426 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1427 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1428 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1429 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1430 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1431 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1432 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1433 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1434 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1435 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1436 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1437 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1438 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1439 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001440 "src/math/exp-avx512f-p5-scalef.c",
1441 "src/math/exp-avx512f-p5.c",
1442 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001443 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001444 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001445]
1446
1447AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001448 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001449 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1450 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1451 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1452 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1453 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1454 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1455 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1456 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1457 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1458 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001459]
1460
1461AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001462 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001463 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1464 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001465 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001466 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1467 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001468 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1469 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001470 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001471 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1472 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1473 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1474 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1475 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1476 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1477 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1478 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1479 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1480 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1481 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1482 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1483 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1484 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1485 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1486 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1487 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1488 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1489 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1490 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1491 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001492 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001493 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1494 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1495 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1496 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1497 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1498 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1499 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1500 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1501 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1502 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1503 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1504 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1505 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1506 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1507 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1508 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1509 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1510 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1511 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1512 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1513 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1514 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1515 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1516 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1517 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1518 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1519 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1520 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1521 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1522 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1523 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1524 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1525 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1526 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1527 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1528 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1529 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1530 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001531]
1532
Marat Dukhan1b354632020-03-23 12:50:22 -07001533INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001534 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001535 "src/xnnpack/argmaxpool.h",
1536 "src/xnnpack/avgpool.h",
1537 "src/xnnpack/clamp.h",
1538 "src/xnnpack/common.h",
1539 "src/xnnpack/conv.h",
1540 "src/xnnpack/dwconv.h",
1541 "src/xnnpack/gavgpool.h",
1542 "src/xnnpack/gemm.h",
1543 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001544 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001545 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001546 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001547 "src/xnnpack/lut.h",
1548 "src/xnnpack/math.h",
1549 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001550 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001551 "src/xnnpack/packx.h",
1552 "src/xnnpack/pad.h",
1553 "src/xnnpack/params.h",
1554 "src/xnnpack/pavgpool.h",
1555 "src/xnnpack/ppmm.h",
1556 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001557 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001558 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001559 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001560 "src/xnnpack/rmax.h",
1561 "src/xnnpack/scalar-utils.h",
1562 "src/xnnpack/spmm.h",
1563 "src/xnnpack/unpool.h",
1564 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001565 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001566 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001567 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001568 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001569 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001570 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001571 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001572]
1573
1574INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001575 "include/xnnpack.h",
1576 "src/xnnpack/allocator.h",
1577 "src/xnnpack/compute.h",
1578 "src/xnnpack/im2col.h",
1579 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001580 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581 "src/xnnpack/operator.h",
1582 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001583 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001584 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001585 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001586 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001587]
1588
Marat Dukhan1b354632020-03-23 12:50:22 -07001589ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001590 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001591]
1592
Marat Dukhan1b354632020-03-23 12:50:22 -07001593MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001594 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595 "include/xnnpack.h",
1596]
1597
Marat Dukhan1b354632020-03-23 12:50:22 -07001598MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001599 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001600 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001601 "src/xnnpack/requantization.h",
1602 "include/xnnpack.h",
1603]
1604
1605OPERATOR_TEST_PARAMS_HDRS = [
1606 "src/xnnpack/params.h",
1607 "src/xnnpack/common.h",
1608]
1609
1610WEIGHTS_PACK_HDRS = [
1611 "src/xnnpack/pack.h",
1612 "src/xnnpack/operator.h",
1613 "src/xnnpack/compute.h",
1614]
1615
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001616LOGGING_COPTS = select({
1617 # No logging in optimized mode
1618 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1619 # Full logging in debug mode
1620 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1621 # Error-only logging in default (fastbuild) mode
1622 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1623})
1624
1625LOGGING_HDRS = [
1626 "src/xnnpack/log.h",
1627]
1628
Marat Dukhan08c4a432019-10-03 09:29:21 -07001629xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001630 name = "tables",
1631 srcs = TABLE_SRCS,
1632 hdrs = INTERNAL_HDRS,
1633 copts = xnnpack_std_copts(),
1634)
1635
1636xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001637 name = "scalar_ukernels",
1638 srcs = SCALAR_UKERNELS,
1639 hdrs = INTERNAL_HDRS,
1640 aarch32_copts = ["-marm"],
1641 copts = xnnpack_std_copts(),
1642 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001643 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001644 "@FP16",
1645 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001646 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001647 ],
1648)
1649
1650xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001651 name = "wasm_ukernels",
1652 hdrs = INTERNAL_HDRS,
1653 copts = xnnpack_std_copts(),
1654 wasm_srcs = WASM_UKERNELS,
1655 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001656 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001657 "@FP16",
1658 "@FXdiv",
1659 "@pthreadpool",
1660 ],
1661)
1662
1663xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001664 name = "psimd_fastmath_ukernels",
1665 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001666 hdrs = INTERNAL_HDRS,
1667 aarch32_copts = [
1668 "-marm",
1669 "-mfpu=neon",
1670 ],
1671 copts = xnnpack_std_copts(),
1672 optimized_copts = [
1673 "-O3",
1674 "-ffast-math",
1675 ],
1676 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001677 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001678 "@FP16",
1679 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001680 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001681 ],
1682)
1683
1684xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001685 name = "psimd_accmath_ukernels",
1686 srcs = PSIMD_ACCMATH_UKERNELS,
1687 hdrs = INTERNAL_HDRS,
1688 aarch32_copts = [
1689 "-marm",
1690 "-mfpu=neon",
1691 ],
1692 copts = xnnpack_std_copts(),
1693 optimized_copts = [
1694 "-O3",
1695 ],
1696 deps = [
1697 ":tables",
1698 "@FP16",
1699 "@psimd",
1700 "@pthreadpool",
1701 ],
1702)
1703
1704xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001705 name = "neon_ukernels",
1706 hdrs = INTERNAL_HDRS,
1707 aarch32_copts = [
1708 "-marm",
1709 "-mfpu=neon",
1710 ],
1711 aarch32_srcs = NEON_UKERNELS,
1712 aarch64_srcs = NEON_UKERNELS,
1713 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001714 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001715 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001716 "@FP16",
1717 "@pthreadpool",
1718 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001719)
1720
1721xnnpack_cc_library(
1722 name = "neonfma_ukernels",
1723 hdrs = INTERNAL_HDRS,
1724 aarch32_copts = [
1725 "-marm",
1726 "-mfpu=neon-vfpv4",
1727 ],
1728 aarch32_srcs = NEONFMA_UKERNELS,
1729 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1730 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001731 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001732 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001733 "@FP16",
1734 "@pthreadpool",
1735 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001736)
1737
1738xnnpack_cc_library(
1739 name = "neonfp16arith_ukernels",
1740 hdrs = INTERNAL_HDRS,
1741 aarch64_copts = ["-march=armv8.2-a+fp16"],
1742 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1743 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001744 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001745 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001746 "@FP16",
1747 "@pthreadpool",
1748 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001749)
1750
1751xnnpack_cc_library(
1752 name = "sse2_ukernels",
1753 hdrs = INTERNAL_HDRS,
1754 copts = xnnpack_std_copts(),
1755 x86_copts = ["-msse2"],
1756 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001757 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001758 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001759 "@FP16",
1760 "@pthreadpool",
1761 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001762)
1763
1764xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001765 name = "ssse3_ukernels",
1766 hdrs = INTERNAL_HDRS,
1767 copts = xnnpack_std_copts(),
1768 x86_copts = ["-mssse3"],
1769 x86_srcs = SSSE3_UKERNELS,
1770 deps = [
1771 ":tables",
1772 "@FP16",
1773 "@pthreadpool",
1774 ],
1775)
1776
1777xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001778 name = "sse41_ukernels",
1779 hdrs = INTERNAL_HDRS,
1780 copts = xnnpack_std_copts(),
1781 x86_copts = ["-msse4.1"],
1782 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001783 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001784 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001785 "@FP16",
1786 "@pthreadpool",
1787 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001788)
1789
1790xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001791 name = "avx_ukernels",
1792 hdrs = INTERNAL_HDRS,
1793 copts = xnnpack_std_copts(),
1794 x86_copts = ["-mavx"],
1795 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001796 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001797 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001798 "@FP16",
1799 "@pthreadpool",
1800 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001801)
1802
1803xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001804 name = "fma3_ukernels",
1805 hdrs = INTERNAL_HDRS,
1806 copts = xnnpack_std_copts(),
1807 x86_copts = [
1808 "-mfma",
1809 ],
1810 x86_srcs = FMA3_UKERNELS,
1811 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001812 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001813 "@FP16",
1814 "@pthreadpool",
1815 ],
1816)
1817
1818xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001819 name = "avx2_ukernels",
1820 hdrs = INTERNAL_HDRS,
1821 copts = xnnpack_std_copts(),
1822 x86_copts = [
1823 "-mfma",
1824 "-mavx2",
1825 ],
1826 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001827 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001828 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001829 "@FP16",
1830 "@pthreadpool",
1831 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001832)
1833
1834xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 name = "avx512f_ukernels",
1836 hdrs = INTERNAL_HDRS,
1837 copts = xnnpack_std_copts(),
1838 x86_copts = ["-mavx512f"],
1839 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001840 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001841 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001842 "@FP16",
1843 "@pthreadpool",
1844 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001845)
1846
1847xnnpack_cc_library(
1848 name = "asm_ukernels",
1849 hdrs = ["src/xnnpack/assembly.h"],
1850 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001851 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001852 aarch64_srcs = AARCH64_ASM_UKERNELS,
1853)
1854
1855xnnpack_aggregate_library(
1856 name = "ukernels",
1857 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001858 ":psimd_fastmath_ukernels",
1859 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001860 ":neon_ukernels",
1861 ":neonfma_ukernels",
1862 ":asm_ukernels",
1863 ],
1864 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001865 ":psimd_fastmath_ukernels",
1866 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867 ":neon_ukernels",
1868 ":neonfma_ukernels",
1869 ":neonfp16arith_ukernels",
1870 ":asm_ukernels",
1871 ],
1872 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001873 wasm_deps = [
1874 ":wasm_ukernels",
1875 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001876 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001877 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001878 ":psimd_fastmath_ukernels",
1879 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001880 ],
1881 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001882 ":psimd_fastmath_ukernels",
1883 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001884 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001885 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001886 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001887 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001888 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001889 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001890 ":avx512f_ukernels",
1891 ],
1892)
1893
1894xnnpack_cc_library(
1895 name = "im2col",
1896 srcs = ["src/im2col.c"],
1897 hdrs = [
1898 "src/xnnpack/common.h",
1899 "src/xnnpack/im2col.h",
1900 ],
1901 copts = xnnpack_std_copts(),
1902)
1903
1904xnnpack_cc_library(
1905 name = "indirection",
1906 srcs = ["src/indirection.c"],
1907 hdrs = INTERNAL_HDRS,
1908 copts = xnnpack_std_copts(),
1909 deps = [
1910 "@FP16",
1911 "@FXdiv",
1912 "@pthreadpool",
1913 ],
1914)
1915
1916xnnpack_cc_library(
1917 name = "operator_run",
1918 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001919 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1920 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001921 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1922 "-Wno-vla",
Marat Dukhan05702cf2020-03-26 15:41:33 -07001923 ] + select({
1924 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1925 "//conditions:default": [],
1926 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001927 deps = [
1928 "@FP16",
1929 "@FXdiv",
1930 "@clog",
1931 "@pthreadpool",
1932 ],
1933)
1934
1935cc_library(
1936 name = "enable_assembly",
1937 defines = select({
1938 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1939 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001940 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001941 }),
1942)
1943
Marat Dukhancf056b22019-10-07 10:26:29 -07001944xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001945 name = "operators",
1946 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001947 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001948 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001949 ],
1950 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001951 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001952 "-Isrc",
1953 "-Iinclude",
1954 ] + select({
1955 ":debug_build": [],
1956 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001957 }) + select({
1958 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1959 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001960 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001961 wasm_srcs = ["src/wasm-stubs.c"],
1962 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001963 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001964 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001965 "@FP16",
1966 "@FXdiv",
1967 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001968 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001969 ],
1970)
1971
1972cc_library(
1973 name = "XNNPACK",
1974 srcs = [
1975 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08001976 "src/runtime.c",
1977 "src/subgraph.c",
1978 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001979 ],
1980 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1981 "-Isrc",
1982 "-Iinclude",
1983 ] + select({
1984 ":debug_build": [],
1985 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001986 }) + select({
1987 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1988 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001989 }),
1990 includes = ["include"],
1991 linkstatic = True,
1992 textual_hdrs = ["include/xnnpack.h"],
1993 visibility = xnnpack_visibility(),
1994 deps = [
1995 ":enable_assembly",
1996 ":ukernels",
1997 ":operator_run",
1998 ":operators",
1999 "@clog",
2000 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002001 ] + select({
2002 ":emscripten": [],
2003 "//conditions:default": ["@cpuinfo"],
2004 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002005)
2006
2007cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002008 name = "xnnpack_operators_nhwc_f32",
2009 srcs = [
2010 "src/init.c",
2011 ],
2012 copts = xnnpack_std_copts() + LOGGING_COPTS + [
2013 "-Isrc",
2014 "-Iinclude",
2015 ] + select({
2016 ":debug_build": [],
2017 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002018 }) + select({
2019 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2020 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002021 }),
2022 defines = [
2023 "XNN_NO_Q8_OPERATORS",
2024 "XNN_NO_U8_OPERATORS",
2025 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002026 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002027 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002028 includes = ["include"],
2029 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002030 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002031 visibility = xnnpack_visibility(),
2032 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002033 ":enable_assembly",
2034 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002035 ":operator_run",
2036 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002037 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002038 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002039 ] + select({
2040 ":emscripten": [],
2041 "//conditions:default": ["@cpuinfo"],
2042 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002043)
2044
Marat Dukhancf056b22019-10-07 10:26:29 -07002045xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002046 name = "bench_utils",
2047 srcs = ["bench/utils.cc"],
2048 hdrs = ["bench/utils.h"],
2049 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002050 deps = [
2051 "@com_google_benchmark//:benchmark",
2052 "@cpuinfo",
2053 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054)
2055
Frank Barchard7e955972019-10-11 10:34:25 -07002056######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002057
2058xnnpack_benchmark(
2059 name = "q8_gemm_bench",
2060 srcs = [
2061 "bench/gemm.h",
2062 "bench/q8-gemm.cc",
2063 "src/xnnpack/AlignedAllocator.h",
2064 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2065 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
2066 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2067)
2068
2069xnnpack_benchmark(
2070 name = "f16_gemm_bench",
2071 srcs = [
2072 "bench/f16-gemm.cc",
2073 "bench/gemm.h",
2074 "src/xnnpack/AlignedAllocator.h",
2075 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2076 copts = ["-Wno-unused-function"],
2077 deps = MICROKERNEL_BENCHMARK_DEPS,
2078)
2079
2080xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002081 name = "f16_spmm_bench",
2082 srcs = [
2083 "bench/f16-spmm.cc",
2084 "bench/gemm.h",
2085 "bench/google/gemm.h",
2086 "src/xnnpack/AlignedAllocator.h",
2087 ] + MICROKERNEL_BENCHMARK_HDRS,
2088 copts = ["-Wno-unused-function"],
2089 deps = MICROKERNEL_BENCHMARK_DEPS,
2090)
2091
2092xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002093 name = "f32_igemm_bench",
2094 srcs = [
2095 "bench/f32-igemm.cc",
2096 "bench/conv.h",
2097 "src/xnnpack/AlignedAllocator.h",
2098 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002099 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002100)
2101
2102xnnpack_benchmark(
2103 name = "f32_conv_hwc_bench",
2104 srcs = [
2105 "bench/f32-conv-hwc.cc",
2106 "bench/dconv.h",
2107 "src/xnnpack/AlignedAllocator.h",
2108 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2109 copts = ["-Wno-unused-function"],
2110 deps = MICROKERNEL_BENCHMARK_DEPS,
2111)
2112
2113xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002114 name = "f32_conv_hwc2spchw_bench",
2115 srcs = [
2116 "bench/f32-conv-hwc2spchw.cc",
2117 "bench/dconv.h",
2118 "src/xnnpack/AlignedAllocator.h",
2119 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2120 copts = ["-Wno-unused-function"],
2121 deps = MICROKERNEL_BENCHMARK_DEPS,
2122)
2123
2124xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002125 name = "f32_dwconv_bench",
2126 srcs = [
2127 "bench/f32-dwconv.cc",
2128 "bench/dwconv.h",
2129 "src/xnnpack/AlignedAllocator.h",
2130 ] + MICROKERNEL_BENCHMARK_HDRS,
2131 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2132)
2133
2134xnnpack_benchmark(
2135 name = "f32_dwconv_spchw_bench",
2136 srcs = [
2137 "bench/f32-dwconv-spchw.cc",
2138 "bench/dwconv.h",
2139 "src/xnnpack/AlignedAllocator.h",
2140 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2141 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2142)
2143
2144xnnpack_benchmark(
2145 name = "f32_gemm_bench",
2146 srcs = [
2147 "bench/f32-gemm.cc",
2148 "bench/gemm.h",
2149 "src/xnnpack/AlignedAllocator.h",
2150 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2151 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002152 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002153)
2154
2155xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002156 name = "f32_raddexpminusmax_bench",
2157 srcs = [
2158 "bench/f32-raddexpminusmax.cc",
2159 "src/xnnpack/AlignedAllocator.h",
2160 ] + MICROKERNEL_BENCHMARK_HDRS,
2161 deps = MICROKERNEL_BENCHMARK_DEPS,
2162)
2163
2164xnnpack_benchmark(
2165 name = "f32_raddextexp_bench",
2166 srcs = [
2167 "bench/f32-raddextexp.cc",
2168 "src/xnnpack/AlignedAllocator.h",
2169 ] + MICROKERNEL_BENCHMARK_HDRS,
2170 deps = MICROKERNEL_BENCHMARK_DEPS,
2171)
2172
2173xnnpack_benchmark(
2174 name = "f32_raddstoreexpminusmax_bench",
2175 srcs = [
2176 "bench/f32-raddstoreexpminusmax.cc",
2177 "src/xnnpack/AlignedAllocator.h",
2178 ] + MICROKERNEL_BENCHMARK_HDRS,
2179 deps = MICROKERNEL_BENCHMARK_DEPS,
2180)
2181
2182xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183 name = "f32_rmax_bench",
2184 srcs = [
2185 "bench/f32-rmax.cc",
2186 "src/xnnpack/AlignedAllocator.h",
2187 ] + MICROKERNEL_BENCHMARK_HDRS,
2188 deps = MICROKERNEL_BENCHMARK_DEPS,
2189)
2190
2191xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002192 name = "f32_sigmoid_bench",
2193 srcs = [
2194 "bench/f32-sigmoid.cc",
2195 "src/xnnpack/AlignedAllocator.h",
2196 ] + MICROKERNEL_BENCHMARK_HDRS,
2197 copts = ["-Wno-unused-function"],
2198 deps = MICROKERNEL_BENCHMARK_DEPS,
2199)
2200
2201xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002202 name = "f32_spmm_bench",
2203 srcs = [
2204 "bench/f32-spmm.cc",
2205 "bench/gemm.h",
2206 "src/xnnpack/AlignedAllocator.h",
2207 ] + MICROKERNEL_BENCHMARK_HDRS,
2208 copts = ["-Wno-unused-function"],
2209 deps = MICROKERNEL_BENCHMARK_DEPS,
2210)
2211
2212xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002213 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002214 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002215 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002216 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002217 copts = ["-Wno-unused-function"] + xnnpack_optional_dnnl_copts(),
2218 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002219)
2220
2221xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002222 name = "f32_vscaleexpminusmax_bench",
2223 srcs = [
2224 "bench/f32-vscaleexpminusmax.cc",
2225 "src/xnnpack/AlignedAllocator.h",
2226 ] + MICROKERNEL_BENCHMARK_HDRS,
2227 deps = MICROKERNEL_BENCHMARK_DEPS,
2228)
2229
2230xnnpack_benchmark(
2231 name = "f32_vscaleextexp_bench",
2232 srcs = [
2233 "bench/f32-vscaleextexp.cc",
2234 "src/xnnpack/AlignedAllocator.h",
2235 ] + MICROKERNEL_BENCHMARK_HDRS,
2236 deps = MICROKERNEL_BENCHMARK_DEPS,
2237)
2238
2239xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002240 name = "f32_im2col_gemm_bench",
2241 srcs = [
2242 "bench/f32-im2col-gemm.cc",
2243 "bench/conv.h",
2244 "src/xnnpack/AlignedAllocator.h",
2245 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2246 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2247)
2248
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002249xnnpack_benchmark(
2250 name = "requantization_bench",
2251 srcs = [
2252 "bench/requantization.cc",
2253 "src/xnnpack/requantization-stubs.h",
2254 "src/xnnpack/AlignedAllocator.h",
2255 ] + MICROKERNEL_BENCHMARK_HDRS,
2256 deps = MICROKERNEL_BENCHMARK_DEPS,
2257)
2258
Marat Dukhan08c4a432019-10-03 09:29:21 -07002259########################### Benchmarks for operators ###########################
2260
2261xnnpack_benchmark(
2262 name = "add_bench",
2263 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002264 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002265)
2266
2267xnnpack_benchmark(
2268 name = "average_pooling_bench",
2269 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002270 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002271 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002272)
2273
2274xnnpack_benchmark(
2275 name = "channel_shuffle_bench",
2276 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002277 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002278)
2279
2280xnnpack_benchmark(
2281 name = "convolution_bench",
2282 srcs = ["bench/convolution.cc"],
2283 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002284 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002285)
2286
2287xnnpack_benchmark(
2288 name = "deconvolution_bench",
2289 srcs = ["bench/deconvolution.cc"],
2290 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002291 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002292)
2293
2294xnnpack_benchmark(
2295 name = "global_average_pooling_bench",
2296 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002297 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002298)
2299
2300xnnpack_benchmark(
2301 name = "max_pooling_bench",
2302 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002303 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002304)
2305
2306xnnpack_benchmark(
2307 name = "sigmoid_bench",
2308 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002309 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002310 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002311)
2312
2313xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002314 name = "prelu_bench",
2315 srcs = ["bench/prelu.cc"],
2316 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002317 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002318)
2319
2320xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002321 name = "softmax_bench",
2322 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002323 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002324 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325)
2326
Marat Dukhanc068bb62019-10-04 13:24:39 -07002327############################# End-to-end benchmarks ############################
2328
2329cc_library(
2330 name = "mobilenet_v1",
2331 srcs = ["models/mobilenet-v1.cc"],
2332 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002333 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002334 linkstatic = True,
2335 deps = [
2336 ":XNNPACK",
2337 "@pthreadpool",
2338 ],
2339)
2340
2341cc_library(
2342 name = "mobilenet_v2",
2343 srcs = ["models/mobilenet-v2.cc"],
2344 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002345 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002346 linkstatic = True,
2347 deps = [
2348 ":XNNPACK",
2349 "@pthreadpool",
2350 ],
2351)
2352
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002353cc_library(
2354 name = "mobilenet_v3_large",
2355 srcs = ["models/mobilenet-v3-large.cc"],
2356 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002357 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002358 linkstatic = True,
2359 deps = [
2360 ":XNNPACK",
2361 "@pthreadpool",
2362 ],
2363)
2364
2365cc_library(
2366 name = "mobilenet_v3_small",
2367 srcs = ["models/mobilenet-v3-small.cc"],
2368 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002369 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002370 linkstatic = True,
2371 deps = [
2372 ":XNNPACK",
2373 "@pthreadpool",
2374 ],
2375)
2376
Marat Dukhanc068bb62019-10-04 13:24:39 -07002377xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002378 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002379 srcs = [
2380 "bench/f32-dwconv-e2e.cc",
2381 "bench/end2end.h",
2382 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002383 copts = ["-Wno-unused-function"],
2384 deps = MICROKERNEL_BENCHMARK_DEPS + [
2385 ":XNNPACK",
2386 ":mobilenet_v1",
2387 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002388 ":mobilenet_v3_large",
2389 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002390 ],
2391)
2392
2393xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002394 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002395 srcs = [
2396 "bench/f32-gemm-e2e.cc",
2397 "bench/end2end.h",
2398 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002399 copts = ["-Wno-unused-function"],
2400 deps = MICROKERNEL_BENCHMARK_DEPS + [
2401 ":XNNPACK",
2402 ":mobilenet_v1",
2403 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002404 ":mobilenet_v3_large",
2405 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002406 ],
2407)
2408
2409xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002410 name = "end2end_bench",
2411 srcs = ["bench/end2end.cc"],
2412 deps = [
2413 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002414 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002415 ":mobilenet_v1",
2416 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002417 ":mobilenet_v3_large",
2418 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002419 "@pthreadpool",
2420 ],
2421)
2422
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002423#################### Accuracy evaluation for math functions ####################
2424
2425xnnpack_benchmark(
2426 name = "f32_exp_eval",
2427 srcs = [
2428 "eval/f32-exp.cc",
2429 "src/xnnpack/AlignedAllocator.h",
2430 ] + ACCURACY_EVAL_HDRS,
2431 deps = ACCURACY_EVAL_DEPS,
2432)
2433
Marat Dukhan515c9772019-10-17 18:07:57 -07002434xnnpack_benchmark(
2435 name = "f32_expminus_eval",
2436 srcs = [
2437 "eval/f32-expminus.cc",
2438 "src/xnnpack/AlignedAllocator.h",
2439 ] + ACCURACY_EVAL_HDRS,
2440 deps = ACCURACY_EVAL_DEPS,
2441)
2442
Marat Dukhan98ba4412019-10-23 02:14:28 -07002443xnnpack_benchmark(
2444 name = "f32_extexp_eval",
2445 srcs = [
2446 "eval/f32-extexp.cc",
2447 "src/xnnpack/AlignedAllocator.h",
2448 ] + ACCURACY_EVAL_HDRS,
2449 deps = ACCURACY_EVAL_DEPS,
2450)
2451
Marat Dukhan346a9e52019-11-15 09:06:30 -08002452xnnpack_benchmark(
2453 name = "f32_sigmoid_eval",
2454 srcs = [
2455 "eval/f32-sigmoid.cc",
2456 "src/xnnpack/AlignedAllocator.h",
2457 ] + ACCURACY_EVAL_HDRS,
2458 deps = ACCURACY_EVAL_DEPS,
2459)
2460
Marat Dukhan08c4a432019-10-03 09:29:21 -07002461######################### Unit tests for micro-kernels #########################
2462
2463xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002464 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002466 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002467 "test/gemm-microkernel-tester.h",
2468 "src/xnnpack/AlignedAllocator.h",
2469 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2470 deps = MICROKERNEL_TEST_DEPS,
2471)
2472
2473xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002474 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002475 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002476 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002477 "test/spmm-microkernel-tester.h",
2478 "src/xnnpack/AlignedAllocator.h",
2479 ] + MICROKERNEL_TEST_HDRS,
2480 deps = MICROKERNEL_TEST_DEPS,
2481)
2482
2483xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002484 name = "f32_argmaxpool_test",
2485 srcs = [
2486 "test/f32-argmaxpool.cc",
2487 "test/argmaxpool-microkernel-tester.h",
2488 "src/xnnpack/AlignedAllocator.h",
2489 ] + MICROKERNEL_TEST_HDRS,
2490 deps = MICROKERNEL_TEST_DEPS,
2491)
2492
2493xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002494 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002495 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002496 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002497 "test/avgpool-microkernel-tester.h",
2498 "src/xnnpack/AlignedAllocator.h",
2499 ] + MICROKERNEL_TEST_HDRS,
2500 deps = MICROKERNEL_TEST_DEPS,
2501)
2502
2503xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002504 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002505 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002506 "test/f32-ibilinear.cc",
2507 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002508 "src/xnnpack/AlignedAllocator.h",
2509 ] + MICROKERNEL_TEST_HDRS,
2510 deps = MICROKERNEL_TEST_DEPS,
2511)
2512
2513xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002514 name = "f32_clamp_test",
2515 srcs = [
2516 "test/f32-clamp.cc",
2517 "test/clamp-microkernel-tester.h",
2518 ] + MICROKERNEL_TEST_HDRS,
2519 deps = MICROKERNEL_TEST_DEPS,
2520)
2521
2522xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002523 name = "f32_igemm_test",
2524 srcs = [
2525 "test/f32-igemm.cc",
2526 "test/gemm-microkernel-tester.h",
2527 "src/xnnpack/AlignedAllocator.h",
2528 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2529 deps = MICROKERNEL_TEST_DEPS,
2530)
2531
2532xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002533 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002534 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002535 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002536 "test/gemm-microkernel-tester.h",
2537 "src/xnnpack/AlignedAllocator.h",
2538 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2539 deps = MICROKERNEL_TEST_DEPS,
2540)
2541
2542xnnpack_unit_test(
2543 name = "f32_conv_hwc_test",
2544 srcs = [
2545 "test/f32-conv-hwc.cc",
2546 "test/conv-hwc-microkernel-tester.h",
2547 "src/xnnpack/AlignedAllocator.h",
2548 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2549 deps = MICROKERNEL_TEST_DEPS,
2550)
2551
2552xnnpack_unit_test(
2553 name = "f32_conv_hwc2spchw_test",
2554 srcs = [
2555 "test/f32-conv-hwc2spchw.cc",
2556 "test/conv-hwc2spchw-microkernel-tester.h",
2557 "src/xnnpack/AlignedAllocator.h",
2558 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2559 deps = MICROKERNEL_TEST_DEPS,
2560)
2561
2562xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002563 name = "f32_dwconv_test",
2564 srcs = [
2565 "test/f32-dwconv.cc",
2566 "test/dwconv-microkernel-tester.h",
2567 "src/xnnpack/AlignedAllocator.h",
2568 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2569 deps = MICROKERNEL_TEST_DEPS,
2570)
2571
2572xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002573 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002574 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002575 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002576 "test/dwconv-microkernel-tester.h",
2577 "src/xnnpack/AlignedAllocator.h",
2578 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2579 deps = MICROKERNEL_TEST_DEPS,
2580)
2581
2582xnnpack_unit_test(
2583 name = "f32_dwconv_spchw_test",
2584 srcs = [
2585 "test/f32-dwconv-spchw.cc",
2586 "test/dwconv-spchw-microkernel-tester.h",
2587 "src/xnnpack/AlignedAllocator.h",
2588 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2589 deps = MICROKERNEL_TEST_DEPS,
2590)
2591
2592xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002593 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002594 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002595 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002596 "test/gavgpool-microkernel-tester.h",
2597 "src/xnnpack/AlignedAllocator.h",
2598 ] + MICROKERNEL_TEST_HDRS,
2599 deps = MICROKERNEL_TEST_DEPS,
2600)
2601
2602xnnpack_unit_test(
2603 name = "f32_gavgpool_spchw_test",
2604 srcs = [
2605 "test/f32-gavgpool-spchw.cc",
2606 "test/gavgpool-spchw-microkernel-tester.h",
2607 "src/xnnpack/AlignedAllocator.h",
2608 ] + MICROKERNEL_TEST_HDRS,
2609 deps = MICROKERNEL_TEST_DEPS,
2610)
2611
2612xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002613 name = "f32_gemm_test",
2614 srcs = [
2615 "test/f32-gemm.cc",
2616 "test/gemm-microkernel-tester.h",
2617 "src/xnnpack/AlignedAllocator.h",
2618 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2619 deps = MICROKERNEL_TEST_DEPS,
2620)
2621
2622xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002623 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002624 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002625 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002626 "test/gemm-microkernel-tester.h",
2627 "src/xnnpack/AlignedAllocator.h",
2628 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2629 deps = MICROKERNEL_TEST_DEPS,
2630)
2631
2632xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002633 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002634 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002635 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002636 "test/gemm-microkernel-tester.h",
2637 "src/xnnpack/AlignedAllocator.h",
2638 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2639 deps = MICROKERNEL_TEST_DEPS,
2640)
2641
2642xnnpack_unit_test(
2643 name = "f32_hswish_test",
2644 srcs = [
2645 "test/f32-hswish.cc",
2646 "test/hswish-microkernel-tester.h",
2647 ] + MICROKERNEL_TEST_HDRS,
2648 deps = MICROKERNEL_TEST_DEPS,
2649)
2650
2651xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002652 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002653 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002654 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002655 "test/maxpool-microkernel-tester.h",
2656 ] + MICROKERNEL_TEST_HDRS,
2657 deps = MICROKERNEL_TEST_DEPS,
2658)
2659
2660xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002661 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002662 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002663 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002664 "test/avgpool-microkernel-tester.h",
2665 "src/xnnpack/AlignedAllocator.h",
2666 ] + MICROKERNEL_TEST_HDRS,
2667 deps = MICROKERNEL_TEST_DEPS,
2668)
2669
2670xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002671 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002672 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002673 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002674 "test/gemm-microkernel-tester.h",
2675 "src/xnnpack/AlignedAllocator.h",
2676 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2677 deps = MICROKERNEL_TEST_DEPS,
2678)
2679
2680xnnpack_unit_test(
2681 name = "f32_prelu_test",
2682 srcs = [
2683 "test/f32-prelu.cc",
2684 "test/prelu-microkernel-tester.h",
2685 "src/xnnpack/AlignedAllocator.h",
2686 ] + MICROKERNEL_TEST_HDRS,
2687 deps = MICROKERNEL_TEST_DEPS,
2688)
2689
2690xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002691 name = "f32_raddexpminusmax_test",
2692 srcs = [
2693 "test/f32-raddexpminusmax.cc",
2694 "test/raddexpminusmax-microkernel-tester.h",
2695 ] + MICROKERNEL_TEST_HDRS,
2696 deps = MICROKERNEL_TEST_DEPS,
2697)
2698
2699xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002700 name = "f32_raddextexp_test",
2701 srcs = [
2702 "test/f32-raddextexp.cc",
2703 "test/raddextexp-microkernel-tester.h",
2704 ] + MICROKERNEL_TEST_HDRS,
2705 deps = MICROKERNEL_TEST_DEPS,
2706)
2707
2708xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002709 name = "f32_raddstoreexpminusmax_test",
2710 srcs = [
2711 "test/f32-raddstoreexpminusmax.cc",
2712 "test/raddstoreexpminusmax-microkernel-tester.h",
2713 ] + MICROKERNEL_TEST_HDRS,
2714 deps = MICROKERNEL_TEST_DEPS,
2715)
2716
2717xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002718 name = "f32_rmax_test",
2719 srcs = [
2720 "test/f32-rmax.cc",
2721 "test/rmax-microkernel-tester.h",
2722 ] + MICROKERNEL_TEST_HDRS,
2723 deps = MICROKERNEL_TEST_DEPS,
2724)
2725
2726xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002727 name = "f32_sigmoid_test",
2728 srcs = [
2729 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002730 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002731 ] + MICROKERNEL_TEST_HDRS,
2732 deps = MICROKERNEL_TEST_DEPS,
2733)
2734
2735xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002736 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002738 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002739 "test/spmm-microkernel-tester.h",
2740 "src/xnnpack/AlignedAllocator.h",
2741 ] + MICROKERNEL_TEST_HDRS,
2742 deps = MICROKERNEL_TEST_DEPS,
2743)
2744
2745xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002746 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002748 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002749 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002750 ] + MICROKERNEL_TEST_HDRS,
2751 deps = MICROKERNEL_TEST_DEPS,
2752)
2753
2754xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002755 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002756 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002757 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002758 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002759 ] + MICROKERNEL_TEST_HDRS,
2760 deps = MICROKERNEL_TEST_DEPS,
2761)
2762
2763xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002764 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002765 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002766 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002767 "test/vbinary-microkernel-tester.h",
2768 ] + MICROKERNEL_TEST_HDRS,
2769 deps = MICROKERNEL_TEST_DEPS,
2770)
2771
2772xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002773 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002774 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002775 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002776 "test/vbinaryc-microkernel-tester.h",
2777 ] + MICROKERNEL_TEST_HDRS,
2778 deps = MICROKERNEL_TEST_DEPS,
2779)
2780
2781xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002782 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002783 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002784 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002785 "test/vbinaryc-microkernel-tester.h",
2786 ] + MICROKERNEL_TEST_HDRS,
2787 deps = MICROKERNEL_TEST_DEPS,
2788)
2789
2790xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002791 name = "f32_vmax_test",
2792 srcs = [
2793 "test/f32-vmax.cc",
2794 "test/vbinary-microkernel-tester.h",
2795 ] + MICROKERNEL_TEST_HDRS,
2796 deps = MICROKERNEL_TEST_DEPS,
2797)
2798
2799xnnpack_unit_test(
2800 name = "f32_vmaxc_test",
2801 srcs = [
2802 "test/f32-vmaxc.cc",
2803 "test/vbinaryc-microkernel-tester.h",
2804 ] + MICROKERNEL_TEST_HDRS,
2805 deps = MICROKERNEL_TEST_DEPS,
2806)
2807
2808xnnpack_unit_test(
2809 name = "f32_vmin_test",
2810 srcs = [
2811 "test/f32-vmin.cc",
2812 "test/vbinary-microkernel-tester.h",
2813 ] + MICROKERNEL_TEST_HDRS,
2814 deps = MICROKERNEL_TEST_DEPS,
2815)
2816
2817xnnpack_unit_test(
2818 name = "f32_vminc_test",
2819 srcs = [
2820 "test/f32-vminc.cc",
2821 "test/vbinaryc-microkernel-tester.h",
2822 ] + MICROKERNEL_TEST_HDRS,
2823 deps = MICROKERNEL_TEST_DEPS,
2824)
2825
2826xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002827 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002828 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002829 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002830 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002831 ] + MICROKERNEL_TEST_HDRS,
2832 deps = MICROKERNEL_TEST_DEPS,
2833)
2834
2835xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002836 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002837 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002838 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002839 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840 ] + MICROKERNEL_TEST_HDRS,
2841 deps = MICROKERNEL_TEST_DEPS,
2842)
2843
2844xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002845 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002846 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002847 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002848 "test/vmulcaddc-microkernel-tester.h",
2849 "src/xnnpack/AlignedAllocator.h",
2850 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2851 deps = MICROKERNEL_TEST_DEPS,
2852)
2853
2854xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002855 name = "f32_vscale_test",
2856 srcs = [
2857 "test/f32-vscale.cc",
2858 "test/vscale-microkernel-tester.h",
2859 ] + MICROKERNEL_TEST_HDRS,
2860 deps = MICROKERNEL_TEST_DEPS,
2861)
2862
2863xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002864 name = "f32_vscaleexpminusmax_test",
2865 srcs = [
2866 "test/f32-vscaleexpminusmax.cc",
2867 "test/vscaleexpminusmax-microkernel-tester.h",
2868 ] + MICROKERNEL_TEST_HDRS,
2869 deps = MICROKERNEL_TEST_DEPS,
2870)
2871
2872xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002873 name = "f32_vscaleextexp_test",
2874 srcs = [
2875 "test/f32-vscaleextexp.cc",
2876 "test/vscaleextexp-microkernel-tester.h",
2877 ] + MICROKERNEL_TEST_HDRS,
2878 deps = MICROKERNEL_TEST_DEPS,
2879)
2880
2881xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002882 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07002883 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002884 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002885 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002886 ] + MICROKERNEL_TEST_HDRS,
2887 deps = MICROKERNEL_TEST_DEPS,
2888)
2889
2890xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002891 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002892 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002893 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002894 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002895 ] + MICROKERNEL_TEST_HDRS,
2896 deps = MICROKERNEL_TEST_DEPS,
2897)
2898
2899xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002900 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002901 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002902 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002903 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002904 ] + MICROKERNEL_TEST_HDRS,
2905 deps = MICROKERNEL_TEST_DEPS,
2906)
2907
2908xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002909 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002910 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002911 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002912 "test/avgpool-microkernel-tester.h",
2913 "src/xnnpack/AlignedAllocator.h",
2914 ] + MICROKERNEL_TEST_HDRS,
2915 deps = MICROKERNEL_TEST_DEPS,
2916)
2917
2918xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002919 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002920 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002921 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002922 "test/gemm-microkernel-tester.h",
2923 "src/xnnpack/AlignedAllocator.h",
2924 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2925 deps = MICROKERNEL_TEST_DEPS,
2926)
2927
2928xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002929 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002931 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002932 "test/dwconv-microkernel-tester.h",
2933 "src/xnnpack/AlignedAllocator.h",
2934 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2935 deps = MICROKERNEL_TEST_DEPS,
2936)
2937
2938xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002939 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002940 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002941 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002942 "test/gavgpool-microkernel-tester.h",
2943 "src/xnnpack/AlignedAllocator.h",
2944 ] + MICROKERNEL_TEST_HDRS,
2945 deps = MICROKERNEL_TEST_DEPS,
2946)
2947
2948xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002949 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002950 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002951 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002952 "test/gemm-microkernel-tester.h",
2953 "src/xnnpack/AlignedAllocator.h",
2954 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2955 deps = MICROKERNEL_TEST_DEPS,
2956)
2957
2958xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002959 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002960 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002961 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002962 "test/vadd-microkernel-tester.h",
2963 ] + MICROKERNEL_TEST_HDRS,
2964 deps = MICROKERNEL_TEST_DEPS,
2965)
2966
2967xnnpack_unit_test(
2968 name = "u8_clamp_test",
2969 srcs = [
2970 "test/u8-clamp.cc",
2971 "test/clamp-microkernel-tester.h",
2972 ] + MICROKERNEL_TEST_HDRS,
2973 deps = MICROKERNEL_TEST_DEPS,
2974)
2975
2976xnnpack_unit_test(
2977 name = "u8_lut32norm_test",
2978 srcs = [
2979 "test/u8-lut32norm.cc",
2980 "test/lut-norm-microkernel-tester.h",
2981 ] + MICROKERNEL_TEST_HDRS,
2982 deps = MICROKERNEL_TEST_DEPS,
2983)
2984
2985xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002986 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002988 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002989 "test/maxpool-microkernel-tester.h",
2990 ] + MICROKERNEL_TEST_HDRS,
2991 deps = MICROKERNEL_TEST_DEPS,
2992)
2993
2994xnnpack_unit_test(
2995 name = "u8_rmax_test",
2996 srcs = [
2997 "test/u8-rmax.cc",
2998 "test/rmax-microkernel-tester.h",
2999 ] + MICROKERNEL_TEST_HDRS,
3000 deps = MICROKERNEL_TEST_DEPS,
3001)
3002
3003xnnpack_unit_test(
3004 name = "x32_packx_test",
3005 srcs = [
3006 "test/x32-packx.cc",
3007 "test/pack-microkernel-tester.h",
3008 "src/xnnpack/AlignedAllocator.h",
3009 ] + MICROKERNEL_TEST_HDRS,
3010 deps = MICROKERNEL_TEST_DEPS,
3011)
3012
3013xnnpack_unit_test(
3014 name = "x32_pad_test",
3015 srcs = [
3016 "test/x32-pad.cc",
3017 "test/pad-microkernel-tester.h",
3018 ] + MICROKERNEL_TEST_HDRS,
3019 deps = MICROKERNEL_TEST_DEPS,
3020)
3021
3022xnnpack_unit_test(
3023 name = "x32_unpool_test",
3024 srcs = [
3025 "test/x32-unpool.cc",
3026 "test/unpool-microkernel-tester.h",
3027 ] + MICROKERNEL_TEST_HDRS,
3028 deps = MICROKERNEL_TEST_DEPS,
3029)
3030
3031xnnpack_unit_test(
3032 name = "x32_zip_test",
3033 srcs = [
3034 "test/x32-zip.cc",
3035 "test/zip-microkernel-tester.h",
3036 ] + MICROKERNEL_TEST_HDRS,
3037 deps = MICROKERNEL_TEST_DEPS,
3038)
3039
3040xnnpack_unit_test(
3041 name = "x8_lut_test",
3042 srcs = [
3043 "test/x8-lut.cc",
3044 "test/lut-microkernel-tester.h",
3045 ] + MICROKERNEL_TEST_HDRS,
3046 deps = MICROKERNEL_TEST_DEPS,
3047)
3048
3049xnnpack_unit_test(
3050 name = "x8_zip_test",
3051 srcs = [
3052 "test/x8-zip.cc",
3053 "test/zip-microkernel-tester.h",
3054 ] + MICROKERNEL_TEST_HDRS,
3055 deps = MICROKERNEL_TEST_DEPS,
3056)
3057
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003058xnnpack_unit_test(
3059 name = "requantization_test",
3060 srcs = [
3061 "src/xnnpack/requantization-stubs.h",
3062 "test/requantization.cc",
3063 "test/requantization-tester.h",
3064 ] + MICROKERNEL_TEST_HDRS,
3065 deps = MICROKERNEL_TEST_DEPS,
3066)
3067
Marat Dukhan20c3b922020-03-10 03:45:06 -07003068########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069
3070xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003071 name = "operator_size_test",
3072 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003073 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003074)
3075
Marat Dukhan20c3b922020-03-10 03:45:06 -07003076xnnpack_binary(
3077 name = "subgraph_size_test",
3078 srcs = ["test/subgraph-size.c"],
3079 deps = [":XNNPACK"],
3080)
3081
3082########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003083
3084xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003085 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003087 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003088 "test/add-operator-tester.h",
3089 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003090 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003091)
3092
3093xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003094 name = "add_nd_test",
3095 srcs = [
3096 "test/add-nd.cc",
3097 "test/binary-elementwise-operator-tester.h",
3098 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003099 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003100)
3101
3102xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003103 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003104 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003105 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106 "test/argmax-pooling-operator-tester.h",
3107 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003108 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003109)
3110
3111xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003112 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003113 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003114 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003115 "test/average-pooling-operator-tester.h",
3116 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003117 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003118)
3119
3120xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003121 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003123 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003124 "test/channel-pad-operator-tester.h",
3125 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003126 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003127)
3128
3129xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003130 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003131 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003132 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003133 "test/channel-shuffle-operator-tester.h",
3134 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003135 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003136)
3137
3138xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003139 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003141 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142 "test/clamp-operator-tester.h",
3143 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003144 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003145)
3146
3147xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003148 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003149 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003150 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003151 "test/convolution-operator-tester.h",
3152 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003153 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003154)
3155
3156xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003157 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003158 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003159 "test/convolution-nchw.cc",
3160 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003162 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003163)
3164
3165xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003166 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003168 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003169 "test/deconvolution-operator-tester.h",
3170 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003171 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003172)
3173
3174xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003175 name = "divide_nd_test",
3176 srcs = [
3177 "test/binary-elementwise-operator-tester.h",
3178 "test/divide-nd.cc",
3179 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003180 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003181)
3182
3183xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003184 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003185 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003186 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 "test/fully-connected-operator-tester.h",
3188 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003189 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190)
3191
3192xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003193 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003194 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003195 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003196 "test/global-average-pooling-operator-tester.h",
3197 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003198 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003199)
3200
3201xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003202 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003204 "test/global-average-pooling-ncw.cc",
3205 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003206 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003207 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003208)
3209
3210xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003211 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003213 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003214 "test/hardswish-operator-tester.h",
3215 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003216 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003217)
3218
3219xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003220 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003221 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003222 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003223 "test/leaky-relu-operator-tester.h",
3224 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003225 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003226)
3227
3228xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003229 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003230 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003231 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003232 "test/max-pooling-operator-tester.h",
3233 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003234 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003235)
3236
3237xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003238 name = "maximum_nd_test",
3239 srcs = [
3240 "test/binary-elementwise-operator-tester.h",
3241 "test/maximum-nd.cc",
3242 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003243 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003244)
3245
3246xnnpack_unit_test(
3247 name = "minimum_nd_test",
3248 srcs = [
3249 "test/binary-elementwise-operator-tester.h",
3250 "test/minimum-nd.cc",
3251 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003252 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003253)
3254
3255xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003256 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003257 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003258 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003259 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003260 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003261 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003262)
3263
3264xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003265 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003266 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003267 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268 "test/prelu-operator-tester.h",
3269 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003270 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271)
3272
3273xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003274 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003275 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003276 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003277 "test/resize-bilinear-operator-tester.h",
3278 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003279 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003280)
3281
3282xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003283 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003284 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003285 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003286 "test/sigmoid-operator-tester.h",
3287 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003288 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003289)
3290
3291xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003292 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003293 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003294 "test/softmax-nc.cc",
3295 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003296 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003297 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003298)
3299
3300xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003301 name = "subtract_nd_test",
3302 srcs = [
3303 "test/binary-elementwise-operator-tester.h",
3304 "test/subtract-nd.cc",
3305 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003306 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003307)
3308
3309xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003310 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003311 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003312 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003313 "test/unpooling-operator-tester.h",
3314 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003315 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003316)
3317
3318############################# Build configurations #############################
3319
Marat Dukhanb8642352019-10-30 15:43:02 -07003320# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003321config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003322 name = "xnn_enable_assembly_explicit_true",
3323 define_values = {"xnn_enable_assembly": "true"},
3324)
3325
3326# Disables usage of assembly kernels.
3327config_setting(
3328 name = "xnn_enable_assembly_explicit_false",
3329 define_values = {"xnn_enable_assembly": "false"},
3330)
3331
Marat Dukhan05702cf2020-03-26 15:41:33 -07003332# Disables usage of HMP-aware optimizations.
3333config_setting(
3334 name = "xnn_enable_hmp_explicit_false",
3335 define_values = {"xnn_enable_hmp": "false"},
3336)
3337
Marat Dukhanb8642352019-10-30 15:43:02 -07003338# Builds with -c dbg
3339config_setting(
3340 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003341 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003342 "compilation_mode": "dbg",
3343 },
3344)
3345
3346# Builds with -c opt
3347config_setting(
3348 name = "optimized_build",
3349 values = {
3350 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003351 },
3352)
3353
3354config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003355 name = "linux_k8",
3356 values = {"cpu": "k8"},
3357)
3358
3359config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003360 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003361 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003362)
3363
3364config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003365 name = "android",
3366 values = {"crosstool_top": "//external:android/crosstool"},
3367)
3368
3369config_setting(
3370 name = "android_armv7",
3371 values = {
3372 "crosstool_top": "//external:android/crosstool",
3373 "cpu": "armeabi-v7a",
3374 },
3375)
3376
3377config_setting(
3378 name = "android_arm64",
3379 values = {
3380 "crosstool_top": "//external:android/crosstool",
3381 "cpu": "arm64-v8a",
3382 },
3383)
3384
3385config_setting(
3386 name = "android_x86",
3387 values = {
3388 "crosstool_top": "//external:android/crosstool",
3389 "cpu": "x86",
3390 },
3391)
3392
3393config_setting(
3394 name = "android_x86_64",
3395 values = {
3396 "crosstool_top": "//external:android/crosstool",
3397 "cpu": "x86_64",
3398 },
3399)
3400
3401config_setting(
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003402 name = "windows_x86",
3403 values = {"cpu": "win_x86"},
3404)
3405
3406config_setting(
3407 name = "windows_x86_64",
3408 values = {"cpu": "win_x64"},
3409)
3410
3411config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003412 name = "macos_x86_64",
3413 values = {
3414 "apple_platform_type": "macos",
3415 "cpu": "darwin",
3416 },
3417)
3418
3419config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003420 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003421 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003422)
3423
3424config_setting(
3425 name = "emscripten_wasm",
3426 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003427 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003428 "cpu": "wasm",
3429 },
3430)
3431
3432config_setting(
3433 name = "emscripten_wasmsimd",
3434 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003435 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003437 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003438 },
3439)
3440
3441config_setting(
3442 name = "emscripten_asmjs",
3443 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003444 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003445 "cpu": "asmjs",
3446 },
3447)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003448
3449config_setting(
3450 name = "ios_armv7",
3451 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003452 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003453 "cpu": "ios_armv7",
3454 },
3455)
3456
3457config_setting(
3458 name = "ios_arm64",
3459 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003460 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003461 "cpu": "ios_arm64",
3462 },
3463)
3464
3465config_setting(
3466 name = "ios_arm64e",
3467 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003468 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003469 "cpu": "ios_arm64e",
3470 },
3471)
3472
3473config_setting(
3474 name = "ios_x86",
3475 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003476 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003477 "cpu": "ios_i386",
3478 },
3479)
3480
3481config_setting(
3482 name = "ios_x86_64",
3483 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003484 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003485 "cpu": "ios_x86_64",
3486 },
3487)
3488
3489config_setting(
3490 name = "watchos_armv7k",
3491 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003492 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003493 "cpu": "watchos_armv7k",
3494 },
3495)
3496
3497config_setting(
3498 name = "watchos_arm64_32",
3499 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003500 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003501 "cpu": "watchos_arm64_32",
3502 },
3503)
3504
3505config_setting(
3506 name = "watchos_x86",
3507 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003508 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003509 "cpu": "watchos_i386",
3510 },
3511)
3512
3513config_setting(
3514 name = "watchos_x86_64",
3515 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003516 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003517 "cpu": "watchos_x86_64",
3518 },
3519)
3520
3521config_setting(
3522 name = "tvos_arm64",
3523 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003524 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003525 "cpu": "tvos_arm64",
3526 },
3527)
3528
3529config_setting(
3530 name = "tvos_x86_64",
3531 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003532 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003533 "cpu": "tvos_x86_64",
3534 },
3535)