blob: 1b2323ced5e985e17841e8deeb8e59c079053d90 [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 Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_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_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
Marat Dukhan08c4a432019-10-03 09:29:21 -070022MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
Marat Dukhan1b354632020-03-23 12:50:22 -070046OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
Marat Dukhanfd8e6892020-01-27 15:25:25 -080072 "src/softmax-nc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080073 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -070085 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
86 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -070087 "src/f32-clamp/gen/scalar-x1.c",
88 "src/f32-clamp/gen/scalar-x2.c",
89 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080090 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080091 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070092 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070093 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070094 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080095 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080096 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -070097 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x4-scalar.c",
99 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
100 "src/f32-dwconv/gen/up1x9-scalar.c",
101 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
102 "src/f32-dwconv/gen/up1x25-scalar.c",
103 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x4-scalar.c",
105 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
106 "src/f32-dwconv/gen/up2x9-scalar.c",
107 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
108 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700109 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
110 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
111 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
112 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700113 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
114 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700115 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
116 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
117 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
118 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700119 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
120 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700121 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700122 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
123 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700124 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
125 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
126 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700127 "src/f32-gemm/gen/1x4-scalar.c",
128 "src/f32-gemm/gen/2x4-scalar.c",
129 "src/f32-gemm/gen/4x2-scalar.c",
130 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700131 "src/f32-gemm/gen/1x4-minmax-scalar.c",
132 "src/f32-gemm/gen/2x4-minmax-scalar.c",
133 "src/f32-gemm/gen/4x2-minmax-scalar.c",
134 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800135 "src/f32-hswish/gen/scalar-x1.c",
136 "src/f32-hswish/gen/scalar-x2.c",
137 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700138 "src/f32-ibilinear/gen/scalar-c1.c",
139 "src/f32-ibilinear/gen/scalar-c2.c",
140 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-igemm/gen/1x4-scalar.c",
142 "src/f32-igemm/gen/2x4-scalar.c",
143 "src/f32-igemm/gen/4x2-scalar.c",
144 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-igemm/gen/1x4-minmax-scalar.c",
146 "src/f32-igemm/gen/2x4-minmax-scalar.c",
147 "src/f32-igemm/gen/4x2-minmax-scalar.c",
148 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700149 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
150 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
151 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700152 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
153 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
154 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
155 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800156 "src/f32-prelu/gen/scalar-2x1.c",
157 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800158 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
159 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
160 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
161 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
162 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
163 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
164 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
165 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
166 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
167 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
168 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
169 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700170 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800171 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
172 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
173 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
174 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
175 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
176 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
177 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
178 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
179 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700180 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
181 "src/f32-spmm/gen/1x1-minmax-scalar.c",
182 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
183 "src/f32-spmm/gen/2x1-minmax-scalar.c",
184 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
185 "src/f32-spmm/gen/4x1-minmax-scalar.c",
186 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
187 "src/f32-spmm/gen/8x1-minmax-scalar.c",
188 "src/f32-spmm/gen/8x2-minmax-scalar.c",
189 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700190 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
191 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
192 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
193 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
194 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
195 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
196 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
197 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
198 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
199 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
200 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
201 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800202 "src/f32-vbinary/gen/vmax-scalar-x1.c",
203 "src/f32-vbinary/gen/vmax-scalar-x2.c",
204 "src/f32-vbinary/gen/vmax-scalar-x4.c",
205 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
206 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
207 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
208 "src/f32-vbinary/gen/vmin-scalar-x1.c",
209 "src/f32-vbinary/gen/vmin-scalar-x2.c",
210 "src/f32-vbinary/gen/vmin-scalar-x4.c",
211 "src/f32-vbinary/gen/vminc-scalar-x1.c",
212 "src/f32-vbinary/gen/vminc-scalar-x2.c",
213 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700214 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
215 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
216 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
217 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
218 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
219 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
220 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
221 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
222 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
223 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
224 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
225 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
226 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
227 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
228 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
229 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700232 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
233 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
234 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800235 "src/math/expminus-scalar-lut2048-p1.c",
236 "src/math/expminus-scalar-lut64-p2.c",
237 "src/math/expminus-scalar-p5.c",
238 "src/math/sigmoid-scalar-lut2048-p1-div.c",
239 "src/math/sigmoid-scalar-lut64-p2-div.c",
240 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700241 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
242 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700243 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700244 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
245 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700246 "src/q8-gemm/2x2-minmax-scalar.c",
247 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700248 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700249 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700250 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700251 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700252 "src/u8-rmax/scalar.c",
253 "src/x32-packx/x2-scalar.c",
254 "src/x32-packx/x3-scalar.c",
255 "src/x32-packx/x4-scalar.c",
256 "src/x32-pad/x2-scalar.c",
257 "src/x32-unpool/scalar.c",
258 "src/x32-zip/x2-scalar.c",
259 "src/x32-zip/x3-scalar.c",
260 "src/x32-zip/x4-scalar.c",
261 "src/x32-zip/xm-scalar.c",
262 "src/x8-lut/scalar.c",
263 "src/x8-zip/x2-scalar.c",
264 "src/x8-zip/x3-scalar.c",
265 "src/x8-zip/x4-scalar.c",
266 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700267 "src/requantization/precise-scalar.c",
268 "src/requantization/fp32-scalar.c",
269 "src/requantization/q31-scalar.c",
270 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700271]
272
Marat Dukhan436ebe62019-12-04 15:10:12 -0800273WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700274 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
275 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700276 "src/f32-clamp/gen/wasm-x1.c",
277 "src/f32-clamp/gen/wasm-x2.c",
278 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700279 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
280 "src/f32-dwconv/gen/up1x4-wasm.c",
281 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
282 "src/f32-dwconv/gen/up1x9-wasm.c",
283 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
284 "src/f32-dwconv/gen/up1x25-wasm.c",
285 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
286 "src/f32-dwconv/gen/up2x4-wasm.c",
287 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
288 "src/f32-dwconv/gen/up2x9-wasm.c",
289 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
290 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700291 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
292 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
293 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
294 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700295 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
296 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700297 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
298 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
299 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
300 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700301 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
302 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700303 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
304 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700305 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
306 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
307 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700308 "src/f32-gemm/gen/1x4-wasm.c",
309 "src/f32-gemm/gen/2x4-wasm.c",
310 "src/f32-gemm/gen/4x2-wasm.c",
311 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700312 "src/f32-gemm/gen/1x4-minmax-wasm.c",
313 "src/f32-gemm/gen/2x4-minmax-wasm.c",
314 "src/f32-gemm/gen/4x2-minmax-wasm.c",
315 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800316 "src/f32-hswish/gen/wasm-x1.c",
317 "src/f32-hswish/gen/wasm-x2.c",
318 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700319 "src/f32-igemm/gen/1x4-wasm.c",
320 "src/f32-igemm/gen/2x4-wasm.c",
321 "src/f32-igemm/gen/4x2-wasm.c",
322 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700323 "src/f32-igemm/gen/1x4-minmax-wasm.c",
324 "src/f32-igemm/gen/2x4-minmax-wasm.c",
325 "src/f32-igemm/gen/4x2-minmax-wasm.c",
326 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700327 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
328 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
329 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700330 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
331 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
332 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
333 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
334 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
335 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
336 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
337 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
338 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
339 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
340 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
341 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800342 "src/f32-vbinary/gen/vmax-wasm-x1.c",
343 "src/f32-vbinary/gen/vmax-wasm-x2.c",
344 "src/f32-vbinary/gen/vmax-wasm-x4.c",
345 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
346 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
347 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
348 "src/f32-vbinary/gen/vmin-wasm-x1.c",
349 "src/f32-vbinary/gen/vmin-wasm-x2.c",
350 "src/f32-vbinary/gen/vmin-wasm-x4.c",
351 "src/f32-vbinary/gen/vminc-wasm-x1.c",
352 "src/f32-vbinary/gen/vminc-wasm-x2.c",
353 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700354 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
355 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
356 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
357 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
358 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
359 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
360 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
361 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
362 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
363 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
364 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
365 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
366 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
367 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
368 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
369 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
370 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
371 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700372 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
373 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
374 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800375]
376
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800377PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800378 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800379 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800380 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700381 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
382 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700383 "src/f32-clamp/gen/psimd-x4.c",
384 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700385 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
386 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
387 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
388 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
389 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
390 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
391 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
392 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
393 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
394 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
395 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
396 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700397 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
398 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700399 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
400 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
401 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
402 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
403 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
404 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
405 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
406 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
407 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
408 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
409 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
410 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
411 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
412 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
413 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
414 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
415 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
416 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
417 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800418 "src/f32-hswish/gen/psimd-x4.c",
419 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700420 "src/f32-ibilinear/gen/psimd-c4.c",
421 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
423 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
424 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
425 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
426 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
427 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
428 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
429 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
430 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
431 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700432 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
433 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
434 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700435 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800436 "src/f32-prelu/gen/psimd-2x4.c",
437 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800438 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700439 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
440 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
441 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
442 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
443 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
444 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
445 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
446 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800447 "src/f32-vbinary/gen/vmax-psimd-x4.c",
448 "src/f32-vbinary/gen/vmax-psimd-x8.c",
449 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
450 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
451 "src/f32-vbinary/gen/vmin-psimd-x4.c",
452 "src/f32-vbinary/gen/vmin-psimd-x8.c",
453 "src/f32-vbinary/gen/vminc-psimd-x4.c",
454 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700455 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
456 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
457 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
458 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
459 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
460 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
461 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
462 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
463 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
464 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
465 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
466 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700467 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
468 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700469 "src/x32-packx/x4-psimd.c",
470 "src/x32-pad/x2-psimd.c",
471 "src/x32-unpool/psimd.c",
472 "src/x32-zip/x2-psimd.c",
473 "src/x32-zip/x3-psimd.c",
474 "src/x32-zip/x4-psimd.c",
475 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700476 "src/requantization/precise-psimd.c",
477 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700478]
479
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800480PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800481 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
482 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
483 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
484 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
485 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
486 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
487 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
488 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
489 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
490 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
491 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
492 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800493 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
494 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
495 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
496 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
497 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
498 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
499 "src/math/sigmoid-psimd-p5-div.c",
500]
501
Marat Dukhan08c4a432019-10-03 09:29:21 -0700502# ISA-specific micro-kernels
503NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700504 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
505 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700506 "src/f32-clamp/gen/neon-x4.c",
507 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700508 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
509 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
510 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
511 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700512 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700513 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
514 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700515 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
516 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
517 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
518 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
519 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
520 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
521 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
522 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
523 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
524 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
525 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
526 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
527 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
528 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
529 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
530 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
531 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
532 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
533 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
534 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
535 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
536 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
537 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
538 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
539 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
540 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
541 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
542 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
543 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
544 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
545 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800546 "src/f32-hswish/gen/neon-x4.c",
547 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700548 "src/f32-ibilinear/gen/neon-c4.c",
549 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700550 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
551 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
552 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
553 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
554 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
555 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
556 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
557 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
558 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
559 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
560 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
561 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
562 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
563 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
564 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
565 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700566 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
567 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
568 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700569 "src/f32-ppmm/gen/4x8-minmax-neon.c",
570 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800571 "src/f32-prelu/gen/neon-2x4.c",
572 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
580 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
581 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
582 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
583 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
584 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
585 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
586 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
587 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
588 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
589 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
590 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
591 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
592 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
593 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
594 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
595 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
596 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700597 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800598 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
601 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
602 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
603 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
604 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
605 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
606 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
607 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
608 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
609 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
610 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
611 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
612 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
613 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
614 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
615 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
616 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700617 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
618 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
619 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
620 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800621 "src/f32-vbinary/gen/vmax-neon-x4.c",
622 "src/f32-vbinary/gen/vmax-neon-x8.c",
623 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
624 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
625 "src/f32-vbinary/gen/vmin-neon-x4.c",
626 "src/f32-vbinary/gen/vmin-neon-x8.c",
627 "src/f32-vbinary/gen/vminc-neon-x4.c",
628 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700629 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
630 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
631 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
632 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
633 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
634 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
635 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
636 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
637 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
638 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700639 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
640 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
641 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
642 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700643 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700644 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
645 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700646 "src/q8-gemm/4x8-minmax-neon.c",
647 "src/q8-gemm/8x8-minmax-neon.c",
648 "src/q8-igemm/4x8-minmax-neon.c",
649 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700650 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700651 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700652 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700653 "src/u8-rmax/neon.c",
654 "src/x32-packx/x4-neon-st4.c",
655 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700656 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700657 "src/x32-zip/x2-neon.c",
658 "src/x32-zip/x3-neon.c",
659 "src/x32-zip/x4-neon.c",
660 "src/x32-zip/xm-neon.c",
661 "src/x8-zip/x2-neon.c",
662 "src/x8-zip/x3-neon.c",
663 "src/x8-zip/x4-neon.c",
664 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800665 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800666 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
667 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
668 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
669 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
670 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
671 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700672 "src/requantization/precise-neon.c",
673 "src/requantization/fp32-neon.c",
674 "src/requantization/q31-neon.c",
675 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700676]
677
678NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700679 "src/f32-ibilinear/gen/neonfma-c4.c",
680 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700681 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
682 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
683 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
684 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
685 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
686 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
687 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
688 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
689 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
690 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
691 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
692 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
693 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
694 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
695 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
696 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
697 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
698 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
699 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
700 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
701 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
702 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
703 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
704 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
705 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
706 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
707 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
708 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
709 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
710 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
711 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800712 "src/f32-hswish/gen/neonfma-x4.c",
713 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700714 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
715 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800716 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
717 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
718 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
719 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
720 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
722 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
723 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
725 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
726 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
727 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
728 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
729 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
730 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
731 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
732 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
733 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
734 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
735 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
736 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
737 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
738 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
739 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800740 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
741 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
742 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
743 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
744 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
745 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
746 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
747 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
748 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
749 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
750 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
751 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
752 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
753 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
754 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
755 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
756 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
757 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
758 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
759 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
768 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
769 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
770 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
771 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
772 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
773 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
774 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
775 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
776 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
777 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
778 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
779 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700794 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
795 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800796 "src/math/exp-neonfma-lut64-p2.c",
797 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800798 "src/math/expminus-neonfma-lut2048-p1.c",
799 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800800 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800801 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
802 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
803 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
804 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
805 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
806 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
807 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
808 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
809 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
810 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
811 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
812 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
813 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
814 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
815 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
816 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
817 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
818 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700819]
820
821AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700822 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
823 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
824 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
825 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
826 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
827 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700828 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
829 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
830 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
831 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
832 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
833 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
834 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
835 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
836 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
837 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
838 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
839 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
840 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
841 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
842 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
843 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
844 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
845 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
846 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
847 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700848 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
849 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
850 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
851 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
852 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
853 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
854 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800855 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
856 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
857 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
858 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
859 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
860 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
861 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
862 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
863 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
864 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
865 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
866 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
867 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
868 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
869 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
870 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
871 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
872 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700873 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
874 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
875 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
876 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
877 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
878 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
879 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
880 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
881 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
882 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
883 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
884 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
885 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
886 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
887 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
888 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
889 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
890 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800891 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
892 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
893 "src/math/sigmoid-neonfma-rr1-p5-div.c",
894 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
895 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
896 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700897]
898
899AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard1f4e4612020-04-13 18:24:54 -0700900 "src/f16-gemm/gen/1x8-neonfp16arith-ld64.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800901 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
902 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
903 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700904 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
905 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
906 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
907 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
908 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
909 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
910 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
911 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700912]
913
914SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700915 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
916 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700917 "src/f32-clamp/gen/sse-x4.c",
918 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800919 "src/f32-dwconv-spchw/3x3p1-sse.c",
920 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700921 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
922 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
923 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
924 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
925 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
926 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
927 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
928 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
929 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
930 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
931 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
932 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700933 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700934 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
935 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700936 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
937 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
938 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
939 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
940 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
941 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
942 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
943 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
944 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
945 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
946 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
947 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
948 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800949 "src/f32-hswish/gen/sse-x4.c",
950 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700951 "src/f32-ibilinear/gen/sse-c4.c",
952 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700953 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
954 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
955 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
956 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
957 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
958 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
959 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700960 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
961 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
962 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700963 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700964 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700965 "src/f32-spmm/gen/4x1-minmax-sse.c",
966 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700967 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
968 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
969 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
970 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
971 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
972 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
973 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
974 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800975 "src/f32-vbinary/gen/vmax-sse-x4.c",
976 "src/f32-vbinary/gen/vmax-sse-x8.c",
977 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
978 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
979 "src/f32-vbinary/gen/vmin-sse-x4.c",
980 "src/f32-vbinary/gen/vmin-sse-x8.c",
981 "src/f32-vbinary/gen/vminc-sse-x4.c",
982 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700983 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
984 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
985 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
986 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
987 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
988 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
989 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
990 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
991 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
992 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
993 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
994 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700995 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
996 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700997 "src/x32-packx/x4-sse.c",
998]
999
1000SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001001 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1002 "src/f32-argmaxpool/4x-sse2-c4.c",
1003 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001004 "src/f32-prelu/gen/sse2-2x4.c",
1005 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001006 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1007 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1008 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1009 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1010 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1012 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1013 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1014 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1015 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1016 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1017 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001018 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001019 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001020 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001021 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001022 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1023 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001024 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1025 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001026 "src/q8-igemm/4x4c2-minmax-sse2.c",
1027 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001028 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1029 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001030 "src/q8-gemm/2x4c8-minmax-sse2.c",
1031 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001032 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001033 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001034 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001035 "src/u8-rmax/sse2.c",
1036 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001037 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001038 "src/x32-zip/x2-sse2.c",
1039 "src/x32-zip/x3-sse2.c",
1040 "src/x32-zip/x4-sse2.c",
1041 "src/x32-zip/xm-sse2.c",
1042 "src/x8-zip/x2-sse2.c",
1043 "src/x8-zip/x3-sse2.c",
1044 "src/x8-zip/x4-sse2.c",
1045 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001046 "src/math/exp-sse2-p5.c",
1047 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001048 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001049 "src/requantization/precise-sse2.c",
1050 "src/requantization/fp32-sse2.c",
1051 "src/requantization/q31-sse2.c",
1052 "src/requantization/gemmlowp-sse2.c",
1053]
1054
1055SSSE3_UKERNELS = [
1056 "src/requantization/precise-ssse3.c",
1057 "src/requantization/q31-ssse3.c",
1058 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001059]
1060
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001061SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001062 "src/f32-prelu/gen/sse41-2x4.c",
1063 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001064 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001065 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001066 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001067 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001068 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1069 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001070 "src/requantization/precise-sse4.c",
1071 "src/requantization/q31-sse4.c",
1072 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001073]
1074
Marat Dukhan08c4a432019-10-03 09:29:21 -07001075AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001076 "src/f32-clamp/gen/avx-x8.c",
1077 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001078 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1079 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1080 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1081 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1082 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1083 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1084 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1085 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1086 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1087 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1088 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1089 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1090 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1091 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1092 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1093 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1094 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1095 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1096 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1097 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1098 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1099 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1100 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1101 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1102 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1103 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1104 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1105 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1106 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1107 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001108 "src/f32-hswish/gen/avx-x8.c",
1109 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001110 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1111 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1112 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1113 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1114 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1115 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1116 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1117 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1118 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001119 "src/f32-prelu/gen/avx-2x8.c",
1120 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001122 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1123 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1124 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1125 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1126 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1127 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1128 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1129 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001130 "src/f32-vbinary/gen/vmax-avx-x8.c",
1131 "src/f32-vbinary/gen/vmax-avx-x16.c",
1132 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1133 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1134 "src/f32-vbinary/gen/vmin-avx-x8.c",
1135 "src/f32-vbinary/gen/vmin-avx-x16.c",
1136 "src/f32-vbinary/gen/vminc-avx-x8.c",
1137 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001138 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1139 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1140 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1141 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1142 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1143 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1144 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1145 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1146 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1147 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1148 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1149 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001150 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001151]
1152
Marat Dukhanfda12b82019-11-21 12:27:59 -08001153FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001154 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1155 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1156 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1157 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1158 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1159 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1160 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1161 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1162 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1163 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1164 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1165 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1166 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1167 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1168 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1169 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1170 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1171 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1172 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1173 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1174 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1175 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1176 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1177 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1178 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1179 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1180 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1181 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1182 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1183 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1184 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1185 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1186 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1187 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1188 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1189 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1190 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1191 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1192 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1193 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001194 "src/f32-hswish/gen/fma3-x8.c",
1195 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001196 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1197 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1198 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1199 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1200 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1201 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1202 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1203 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1204 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1205 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1206 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1207 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1208 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1209 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001210]
1211
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001212AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001213 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1214 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1215 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1216 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1217 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1218 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1219 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1220 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1221 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1222 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1223 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1224 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1225 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1226 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1227 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1228 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1229 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1230 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1231 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1232 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1233 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1234 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1235 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1236 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1237 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1238 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1239 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1240 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1241 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1242 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1243 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1244 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1245 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1246 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1247 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1248 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001249 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1250 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1251 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1252 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1253 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1254 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1255 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1256 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1257 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1258 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1259 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1260 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1261 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1262 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1263 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1264 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1265 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1266 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1267 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1268 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1269 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1270 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1271 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1272 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1273 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1274 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1275 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1276 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1277 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1278 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001279 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1280 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1281 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1282 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1283 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1284 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1285 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1286 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1287 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1288 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1289 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1290 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1291 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1292 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1293 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1294 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1295 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1296 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1297 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1298 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1299 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1300 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1301 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1302 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001303 "src/math/exp-avx2-p5.c",
1304 "src/math/exp-avx2-perm-p3.c",
1305 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001306 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001307 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001308 "src/math/sigmoid-avx2-rr2-p5-div.c",
1309 "src/math/sigmoid-avx2-rr1-p5-div.c",
1310 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1311 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1312 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1313 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001314]
1315
Marat Dukhan08c4a432019-10-03 09:29:21 -07001316AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001317 "src/f32-clamp/gen/avx512f-x16.c",
1318 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001319 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1320 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1321 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1322 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1323 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1324 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1325 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1326 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1327 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1328 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1329 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1330 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1331 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1332 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1333 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1334 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1335 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1336 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1337 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1338 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1339 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1340 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1341 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1342 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001343 "src/f32-hswish/gen/avx512f-x16.c",
1344 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001345 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1346 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1347 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1348 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1349 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1350 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001351 "src/f32-prelu/gen/avx512f-2x16.c",
1352 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001353 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1354 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1355 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1356 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1357 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1358 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1359 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1360 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1361 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1362 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1363 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1364 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1365 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1366 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1367 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1368 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1369 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1370 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1371 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1372 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1373 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1374 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1375 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1376 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1377 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1378 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1379 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1380 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1381 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1382 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1383 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1384 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1385 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1386 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1387 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1388 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001389 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001390 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1391 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1392 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1393 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1394 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1395 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1396 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1397 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001398 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1399 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1400 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1401 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1402 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1403 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1404 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1405 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001406 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1407 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1408 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1409 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1410 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1411 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1412 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1413 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1414 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1415 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1416 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1417 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001418 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001419 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1420 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1421 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1422 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1423 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1424 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1425 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1426 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1427 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1428 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1429 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1430 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1431 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1432 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1433 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1434 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1435 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1436 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1437 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1438 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1439 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1440 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1441 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1442 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001443 "src/math/exp-avx512f-p5-scalef.c",
1444 "src/math/exp-avx512f-p5.c",
1445 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001446 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001447 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001448]
1449
1450AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001451 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001452 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1453 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1454 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1455 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1456 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1457 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1458 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1459 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1460 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1461 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001462]
1463
1464AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001465 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001466 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1467 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001468 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001469 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1470 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001471 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1472 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1473 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
1474 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1475 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1476 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001477 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1478 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001479 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001480 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1481 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1482 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1483 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1484 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1485 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1486 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1487 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1488 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1489 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1490 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1491 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1492 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1493 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1494 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1495 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1496 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1497 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1498 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1499 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1500 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001501 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001502 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1503 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1504 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1505 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1506 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1507 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1508 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1509 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1510 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1511 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1512 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1513 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1514 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1515 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1516 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1517 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1518 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1519 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1520 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1521 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1522 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1523 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1524 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1525 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1526 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1527 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1528 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1529 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1530 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1531 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1532 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1533 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1534 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1535 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1536 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1537 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1538 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1539 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001540]
1541
Marat Dukhan1b354632020-03-23 12:50:22 -07001542INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001543 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001544 "src/xnnpack/argmaxpool.h",
1545 "src/xnnpack/avgpool.h",
1546 "src/xnnpack/clamp.h",
1547 "src/xnnpack/common.h",
1548 "src/xnnpack/conv.h",
1549 "src/xnnpack/dwconv.h",
1550 "src/xnnpack/gavgpool.h",
1551 "src/xnnpack/gemm.h",
1552 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001553 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001554 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001555 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001556 "src/xnnpack/lut.h",
1557 "src/xnnpack/math.h",
1558 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001559 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001560 "src/xnnpack/packx.h",
1561 "src/xnnpack/pad.h",
1562 "src/xnnpack/params.h",
1563 "src/xnnpack/pavgpool.h",
1564 "src/xnnpack/ppmm.h",
1565 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001566 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001567 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001568 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001569 "src/xnnpack/rmax.h",
1570 "src/xnnpack/scalar-utils.h",
1571 "src/xnnpack/spmm.h",
1572 "src/xnnpack/unpool.h",
1573 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001574 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001575 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001576 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001577 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001578 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001579 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001580 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001581]
1582
1583INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001584 "include/xnnpack.h",
1585 "src/xnnpack/allocator.h",
1586 "src/xnnpack/compute.h",
1587 "src/xnnpack/im2col.h",
1588 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001589 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001590 "src/xnnpack/operator.h",
1591 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001592 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001593 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001594 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001595 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001596]
1597
Marat Dukhan1b354632020-03-23 12:50:22 -07001598ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001599 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001600]
1601
Marat Dukhan1b354632020-03-23 12:50:22 -07001602MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001603 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604 "include/xnnpack.h",
1605]
1606
Marat Dukhan1b354632020-03-23 12:50:22 -07001607MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001608 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001609 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001610 "src/xnnpack/requantization.h",
1611 "include/xnnpack.h",
1612]
1613
1614OPERATOR_TEST_PARAMS_HDRS = [
1615 "src/xnnpack/params.h",
1616 "src/xnnpack/common.h",
1617]
1618
1619WEIGHTS_PACK_HDRS = [
1620 "src/xnnpack/pack.h",
1621 "src/xnnpack/operator.h",
1622 "src/xnnpack/compute.h",
1623]
1624
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001625LOGGING_COPTS = select({
1626 # No logging in optimized mode
1627 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1628 # Full logging in debug mode
1629 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1630 # Error-only logging in default (fastbuild) mode
1631 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1632})
1633
1634LOGGING_HDRS = [
1635 "src/xnnpack/log.h",
1636]
1637
Marat Dukhan08c4a432019-10-03 09:29:21 -07001638xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001639 name = "tables",
1640 srcs = TABLE_SRCS,
1641 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001642 gcc_copts = xnnpack_gcc_std_copts(),
1643 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001644)
1645
1646xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001647 name = "scalar_ukernels",
1648 srcs = SCALAR_UKERNELS,
1649 hdrs = INTERNAL_HDRS,
1650 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001651 gcc_copts = xnnpack_gcc_std_copts(),
1652 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001653 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001654 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001655 "@FP16",
1656 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001657 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001658 ],
1659)
1660
1661xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001662 name = "wasm_ukernels",
1663 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001664 gcc_copts = xnnpack_gcc_std_copts(),
1665 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001666 wasm_srcs = WASM_UKERNELS,
1667 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001668 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001669 "@FP16",
1670 "@FXdiv",
1671 "@pthreadpool",
1672 ],
1673)
1674
1675xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001676 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001677 hdrs = INTERNAL_HDRS,
1678 aarch32_copts = [
1679 "-marm",
1680 "-mfpu=neon",
1681 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001682 gcc_copts = xnnpack_gcc_std_copts(),
1683 gcc_x86_copts = ["-msse2"],
1684 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001685 optimized_copts = [
1686 "-O3",
1687 "-ffast-math",
1688 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001689 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001690 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001691 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001692 "@FP16",
1693 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001694 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001695 ],
1696)
1697
1698xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001699 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001700 hdrs = INTERNAL_HDRS,
1701 aarch32_copts = [
1702 "-marm",
1703 "-mfpu=neon",
1704 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001705 gcc_copts = xnnpack_gcc_std_copts(),
1706 gcc_x86_copts = ["-msse2"],
1707 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001708 optimized_copts = [
1709 "-O3",
1710 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001711 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001712 deps = [
1713 ":tables",
1714 "@FP16",
1715 "@psimd",
1716 "@pthreadpool",
1717 ],
1718)
1719
1720xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721 name = "neon_ukernels",
1722 hdrs = INTERNAL_HDRS,
1723 aarch32_copts = [
1724 "-marm",
1725 "-mfpu=neon",
1726 ],
1727 aarch32_srcs = NEON_UKERNELS,
1728 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001729 gcc_copts = xnnpack_gcc_std_copts(),
1730 msvc_copts = xnnpack_msvc_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 = "neonfma_ukernels",
1740 hdrs = INTERNAL_HDRS,
1741 aarch32_copts = [
1742 "-marm",
1743 "-mfpu=neon-vfpv4",
1744 ],
1745 aarch32_srcs = NEONFMA_UKERNELS,
1746 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001747 gcc_copts = xnnpack_gcc_std_copts(),
1748 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001749 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001750 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001751 "@FP16",
1752 "@pthreadpool",
1753 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001754)
1755
1756xnnpack_cc_library(
1757 name = "neonfp16arith_ukernels",
1758 hdrs = INTERNAL_HDRS,
1759 aarch64_copts = ["-march=armv8.2-a+fp16"],
1760 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001761 gcc_copts = xnnpack_gcc_std_copts(),
1762 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001763 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001764 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001765 "@FP16",
1766 "@pthreadpool",
1767 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001768)
1769
1770xnnpack_cc_library(
1771 name = "sse2_ukernels",
1772 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001773 gcc_copts = xnnpack_gcc_std_copts(),
1774 gcc_x86_copts = ["-msse2"],
1775 msvc_copts = xnnpack_msvc_std_copts(),
1776 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001778 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001779 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001780 "@FP16",
1781 "@pthreadpool",
1782 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001783)
1784
1785xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001786 name = "ssse3_ukernels",
1787 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001788 gcc_copts = xnnpack_gcc_std_copts(),
1789 gcc_x86_copts = ["-mssse3"],
1790 msvc_copts = xnnpack_msvc_std_copts(),
1791 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001792 x86_srcs = SSSE3_UKERNELS,
1793 deps = [
1794 ":tables",
1795 "@FP16",
1796 "@pthreadpool",
1797 ],
1798)
1799
1800xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001801 name = "sse41_ukernels",
1802 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001803 gcc_copts = xnnpack_gcc_std_copts(),
1804 gcc_x86_copts = ["-msse4.1"],
1805 msvc_copts = xnnpack_msvc_std_copts(),
1806 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001807 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001808 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001809 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001810 "@FP16",
1811 "@pthreadpool",
1812 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001813)
1814
1815xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001816 name = "avx_ukernels",
1817 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001818 gcc_copts = xnnpack_gcc_std_copts(),
1819 gcc_x86_copts = ["-mavx"],
1820 msvc_copts = xnnpack_msvc_std_copts(),
1821 msvc_x86_32_copts = ["/arch:AVX"],
1822 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001823 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001824 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001825 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001826 "@FP16",
1827 "@pthreadpool",
1828 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001829)
1830
1831xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001832 name = "fma3_ukernels",
1833 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001834 gcc_copts = xnnpack_gcc_std_copts(),
1835 gcc_x86_copts = ["-mfma"],
1836 msvc_copts = xnnpack_msvc_std_copts(),
1837 msvc_x86_32_copts = ["/arch:AVX"],
1838 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001839 x86_srcs = FMA3_UKERNELS,
1840 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001841 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001842 "@FP16",
1843 "@pthreadpool",
1844 ],
1845)
1846
1847xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001848 name = "avx2_ukernels",
1849 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001850 gcc_copts = xnnpack_gcc_std_copts(),
1851 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001852 "-mfma",
1853 "-mavx2",
1854 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001855 msvc_copts = xnnpack_msvc_std_copts(),
1856 msvc_x86_32_copts = ["/arch:AVX2"],
1857 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001858 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001859 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001860 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001861 "@FP16",
1862 "@pthreadpool",
1863 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001864)
1865
1866xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867 name = "avx512f_ukernels",
1868 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001869 gcc_copts = xnnpack_gcc_std_copts(),
1870 gcc_x86_copts = ["-mavx512f"],
1871 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1872 msvc_copts = xnnpack_msvc_std_copts(),
1873 msvc_x86_32_copts = ["/arch:AVX512"],
1874 msvc_x86_64_copts = ["/arch:AVX512"],
1875 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001876 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001877 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001878 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001879 "@FP16",
1880 "@pthreadpool",
1881 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001882)
1883
1884xnnpack_cc_library(
1885 name = "asm_ukernels",
1886 hdrs = ["src/xnnpack/assembly.h"],
1887 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001888 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001889 aarch64_srcs = AARCH64_ASM_UKERNELS,
1890)
1891
1892xnnpack_aggregate_library(
1893 name = "ukernels",
1894 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001895 ":neon_ukernels",
1896 ":neonfma_ukernels",
1897 ":asm_ukernels",
1898 ],
1899 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001900 ":neon_ukernels",
1901 ":neonfma_ukernels",
1902 ":neonfp16arith_ukernels",
1903 ":asm_ukernels",
1904 ],
1905 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07001906 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001907 ":psimd_fastmath_ukernels",
1908 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001909 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001910 wasm_deps = [":wasm_ukernels"],
1911 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001912 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001913 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001914 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001915 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001916 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001917 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001918 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001919 ":avx512f_ukernels",
1920 ],
1921)
1922
1923xnnpack_cc_library(
1924 name = "im2col",
1925 srcs = ["src/im2col.c"],
1926 hdrs = [
1927 "src/xnnpack/common.h",
1928 "src/xnnpack/im2col.h",
1929 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001930 gcc_copts = xnnpack_gcc_std_copts(),
1931 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001932)
1933
1934xnnpack_cc_library(
1935 name = "indirection",
1936 srcs = ["src/indirection.c"],
1937 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001938 gcc_copts = xnnpack_gcc_std_copts(),
1939 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001940 deps = [
1941 "@FP16",
1942 "@FXdiv",
1943 "@pthreadpool",
1944 ],
1945)
1946
1947xnnpack_cc_library(
1948 name = "operator_run",
1949 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001950 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001951 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07001952 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1953 "//conditions:default": [],
1954 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07001955 gcc_copts = xnnpack_gcc_std_copts(),
1956 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001957 deps = [
1958 "@FP16",
1959 "@FXdiv",
1960 "@clog",
1961 "@pthreadpool",
1962 ],
1963)
1964
1965cc_library(
1966 name = "enable_assembly",
1967 defines = select({
1968 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1969 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001970 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001971 }),
1972)
1973
Marat Dukhancf056b22019-10-07 10:26:29 -07001974xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001975 name = "operators",
1976 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001977 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001978 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001979 ],
1980 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001981 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001982 "-Isrc",
1983 "-Iinclude",
1984 ] + select({
1985 ":debug_build": [],
1986 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001987 }) + select({
1988 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1989 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001990 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07001991 gcc_copts = xnnpack_gcc_std_copts(),
1992 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhancf056b22019-10-07 10:26:29 -07001993 wasm_srcs = ["src/wasm-stubs.c"],
1994 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001995 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001996 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997 "@FP16",
1998 "@FXdiv",
1999 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002000 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002001 ],
2002)
2003
Marat Dukhan10a38082020-04-17 03:58:35 -07002004xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002005 name = "XNNPACK",
2006 srcs = [
2007 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002008 "src/runtime.c",
2009 "src/subgraph.c",
2010 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002011 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002012 hdrs = ["include/xnnpack.h"],
2013 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002014 "-Isrc",
2015 "-Iinclude",
2016 ] + select({
2017 ":debug_build": [],
2018 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002019 }) + select({
2020 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2021 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002022 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002023 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002024 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002025 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002026 visibility = xnnpack_visibility(),
2027 deps = [
2028 ":enable_assembly",
2029 ":ukernels",
2030 ":operator_run",
2031 ":operators",
2032 "@clog",
2033 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002034 ] + select({
2035 ":emscripten": [],
2036 "//conditions:default": ["@cpuinfo"],
2037 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002038)
2039
Marat Dukhan10a38082020-04-17 03:58:35 -07002040xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002041 name = "xnnpack_operators_nhwc_f32",
2042 srcs = [
2043 "src/init.c",
2044 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002045 hdrs = ["include/xnnpack.h"],
2046 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002047 "-Isrc",
2048 "-Iinclude",
2049 ] + select({
2050 ":debug_build": [],
2051 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002052 }) + select({
2053 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2054 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002055 }),
2056 defines = [
2057 "XNN_NO_Q8_OPERATORS",
2058 "XNN_NO_U8_OPERATORS",
2059 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002060 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002061 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002062 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002063 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002064 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002065 visibility = xnnpack_visibility(),
2066 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002067 ":enable_assembly",
2068 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002069 ":operator_run",
2070 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002071 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002073 ] + select({
2074 ":emscripten": [],
2075 "//conditions:default": ["@cpuinfo"],
2076 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002077)
2078
Marat Dukhancf056b22019-10-07 10:26:29 -07002079xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002080 name = "bench_utils",
2081 srcs = ["bench/utils.cc"],
2082 hdrs = ["bench/utils.h"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002083 gcc_copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002084 deps = [
2085 "@com_google_benchmark//:benchmark",
2086 "@cpuinfo",
2087 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002088)
2089
Frank Barchard7e955972019-10-11 10:34:25 -07002090######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002091
2092xnnpack_benchmark(
2093 name = "q8_gemm_bench",
2094 srcs = [
2095 "bench/gemm.h",
2096 "bench/q8-gemm.cc",
2097 "src/xnnpack/AlignedAllocator.h",
2098 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002099 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
2100 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002101 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2102)
2103
2104xnnpack_benchmark(
2105 name = "f16_gemm_bench",
2106 srcs = [
2107 "bench/f16-gemm.cc",
2108 "bench/gemm.h",
2109 "src/xnnpack/AlignedAllocator.h",
2110 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002111 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002112 deps = MICROKERNEL_BENCHMARK_DEPS,
2113)
2114
2115xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002116 name = "f16_spmm_bench",
2117 srcs = [
2118 "bench/f16-spmm.cc",
2119 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002120 "src/xnnpack/AlignedAllocator.h",
2121 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002122 gcc_copts = ["-Wno-unused-function"],
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002123 deps = MICROKERNEL_BENCHMARK_DEPS,
2124)
2125
2126xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002127 name = "f32_igemm_bench",
2128 srcs = [
2129 "bench/f32-igemm.cc",
2130 "bench/conv.h",
2131 "src/xnnpack/AlignedAllocator.h",
2132 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002133 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002134)
2135
2136xnnpack_benchmark(
2137 name = "f32_conv_hwc_bench",
2138 srcs = [
2139 "bench/f32-conv-hwc.cc",
2140 "bench/dconv.h",
2141 "src/xnnpack/AlignedAllocator.h",
2142 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002143 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002144 deps = MICROKERNEL_BENCHMARK_DEPS,
2145)
2146
2147xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002148 name = "f32_conv_hwc2spchw_bench",
2149 srcs = [
2150 "bench/f32-conv-hwc2spchw.cc",
2151 "bench/dconv.h",
2152 "src/xnnpack/AlignedAllocator.h",
2153 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002154 gcc_copts = ["-Wno-unused-function"],
Erich Elsen563df5f2019-10-23 08:02:21 -07002155 deps = MICROKERNEL_BENCHMARK_DEPS,
2156)
2157
2158xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002159 name = "f32_dwconv_bench",
2160 srcs = [
2161 "bench/f32-dwconv.cc",
2162 "bench/dwconv.h",
2163 "src/xnnpack/AlignedAllocator.h",
2164 ] + MICROKERNEL_BENCHMARK_HDRS,
2165 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2166)
2167
2168xnnpack_benchmark(
2169 name = "f32_dwconv_spchw_bench",
2170 srcs = [
2171 "bench/f32-dwconv-spchw.cc",
2172 "bench/dwconv.h",
2173 "src/xnnpack/AlignedAllocator.h",
2174 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2175 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2176)
2177
2178xnnpack_benchmark(
2179 name = "f32_gemm_bench",
2180 srcs = [
2181 "bench/f32-gemm.cc",
2182 "bench/gemm.h",
2183 "src/xnnpack/AlignedAllocator.h",
2184 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002185 copts = xnnpack_optional_ruy_copts(),
2186 gcc_copts = ["-Wno-unused-function"],
Frank Barchard7e955972019-10-11 10:34:25 -07002187 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002188)
2189
2190xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002191 name = "f32_raddexpminusmax_bench",
2192 srcs = [
2193 "bench/f32-raddexpminusmax.cc",
2194 "src/xnnpack/AlignedAllocator.h",
2195 ] + MICROKERNEL_BENCHMARK_HDRS,
2196 deps = MICROKERNEL_BENCHMARK_DEPS,
2197)
2198
2199xnnpack_benchmark(
2200 name = "f32_raddextexp_bench",
2201 srcs = [
2202 "bench/f32-raddextexp.cc",
2203 "src/xnnpack/AlignedAllocator.h",
2204 ] + MICROKERNEL_BENCHMARK_HDRS,
2205 deps = MICROKERNEL_BENCHMARK_DEPS,
2206)
2207
2208xnnpack_benchmark(
2209 name = "f32_raddstoreexpminusmax_bench",
2210 srcs = [
2211 "bench/f32-raddstoreexpminusmax.cc",
2212 "src/xnnpack/AlignedAllocator.h",
2213 ] + MICROKERNEL_BENCHMARK_HDRS,
2214 deps = MICROKERNEL_BENCHMARK_DEPS,
2215)
2216
2217xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002218 name = "f32_rmax_bench",
2219 srcs = [
2220 "bench/f32-rmax.cc",
2221 "src/xnnpack/AlignedAllocator.h",
2222 ] + MICROKERNEL_BENCHMARK_HDRS,
2223 deps = MICROKERNEL_BENCHMARK_DEPS,
2224)
2225
2226xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002227 name = "f32_sigmoid_bench",
2228 srcs = [
2229 "bench/f32-sigmoid.cc",
2230 "src/xnnpack/AlignedAllocator.h",
2231 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002232 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan14bec502019-11-18 11:35:31 -08002233 deps = MICROKERNEL_BENCHMARK_DEPS,
2234)
2235
2236xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002237 name = "f32_spmm_bench",
2238 srcs = [
2239 "bench/f32-spmm.cc",
2240 "bench/gemm.h",
2241 "src/xnnpack/AlignedAllocator.h",
2242 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002243 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244 deps = MICROKERNEL_BENCHMARK_DEPS,
2245)
2246
2247xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002248 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002249 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002250 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002251 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002252 copts = xnnpack_optional_dnnl_copts(),
2253 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002254 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002255)
2256
2257xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002258 name = "f32_vscaleexpminusmax_bench",
2259 srcs = [
2260 "bench/f32-vscaleexpminusmax.cc",
2261 "src/xnnpack/AlignedAllocator.h",
2262 ] + MICROKERNEL_BENCHMARK_HDRS,
2263 deps = MICROKERNEL_BENCHMARK_DEPS,
2264)
2265
2266xnnpack_benchmark(
2267 name = "f32_vscaleextexp_bench",
2268 srcs = [
2269 "bench/f32-vscaleextexp.cc",
2270 "src/xnnpack/AlignedAllocator.h",
2271 ] + MICROKERNEL_BENCHMARK_HDRS,
2272 deps = MICROKERNEL_BENCHMARK_DEPS,
2273)
2274
2275xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002276 name = "f32_im2col_gemm_bench",
2277 srcs = [
2278 "bench/f32-im2col-gemm.cc",
2279 "bench/conv.h",
2280 "src/xnnpack/AlignedAllocator.h",
2281 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2282 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2283)
2284
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002285xnnpack_benchmark(
2286 name = "requantization_bench",
2287 srcs = [
2288 "bench/requantization.cc",
2289 "src/xnnpack/requantization-stubs.h",
2290 "src/xnnpack/AlignedAllocator.h",
2291 ] + MICROKERNEL_BENCHMARK_HDRS,
2292 deps = MICROKERNEL_BENCHMARK_DEPS,
2293)
2294
Marat Dukhan08c4a432019-10-03 09:29:21 -07002295########################### Benchmarks for operators ###########################
2296
2297xnnpack_benchmark(
2298 name = "add_bench",
2299 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002300 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002301)
2302
2303xnnpack_benchmark(
2304 name = "average_pooling_bench",
2305 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002306 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002307 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002308)
2309
2310xnnpack_benchmark(
2311 name = "channel_shuffle_bench",
2312 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002313 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002314)
2315
2316xnnpack_benchmark(
2317 name = "convolution_bench",
2318 srcs = ["bench/convolution.cc"],
2319 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002320 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002321)
2322
2323xnnpack_benchmark(
2324 name = "deconvolution_bench",
2325 srcs = ["bench/deconvolution.cc"],
2326 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002327 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002328)
2329
2330xnnpack_benchmark(
2331 name = "global_average_pooling_bench",
2332 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002333 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002334)
2335
2336xnnpack_benchmark(
2337 name = "max_pooling_bench",
2338 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002339 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340)
2341
2342xnnpack_benchmark(
2343 name = "sigmoid_bench",
2344 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002345 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002346 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002347)
2348
2349xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002350 name = "prelu_bench",
2351 srcs = ["bench/prelu.cc"],
2352 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002353 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002354)
2355
2356xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002357 name = "softmax_bench",
2358 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002359 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002360 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002361)
2362
Marat Dukhanc068bb62019-10-04 13:24:39 -07002363############################# End-to-end benchmarks ############################
2364
2365cc_library(
2366 name = "mobilenet_v1",
2367 srcs = ["models/mobilenet-v1.cc"],
2368 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002369 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002370 linkstatic = True,
2371 deps = [
2372 ":XNNPACK",
2373 "@pthreadpool",
2374 ],
2375)
2376
2377cc_library(
2378 name = "mobilenet_v2",
2379 srcs = ["models/mobilenet-v2.cc"],
2380 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002381 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002382 linkstatic = True,
2383 deps = [
2384 ":XNNPACK",
2385 "@pthreadpool",
2386 ],
2387)
2388
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002389cc_library(
2390 name = "mobilenet_v3_large",
2391 srcs = ["models/mobilenet-v3-large.cc"],
2392 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002393 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002394 linkstatic = True,
2395 deps = [
2396 ":XNNPACK",
2397 "@pthreadpool",
2398 ],
2399)
2400
2401cc_library(
2402 name = "mobilenet_v3_small",
2403 srcs = ["models/mobilenet-v3-small.cc"],
2404 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002405 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002406 linkstatic = True,
2407 deps = [
2408 ":XNNPACK",
2409 "@pthreadpool",
2410 ],
2411)
2412
Marat Dukhanc068bb62019-10-04 13:24:39 -07002413xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002414 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002415 srcs = [
2416 "bench/f32-dwconv-e2e.cc",
2417 "bench/end2end.h",
2418 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002419 gcc_copts = ["-Wno-unused-function"],
Marat Dukhanef4416e2019-10-31 13:44:40 -07002420 deps = MICROKERNEL_BENCHMARK_DEPS + [
2421 ":XNNPACK",
2422 ":mobilenet_v1",
2423 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002424 ":mobilenet_v3_large",
2425 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002426 ],
2427)
2428
2429xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002430 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002431 srcs = [
2432 "bench/f32-gemm-e2e.cc",
2433 "bench/end2end.h",
2434 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002435 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan5f18d262019-10-31 10:24:14 -07002436 deps = MICROKERNEL_BENCHMARK_DEPS + [
2437 ":XNNPACK",
2438 ":mobilenet_v1",
2439 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002440 ":mobilenet_v3_large",
2441 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002442 ],
2443)
2444
2445xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002446 name = "end2end_bench",
2447 srcs = ["bench/end2end.cc"],
2448 deps = [
2449 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002450 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002451 ":mobilenet_v1",
2452 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002453 ":mobilenet_v3_large",
2454 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002455 "@pthreadpool",
2456 ],
2457)
2458
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002459#################### Accuracy evaluation for math functions ####################
2460
2461xnnpack_benchmark(
2462 name = "f32_exp_eval",
2463 srcs = [
2464 "eval/f32-exp.cc",
2465 "src/xnnpack/AlignedAllocator.h",
2466 ] + ACCURACY_EVAL_HDRS,
2467 deps = ACCURACY_EVAL_DEPS,
2468)
2469
Marat Dukhan515c9772019-10-17 18:07:57 -07002470xnnpack_benchmark(
2471 name = "f32_expminus_eval",
2472 srcs = [
2473 "eval/f32-expminus.cc",
2474 "src/xnnpack/AlignedAllocator.h",
2475 ] + ACCURACY_EVAL_HDRS,
2476 deps = ACCURACY_EVAL_DEPS,
2477)
2478
Marat Dukhan98ba4412019-10-23 02:14:28 -07002479xnnpack_benchmark(
2480 name = "f32_extexp_eval",
2481 srcs = [
2482 "eval/f32-extexp.cc",
2483 "src/xnnpack/AlignedAllocator.h",
2484 ] + ACCURACY_EVAL_HDRS,
2485 deps = ACCURACY_EVAL_DEPS,
2486)
2487
Marat Dukhan346a9e52019-11-15 09:06:30 -08002488xnnpack_benchmark(
2489 name = "f32_sigmoid_eval",
2490 srcs = [
2491 "eval/f32-sigmoid.cc",
2492 "src/xnnpack/AlignedAllocator.h",
2493 ] + ACCURACY_EVAL_HDRS,
2494 deps = ACCURACY_EVAL_DEPS,
2495)
2496
Marat Dukhan08c4a432019-10-03 09:29:21 -07002497######################### Unit tests for micro-kernels #########################
2498
2499xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002500 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002501 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002502 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002503 "test/gemm-microkernel-tester.h",
2504 "src/xnnpack/AlignedAllocator.h",
2505 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2506 deps = MICROKERNEL_TEST_DEPS,
2507)
2508
2509xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002510 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002511 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002512 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002513 "test/spmm-microkernel-tester.h",
2514 "src/xnnpack/AlignedAllocator.h",
2515 ] + MICROKERNEL_TEST_HDRS,
2516 deps = MICROKERNEL_TEST_DEPS,
2517)
2518
2519xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002520 name = "f32_argmaxpool_test",
2521 srcs = [
2522 "test/f32-argmaxpool.cc",
2523 "test/argmaxpool-microkernel-tester.h",
2524 "src/xnnpack/AlignedAllocator.h",
2525 ] + MICROKERNEL_TEST_HDRS,
2526 deps = MICROKERNEL_TEST_DEPS,
2527)
2528
2529xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002530 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002531 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002532 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002533 "test/avgpool-microkernel-tester.h",
2534 "src/xnnpack/AlignedAllocator.h",
2535 ] + MICROKERNEL_TEST_HDRS,
2536 deps = MICROKERNEL_TEST_DEPS,
2537)
2538
2539xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002540 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002541 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002542 "test/f32-ibilinear.cc",
2543 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002544 "src/xnnpack/AlignedAllocator.h",
2545 ] + MICROKERNEL_TEST_HDRS,
2546 deps = MICROKERNEL_TEST_DEPS,
2547)
2548
2549xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002550 name = "f32_clamp_test",
2551 srcs = [
2552 "test/f32-clamp.cc",
2553 "test/clamp-microkernel-tester.h",
2554 ] + MICROKERNEL_TEST_HDRS,
2555 deps = MICROKERNEL_TEST_DEPS,
2556)
2557
2558xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002559 name = "f32_igemm_test",
2560 srcs = [
2561 "test/f32-igemm.cc",
2562 "test/gemm-microkernel-tester.h",
2563 "src/xnnpack/AlignedAllocator.h",
2564 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2565 deps = MICROKERNEL_TEST_DEPS,
2566)
2567
2568xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002569 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002570 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002571 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002572 "test/gemm-microkernel-tester.h",
2573 "src/xnnpack/AlignedAllocator.h",
2574 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002575 # Work-around for "too many sections" error
2576 mingw_copts = ["-Wa,-mbig-obj"],
2577 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002578 deps = MICROKERNEL_TEST_DEPS,
2579)
2580
2581xnnpack_unit_test(
2582 name = "f32_conv_hwc_test",
2583 srcs = [
2584 "test/f32-conv-hwc.cc",
2585 "test/conv-hwc-microkernel-tester.h",
2586 "src/xnnpack/AlignedAllocator.h",
2587 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2588 deps = MICROKERNEL_TEST_DEPS,
2589)
2590
2591xnnpack_unit_test(
2592 name = "f32_conv_hwc2spchw_test",
2593 srcs = [
2594 "test/f32-conv-hwc2spchw.cc",
2595 "test/conv-hwc2spchw-microkernel-tester.h",
2596 "src/xnnpack/AlignedAllocator.h",
2597 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2598 deps = MICROKERNEL_TEST_DEPS,
2599)
2600
2601xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002602 name = "f32_dwconv_test",
2603 srcs = [
2604 "test/f32-dwconv.cc",
2605 "test/dwconv-microkernel-tester.h",
2606 "src/xnnpack/AlignedAllocator.h",
2607 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2608 deps = MICROKERNEL_TEST_DEPS,
2609)
2610
2611xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002612 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002613 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002614 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002615 "test/dwconv-microkernel-tester.h",
2616 "src/xnnpack/AlignedAllocator.h",
2617 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2618 deps = MICROKERNEL_TEST_DEPS,
2619)
2620
2621xnnpack_unit_test(
2622 name = "f32_dwconv_spchw_test",
2623 srcs = [
2624 "test/f32-dwconv-spchw.cc",
2625 "test/dwconv-spchw-microkernel-tester.h",
2626 "src/xnnpack/AlignedAllocator.h",
2627 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2628 deps = MICROKERNEL_TEST_DEPS,
2629)
2630
2631xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002632 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002633 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002634 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002635 "test/gavgpool-microkernel-tester.h",
2636 "src/xnnpack/AlignedAllocator.h",
2637 ] + MICROKERNEL_TEST_HDRS,
2638 deps = MICROKERNEL_TEST_DEPS,
2639)
2640
2641xnnpack_unit_test(
2642 name = "f32_gavgpool_spchw_test",
2643 srcs = [
2644 "test/f32-gavgpool-spchw.cc",
2645 "test/gavgpool-spchw-microkernel-tester.h",
2646 "src/xnnpack/AlignedAllocator.h",
2647 ] + MICROKERNEL_TEST_HDRS,
2648 deps = MICROKERNEL_TEST_DEPS,
2649)
2650
2651xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002652 name = "f32_gemm_test",
2653 srcs = [
2654 "test/f32-gemm.cc",
2655 "test/gemm-microkernel-tester.h",
2656 "src/xnnpack/AlignedAllocator.h",
2657 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2658 deps = MICROKERNEL_TEST_DEPS,
2659)
2660
2661xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002662 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002663 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002664 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002665 "test/gemm-microkernel-tester.h",
2666 "src/xnnpack/AlignedAllocator.h",
2667 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2668 deps = MICROKERNEL_TEST_DEPS,
2669)
2670
2671xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002672 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002673 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002674 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002675 "test/gemm-microkernel-tester.h",
2676 "src/xnnpack/AlignedAllocator.h",
2677 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2678 deps = MICROKERNEL_TEST_DEPS,
2679)
2680
2681xnnpack_unit_test(
2682 name = "f32_hswish_test",
2683 srcs = [
2684 "test/f32-hswish.cc",
2685 "test/hswish-microkernel-tester.h",
2686 ] + MICROKERNEL_TEST_HDRS,
2687 deps = MICROKERNEL_TEST_DEPS,
2688)
2689
2690xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002691 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002692 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002693 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002694 "test/maxpool-microkernel-tester.h",
2695 ] + MICROKERNEL_TEST_HDRS,
2696 deps = MICROKERNEL_TEST_DEPS,
2697)
2698
2699xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002700 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002701 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002702 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002703 "test/avgpool-microkernel-tester.h",
2704 "src/xnnpack/AlignedAllocator.h",
2705 ] + MICROKERNEL_TEST_HDRS,
2706 deps = MICROKERNEL_TEST_DEPS,
2707)
2708
2709xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002710 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002711 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002712 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002713 "test/gemm-microkernel-tester.h",
2714 "src/xnnpack/AlignedAllocator.h",
2715 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2716 deps = MICROKERNEL_TEST_DEPS,
2717)
2718
2719xnnpack_unit_test(
2720 name = "f32_prelu_test",
2721 srcs = [
2722 "test/f32-prelu.cc",
2723 "test/prelu-microkernel-tester.h",
2724 "src/xnnpack/AlignedAllocator.h",
2725 ] + MICROKERNEL_TEST_HDRS,
2726 deps = MICROKERNEL_TEST_DEPS,
2727)
2728
2729xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002730 name = "f32_raddexpminusmax_test",
2731 srcs = [
2732 "test/f32-raddexpminusmax.cc",
2733 "test/raddexpminusmax-microkernel-tester.h",
2734 ] + MICROKERNEL_TEST_HDRS,
2735 deps = MICROKERNEL_TEST_DEPS,
2736)
2737
2738xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002739 name = "f32_raddextexp_test",
2740 srcs = [
2741 "test/f32-raddextexp.cc",
2742 "test/raddextexp-microkernel-tester.h",
2743 ] + MICROKERNEL_TEST_HDRS,
2744 deps = MICROKERNEL_TEST_DEPS,
2745)
2746
2747xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002748 name = "f32_raddstoreexpminusmax_test",
2749 srcs = [
2750 "test/f32-raddstoreexpminusmax.cc",
2751 "test/raddstoreexpminusmax-microkernel-tester.h",
2752 ] + MICROKERNEL_TEST_HDRS,
2753 deps = MICROKERNEL_TEST_DEPS,
2754)
2755
2756xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002757 name = "f32_rmax_test",
2758 srcs = [
2759 "test/f32-rmax.cc",
2760 "test/rmax-microkernel-tester.h",
2761 ] + MICROKERNEL_TEST_HDRS,
2762 deps = MICROKERNEL_TEST_DEPS,
2763)
2764
2765xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002766 name = "f32_sigmoid_test",
2767 srcs = [
2768 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002769 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002770 ] + MICROKERNEL_TEST_HDRS,
2771 deps = MICROKERNEL_TEST_DEPS,
2772)
2773
2774xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002775 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002776 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002777 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002778 "test/spmm-microkernel-tester.h",
2779 "src/xnnpack/AlignedAllocator.h",
2780 ] + MICROKERNEL_TEST_HDRS,
2781 deps = MICROKERNEL_TEST_DEPS,
2782)
2783
2784xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002785 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002786 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002787 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002788 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002789 ] + MICROKERNEL_TEST_HDRS,
2790 deps = MICROKERNEL_TEST_DEPS,
2791)
2792
2793xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002794 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002795 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002796 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002797 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002798 ] + MICROKERNEL_TEST_HDRS,
2799 deps = MICROKERNEL_TEST_DEPS,
2800)
2801
2802xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002803 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002804 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002805 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002806 "test/vbinary-microkernel-tester.h",
2807 ] + MICROKERNEL_TEST_HDRS,
2808 deps = MICROKERNEL_TEST_DEPS,
2809)
2810
2811xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002812 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002813 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002814 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002815 "test/vbinaryc-microkernel-tester.h",
2816 ] + MICROKERNEL_TEST_HDRS,
2817 deps = MICROKERNEL_TEST_DEPS,
2818)
2819
2820xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002821 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002822 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002823 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002824 "test/vbinaryc-microkernel-tester.h",
2825 ] + MICROKERNEL_TEST_HDRS,
2826 deps = MICROKERNEL_TEST_DEPS,
2827)
2828
2829xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002830 name = "f32_vmax_test",
2831 srcs = [
2832 "test/f32-vmax.cc",
2833 "test/vbinary-microkernel-tester.h",
2834 ] + MICROKERNEL_TEST_HDRS,
2835 deps = MICROKERNEL_TEST_DEPS,
2836)
2837
2838xnnpack_unit_test(
2839 name = "f32_vmaxc_test",
2840 srcs = [
2841 "test/f32-vmaxc.cc",
2842 "test/vbinaryc-microkernel-tester.h",
2843 ] + MICROKERNEL_TEST_HDRS,
2844 deps = MICROKERNEL_TEST_DEPS,
2845)
2846
2847xnnpack_unit_test(
2848 name = "f32_vmin_test",
2849 srcs = [
2850 "test/f32-vmin.cc",
2851 "test/vbinary-microkernel-tester.h",
2852 ] + MICROKERNEL_TEST_HDRS,
2853 deps = MICROKERNEL_TEST_DEPS,
2854)
2855
2856xnnpack_unit_test(
2857 name = "f32_vminc_test",
2858 srcs = [
2859 "test/f32-vminc.cc",
2860 "test/vbinaryc-microkernel-tester.h",
2861 ] + MICROKERNEL_TEST_HDRS,
2862 deps = MICROKERNEL_TEST_DEPS,
2863)
2864
2865xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002866 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002868 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002869 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002870 ] + MICROKERNEL_TEST_HDRS,
2871 deps = MICROKERNEL_TEST_DEPS,
2872)
2873
2874xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002875 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002876 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002877 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002878 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002879 ] + MICROKERNEL_TEST_HDRS,
2880 deps = MICROKERNEL_TEST_DEPS,
2881)
2882
2883xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002884 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002886 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002887 "test/vmulcaddc-microkernel-tester.h",
2888 "src/xnnpack/AlignedAllocator.h",
2889 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2890 deps = MICROKERNEL_TEST_DEPS,
2891)
2892
2893xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002894 name = "f32_vscale_test",
2895 srcs = [
2896 "test/f32-vscale.cc",
2897 "test/vscale-microkernel-tester.h",
2898 ] + MICROKERNEL_TEST_HDRS,
2899 deps = MICROKERNEL_TEST_DEPS,
2900)
2901
2902xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002903 name = "f32_vscaleexpminusmax_test",
2904 srcs = [
2905 "test/f32-vscaleexpminusmax.cc",
2906 "test/vscaleexpminusmax-microkernel-tester.h",
2907 ] + MICROKERNEL_TEST_HDRS,
2908 deps = MICROKERNEL_TEST_DEPS,
2909)
2910
2911xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002912 name = "f32_vscaleextexp_test",
2913 srcs = [
2914 "test/f32-vscaleextexp.cc",
2915 "test/vscaleextexp-microkernel-tester.h",
2916 ] + MICROKERNEL_TEST_HDRS,
2917 deps = MICROKERNEL_TEST_DEPS,
2918)
2919
2920xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002921 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07002922 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002923 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002924 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002925 ] + MICROKERNEL_TEST_HDRS,
2926 deps = MICROKERNEL_TEST_DEPS,
2927)
2928
2929xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002930 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002931 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002932 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002933 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002934 ] + MICROKERNEL_TEST_HDRS,
2935 deps = MICROKERNEL_TEST_DEPS,
2936)
2937
2938xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002939 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002940 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002941 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002942 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002943 ] + MICROKERNEL_TEST_HDRS,
2944 deps = MICROKERNEL_TEST_DEPS,
2945)
2946
2947xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002948 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002949 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002950 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002951 "test/avgpool-microkernel-tester.h",
2952 "src/xnnpack/AlignedAllocator.h",
2953 ] + MICROKERNEL_TEST_HDRS,
2954 deps = MICROKERNEL_TEST_DEPS,
2955)
2956
2957xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002958 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002959 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002960 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 "test/gemm-microkernel-tester.h",
2962 "src/xnnpack/AlignedAllocator.h",
2963 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2964 deps = MICROKERNEL_TEST_DEPS,
2965)
2966
2967xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002968 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002969 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002970 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002971 "test/dwconv-microkernel-tester.h",
2972 "src/xnnpack/AlignedAllocator.h",
2973 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2974 deps = MICROKERNEL_TEST_DEPS,
2975)
2976
2977xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002978 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002980 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002981 "test/gavgpool-microkernel-tester.h",
2982 "src/xnnpack/AlignedAllocator.h",
2983 ] + MICROKERNEL_TEST_HDRS,
2984 deps = MICROKERNEL_TEST_DEPS,
2985)
2986
2987xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002988 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002989 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002990 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 "test/gemm-microkernel-tester.h",
2992 "src/xnnpack/AlignedAllocator.h",
2993 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2994 deps = MICROKERNEL_TEST_DEPS,
2995)
2996
2997xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002998 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003000 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003001 "test/vadd-microkernel-tester.h",
3002 ] + MICROKERNEL_TEST_HDRS,
3003 deps = MICROKERNEL_TEST_DEPS,
3004)
3005
3006xnnpack_unit_test(
3007 name = "u8_clamp_test",
3008 srcs = [
3009 "test/u8-clamp.cc",
3010 "test/clamp-microkernel-tester.h",
3011 ] + MICROKERNEL_TEST_HDRS,
3012 deps = MICROKERNEL_TEST_DEPS,
3013)
3014
3015xnnpack_unit_test(
3016 name = "u8_lut32norm_test",
3017 srcs = [
3018 "test/u8-lut32norm.cc",
3019 "test/lut-norm-microkernel-tester.h",
3020 ] + MICROKERNEL_TEST_HDRS,
3021 deps = MICROKERNEL_TEST_DEPS,
3022)
3023
3024xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003025 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003026 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003027 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003028 "test/maxpool-microkernel-tester.h",
3029 ] + MICROKERNEL_TEST_HDRS,
3030 deps = MICROKERNEL_TEST_DEPS,
3031)
3032
3033xnnpack_unit_test(
3034 name = "u8_rmax_test",
3035 srcs = [
3036 "test/u8-rmax.cc",
3037 "test/rmax-microkernel-tester.h",
3038 ] + MICROKERNEL_TEST_HDRS,
3039 deps = MICROKERNEL_TEST_DEPS,
3040)
3041
3042xnnpack_unit_test(
3043 name = "x32_packx_test",
3044 srcs = [
3045 "test/x32-packx.cc",
3046 "test/pack-microkernel-tester.h",
3047 "src/xnnpack/AlignedAllocator.h",
3048 ] + MICROKERNEL_TEST_HDRS,
3049 deps = MICROKERNEL_TEST_DEPS,
3050)
3051
3052xnnpack_unit_test(
3053 name = "x32_pad_test",
3054 srcs = [
3055 "test/x32-pad.cc",
3056 "test/pad-microkernel-tester.h",
3057 ] + MICROKERNEL_TEST_HDRS,
3058 deps = MICROKERNEL_TEST_DEPS,
3059)
3060
3061xnnpack_unit_test(
3062 name = "x32_unpool_test",
3063 srcs = [
3064 "test/x32-unpool.cc",
3065 "test/unpool-microkernel-tester.h",
3066 ] + MICROKERNEL_TEST_HDRS,
3067 deps = MICROKERNEL_TEST_DEPS,
3068)
3069
3070xnnpack_unit_test(
3071 name = "x32_zip_test",
3072 srcs = [
3073 "test/x32-zip.cc",
3074 "test/zip-microkernel-tester.h",
3075 ] + MICROKERNEL_TEST_HDRS,
3076 deps = MICROKERNEL_TEST_DEPS,
3077)
3078
3079xnnpack_unit_test(
3080 name = "x8_lut_test",
3081 srcs = [
3082 "test/x8-lut.cc",
3083 "test/lut-microkernel-tester.h",
3084 ] + MICROKERNEL_TEST_HDRS,
3085 deps = MICROKERNEL_TEST_DEPS,
3086)
3087
3088xnnpack_unit_test(
3089 name = "x8_zip_test",
3090 srcs = [
3091 "test/x8-zip.cc",
3092 "test/zip-microkernel-tester.h",
3093 ] + MICROKERNEL_TEST_HDRS,
3094 deps = MICROKERNEL_TEST_DEPS,
3095)
3096
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003097xnnpack_unit_test(
3098 name = "requantization_test",
3099 srcs = [
3100 "src/xnnpack/requantization-stubs.h",
3101 "test/requantization.cc",
3102 "test/requantization-tester.h",
3103 ] + MICROKERNEL_TEST_HDRS,
3104 deps = MICROKERNEL_TEST_DEPS,
3105)
3106
Marat Dukhan20c3b922020-03-10 03:45:06 -07003107########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003108
3109xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003110 name = "operator_size_test",
3111 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003112 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003113)
3114
Marat Dukhan20c3b922020-03-10 03:45:06 -07003115xnnpack_binary(
3116 name = "subgraph_size_test",
3117 srcs = ["test/subgraph-size.c"],
3118 deps = [":XNNPACK"],
3119)
3120
3121########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122
3123xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003124 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003125 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003126 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003127 "test/add-operator-tester.h",
3128 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003129 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130)
3131
3132xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003133 name = "add_nd_test",
3134 srcs = [
3135 "test/add-nd.cc",
3136 "test/binary-elementwise-operator-tester.h",
3137 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003138 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003139)
3140
3141xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003142 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003143 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003144 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003145 "test/argmax-pooling-operator-tester.h",
3146 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003147 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003148)
3149
3150xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003151 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003152 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003153 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003154 "test/average-pooling-operator-tester.h",
3155 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003156 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003157)
3158
3159xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003160 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003162 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003163 "test/channel-pad-operator-tester.h",
3164 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003165 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166)
3167
3168xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003169 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003171 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003172 "test/channel-shuffle-operator-tester.h",
3173 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003174 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003175)
3176
3177xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003178 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003179 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003180 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181 "test/clamp-operator-tester.h",
3182 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003183 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184)
3185
3186xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003187 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003188 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003189 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190 "test/convolution-operator-tester.h",
3191 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003192 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193)
3194
3195xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003196 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003197 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003198 "test/convolution-nchw.cc",
3199 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003200 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003201 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003202)
3203
3204xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003205 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003206 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003207 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003208 "test/deconvolution-operator-tester.h",
3209 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003210 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003211)
3212
3213xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003214 name = "divide_nd_test",
3215 srcs = [
3216 "test/binary-elementwise-operator-tester.h",
3217 "test/divide-nd.cc",
3218 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003219 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003220)
3221
3222xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003223 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003224 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003225 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003226 "test/fully-connected-operator-tester.h",
3227 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003228 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003229)
3230
3231xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003232 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003233 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003234 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003235 "test/global-average-pooling-operator-tester.h",
3236 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003237 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003238)
3239
3240xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003241 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003242 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003243 "test/global-average-pooling-ncw.cc",
3244 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003246 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003247)
3248
3249xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003250 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003251 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003252 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 "test/hardswish-operator-tester.h",
3254 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003255 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003256)
3257
3258xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003259 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003260 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003261 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003262 "test/leaky-relu-operator-tester.h",
3263 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003264 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265)
3266
3267xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003268 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003269 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003270 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271 "test/max-pooling-operator-tester.h",
3272 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003273 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003274)
3275
3276xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003277 name = "maximum_nd_test",
3278 srcs = [
3279 "test/binary-elementwise-operator-tester.h",
3280 "test/maximum-nd.cc",
3281 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003282 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003283)
3284
3285xnnpack_unit_test(
3286 name = "minimum_nd_test",
3287 srcs = [
3288 "test/binary-elementwise-operator-tester.h",
3289 "test/minimum-nd.cc",
3290 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003291 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003292)
3293
3294xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003295 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003296 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003297 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003298 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003299 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003300 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003301)
3302
3303xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003304 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003305 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003306 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003307 "test/prelu-operator-tester.h",
3308 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003309 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003310)
3311
3312xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003313 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003314 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003315 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003316 "test/resize-bilinear-operator-tester.h",
3317 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003318 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003319)
3320
3321xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003322 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003323 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003324 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003325 "test/sigmoid-operator-tester.h",
3326 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003327 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003328)
3329
3330xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003331 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003332 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003333 "test/softmax-nc.cc",
3334 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003335 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003336 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003337)
3338
3339xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003340 name = "subtract_nd_test",
3341 srcs = [
3342 "test/binary-elementwise-operator-tester.h",
3343 "test/subtract-nd.cc",
3344 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003345 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003346)
3347
3348xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003349 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003351 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003352 "test/unpooling-operator-tester.h",
3353 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003354 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003355)
3356
3357############################# Build configurations #############################
3358
Marat Dukhanb8642352019-10-30 15:43:02 -07003359# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003360config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003361 name = "xnn_enable_assembly_explicit_true",
3362 define_values = {"xnn_enable_assembly": "true"},
3363)
3364
3365# Disables usage of assembly kernels.
3366config_setting(
3367 name = "xnn_enable_assembly_explicit_false",
3368 define_values = {"xnn_enable_assembly": "false"},
3369)
3370
Marat Dukhan05702cf2020-03-26 15:41:33 -07003371# Disables usage of HMP-aware optimizations.
3372config_setting(
3373 name = "xnn_enable_hmp_explicit_false",
3374 define_values = {"xnn_enable_hmp": "false"},
3375)
3376
Marat Dukhanb8642352019-10-30 15:43:02 -07003377# Builds with -c dbg
3378config_setting(
3379 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003380 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003381 "compilation_mode": "dbg",
3382 },
3383)
3384
3385# Builds with -c opt
3386config_setting(
3387 name = "optimized_build",
3388 values = {
3389 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003390 },
3391)
3392
3393config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003394 name = "linux_k8",
3395 values = {"cpu": "k8"},
3396)
3397
3398config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003399 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003400 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003401)
3402
3403config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003404 name = "linux_armhf",
3405 values = {"cpu": "armhf"},
3406)
3407
3408config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003409 name = "android",
3410 values = {"crosstool_top": "//external:android/crosstool"},
3411)
3412
3413config_setting(
3414 name = "android_armv7",
3415 values = {
3416 "crosstool_top": "//external:android/crosstool",
3417 "cpu": "armeabi-v7a",
3418 },
3419)
3420
3421config_setting(
3422 name = "android_arm64",
3423 values = {
3424 "crosstool_top": "//external:android/crosstool",
3425 "cpu": "arm64-v8a",
3426 },
3427)
3428
3429config_setting(
3430 name = "android_x86",
3431 values = {
3432 "crosstool_top": "//external:android/crosstool",
3433 "cpu": "x86",
3434 },
3435)
3436
3437config_setting(
3438 name = "android_x86_64",
3439 values = {
3440 "crosstool_top": "//external:android/crosstool",
3441 "cpu": "x86_64",
3442 },
3443)
3444
3445config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003446 name = "windows_x86_64",
3447 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003448)
3449
3450config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003451 name = "windows_x86_64_clang",
3452 values = {
3453 "compiler": "clang-cl",
3454 "cpu": "x64_windows",
3455 },
3456)
3457
3458config_setting(
3459 name = "windows_x86_64_mingw",
3460 values = {
3461 "compiler": "mingw-gcc",
3462 "cpu": "x64_windows",
3463 },
3464)
3465
3466config_setting(
3467 name = "windows_x86_64_msys",
3468 values = {
3469 "compiler": "msys-gcc",
3470 "cpu": "x64_windows",
3471 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003472)
3473
3474config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003475 name = "macos_x86_64",
3476 values = {
3477 "apple_platform_type": "macos",
3478 "cpu": "darwin",
3479 },
3480)
3481
3482config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003483 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003484 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003485)
3486
3487config_setting(
3488 name = "emscripten_wasm",
3489 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003490 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003491 "cpu": "wasm",
3492 },
3493)
3494
3495config_setting(
3496 name = "emscripten_wasmsimd",
3497 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003498 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003499 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003500 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003501 },
3502)
3503
3504config_setting(
3505 name = "emscripten_asmjs",
3506 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003507 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 "cpu": "asmjs",
3509 },
3510)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003511
3512config_setting(
3513 name = "ios_armv7",
3514 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003515 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003516 "cpu": "ios_armv7",
3517 },
3518)
3519
3520config_setting(
3521 name = "ios_arm64",
3522 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003523 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003524 "cpu": "ios_arm64",
3525 },
3526)
3527
3528config_setting(
3529 name = "ios_arm64e",
3530 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003531 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003532 "cpu": "ios_arm64e",
3533 },
3534)
3535
3536config_setting(
3537 name = "ios_x86",
3538 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003539 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003540 "cpu": "ios_i386",
3541 },
3542)
3543
3544config_setting(
3545 name = "ios_x86_64",
3546 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003547 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003548 "cpu": "ios_x86_64",
3549 },
3550)
3551
3552config_setting(
3553 name = "watchos_armv7k",
3554 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003555 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003556 "cpu": "watchos_armv7k",
3557 },
3558)
3559
3560config_setting(
3561 name = "watchos_arm64_32",
3562 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003563 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003564 "cpu": "watchos_arm64_32",
3565 },
3566)
3567
3568config_setting(
3569 name = "watchos_x86",
3570 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003571 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003572 "cpu": "watchos_i386",
3573 },
3574)
3575
3576config_setting(
3577 name = "watchos_x86_64",
3578 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003579 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003580 "cpu": "watchos_x86_64",
3581 },
3582)
3583
3584config_setting(
3585 name = "tvos_arm64",
3586 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003587 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003588 "cpu": "tvos_arm64",
3589 },
3590)
3591
3592config_setting(
3593 name = "tvos_x86_64",
3594 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003595 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003596 "cpu": "tvos_x86_64",
3597 },
3598)