blob: 5719969b0d8a31a74049346402daef2ea148d6f1 [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 Dukhane8265432020-04-28 18:42:59 -070053 "src/operators/add-nc.c",
54 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
57 "src/operators/channel-pad-nc.c",
58 "src/operators/channel-shuffle-nc.c",
59 "src/operators/clamp-nc.c",
60 "src/operators/convolution-nchw.c",
61 "src/operators/convolution-nhwc.c",
62 "src/operators/deconvolution-nhwc.c",
63 "src/operators/fully-connected-nc.c",
64 "src/operators/global-average-pooling-ncw.c",
65 "src/operators/global-average-pooling-nwc.c",
66 "src/operators/hardswish-nc.c",
67 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
70 "src/operators/resize-bilinear-nhwc.c",
71 "src/operators/sigmoid-nc.c",
72 "src/operators/softmax-nc.c",
73 "src/operators/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 Dukhanf5425ea2020-04-24 01:46:00 -0700508 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
509 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
510 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
511 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700512 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
513 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
514 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
515 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700516 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
517 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
518 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
519 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700520 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700521 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
522 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700523 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
524 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
525 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
526 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
527 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
528 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
529 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
530 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
531 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
532 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
533 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
534 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
535 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
536 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
537 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
538 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
539 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
540 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
541 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
542 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
543 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
544 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
545 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
546 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
547 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
548 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
549 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
550 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
551 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
552 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
553 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800554 "src/f32-hswish/gen/neon-x4.c",
555 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700556 "src/f32-ibilinear/gen/neon-c4.c",
557 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700558 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
559 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
560 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
561 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
562 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
563 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
564 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
565 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
566 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
567 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
568 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
569 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
570 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
571 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
572 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
573 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700574 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
575 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
576 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-ppmm/gen/4x8-minmax-neon.c",
578 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800579 "src/f32-prelu/gen/neon-2x4.c",
580 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800581 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
582 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
583 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
584 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
585 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
586 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
587 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
588 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
589 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
590 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
591 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
592 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
593 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
594 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
595 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
596 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
597 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
598 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
599 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
600 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
601 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
602 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
603 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
604 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700605 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800606 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800607 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
608 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
609 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
610 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
611 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
612 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
613 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
614 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
615 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
616 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
617 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
618 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
619 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
620 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
621 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
622 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
623 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
624 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700625 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
626 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
627 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
628 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800629 "src/f32-vbinary/gen/vmax-neon-x4.c",
630 "src/f32-vbinary/gen/vmax-neon-x8.c",
631 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
632 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
633 "src/f32-vbinary/gen/vmin-neon-x4.c",
634 "src/f32-vbinary/gen/vmin-neon-x8.c",
635 "src/f32-vbinary/gen/vminc-neon-x4.c",
636 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700637 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
638 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
639 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
640 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
641 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
642 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
643 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
644 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
645 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
646 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700647 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
648 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
649 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
650 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700651 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700652 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
653 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700654 "src/q8-gemm/4x8-minmax-neon.c",
655 "src/q8-gemm/8x8-minmax-neon.c",
656 "src/q8-igemm/4x8-minmax-neon.c",
657 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700658 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700659 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700660 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700661 "src/u8-rmax/neon.c",
662 "src/x32-packx/x4-neon-st4.c",
663 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700664 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700665 "src/x32-zip/x2-neon.c",
666 "src/x32-zip/x3-neon.c",
667 "src/x32-zip/x4-neon.c",
668 "src/x32-zip/xm-neon.c",
669 "src/x8-zip/x2-neon.c",
670 "src/x8-zip/x3-neon.c",
671 "src/x8-zip/x4-neon.c",
672 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800673 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800674 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
675 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
676 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
677 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
678 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
679 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700680 "src/requantization/precise-neon.c",
681 "src/requantization/fp32-neon.c",
682 "src/requantization/q31-neon.c",
683 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700684]
685
686NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700687 "src/f32-ibilinear/gen/neonfma-c4.c",
688 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700689 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
690 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
691 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
692 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
693 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
694 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
695 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
696 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
697 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700698 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
699 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
700 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
701 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700702 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
703 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
704 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
705 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700706 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
707 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
708 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
709 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700710 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
711 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
712 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
713 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
714 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
715 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
716 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
717 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
718 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
719 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
720 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
721 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
722 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
723 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
724 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
725 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
726 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
727 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800728 "src/f32-hswish/gen/neonfma-x4.c",
729 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700730 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
731 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800732 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
733 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
734 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
735 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
736 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
737 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
738 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
739 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
740 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
741 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
742 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
743 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
744 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
745 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
746 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
747 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
748 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
749 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
750 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
751 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
752 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
753 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
754 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
755 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800756 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
757 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
758 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
759 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
768 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
769 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
770 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
771 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
772 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
773 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
774 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
775 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
776 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
777 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
778 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
779 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700810 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
811 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800812 "src/math/exp-neonfma-lut64-p2.c",
813 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800814 "src/math/expminus-neonfma-lut2048-p1.c",
815 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800816 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800817 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
818 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
819 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
820 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
821 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
822 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
823 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
824 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
825 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
826 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
827 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
828 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
829 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
830 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
831 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
832 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
833 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
834 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700835]
836
837AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700838 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
839 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
840 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
841 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
842 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
843 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700844 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
845 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
846 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
847 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
848 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
849 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
850 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
851 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
852 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
853 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
854 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
855 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
856 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
857 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
858 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
859 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
860 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
861 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
862 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
863 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700864 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
865 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
866 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
867 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
868 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
869 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
870 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800871 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
872 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
873 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
874 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
875 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
876 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
877 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
878 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
879 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
880 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
881 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
882 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
883 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
884 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
885 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
886 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
887 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
888 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700889 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
890 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
891 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
892 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
893 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
894 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
895 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
896 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
897 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
898 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
899 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
900 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
901 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
902 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
903 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
904 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
905 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
906 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800907 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
908 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
909 "src/math/sigmoid-neonfma-rr1-p5-div.c",
910 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
911 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
912 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700913]
914
915AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard83ea9392020-04-20 20:15:14 -0700916 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
917 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
918 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
919 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
920 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
921 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
922 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
923 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700924 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
925 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
926 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
927 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
928 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
929 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
930 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
931 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700932]
933
934SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700935 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
936 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700937 "src/f32-clamp/gen/sse-x4.c",
938 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800939 "src/f32-dwconv-spchw/3x3p1-sse.c",
940 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700941 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
942 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
943 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
944 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
945 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
946 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
947 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
948 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
949 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
950 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
951 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
952 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700953 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700954 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
955 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700956 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
957 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
958 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
959 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
960 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
961 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
962 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
963 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
964 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
965 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
966 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
967 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
968 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800969 "src/f32-hswish/gen/sse-x4.c",
970 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700971 "src/f32-ibilinear/gen/sse-c4.c",
972 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700973 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
974 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
975 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
976 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
977 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
978 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
979 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700980 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
981 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
982 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700983 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700984 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700985 "src/f32-spmm/gen/4x1-minmax-sse.c",
986 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700987 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
988 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
989 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
990 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
991 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
992 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
993 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
994 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800995 "src/f32-vbinary/gen/vmax-sse-x4.c",
996 "src/f32-vbinary/gen/vmax-sse-x8.c",
997 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
998 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
999 "src/f32-vbinary/gen/vmin-sse-x4.c",
1000 "src/f32-vbinary/gen/vmin-sse-x8.c",
1001 "src/f32-vbinary/gen/vminc-sse-x4.c",
1002 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001003 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1004 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1005 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1006 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1007 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1008 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1009 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1010 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1011 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1012 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1013 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1014 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001015 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1016 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001017 "src/x32-packx/x4-sse.c",
1018]
1019
1020SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001021 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1022 "src/f32-argmaxpool/4x-sse2-c4.c",
1023 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001024 "src/f32-prelu/gen/sse2-2x4.c",
1025 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001026 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1027 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1028 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1029 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1030 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1031 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1032 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1033 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1034 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1035 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1036 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1037 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001038 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001039 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001040 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001041 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001042 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1043 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001044 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1045 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001046 "src/q8-igemm/4x4c2-minmax-sse2.c",
1047 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001048 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1049 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001050 "src/q8-gemm/2x4c8-minmax-sse2.c",
1051 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001052 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001053 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001054 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001055 "src/u8-rmax/sse2.c",
1056 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001057 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001058 "src/x32-zip/x2-sse2.c",
1059 "src/x32-zip/x3-sse2.c",
1060 "src/x32-zip/x4-sse2.c",
1061 "src/x32-zip/xm-sse2.c",
1062 "src/x8-zip/x2-sse2.c",
1063 "src/x8-zip/x3-sse2.c",
1064 "src/x8-zip/x4-sse2.c",
1065 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001066 "src/math/exp-sse2-p5.c",
1067 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001068 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001069 "src/requantization/precise-sse2.c",
1070 "src/requantization/fp32-sse2.c",
1071 "src/requantization/q31-sse2.c",
1072 "src/requantization/gemmlowp-sse2.c",
1073]
1074
1075SSSE3_UKERNELS = [
1076 "src/requantization/precise-ssse3.c",
1077 "src/requantization/q31-ssse3.c",
1078 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001079]
1080
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001081SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001082 "src/f32-prelu/gen/sse41-2x4.c",
1083 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001084 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001085 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001086 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001087 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001088 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1089 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001090 "src/requantization/precise-sse4.c",
1091 "src/requantization/q31-sse4.c",
1092 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001093]
1094
Marat Dukhan08c4a432019-10-03 09:29:21 -07001095AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001096 "src/f32-clamp/gen/avx-x8.c",
1097 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001098 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1099 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1100 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1101 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1102 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1103 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1104 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1105 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1106 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1107 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1108 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1109 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1110 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1111 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1112 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1113 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1114 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1115 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1116 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1117 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1118 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1119 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1120 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1121 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1122 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1123 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1124 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1125 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1126 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1127 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001128 "src/f32-hswish/gen/avx-x8.c",
1129 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001130 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1131 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1132 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1133 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1134 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1135 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1136 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1137 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1138 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001139 "src/f32-prelu/gen/avx-2x8.c",
1140 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001141 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001142 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1143 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1144 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1145 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1146 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1147 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1148 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1149 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001150 "src/f32-vbinary/gen/vmax-avx-x8.c",
1151 "src/f32-vbinary/gen/vmax-avx-x16.c",
1152 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1153 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1154 "src/f32-vbinary/gen/vmin-avx-x8.c",
1155 "src/f32-vbinary/gen/vmin-avx-x16.c",
1156 "src/f32-vbinary/gen/vminc-avx-x8.c",
1157 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001158 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1159 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1160 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1161 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1162 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1163 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1164 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1165 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1166 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1167 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1168 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1169 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001170 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001171]
1172
Marat Dukhanfda12b82019-11-21 12:27:59 -08001173FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001174 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1175 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1176 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1177 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1178 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1179 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1180 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1181 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1182 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1183 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1184 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1185 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1186 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1187 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1188 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1189 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1190 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1191 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1192 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1193 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1194 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1195 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1196 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1197 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1198 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1199 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1200 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1201 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1202 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1203 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1204 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1205 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1206 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1207 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1208 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1209 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1210 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1211 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1212 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1213 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001214 "src/f32-hswish/gen/fma3-x8.c",
1215 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001216 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1217 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1218 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1219 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1220 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1221 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1222 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1223 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1224 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1225 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1226 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1227 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1228 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1229 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001230]
1231
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001232AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001233 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1234 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1235 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1236 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1237 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1238 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1239 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1240 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1241 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1242 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1243 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1244 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1245 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1246 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1247 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1248 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1249 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1250 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1251 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1252 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1253 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1254 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1255 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1256 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1257 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1258 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1259 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1260 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1261 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1262 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1263 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1264 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1265 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1266 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1267 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1268 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001269 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1270 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1271 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1272 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1273 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1274 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1275 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1276 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1277 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1278 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1279 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1280 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1281 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1282 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1283 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1284 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1285 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1286 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1287 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1288 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1289 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1290 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1291 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1292 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1293 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1294 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1295 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1296 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1297 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1298 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001299 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1300 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1301 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1302 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1303 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1304 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1305 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1306 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1307 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1308 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1309 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1310 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1311 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1312 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1313 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1314 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1315 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1316 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1317 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1318 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1319 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1320 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1321 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1322 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001323 "src/math/exp-avx2-p5.c",
1324 "src/math/exp-avx2-perm-p3.c",
1325 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001326 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001327 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001328 "src/math/sigmoid-avx2-rr2-p5-div.c",
1329 "src/math/sigmoid-avx2-rr1-p5-div.c",
1330 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1331 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1332 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1333 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001334]
1335
Marat Dukhan08c4a432019-10-03 09:29:21 -07001336AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001337 "src/f32-clamp/gen/avx512f-x16.c",
1338 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001339 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1340 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1341 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1342 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1343 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1344 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1345 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1346 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1347 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1348 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1349 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1350 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1351 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1352 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1353 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1354 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1355 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1356 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1357 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1358 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1359 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1360 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1361 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1362 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001363 "src/f32-hswish/gen/avx512f-x16.c",
1364 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001365 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1366 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1367 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1368 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1369 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1370 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001371 "src/f32-prelu/gen/avx512f-2x16.c",
1372 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001373 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1374 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1375 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1376 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1377 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1378 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1379 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1380 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1381 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1382 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1383 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1384 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1385 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1386 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1387 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1388 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1389 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1390 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1391 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1392 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1393 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1394 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1395 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1396 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1397 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1398 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1399 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1400 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1401 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1402 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1403 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1404 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1405 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1406 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1407 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1408 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001409 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001410 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1411 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1412 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1413 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1414 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1415 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1416 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1417 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001418 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1419 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1420 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1421 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1422 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1423 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1424 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1425 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001426 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1427 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1428 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1429 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1430 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1431 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1432 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1433 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1434 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1435 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1436 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1437 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001438 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001439 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1440 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1441 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1442 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1443 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1444 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1445 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1446 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1447 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1448 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1449 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1450 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1451 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1452 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1453 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1454 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1455 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1456 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1457 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1458 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1459 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1460 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1461 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1462 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001463 "src/math/exp-avx512f-p5-scalef.c",
1464 "src/math/exp-avx512f-p5.c",
1465 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001466 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001467 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001468]
1469
1470AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001471 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001472 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1473 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1474 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1475 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1476 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1477 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1478 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1479 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1480 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1481 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001482]
1483
1484AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001485 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001486 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1487 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001488 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001489 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1490 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001491 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1492 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1493 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001494 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001495 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1496 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1497 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001498 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001499 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1500 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001501 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001502 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1503 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1504 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1505 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1506 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1507 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1508 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1509 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1510 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1511 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1512 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1513 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1514 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1515 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1516 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1517 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1518 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1519 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1520 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1521 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1522 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001523 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001524 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1525 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1526 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1527 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1528 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1529 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1530 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1531 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1532 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1533 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1534 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1535 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1536 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1537 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1538 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1539 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1540 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1541 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1542 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1543 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1544 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1545 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1546 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1547 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1548 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1549 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1550 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1551 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1552 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1553 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1554 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1555 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1556 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1557 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1558 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1559 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1560 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1561 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001562]
1563
Marat Dukhan1b354632020-03-23 12:50:22 -07001564INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001565 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001566 "src/xnnpack/argmaxpool.h",
1567 "src/xnnpack/avgpool.h",
1568 "src/xnnpack/clamp.h",
1569 "src/xnnpack/common.h",
1570 "src/xnnpack/conv.h",
1571 "src/xnnpack/dwconv.h",
1572 "src/xnnpack/gavgpool.h",
1573 "src/xnnpack/gemm.h",
1574 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001575 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001576 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001577 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001578 "src/xnnpack/lut.h",
1579 "src/xnnpack/math.h",
1580 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001581 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/xnnpack/packx.h",
1583 "src/xnnpack/pad.h",
1584 "src/xnnpack/params.h",
1585 "src/xnnpack/pavgpool.h",
1586 "src/xnnpack/ppmm.h",
1587 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001588 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001589 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001590 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001591 "src/xnnpack/rmax.h",
1592 "src/xnnpack/scalar-utils.h",
1593 "src/xnnpack/spmm.h",
1594 "src/xnnpack/unpool.h",
1595 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001596 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001597 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001598 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001599 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001600 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001601 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001602 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001603]
1604
1605INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001606 "include/xnnpack.h",
1607 "src/xnnpack/allocator.h",
1608 "src/xnnpack/compute.h",
1609 "src/xnnpack/im2col.h",
1610 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001611 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001612 "src/xnnpack/operator.h",
1613 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001614 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001615 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001616 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001617 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001618]
1619
Marat Dukhan1b354632020-03-23 12:50:22 -07001620ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001621 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001622]
1623
Marat Dukhan1b354632020-03-23 12:50:22 -07001624MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001625 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001626 "include/xnnpack.h",
1627]
1628
Marat Dukhan1b354632020-03-23 12:50:22 -07001629MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001630 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001631 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001632 "src/xnnpack/requantization.h",
1633 "include/xnnpack.h",
1634]
1635
1636OPERATOR_TEST_PARAMS_HDRS = [
1637 "src/xnnpack/params.h",
1638 "src/xnnpack/common.h",
1639]
1640
1641WEIGHTS_PACK_HDRS = [
1642 "src/xnnpack/pack.h",
1643 "src/xnnpack/operator.h",
1644 "src/xnnpack/compute.h",
1645]
1646
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001647LOGGING_COPTS = select({
1648 # No logging in optimized mode
1649 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1650 # Full logging in debug mode
1651 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1652 # Error-only logging in default (fastbuild) mode
1653 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1654})
1655
1656LOGGING_HDRS = [
1657 "src/xnnpack/log.h",
1658]
1659
Marat Dukhan08c4a432019-10-03 09:29:21 -07001660xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001661 name = "tables",
1662 srcs = TABLE_SRCS,
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 Dukhan3a77ea72019-12-23 12:10:24 -08001666)
1667
1668xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001669 name = "scalar_ukernels",
1670 srcs = SCALAR_UKERNELS,
1671 hdrs = INTERNAL_HDRS,
1672 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001673 gcc_copts = xnnpack_gcc_std_copts(),
1674 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001675 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001676 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001677 "@FP16",
1678 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001679 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001680 ],
1681)
1682
1683xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001684 name = "wasm_ukernels",
1685 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001686 gcc_copts = xnnpack_gcc_std_copts(),
1687 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001688 wasm_srcs = WASM_UKERNELS,
1689 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001690 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001691 "@FP16",
1692 "@FXdiv",
1693 "@pthreadpool",
1694 ],
1695)
1696
1697xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001698 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001699 hdrs = INTERNAL_HDRS,
1700 aarch32_copts = [
1701 "-marm",
1702 "-mfpu=neon",
1703 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001704 gcc_copts = xnnpack_gcc_std_copts(),
1705 gcc_x86_copts = ["-msse2"],
1706 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001707 optimized_copts = [
1708 "-O3",
1709 "-ffast-math",
1710 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001711 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001712 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001713 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001714 "@FP16",
1715 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001716 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001717 ],
1718)
1719
1720xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001721 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001722 hdrs = INTERNAL_HDRS,
1723 aarch32_copts = [
1724 "-marm",
1725 "-mfpu=neon",
1726 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001727 gcc_copts = xnnpack_gcc_std_copts(),
1728 gcc_x86_copts = ["-msse2"],
1729 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001730 optimized_copts = [
1731 "-O3",
1732 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001733 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001734 deps = [
1735 ":tables",
1736 "@FP16",
1737 "@psimd",
1738 "@pthreadpool",
1739 ],
1740)
1741
1742xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001743 name = "neon_ukernels",
1744 hdrs = INTERNAL_HDRS,
1745 aarch32_copts = [
1746 "-marm",
1747 "-mfpu=neon",
1748 ],
1749 aarch32_srcs = NEON_UKERNELS,
1750 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001751 gcc_copts = xnnpack_gcc_std_copts(),
1752 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001753 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001754 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001755 "@FP16",
1756 "@pthreadpool",
1757 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001758)
1759
1760xnnpack_cc_library(
1761 name = "neonfma_ukernels",
1762 hdrs = INTERNAL_HDRS,
1763 aarch32_copts = [
1764 "-marm",
1765 "-mfpu=neon-vfpv4",
1766 ],
1767 aarch32_srcs = NEONFMA_UKERNELS,
1768 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001769 gcc_copts = xnnpack_gcc_std_copts(),
1770 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001771 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001772 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001773 "@FP16",
1774 "@pthreadpool",
1775 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001776)
1777
1778xnnpack_cc_library(
1779 name = "neonfp16arith_ukernels",
1780 hdrs = INTERNAL_HDRS,
1781 aarch64_copts = ["-march=armv8.2-a+fp16"],
1782 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001783 gcc_copts = xnnpack_gcc_std_copts(),
1784 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001785 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001786 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001787 "@FP16",
1788 "@pthreadpool",
1789 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001790)
1791
1792xnnpack_cc_library(
1793 name = "sse2_ukernels",
1794 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001795 gcc_copts = xnnpack_gcc_std_copts(),
1796 gcc_x86_copts = ["-msse2"],
1797 msvc_copts = xnnpack_msvc_std_copts(),
1798 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001799 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001800 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001801 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001802 "@FP16",
1803 "@pthreadpool",
1804 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001805)
1806
1807xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001808 name = "ssse3_ukernels",
1809 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001810 gcc_copts = xnnpack_gcc_std_copts(),
1811 gcc_x86_copts = ["-mssse3"],
1812 msvc_copts = xnnpack_msvc_std_copts(),
1813 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001814 x86_srcs = SSSE3_UKERNELS,
1815 deps = [
1816 ":tables",
1817 "@FP16",
1818 "@pthreadpool",
1819 ],
1820)
1821
1822xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001823 name = "sse41_ukernels",
1824 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001825 gcc_copts = xnnpack_gcc_std_copts(),
1826 gcc_x86_copts = ["-msse4.1"],
1827 msvc_copts = xnnpack_msvc_std_copts(),
1828 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001829 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001830 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001831 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001832 "@FP16",
1833 "@pthreadpool",
1834 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001835)
1836
1837xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838 name = "avx_ukernels",
1839 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001840 gcc_copts = xnnpack_gcc_std_copts(),
1841 gcc_x86_copts = ["-mavx"],
1842 msvc_copts = xnnpack_msvc_std_copts(),
1843 msvc_x86_32_copts = ["/arch:AVX"],
1844 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001845 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001846 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001847 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001848 "@FP16",
1849 "@pthreadpool",
1850 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001851)
1852
1853xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001854 name = "fma3_ukernels",
1855 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001856 gcc_copts = xnnpack_gcc_std_copts(),
1857 gcc_x86_copts = ["-mfma"],
1858 msvc_copts = xnnpack_msvc_std_copts(),
1859 msvc_x86_32_copts = ["/arch:AVX"],
1860 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001861 x86_srcs = FMA3_UKERNELS,
1862 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001863 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001864 "@FP16",
1865 "@pthreadpool",
1866 ],
1867)
1868
1869xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001870 name = "avx2_ukernels",
1871 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001872 gcc_copts = xnnpack_gcc_std_copts(),
1873 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001874 "-mfma",
1875 "-mavx2",
1876 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001877 msvc_copts = xnnpack_msvc_std_copts(),
1878 msvc_x86_32_copts = ["/arch:AVX2"],
1879 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001880 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001881 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001882 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001883 "@FP16",
1884 "@pthreadpool",
1885 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001886)
1887
1888xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001889 name = "avx512f_ukernels",
1890 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001891 gcc_copts = xnnpack_gcc_std_copts(),
1892 gcc_x86_copts = ["-mavx512f"],
1893 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1894 msvc_copts = xnnpack_msvc_std_copts(),
1895 msvc_x86_32_copts = ["/arch:AVX512"],
1896 msvc_x86_64_copts = ["/arch:AVX512"],
1897 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001898 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001899 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001900 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001901 "@FP16",
1902 "@pthreadpool",
1903 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001904)
1905
1906xnnpack_cc_library(
1907 name = "asm_ukernels",
1908 hdrs = ["src/xnnpack/assembly.h"],
1909 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001910 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001911 aarch64_srcs = AARCH64_ASM_UKERNELS,
1912)
1913
1914xnnpack_aggregate_library(
1915 name = "ukernels",
1916 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001917 ":neon_ukernels",
1918 ":neonfma_ukernels",
1919 ":asm_ukernels",
1920 ],
1921 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001922 ":neon_ukernels",
1923 ":neonfma_ukernels",
1924 ":neonfp16arith_ukernels",
1925 ":asm_ukernels",
1926 ],
1927 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07001928 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001929 ":psimd_fastmath_ukernels",
1930 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001931 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001932 wasm_deps = [":wasm_ukernels"],
1933 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001934 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001935 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001936 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001937 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001938 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001939 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001940 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001941 ":avx512f_ukernels",
1942 ],
1943)
1944
1945xnnpack_cc_library(
1946 name = "im2col",
1947 srcs = ["src/im2col.c"],
1948 hdrs = [
1949 "src/xnnpack/common.h",
1950 "src/xnnpack/im2col.h",
1951 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001952 gcc_copts = xnnpack_gcc_std_copts(),
1953 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001954)
1955
1956xnnpack_cc_library(
1957 name = "indirection",
1958 srcs = ["src/indirection.c"],
1959 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001960 gcc_copts = xnnpack_gcc_std_copts(),
1961 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001962 deps = [
1963 "@FP16",
1964 "@FXdiv",
1965 "@pthreadpool",
1966 ],
1967)
1968
1969xnnpack_cc_library(
1970 name = "operator_run",
1971 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001972 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001973 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07001974 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1975 "//conditions:default": [],
1976 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07001977 gcc_copts = xnnpack_gcc_std_copts(),
1978 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001979 deps = [
1980 "@FP16",
1981 "@FXdiv",
1982 "@clog",
1983 "@pthreadpool",
1984 ],
1985)
1986
1987cc_library(
1988 name = "enable_assembly",
1989 defines = select({
1990 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1991 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001992 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001993 }),
1994)
1995
Marat Dukhancf056b22019-10-07 10:26:29 -07001996xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997 name = "operators",
1998 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001999 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002000 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002001 ],
2002 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002003 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002004 "-Isrc",
2005 "-Iinclude",
2006 ] + select({
2007 ":debug_build": [],
2008 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002009 }) + select({
2010 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2011 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002012 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002013 gcc_copts = xnnpack_gcc_std_copts(),
2014 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002015 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002016 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002017 "@FP16",
2018 "@FXdiv",
2019 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002020 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002021 ],
2022)
2023
Marat Dukhan10a38082020-04-17 03:58:35 -07002024xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002025 name = "XNNPACK",
2026 srcs = [
2027 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002028 "src/runtime.c",
2029 "src/subgraph.c",
2030 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002031 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002032 hdrs = ["include/xnnpack.h"],
2033 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002034 "-Isrc",
2035 "-Iinclude",
2036 ] + select({
2037 ":debug_build": [],
2038 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002039 }) + select({
2040 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2041 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002042 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002043 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002044 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002045 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002046 visibility = xnnpack_visibility(),
2047 deps = [
2048 ":enable_assembly",
2049 ":ukernels",
2050 ":operator_run",
2051 ":operators",
2052 "@clog",
2053 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002054 ] + select({
2055 ":emscripten": [],
2056 "//conditions:default": ["@cpuinfo"],
2057 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002058)
2059
Marat Dukhan10a38082020-04-17 03:58:35 -07002060xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002061 name = "xnnpack_operators_nhwc_f32",
2062 srcs = [
2063 "src/init.c",
2064 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002065 hdrs = ["include/xnnpack.h"],
2066 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002067 "-Isrc",
2068 "-Iinclude",
2069 ] + select({
2070 ":debug_build": [],
2071 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002072 }) + select({
2073 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2074 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002075 }),
2076 defines = [
2077 "XNN_NO_Q8_OPERATORS",
2078 "XNN_NO_U8_OPERATORS",
2079 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002080 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002081 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002082 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002083 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002084 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002085 visibility = xnnpack_visibility(),
2086 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002087 ":enable_assembly",
2088 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002089 ":operator_run",
2090 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002091 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002092 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002093 ] + select({
2094 ":emscripten": [],
2095 "//conditions:default": ["@cpuinfo"],
2096 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002097)
2098
Marat Dukhancf056b22019-10-07 10:26:29 -07002099xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002100 name = "bench_utils",
2101 srcs = ["bench/utils.cc"],
2102 hdrs = ["bench/utils.h"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002103 gcc_copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002104 deps = [
2105 "@com_google_benchmark//:benchmark",
2106 "@cpuinfo",
2107 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002108)
2109
Frank Barchard7e955972019-10-11 10:34:25 -07002110######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002111
2112xnnpack_benchmark(
2113 name = "q8_gemm_bench",
2114 srcs = [
2115 "bench/gemm.h",
2116 "bench/q8-gemm.cc",
2117 "src/xnnpack/AlignedAllocator.h",
2118 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002119 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
2120 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002121 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2122)
2123
2124xnnpack_benchmark(
2125 name = "f16_gemm_bench",
2126 srcs = [
2127 "bench/f16-gemm.cc",
2128 "bench/gemm.h",
2129 "src/xnnpack/AlignedAllocator.h",
2130 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002131 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002132 deps = MICROKERNEL_BENCHMARK_DEPS,
2133)
2134
2135xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002136 name = "f16_spmm_bench",
2137 srcs = [
2138 "bench/f16-spmm.cc",
2139 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002140 "src/xnnpack/AlignedAllocator.h",
2141 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002142 gcc_copts = ["-Wno-unused-function"],
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002143 deps = MICROKERNEL_BENCHMARK_DEPS,
2144)
2145
2146xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002147 name = "f32_igemm_bench",
2148 srcs = [
2149 "bench/f32-igemm.cc",
2150 "bench/conv.h",
2151 "src/xnnpack/AlignedAllocator.h",
2152 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002153 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002154)
2155
2156xnnpack_benchmark(
2157 name = "f32_conv_hwc_bench",
2158 srcs = [
2159 "bench/f32-conv-hwc.cc",
2160 "bench/dconv.h",
2161 "src/xnnpack/AlignedAllocator.h",
2162 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002163 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002164 deps = MICROKERNEL_BENCHMARK_DEPS,
2165)
2166
2167xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002168 name = "f32_conv_hwc2spchw_bench",
2169 srcs = [
2170 "bench/f32-conv-hwc2spchw.cc",
2171 "bench/dconv.h",
2172 "src/xnnpack/AlignedAllocator.h",
2173 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002174 gcc_copts = ["-Wno-unused-function"],
Erich Elsen563df5f2019-10-23 08:02:21 -07002175 deps = MICROKERNEL_BENCHMARK_DEPS,
2176)
2177
2178xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002179 name = "f32_dwconv_bench",
2180 srcs = [
2181 "bench/f32-dwconv.cc",
2182 "bench/dwconv.h",
2183 "src/xnnpack/AlignedAllocator.h",
2184 ] + MICROKERNEL_BENCHMARK_HDRS,
2185 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2186)
2187
2188xnnpack_benchmark(
2189 name = "f32_dwconv_spchw_bench",
2190 srcs = [
2191 "bench/f32-dwconv-spchw.cc",
2192 "bench/dwconv.h",
2193 "src/xnnpack/AlignedAllocator.h",
2194 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2195 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2196)
2197
2198xnnpack_benchmark(
2199 name = "f32_gemm_bench",
2200 srcs = [
2201 "bench/f32-gemm.cc",
2202 "bench/gemm.h",
2203 "src/xnnpack/AlignedAllocator.h",
2204 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002205 copts = xnnpack_optional_ruy_copts(),
2206 gcc_copts = ["-Wno-unused-function"],
Frank Barchard7e955972019-10-11 10:34:25 -07002207 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002208)
2209
2210xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002211 name = "f32_raddexpminusmax_bench",
2212 srcs = [
2213 "bench/f32-raddexpminusmax.cc",
2214 "src/xnnpack/AlignedAllocator.h",
2215 ] + MICROKERNEL_BENCHMARK_HDRS,
2216 deps = MICROKERNEL_BENCHMARK_DEPS,
2217)
2218
2219xnnpack_benchmark(
2220 name = "f32_raddextexp_bench",
2221 srcs = [
2222 "bench/f32-raddextexp.cc",
2223 "src/xnnpack/AlignedAllocator.h",
2224 ] + MICROKERNEL_BENCHMARK_HDRS,
2225 deps = MICROKERNEL_BENCHMARK_DEPS,
2226)
2227
2228xnnpack_benchmark(
2229 name = "f32_raddstoreexpminusmax_bench",
2230 srcs = [
2231 "bench/f32-raddstoreexpminusmax.cc",
2232 "src/xnnpack/AlignedAllocator.h",
2233 ] + MICROKERNEL_BENCHMARK_HDRS,
2234 deps = MICROKERNEL_BENCHMARK_DEPS,
2235)
2236
2237xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238 name = "f32_rmax_bench",
2239 srcs = [
2240 "bench/f32-rmax.cc",
2241 "src/xnnpack/AlignedAllocator.h",
2242 ] + MICROKERNEL_BENCHMARK_HDRS,
2243 deps = MICROKERNEL_BENCHMARK_DEPS,
2244)
2245
2246xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002247 name = "f32_sigmoid_bench",
2248 srcs = [
2249 "bench/f32-sigmoid.cc",
2250 "src/xnnpack/AlignedAllocator.h",
2251 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002252 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan14bec502019-11-18 11:35:31 -08002253 deps = MICROKERNEL_BENCHMARK_DEPS,
2254)
2255
2256xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002257 name = "f32_spmm_bench",
2258 srcs = [
2259 "bench/f32-spmm.cc",
2260 "bench/gemm.h",
2261 "src/xnnpack/AlignedAllocator.h",
2262 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002263 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002264 deps = MICROKERNEL_BENCHMARK_DEPS,
2265)
2266
2267xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002268 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002269 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002270 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002271 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002272 copts = xnnpack_optional_dnnl_copts(),
2273 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002274 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002275)
2276
2277xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002278 name = "f32_vscaleexpminusmax_bench",
2279 srcs = [
2280 "bench/f32-vscaleexpminusmax.cc",
2281 "src/xnnpack/AlignedAllocator.h",
2282 ] + MICROKERNEL_BENCHMARK_HDRS,
2283 deps = MICROKERNEL_BENCHMARK_DEPS,
2284)
2285
2286xnnpack_benchmark(
2287 name = "f32_vscaleextexp_bench",
2288 srcs = [
2289 "bench/f32-vscaleextexp.cc",
2290 "src/xnnpack/AlignedAllocator.h",
2291 ] + MICROKERNEL_BENCHMARK_HDRS,
2292 deps = MICROKERNEL_BENCHMARK_DEPS,
2293)
2294
2295xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002296 name = "f32_im2col_gemm_bench",
2297 srcs = [
2298 "bench/f32-im2col-gemm.cc",
2299 "bench/conv.h",
2300 "src/xnnpack/AlignedAllocator.h",
2301 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2302 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2303)
2304
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002305xnnpack_benchmark(
2306 name = "requantization_bench",
2307 srcs = [
2308 "bench/requantization.cc",
2309 "src/xnnpack/requantization-stubs.h",
2310 "src/xnnpack/AlignedAllocator.h",
2311 ] + MICROKERNEL_BENCHMARK_HDRS,
2312 deps = MICROKERNEL_BENCHMARK_DEPS,
2313)
2314
Marat Dukhan08c4a432019-10-03 09:29:21 -07002315########################### Benchmarks for operators ###########################
2316
2317xnnpack_benchmark(
2318 name = "add_bench",
2319 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002320 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002321)
2322
2323xnnpack_benchmark(
2324 name = "average_pooling_bench",
2325 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002326 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002327 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002328 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002329)
2330
2331xnnpack_benchmark(
2332 name = "channel_shuffle_bench",
2333 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002334 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002335)
2336
2337xnnpack_benchmark(
2338 name = "convolution_bench",
2339 srcs = ["bench/convolution.cc"],
2340 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002341 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002342 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002343)
2344
2345xnnpack_benchmark(
2346 name = "deconvolution_bench",
2347 srcs = ["bench/deconvolution.cc"],
2348 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002349 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002350 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002351)
2352
2353xnnpack_benchmark(
2354 name = "global_average_pooling_bench",
2355 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002356 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357)
2358
2359xnnpack_benchmark(
2360 name = "max_pooling_bench",
2361 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002362 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002363)
2364
2365xnnpack_benchmark(
2366 name = "sigmoid_bench",
2367 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002368 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002369 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002370 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002371)
2372
2373xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002374 name = "prelu_bench",
2375 srcs = ["bench/prelu.cc"],
2376 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002377 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002378 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002379)
2380
2381xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002382 name = "softmax_bench",
2383 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002384 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002385 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002386 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002387)
2388
Marat Dukhanc068bb62019-10-04 13:24:39 -07002389############################# End-to-end benchmarks ############################
2390
2391cc_library(
2392 name = "mobilenet_v1",
2393 srcs = ["models/mobilenet-v1.cc"],
2394 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002395 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002396 linkstatic = True,
2397 deps = [
2398 ":XNNPACK",
2399 "@pthreadpool",
2400 ],
2401)
2402
2403cc_library(
2404 name = "mobilenet_v2",
2405 srcs = ["models/mobilenet-v2.cc"],
2406 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002407 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002408 linkstatic = True,
2409 deps = [
2410 ":XNNPACK",
2411 "@pthreadpool",
2412 ],
2413)
2414
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002415cc_library(
2416 name = "mobilenet_v3_large",
2417 srcs = ["models/mobilenet-v3-large.cc"],
2418 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002419 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002420 linkstatic = True,
2421 deps = [
2422 ":XNNPACK",
2423 "@pthreadpool",
2424 ],
2425)
2426
2427cc_library(
2428 name = "mobilenet_v3_small",
2429 srcs = ["models/mobilenet-v3-small.cc"],
2430 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002431 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002432 linkstatic = True,
2433 deps = [
2434 ":XNNPACK",
2435 "@pthreadpool",
2436 ],
2437)
2438
Marat Dukhanc068bb62019-10-04 13:24:39 -07002439xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002440 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002441 srcs = [
2442 "bench/f32-dwconv-e2e.cc",
2443 "bench/end2end.h",
2444 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002445 gcc_copts = ["-Wno-unused-function"],
Marat Dukhanef4416e2019-10-31 13:44:40 -07002446 deps = MICROKERNEL_BENCHMARK_DEPS + [
2447 ":XNNPACK",
2448 ":mobilenet_v1",
2449 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002450 ":mobilenet_v3_large",
2451 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002452 ],
2453)
2454
2455xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002456 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002457 srcs = [
2458 "bench/f32-gemm-e2e.cc",
2459 "bench/end2end.h",
2460 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002461 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan5f18d262019-10-31 10:24:14 -07002462 deps = MICROKERNEL_BENCHMARK_DEPS + [
2463 ":XNNPACK",
2464 ":mobilenet_v1",
2465 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002466 ":mobilenet_v3_large",
2467 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002468 ],
2469)
2470
2471xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002472 name = "end2end_bench",
2473 srcs = ["bench/end2end.cc"],
2474 deps = [
2475 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002476 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002477 ":mobilenet_v1",
2478 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002479 ":mobilenet_v3_large",
2480 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002481 "@pthreadpool",
2482 ],
2483)
2484
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002485#################### Accuracy evaluation for math functions ####################
2486
2487xnnpack_benchmark(
2488 name = "f32_exp_eval",
2489 srcs = [
2490 "eval/f32-exp.cc",
2491 "src/xnnpack/AlignedAllocator.h",
2492 ] + ACCURACY_EVAL_HDRS,
2493 deps = ACCURACY_EVAL_DEPS,
2494)
2495
Marat Dukhan515c9772019-10-17 18:07:57 -07002496xnnpack_benchmark(
2497 name = "f32_expminus_eval",
2498 srcs = [
2499 "eval/f32-expminus.cc",
2500 "src/xnnpack/AlignedAllocator.h",
2501 ] + ACCURACY_EVAL_HDRS,
2502 deps = ACCURACY_EVAL_DEPS,
2503)
2504
Marat Dukhan98ba4412019-10-23 02:14:28 -07002505xnnpack_benchmark(
2506 name = "f32_extexp_eval",
2507 srcs = [
2508 "eval/f32-extexp.cc",
2509 "src/xnnpack/AlignedAllocator.h",
2510 ] + ACCURACY_EVAL_HDRS,
2511 deps = ACCURACY_EVAL_DEPS,
2512)
2513
Marat Dukhan346a9e52019-11-15 09:06:30 -08002514xnnpack_benchmark(
2515 name = "f32_sigmoid_eval",
2516 srcs = [
2517 "eval/f32-sigmoid.cc",
2518 "src/xnnpack/AlignedAllocator.h",
2519 ] + ACCURACY_EVAL_HDRS,
2520 deps = ACCURACY_EVAL_DEPS,
2521)
2522
Marat Dukhan08c4a432019-10-03 09:29:21 -07002523######################### Unit tests for micro-kernels #########################
2524
2525xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002526 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002527 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002528 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002529 "test/gemm-microkernel-tester.h",
2530 "src/xnnpack/AlignedAllocator.h",
2531 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2532 deps = MICROKERNEL_TEST_DEPS,
2533)
2534
2535xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002536 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002537 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002538 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002539 "test/spmm-microkernel-tester.h",
2540 "src/xnnpack/AlignedAllocator.h",
2541 ] + MICROKERNEL_TEST_HDRS,
2542 deps = MICROKERNEL_TEST_DEPS,
2543)
2544
2545xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002546 name = "f32_argmaxpool_test",
2547 srcs = [
2548 "test/f32-argmaxpool.cc",
2549 "test/argmaxpool-microkernel-tester.h",
2550 "src/xnnpack/AlignedAllocator.h",
2551 ] + MICROKERNEL_TEST_HDRS,
2552 deps = MICROKERNEL_TEST_DEPS,
2553)
2554
2555xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002556 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002557 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002558 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002559 "test/avgpool-microkernel-tester.h",
2560 "src/xnnpack/AlignedAllocator.h",
2561 ] + MICROKERNEL_TEST_HDRS,
2562 deps = MICROKERNEL_TEST_DEPS,
2563)
2564
2565xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002566 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002567 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002568 "test/f32-ibilinear.cc",
2569 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002570 "src/xnnpack/AlignedAllocator.h",
2571 ] + MICROKERNEL_TEST_HDRS,
2572 deps = MICROKERNEL_TEST_DEPS,
2573)
2574
2575xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002576 name = "f32_clamp_test",
2577 srcs = [
2578 "test/f32-clamp.cc",
2579 "test/clamp-microkernel-tester.h",
2580 ] + MICROKERNEL_TEST_HDRS,
2581 deps = MICROKERNEL_TEST_DEPS,
2582)
2583
2584xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002585 name = "f32_igemm_test",
2586 srcs = [
2587 "test/f32-igemm.cc",
2588 "test/gemm-microkernel-tester.h",
2589 "src/xnnpack/AlignedAllocator.h",
2590 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2591 deps = MICROKERNEL_TEST_DEPS,
2592)
2593
2594xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002595 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002596 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002597 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002598 "test/gemm-microkernel-tester.h",
2599 "src/xnnpack/AlignedAllocator.h",
2600 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002601 # Work-around for "too many sections" error
2602 mingw_copts = ["-Wa,-mbig-obj"],
2603 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002604 deps = MICROKERNEL_TEST_DEPS,
2605)
2606
2607xnnpack_unit_test(
2608 name = "f32_conv_hwc_test",
2609 srcs = [
2610 "test/f32-conv-hwc.cc",
2611 "test/conv-hwc-microkernel-tester.h",
2612 "src/xnnpack/AlignedAllocator.h",
2613 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2614 deps = MICROKERNEL_TEST_DEPS,
2615)
2616
2617xnnpack_unit_test(
2618 name = "f32_conv_hwc2spchw_test",
2619 srcs = [
2620 "test/f32-conv-hwc2spchw.cc",
2621 "test/conv-hwc2spchw-microkernel-tester.h",
2622 "src/xnnpack/AlignedAllocator.h",
2623 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2624 deps = MICROKERNEL_TEST_DEPS,
2625)
2626
2627xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002628 name = "f32_dwconv_test",
2629 srcs = [
2630 "test/f32-dwconv.cc",
2631 "test/dwconv-microkernel-tester.h",
2632 "src/xnnpack/AlignedAllocator.h",
2633 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2634 deps = MICROKERNEL_TEST_DEPS,
2635)
2636
2637xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002638 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002639 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002640 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002641 "test/dwconv-microkernel-tester.h",
2642 "src/xnnpack/AlignedAllocator.h",
2643 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2644 deps = MICROKERNEL_TEST_DEPS,
2645)
2646
2647xnnpack_unit_test(
2648 name = "f32_dwconv_spchw_test",
2649 srcs = [
2650 "test/f32-dwconv-spchw.cc",
2651 "test/dwconv-spchw-microkernel-tester.h",
2652 "src/xnnpack/AlignedAllocator.h",
2653 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2654 deps = MICROKERNEL_TEST_DEPS,
2655)
2656
2657xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002658 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002659 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002660 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002661 "test/gavgpool-microkernel-tester.h",
2662 "src/xnnpack/AlignedAllocator.h",
2663 ] + MICROKERNEL_TEST_HDRS,
2664 deps = MICROKERNEL_TEST_DEPS,
2665)
2666
2667xnnpack_unit_test(
2668 name = "f32_gavgpool_spchw_test",
2669 srcs = [
2670 "test/f32-gavgpool-spchw.cc",
2671 "test/gavgpool-spchw-microkernel-tester.h",
2672 "src/xnnpack/AlignedAllocator.h",
2673 ] + MICROKERNEL_TEST_HDRS,
2674 deps = MICROKERNEL_TEST_DEPS,
2675)
2676
2677xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002678 name = "f32_gemm_test",
2679 srcs = [
2680 "test/f32-gemm.cc",
2681 "test/gemm-microkernel-tester.h",
2682 "src/xnnpack/AlignedAllocator.h",
2683 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2684 deps = MICROKERNEL_TEST_DEPS,
2685)
2686
2687xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002688 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002689 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002690 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002691 "test/gemm-microkernel-tester.h",
2692 "src/xnnpack/AlignedAllocator.h",
2693 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2694 deps = MICROKERNEL_TEST_DEPS,
2695)
2696
2697xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002698 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002699 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002700 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002701 "test/gemm-microkernel-tester.h",
2702 "src/xnnpack/AlignedAllocator.h",
2703 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2704 deps = MICROKERNEL_TEST_DEPS,
2705)
2706
2707xnnpack_unit_test(
2708 name = "f32_hswish_test",
2709 srcs = [
2710 "test/f32-hswish.cc",
2711 "test/hswish-microkernel-tester.h",
2712 ] + MICROKERNEL_TEST_HDRS,
2713 deps = MICROKERNEL_TEST_DEPS,
2714)
2715
2716xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002717 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002718 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002719 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002720 "test/maxpool-microkernel-tester.h",
2721 ] + MICROKERNEL_TEST_HDRS,
2722 deps = MICROKERNEL_TEST_DEPS,
2723)
2724
2725xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002726 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002727 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002728 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002729 "test/avgpool-microkernel-tester.h",
2730 "src/xnnpack/AlignedAllocator.h",
2731 ] + MICROKERNEL_TEST_HDRS,
2732 deps = MICROKERNEL_TEST_DEPS,
2733)
2734
2735xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002736 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002738 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002739 "test/gemm-microkernel-tester.h",
2740 "src/xnnpack/AlignedAllocator.h",
2741 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2742 deps = MICROKERNEL_TEST_DEPS,
2743)
2744
2745xnnpack_unit_test(
2746 name = "f32_prelu_test",
2747 srcs = [
2748 "test/f32-prelu.cc",
2749 "test/prelu-microkernel-tester.h",
2750 "src/xnnpack/AlignedAllocator.h",
2751 ] + MICROKERNEL_TEST_HDRS,
2752 deps = MICROKERNEL_TEST_DEPS,
2753)
2754
2755xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002756 name = "f32_raddexpminusmax_test",
2757 srcs = [
2758 "test/f32-raddexpminusmax.cc",
2759 "test/raddexpminusmax-microkernel-tester.h",
2760 ] + MICROKERNEL_TEST_HDRS,
2761 deps = MICROKERNEL_TEST_DEPS,
2762)
2763
2764xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002765 name = "f32_raddextexp_test",
2766 srcs = [
2767 "test/f32-raddextexp.cc",
2768 "test/raddextexp-microkernel-tester.h",
2769 ] + MICROKERNEL_TEST_HDRS,
2770 deps = MICROKERNEL_TEST_DEPS,
2771)
2772
2773xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002774 name = "f32_raddstoreexpminusmax_test",
2775 srcs = [
2776 "test/f32-raddstoreexpminusmax.cc",
2777 "test/raddstoreexpminusmax-microkernel-tester.h",
2778 ] + MICROKERNEL_TEST_HDRS,
2779 deps = MICROKERNEL_TEST_DEPS,
2780)
2781
2782xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002783 name = "f32_rmax_test",
2784 srcs = [
2785 "test/f32-rmax.cc",
2786 "test/rmax-microkernel-tester.h",
2787 ] + MICROKERNEL_TEST_HDRS,
2788 deps = MICROKERNEL_TEST_DEPS,
2789)
2790
2791xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002792 name = "f32_sigmoid_test",
2793 srcs = [
2794 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002795 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002796 ] + MICROKERNEL_TEST_HDRS,
2797 deps = MICROKERNEL_TEST_DEPS,
2798)
2799
2800xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002801 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002802 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002803 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002804 "test/spmm-microkernel-tester.h",
2805 "src/xnnpack/AlignedAllocator.h",
2806 ] + MICROKERNEL_TEST_HDRS,
2807 deps = MICROKERNEL_TEST_DEPS,
2808)
2809
2810xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002811 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002812 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002813 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002814 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002815 ] + MICROKERNEL_TEST_HDRS,
2816 deps = MICROKERNEL_TEST_DEPS,
2817)
2818
2819xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002820 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002821 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002822 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002823 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002824 ] + MICROKERNEL_TEST_HDRS,
2825 deps = MICROKERNEL_TEST_DEPS,
2826)
2827
2828xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002829 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002830 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002831 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002832 "test/vbinary-microkernel-tester.h",
2833 ] + MICROKERNEL_TEST_HDRS,
2834 deps = MICROKERNEL_TEST_DEPS,
2835)
2836
2837xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002838 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002839 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002840 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002841 "test/vbinaryc-microkernel-tester.h",
2842 ] + MICROKERNEL_TEST_HDRS,
2843 deps = MICROKERNEL_TEST_DEPS,
2844)
2845
2846xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002847 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002848 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002849 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002850 "test/vbinaryc-microkernel-tester.h",
2851 ] + MICROKERNEL_TEST_HDRS,
2852 deps = MICROKERNEL_TEST_DEPS,
2853)
2854
2855xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002856 name = "f32_vmax_test",
2857 srcs = [
2858 "test/f32-vmax.cc",
2859 "test/vbinary-microkernel-tester.h",
2860 ] + MICROKERNEL_TEST_HDRS,
2861 deps = MICROKERNEL_TEST_DEPS,
2862)
2863
2864xnnpack_unit_test(
2865 name = "f32_vmaxc_test",
2866 srcs = [
2867 "test/f32-vmaxc.cc",
2868 "test/vbinaryc-microkernel-tester.h",
2869 ] + MICROKERNEL_TEST_HDRS,
2870 deps = MICROKERNEL_TEST_DEPS,
2871)
2872
2873xnnpack_unit_test(
2874 name = "f32_vmin_test",
2875 srcs = [
2876 "test/f32-vmin.cc",
2877 "test/vbinary-microkernel-tester.h",
2878 ] + MICROKERNEL_TEST_HDRS,
2879 deps = MICROKERNEL_TEST_DEPS,
2880)
2881
2882xnnpack_unit_test(
2883 name = "f32_vminc_test",
2884 srcs = [
2885 "test/f32-vminc.cc",
2886 "test/vbinaryc-microkernel-tester.h",
2887 ] + MICROKERNEL_TEST_HDRS,
2888 deps = MICROKERNEL_TEST_DEPS,
2889)
2890
2891xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002892 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002893 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002894 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002895 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002896 ] + MICROKERNEL_TEST_HDRS,
2897 deps = MICROKERNEL_TEST_DEPS,
2898)
2899
2900xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002901 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002902 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002903 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002904 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002905 ] + MICROKERNEL_TEST_HDRS,
2906 deps = MICROKERNEL_TEST_DEPS,
2907)
2908
2909xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002910 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002912 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002913 "test/vmulcaddc-microkernel-tester.h",
2914 "src/xnnpack/AlignedAllocator.h",
2915 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2916 deps = MICROKERNEL_TEST_DEPS,
2917)
2918
2919xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002920 name = "f32_vscale_test",
2921 srcs = [
2922 "test/f32-vscale.cc",
2923 "test/vscale-microkernel-tester.h",
2924 ] + MICROKERNEL_TEST_HDRS,
2925 deps = MICROKERNEL_TEST_DEPS,
2926)
2927
2928xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002929 name = "f32_vscaleexpminusmax_test",
2930 srcs = [
2931 "test/f32-vscaleexpminusmax.cc",
2932 "test/vscaleexpminusmax-microkernel-tester.h",
2933 ] + MICROKERNEL_TEST_HDRS,
2934 deps = MICROKERNEL_TEST_DEPS,
2935)
2936
2937xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002938 name = "f32_vscaleextexp_test",
2939 srcs = [
2940 "test/f32-vscaleextexp.cc",
2941 "test/vscaleextexp-microkernel-tester.h",
2942 ] + MICROKERNEL_TEST_HDRS,
2943 deps = MICROKERNEL_TEST_DEPS,
2944)
2945
2946xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002947 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07002948 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002949 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002950 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002951 ] + MICROKERNEL_TEST_HDRS,
2952 deps = MICROKERNEL_TEST_DEPS,
2953)
2954
2955xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002956 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002957 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002958 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002959 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002960 ] + MICROKERNEL_TEST_HDRS,
2961 deps = MICROKERNEL_TEST_DEPS,
2962)
2963
2964xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002965 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002966 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002967 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002968 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002969 ] + MICROKERNEL_TEST_HDRS,
2970 deps = MICROKERNEL_TEST_DEPS,
2971)
2972
2973xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002974 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002975 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002976 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977 "test/avgpool-microkernel-tester.h",
2978 "src/xnnpack/AlignedAllocator.h",
2979 ] + MICROKERNEL_TEST_HDRS,
2980 deps = MICROKERNEL_TEST_DEPS,
2981)
2982
2983xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002984 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002986 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 "test/gemm-microkernel-tester.h",
2988 "src/xnnpack/AlignedAllocator.h",
2989 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2990 deps = MICROKERNEL_TEST_DEPS,
2991)
2992
2993xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002994 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002996 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002997 "test/dwconv-microkernel-tester.h",
2998 "src/xnnpack/AlignedAllocator.h",
2999 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3000 deps = MICROKERNEL_TEST_DEPS,
3001)
3002
3003xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003004 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003006 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007 "test/gavgpool-microkernel-tester.h",
3008 "src/xnnpack/AlignedAllocator.h",
3009 ] + MICROKERNEL_TEST_HDRS,
3010 deps = MICROKERNEL_TEST_DEPS,
3011)
3012
3013xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003014 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003015 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003016 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003017 "test/gemm-microkernel-tester.h",
3018 "src/xnnpack/AlignedAllocator.h",
3019 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3020 deps = MICROKERNEL_TEST_DEPS,
3021)
3022
3023xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003024 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003026 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003027 "test/vadd-microkernel-tester.h",
3028 ] + MICROKERNEL_TEST_HDRS,
3029 deps = MICROKERNEL_TEST_DEPS,
3030)
3031
3032xnnpack_unit_test(
3033 name = "u8_clamp_test",
3034 srcs = [
3035 "test/u8-clamp.cc",
3036 "test/clamp-microkernel-tester.h",
3037 ] + MICROKERNEL_TEST_HDRS,
3038 deps = MICROKERNEL_TEST_DEPS,
3039)
3040
3041xnnpack_unit_test(
3042 name = "u8_lut32norm_test",
3043 srcs = [
3044 "test/u8-lut32norm.cc",
3045 "test/lut-norm-microkernel-tester.h",
3046 ] + MICROKERNEL_TEST_HDRS,
3047 deps = MICROKERNEL_TEST_DEPS,
3048)
3049
3050xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003051 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003052 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003053 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003054 "test/maxpool-microkernel-tester.h",
3055 ] + MICROKERNEL_TEST_HDRS,
3056 deps = MICROKERNEL_TEST_DEPS,
3057)
3058
3059xnnpack_unit_test(
3060 name = "u8_rmax_test",
3061 srcs = [
3062 "test/u8-rmax.cc",
3063 "test/rmax-microkernel-tester.h",
3064 ] + MICROKERNEL_TEST_HDRS,
3065 deps = MICROKERNEL_TEST_DEPS,
3066)
3067
3068xnnpack_unit_test(
3069 name = "x32_packx_test",
3070 srcs = [
3071 "test/x32-packx.cc",
3072 "test/pack-microkernel-tester.h",
3073 "src/xnnpack/AlignedAllocator.h",
3074 ] + MICROKERNEL_TEST_HDRS,
3075 deps = MICROKERNEL_TEST_DEPS,
3076)
3077
3078xnnpack_unit_test(
3079 name = "x32_pad_test",
3080 srcs = [
3081 "test/x32-pad.cc",
3082 "test/pad-microkernel-tester.h",
3083 ] + MICROKERNEL_TEST_HDRS,
3084 deps = MICROKERNEL_TEST_DEPS,
3085)
3086
3087xnnpack_unit_test(
3088 name = "x32_unpool_test",
3089 srcs = [
3090 "test/x32-unpool.cc",
3091 "test/unpool-microkernel-tester.h",
3092 ] + MICROKERNEL_TEST_HDRS,
3093 deps = MICROKERNEL_TEST_DEPS,
3094)
3095
3096xnnpack_unit_test(
3097 name = "x32_zip_test",
3098 srcs = [
3099 "test/x32-zip.cc",
3100 "test/zip-microkernel-tester.h",
3101 ] + MICROKERNEL_TEST_HDRS,
3102 deps = MICROKERNEL_TEST_DEPS,
3103)
3104
3105xnnpack_unit_test(
3106 name = "x8_lut_test",
3107 srcs = [
3108 "test/x8-lut.cc",
3109 "test/lut-microkernel-tester.h",
3110 ] + MICROKERNEL_TEST_HDRS,
3111 deps = MICROKERNEL_TEST_DEPS,
3112)
3113
3114xnnpack_unit_test(
3115 name = "x8_zip_test",
3116 srcs = [
3117 "test/x8-zip.cc",
3118 "test/zip-microkernel-tester.h",
3119 ] + MICROKERNEL_TEST_HDRS,
3120 deps = MICROKERNEL_TEST_DEPS,
3121)
3122
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003123xnnpack_unit_test(
3124 name = "requantization_test",
3125 srcs = [
3126 "src/xnnpack/requantization-stubs.h",
3127 "test/requantization.cc",
3128 "test/requantization-tester.h",
3129 ] + MICROKERNEL_TEST_HDRS,
3130 deps = MICROKERNEL_TEST_DEPS,
3131)
3132
Marat Dukhan20c3b922020-03-10 03:45:06 -07003133########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003134
3135xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003136 name = "operator_size_test",
3137 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003138 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003139)
3140
Marat Dukhan20c3b922020-03-10 03:45:06 -07003141xnnpack_binary(
3142 name = "subgraph_size_test",
3143 srcs = ["test/subgraph-size.c"],
3144 deps = [":XNNPACK"],
3145)
3146
3147########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003148
3149xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003150 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003151 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003152 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003153 "test/add-operator-tester.h",
3154 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003155 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156)
3157
3158xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003159 name = "add_nd_test",
3160 srcs = [
3161 "test/add-nd.cc",
3162 "test/binary-elementwise-operator-tester.h",
3163 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003164 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003165)
3166
3167xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003168 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003169 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003170 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003171 "test/argmax-pooling-operator-tester.h",
3172 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003173 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003174)
3175
3176xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003177 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003178 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003179 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180 "test/average-pooling-operator-tester.h",
3181 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003182 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003183)
3184
3185xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003186 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003188 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189 "test/channel-pad-operator-tester.h",
3190 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003191 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003192)
3193
3194xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003195 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003196 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003197 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003198 "test/channel-shuffle-operator-tester.h",
3199 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003200 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003201)
3202
3203xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003204 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003205 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003206 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 "test/clamp-operator-tester.h",
3208 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003209 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210)
3211
3212xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003213 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003214 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003215 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003216 "test/convolution-operator-tester.h",
3217 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003218 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219)
3220
3221xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003222 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003223 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003224 "test/convolution-nchw.cc",
3225 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003226 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003227 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003228)
3229
3230xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003231 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003232 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003233 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003234 "test/deconvolution-operator-tester.h",
3235 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003236 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237)
3238
3239xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003240 name = "divide_nd_test",
3241 srcs = [
3242 "test/binary-elementwise-operator-tester.h",
3243 "test/divide-nd.cc",
3244 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003245 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003246)
3247
3248xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003249 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003251 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003252 "test/fully-connected-operator-tester.h",
3253 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003254 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003255)
3256
3257xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003258 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003259 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003260 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003261 "test/global-average-pooling-operator-tester.h",
3262 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003263 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003264)
3265
3266xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003267 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003269 "test/global-average-pooling-ncw.cc",
3270 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003272 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003273)
3274
3275xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003276 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003277 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003278 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003279 "test/hardswish-operator-tester.h",
3280 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003281 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003282)
3283
3284xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003285 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003286 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003287 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003288 "test/leaky-relu-operator-tester.h",
3289 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003290 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003291)
3292
3293xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003294 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003295 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003296 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003297 "test/max-pooling-operator-tester.h",
3298 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003299 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003300)
3301
3302xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003303 name = "maximum_nd_test",
3304 srcs = [
3305 "test/binary-elementwise-operator-tester.h",
3306 "test/maximum-nd.cc",
3307 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003308 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003309)
3310
3311xnnpack_unit_test(
3312 name = "minimum_nd_test",
3313 srcs = [
3314 "test/binary-elementwise-operator-tester.h",
3315 "test/minimum-nd.cc",
3316 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003317 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003318)
3319
3320xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003321 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003322 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003323 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003324 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003325 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003326 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003327)
3328
3329xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003330 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003331 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003332 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003333 "test/prelu-operator-tester.h",
3334 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003335 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003336)
3337
3338xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003339 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003340 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003341 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003342 "test/resize-bilinear-operator-tester.h",
3343 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003344 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003345)
3346
3347xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003348 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003349 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003350 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003351 "test/sigmoid-operator-tester.h",
3352 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003353 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003354)
3355
3356xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003357 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003358 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003359 "test/softmax-nc.cc",
3360 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003361 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003362 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003363)
3364
3365xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003366 name = "subtract_nd_test",
3367 srcs = [
3368 "test/binary-elementwise-operator-tester.h",
3369 "test/subtract-nd.cc",
3370 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003371 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003372)
3373
3374xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003375 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003376 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003377 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003378 "test/unpooling-operator-tester.h",
3379 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003380 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003381)
3382
3383############################# Build configurations #############################
3384
Marat Dukhanb8642352019-10-30 15:43:02 -07003385# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003386config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003387 name = "xnn_enable_assembly_explicit_true",
3388 define_values = {"xnn_enable_assembly": "true"},
3389)
3390
3391# Disables usage of assembly kernels.
3392config_setting(
3393 name = "xnn_enable_assembly_explicit_false",
3394 define_values = {"xnn_enable_assembly": "false"},
3395)
3396
Marat Dukhan05702cf2020-03-26 15:41:33 -07003397# Disables usage of HMP-aware optimizations.
3398config_setting(
3399 name = "xnn_enable_hmp_explicit_false",
3400 define_values = {"xnn_enable_hmp": "false"},
3401)
3402
Marat Dukhanb8642352019-10-30 15:43:02 -07003403# Builds with -c dbg
3404config_setting(
3405 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003407 "compilation_mode": "dbg",
3408 },
3409)
3410
3411# Builds with -c opt
3412config_setting(
3413 name = "optimized_build",
3414 values = {
3415 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003416 },
3417)
3418
3419config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003420 name = "linux_k8",
3421 values = {"cpu": "k8"},
3422)
3423
3424config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003425 name = "linux_arm",
3426 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003427)
3428
3429config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003430 name = "linux_armhf",
3431 values = {"cpu": "armhf"},
3432)
3433
3434config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003435 name = "linux_aarch64",
3436 values = {"cpu": "aarch64"},
3437)
3438
3439config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003440 name = "android",
3441 values = {"crosstool_top": "//external:android/crosstool"},
3442)
3443
3444config_setting(
3445 name = "android_armv7",
3446 values = {
3447 "crosstool_top": "//external:android/crosstool",
3448 "cpu": "armeabi-v7a",
3449 },
3450)
3451
3452config_setting(
3453 name = "android_arm64",
3454 values = {
3455 "crosstool_top": "//external:android/crosstool",
3456 "cpu": "arm64-v8a",
3457 },
3458)
3459
3460config_setting(
3461 name = "android_x86",
3462 values = {
3463 "crosstool_top": "//external:android/crosstool",
3464 "cpu": "x86",
3465 },
3466)
3467
3468config_setting(
3469 name = "android_x86_64",
3470 values = {
3471 "crosstool_top": "//external:android/crosstool",
3472 "cpu": "x86_64",
3473 },
3474)
3475
3476config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003477 name = "windows_x86_64",
3478 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003479)
3480
3481config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003482 name = "windows_x86_64_clang",
3483 values = {
3484 "compiler": "clang-cl",
3485 "cpu": "x64_windows",
3486 },
3487)
3488
3489config_setting(
3490 name = "windows_x86_64_mingw",
3491 values = {
3492 "compiler": "mingw-gcc",
3493 "cpu": "x64_windows",
3494 },
3495)
3496
3497config_setting(
3498 name = "windows_x86_64_msys",
3499 values = {
3500 "compiler": "msys-gcc",
3501 "cpu": "x64_windows",
3502 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003503)
3504
3505config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003506 name = "macos_x86_64",
3507 values = {
3508 "apple_platform_type": "macos",
3509 "cpu": "darwin",
3510 },
3511)
3512
3513config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003514 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003515 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003516)
3517
3518config_setting(
3519 name = "emscripten_wasm",
3520 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003521 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003522 "cpu": "wasm",
3523 },
3524)
3525
3526config_setting(
3527 name = "emscripten_wasmsimd",
3528 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003529 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003531 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003532 },
3533)
3534
3535config_setting(
3536 name = "emscripten_asmjs",
3537 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003538 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539 "cpu": "asmjs",
3540 },
3541)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003542
3543config_setting(
3544 name = "ios_armv7",
3545 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003546 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003547 "cpu": "ios_armv7",
3548 },
3549)
3550
3551config_setting(
3552 name = "ios_arm64",
3553 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003554 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003555 "cpu": "ios_arm64",
3556 },
3557)
3558
3559config_setting(
3560 name = "ios_arm64e",
3561 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003562 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003563 "cpu": "ios_arm64e",
3564 },
3565)
3566
3567config_setting(
3568 name = "ios_x86",
3569 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003570 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003571 "cpu": "ios_i386",
3572 },
3573)
3574
3575config_setting(
3576 name = "ios_x86_64",
3577 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003578 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003579 "cpu": "ios_x86_64",
3580 },
3581)
3582
3583config_setting(
3584 name = "watchos_armv7k",
3585 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003586 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003587 "cpu": "watchos_armv7k",
3588 },
3589)
3590
3591config_setting(
3592 name = "watchos_arm64_32",
3593 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003594 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003595 "cpu": "watchos_arm64_32",
3596 },
3597)
3598
3599config_setting(
3600 name = "watchos_x86",
3601 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003602 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003603 "cpu": "watchos_i386",
3604 },
3605)
3606
3607config_setting(
3608 name = "watchos_x86_64",
3609 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003610 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003611 "cpu": "watchos_x86_64",
3612 },
3613)
3614
3615config_setting(
3616 name = "tvos_arm64",
3617 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003618 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003619 "cpu": "tvos_arm64",
3620 },
3621)
3622
3623config_setting(
3624 name = "tvos_x86_64",
3625 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003626 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003627 "cpu": "tvos_x86_64",
3628 },
3629)