blob: ce0468e6d55c76987cbca4c059825dad768847dc [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",
Frank Barchardb0e4fae2020-05-04 15:27:51 -0700924 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
925 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
926 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
927 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700928 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
929 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
930 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
931 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
932 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
933 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
934 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
935 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700936]
937
938SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700939 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
940 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700941 "src/f32-clamp/gen/sse-x4.c",
942 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800943 "src/f32-dwconv-spchw/3x3p1-sse.c",
944 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700945 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
946 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
947 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
948 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
949 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
950 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
951 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
952 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
953 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
954 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
955 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
956 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700957 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700958 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
959 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700960 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
961 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
962 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
963 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
964 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
965 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
966 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
967 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
968 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
969 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
970 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
971 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
972 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800973 "src/f32-hswish/gen/sse-x4.c",
974 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700975 "src/f32-ibilinear/gen/sse-c4.c",
976 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700977 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
978 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
979 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
980 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
981 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
982 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
983 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700984 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
985 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
986 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700987 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700988 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700989 "src/f32-spmm/gen/4x1-minmax-sse.c",
990 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700991 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
992 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
993 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
994 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
995 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
996 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
997 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
998 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800999 "src/f32-vbinary/gen/vmax-sse-x4.c",
1000 "src/f32-vbinary/gen/vmax-sse-x8.c",
1001 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1002 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1003 "src/f32-vbinary/gen/vmin-sse-x4.c",
1004 "src/f32-vbinary/gen/vmin-sse-x8.c",
1005 "src/f32-vbinary/gen/vminc-sse-x4.c",
1006 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001007 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1008 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1009 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1010 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1011 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1012 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1013 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1014 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1015 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1016 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1017 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1018 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001019 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1020 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001021 "src/x32-packx/x4-sse.c",
1022]
1023
1024SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001025 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1026 "src/f32-argmaxpool/4x-sse2-c4.c",
1027 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001028 "src/f32-prelu/gen/sse2-2x4.c",
1029 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001030 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1031 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1032 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1033 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1034 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1035 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1036 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1037 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1038 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1039 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1040 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1041 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001042 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001043 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001044 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001045 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001046 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1047 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001048 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1049 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001050 "src/q8-igemm/4x4c2-minmax-sse2.c",
1051 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001052 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1053 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001054 "src/q8-gemm/2x4c8-minmax-sse2.c",
1055 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001056 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001057 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001058 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001059 "src/u8-rmax/sse2.c",
1060 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001061 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001062 "src/x32-zip/x2-sse2.c",
1063 "src/x32-zip/x3-sse2.c",
1064 "src/x32-zip/x4-sse2.c",
1065 "src/x32-zip/xm-sse2.c",
1066 "src/x8-zip/x2-sse2.c",
1067 "src/x8-zip/x3-sse2.c",
1068 "src/x8-zip/x4-sse2.c",
1069 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001070 "src/math/exp-sse2-p5.c",
1071 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001072 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001073 "src/requantization/precise-sse2.c",
1074 "src/requantization/fp32-sse2.c",
1075 "src/requantization/q31-sse2.c",
1076 "src/requantization/gemmlowp-sse2.c",
1077]
1078
1079SSSE3_UKERNELS = [
1080 "src/requantization/precise-ssse3.c",
1081 "src/requantization/q31-ssse3.c",
1082 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001083]
1084
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001085SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001086 "src/f32-prelu/gen/sse41-2x4.c",
1087 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001088 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001089 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001090 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001091 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001092 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1093 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001094 "src/requantization/precise-sse4.c",
1095 "src/requantization/q31-sse4.c",
1096 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001097]
1098
Marat Dukhan08c4a432019-10-03 09:29:21 -07001099AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001100 "src/f32-clamp/gen/avx-x8.c",
1101 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001102 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1103 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1104 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1105 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1106 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1107 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1108 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1109 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1110 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1111 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1112 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1113 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1114 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1115 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1116 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1117 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1118 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1119 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1120 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1121 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1122 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1123 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1124 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1125 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1126 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1127 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1128 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1129 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1130 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1131 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001132 "src/f32-hswish/gen/avx-x8.c",
1133 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001134 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1135 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1136 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1137 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1138 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1139 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1140 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1141 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1142 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001143 "src/f32-prelu/gen/avx-2x8.c",
1144 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001145 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001146 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1147 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1148 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1149 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1150 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1151 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1152 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1153 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001154 "src/f32-vbinary/gen/vmax-avx-x8.c",
1155 "src/f32-vbinary/gen/vmax-avx-x16.c",
1156 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1157 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1158 "src/f32-vbinary/gen/vmin-avx-x8.c",
1159 "src/f32-vbinary/gen/vmin-avx-x16.c",
1160 "src/f32-vbinary/gen/vminc-avx-x8.c",
1161 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001162 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1163 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1164 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1165 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1166 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1167 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1168 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1169 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1170 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1171 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1172 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1173 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001174 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001175]
1176
Marat Dukhanfda12b82019-11-21 12:27:59 -08001177FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001178 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1179 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1180 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1181 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1182 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1183 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1184 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1185 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1186 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1187 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1188 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1189 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1190 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1191 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1192 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1193 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1194 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1195 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1196 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1197 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1198 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1199 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1200 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1201 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1202 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1203 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1204 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1205 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1206 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1207 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1208 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1209 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1210 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1211 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1212 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1213 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1214 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1215 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1216 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1217 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001218 "src/f32-hswish/gen/fma3-x8.c",
1219 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001220 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1221 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1222 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1223 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1224 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1225 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1226 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1227 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1228 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1229 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1230 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1231 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1232 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1233 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001234]
1235
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001236AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001237 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1238 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1239 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1240 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1241 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1242 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1243 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1244 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1245 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1246 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1247 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1248 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1249 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1250 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1251 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1252 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1253 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1254 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1255 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1256 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1257 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1258 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1259 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1260 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1261 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1262 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1263 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1264 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1265 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1266 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1267 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1268 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1269 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1270 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1271 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1272 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001273 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1274 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1275 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1276 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1277 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1278 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1279 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1280 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1281 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1282 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1283 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1284 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1285 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1286 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1287 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1288 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1289 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1290 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1291 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1292 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1293 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1294 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1295 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1296 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1297 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1298 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1299 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1300 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1301 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1302 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001303 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1304 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1305 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1306 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1307 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1308 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1309 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1310 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1311 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1312 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1313 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1314 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1315 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1316 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1317 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1318 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1319 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1320 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1321 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1322 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1323 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1324 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1325 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1326 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001327 "src/math/exp-avx2-p5.c",
1328 "src/math/exp-avx2-perm-p3.c",
1329 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001330 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001331 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001332 "src/math/sigmoid-avx2-rr2-p5-div.c",
1333 "src/math/sigmoid-avx2-rr1-p5-div.c",
1334 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1335 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1336 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1337 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001338]
1339
Marat Dukhan08c4a432019-10-03 09:29:21 -07001340AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001341 "src/f32-clamp/gen/avx512f-x16.c",
1342 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001343 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1344 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1345 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1346 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1347 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1348 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1349 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1350 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1351 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1352 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1353 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1354 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1355 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1356 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1357 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1358 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1359 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1360 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1361 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1362 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1363 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1364 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1365 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1366 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001367 "src/f32-hswish/gen/avx512f-x16.c",
1368 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001369 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1370 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1371 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1372 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1373 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1374 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001375 "src/f32-prelu/gen/avx512f-2x16.c",
1376 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001377 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1378 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1379 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1380 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1381 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1382 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1383 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1384 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1385 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1386 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1387 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1388 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1389 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1390 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1391 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1392 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1393 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1394 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1395 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1396 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1397 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1398 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1399 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1400 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1401 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1402 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1403 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1404 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1405 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1406 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1407 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1408 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1409 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1410 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1411 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1412 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001413 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001414 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1415 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1416 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1417 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1418 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1419 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1420 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1421 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001422 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1423 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1424 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1425 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1426 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1427 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1428 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1429 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001430 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1431 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1432 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1433 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1434 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1435 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1436 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1437 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1438 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1439 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1440 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1441 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001442 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001443 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1444 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1445 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1446 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1447 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1448 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1449 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1450 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1451 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1452 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1453 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1454 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1455 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1456 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1457 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1458 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1459 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1460 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1461 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1462 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1463 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1464 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1465 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1466 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001467 "src/math/exp-avx512f-p5-scalef.c",
1468 "src/math/exp-avx512f-p5.c",
1469 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001470 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001471 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001472]
1473
1474AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001475 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001476 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1477 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1478 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1479 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1480 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1481 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1482 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1483 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1484 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1485 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001486]
1487
1488AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001489 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001490 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1491 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001492 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001493 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1494 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001495 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1496 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1497 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001498 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001499 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1500 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1501 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001502 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001503 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1504 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001505 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001506 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1507 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1508 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1509 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1510 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1511 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1512 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1513 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1514 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1515 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1516 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1517 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1518 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1519 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1520 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1521 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1522 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1523 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1524 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1525 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1526 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001527 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001528 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1529 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1530 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1531 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1532 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1533 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1534 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1535 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1536 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1537 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1538 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1539 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1540 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1541 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1542 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1543 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1544 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1545 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1546 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1547 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1548 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1549 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1550 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1551 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1552 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1553 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1554 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1555 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1556 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1557 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1558 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1559 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1560 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1561 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1562 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1563 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1564 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1565 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001566]
1567
Marat Dukhan1b354632020-03-23 12:50:22 -07001568INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001569 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001570 "src/xnnpack/argmaxpool.h",
1571 "src/xnnpack/avgpool.h",
1572 "src/xnnpack/clamp.h",
1573 "src/xnnpack/common.h",
1574 "src/xnnpack/conv.h",
1575 "src/xnnpack/dwconv.h",
1576 "src/xnnpack/gavgpool.h",
1577 "src/xnnpack/gemm.h",
1578 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001579 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001580 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001581 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/xnnpack/lut.h",
1583 "src/xnnpack/math.h",
1584 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001585 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001586 "src/xnnpack/packx.h",
1587 "src/xnnpack/pad.h",
1588 "src/xnnpack/params.h",
1589 "src/xnnpack/pavgpool.h",
1590 "src/xnnpack/ppmm.h",
1591 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001592 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001593 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001594 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595 "src/xnnpack/rmax.h",
1596 "src/xnnpack/scalar-utils.h",
1597 "src/xnnpack/spmm.h",
1598 "src/xnnpack/unpool.h",
1599 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001600 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001601 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001602 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001603 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001604 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001605 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001606 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001607]
1608
1609INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001610 "include/xnnpack.h",
1611 "src/xnnpack/allocator.h",
1612 "src/xnnpack/compute.h",
1613 "src/xnnpack/im2col.h",
1614 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001615 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001616 "src/xnnpack/operator.h",
1617 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001618 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001619 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001620 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001621 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001622]
1623
Marat Dukhan1b354632020-03-23 12:50:22 -07001624ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001625 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001626]
1627
Marat Dukhan1b354632020-03-23 12:50:22 -07001628MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001629 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001630 "include/xnnpack.h",
1631]
1632
Marat Dukhan1b354632020-03-23 12:50:22 -07001633MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001634 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001635 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001636 "src/xnnpack/requantization.h",
1637 "include/xnnpack.h",
1638]
1639
1640OPERATOR_TEST_PARAMS_HDRS = [
1641 "src/xnnpack/params.h",
1642 "src/xnnpack/common.h",
1643]
1644
1645WEIGHTS_PACK_HDRS = [
1646 "src/xnnpack/pack.h",
1647 "src/xnnpack/operator.h",
1648 "src/xnnpack/compute.h",
1649]
1650
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001651LOGGING_COPTS = select({
1652 # No logging in optimized mode
1653 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1654 # Full logging in debug mode
1655 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1656 # Error-only logging in default (fastbuild) mode
1657 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1658})
1659
1660LOGGING_HDRS = [
1661 "src/xnnpack/log.h",
1662]
1663
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001665 name = "tables",
1666 srcs = TABLE_SRCS,
1667 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001668 gcc_copts = xnnpack_gcc_std_copts(),
1669 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001670)
1671
1672xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001673 name = "scalar_ukernels",
1674 srcs = SCALAR_UKERNELS,
1675 hdrs = INTERNAL_HDRS,
1676 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001677 gcc_copts = xnnpack_gcc_std_copts(),
1678 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001679 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001680 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001681 "@FP16",
1682 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001683 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001684 ],
1685)
1686
1687xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001688 name = "wasm_ukernels",
1689 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001690 gcc_copts = xnnpack_gcc_std_copts(),
1691 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001692 wasm_srcs = WASM_UKERNELS,
1693 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001694 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001695 "@FP16",
1696 "@FXdiv",
1697 "@pthreadpool",
1698 ],
1699)
1700
1701xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001702 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001703 hdrs = INTERNAL_HDRS,
1704 aarch32_copts = [
1705 "-marm",
1706 "-mfpu=neon",
1707 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001708 gcc_copts = xnnpack_gcc_std_copts(),
1709 gcc_x86_copts = ["-msse2"],
1710 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001711 optimized_copts = [
1712 "-O3",
1713 "-ffast-math",
1714 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001715 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001716 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001717 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001718 "@FP16",
1719 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001720 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721 ],
1722)
1723
1724xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001725 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001726 hdrs = INTERNAL_HDRS,
1727 aarch32_copts = [
1728 "-marm",
1729 "-mfpu=neon",
1730 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001731 gcc_copts = xnnpack_gcc_std_copts(),
1732 gcc_x86_copts = ["-msse2"],
1733 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001734 optimized_copts = [
1735 "-O3",
1736 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001737 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001738 deps = [
1739 ":tables",
1740 "@FP16",
1741 "@psimd",
1742 "@pthreadpool",
1743 ],
1744)
1745
1746xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747 name = "neon_ukernels",
1748 hdrs = INTERNAL_HDRS,
1749 aarch32_copts = [
1750 "-marm",
1751 "-mfpu=neon",
1752 ],
1753 aarch32_srcs = NEON_UKERNELS,
1754 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001755 gcc_copts = xnnpack_gcc_std_copts(),
1756 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001757 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001758 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001759 "@FP16",
1760 "@pthreadpool",
1761 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001762)
1763
1764xnnpack_cc_library(
1765 name = "neonfma_ukernels",
1766 hdrs = INTERNAL_HDRS,
1767 aarch32_copts = [
1768 "-marm",
1769 "-mfpu=neon-vfpv4",
1770 ],
1771 aarch32_srcs = NEONFMA_UKERNELS,
1772 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001773 gcc_copts = xnnpack_gcc_std_copts(),
1774 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001775 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001776 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001777 "@FP16",
1778 "@pthreadpool",
1779 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001780)
1781
1782xnnpack_cc_library(
1783 name = "neonfp16arith_ukernels",
1784 hdrs = INTERNAL_HDRS,
1785 aarch64_copts = ["-march=armv8.2-a+fp16"],
1786 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001787 gcc_copts = xnnpack_gcc_std_copts(),
1788 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001789 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001790 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001791 "@FP16",
1792 "@pthreadpool",
1793 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001794)
1795
1796xnnpack_cc_library(
1797 name = "sse2_ukernels",
1798 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001799 gcc_copts = xnnpack_gcc_std_copts(),
1800 gcc_x86_copts = ["-msse2"],
1801 msvc_copts = xnnpack_msvc_std_copts(),
1802 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001803 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001804 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001805 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001806 "@FP16",
1807 "@pthreadpool",
1808 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001809)
1810
1811xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001812 name = "ssse3_ukernels",
1813 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001814 gcc_copts = xnnpack_gcc_std_copts(),
1815 gcc_x86_copts = ["-mssse3"],
1816 msvc_copts = xnnpack_msvc_std_copts(),
1817 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001818 x86_srcs = SSSE3_UKERNELS,
1819 deps = [
1820 ":tables",
1821 "@FP16",
1822 "@pthreadpool",
1823 ],
1824)
1825
1826xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001827 name = "sse41_ukernels",
1828 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001829 gcc_copts = xnnpack_gcc_std_copts(),
1830 gcc_x86_copts = ["-msse4.1"],
1831 msvc_copts = xnnpack_msvc_std_copts(),
1832 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001833 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001834 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001835 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001836 "@FP16",
1837 "@pthreadpool",
1838 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001839)
1840
1841xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001842 name = "avx_ukernels",
1843 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001844 gcc_copts = xnnpack_gcc_std_copts(),
1845 gcc_x86_copts = ["-mavx"],
1846 msvc_copts = xnnpack_msvc_std_copts(),
1847 msvc_x86_32_copts = ["/arch:AVX"],
1848 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001849 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001850 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001851 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001852 "@FP16",
1853 "@pthreadpool",
1854 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001855)
1856
1857xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001858 name = "fma3_ukernels",
1859 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001860 gcc_copts = xnnpack_gcc_std_copts(),
1861 gcc_x86_copts = ["-mfma"],
1862 msvc_copts = xnnpack_msvc_std_copts(),
1863 msvc_x86_32_copts = ["/arch:AVX"],
1864 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001865 x86_srcs = FMA3_UKERNELS,
1866 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001867 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001868 "@FP16",
1869 "@pthreadpool",
1870 ],
1871)
1872
1873xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001874 name = "avx2_ukernels",
1875 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001876 gcc_copts = xnnpack_gcc_std_copts(),
1877 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001878 "-mfma",
1879 "-mavx2",
1880 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001881 msvc_copts = xnnpack_msvc_std_copts(),
1882 msvc_x86_32_copts = ["/arch:AVX2"],
1883 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001884 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001885 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001886 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001887 "@FP16",
1888 "@pthreadpool",
1889 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001890)
1891
1892xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001893 name = "avx512f_ukernels",
1894 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001895 gcc_copts = xnnpack_gcc_std_copts(),
1896 gcc_x86_copts = ["-mavx512f"],
1897 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1898 msvc_copts = xnnpack_msvc_std_copts(),
1899 msvc_x86_32_copts = ["/arch:AVX512"],
1900 msvc_x86_64_copts = ["/arch:AVX512"],
1901 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001902 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001903 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001904 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001905 "@FP16",
1906 "@pthreadpool",
1907 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001908)
1909
1910xnnpack_cc_library(
1911 name = "asm_ukernels",
1912 hdrs = ["src/xnnpack/assembly.h"],
1913 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001914 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001915 aarch64_srcs = AARCH64_ASM_UKERNELS,
1916)
1917
1918xnnpack_aggregate_library(
1919 name = "ukernels",
1920 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001921 ":neon_ukernels",
1922 ":neonfma_ukernels",
1923 ":asm_ukernels",
1924 ],
1925 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001926 ":neon_ukernels",
1927 ":neonfma_ukernels",
1928 ":neonfp16arith_ukernels",
1929 ":asm_ukernels",
1930 ],
1931 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07001932 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001933 ":psimd_fastmath_ukernels",
1934 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001935 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001936 wasm_deps = [":wasm_ukernels"],
1937 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001938 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001939 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001940 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001941 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001942 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001943 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001944 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001945 ":avx512f_ukernels",
1946 ],
1947)
1948
1949xnnpack_cc_library(
1950 name = "im2col",
1951 srcs = ["src/im2col.c"],
1952 hdrs = [
1953 "src/xnnpack/common.h",
1954 "src/xnnpack/im2col.h",
1955 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001956 gcc_copts = xnnpack_gcc_std_copts(),
1957 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001958)
1959
1960xnnpack_cc_library(
1961 name = "indirection",
1962 srcs = ["src/indirection.c"],
1963 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001964 gcc_copts = xnnpack_gcc_std_copts(),
1965 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001966 deps = [
1967 "@FP16",
1968 "@FXdiv",
1969 "@pthreadpool",
1970 ],
1971)
1972
1973xnnpack_cc_library(
1974 name = "operator_run",
1975 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001976 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001977 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07001978 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1979 "//conditions:default": [],
1980 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07001981 gcc_copts = xnnpack_gcc_std_copts(),
1982 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001983 deps = [
1984 "@FP16",
1985 "@FXdiv",
1986 "@clog",
1987 "@pthreadpool",
1988 ],
1989)
1990
1991cc_library(
1992 name = "enable_assembly",
1993 defines = select({
1994 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1995 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001996 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997 }),
1998)
1999
Marat Dukhancf056b22019-10-07 10:26:29 -07002000xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002001 name = "operators",
2002 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002003 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002004 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002005 ],
2006 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002007 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002008 "-Isrc",
2009 "-Iinclude",
2010 ] + select({
2011 ":debug_build": [],
2012 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002013 }) + select({
2014 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2015 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002016 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002017 gcc_copts = xnnpack_gcc_std_copts(),
2018 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002019 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002020 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002021 "@FP16",
2022 "@FXdiv",
2023 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002024 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002025 ],
2026)
2027
Marat Dukhan10a38082020-04-17 03:58:35 -07002028xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002029 name = "XNNPACK",
2030 srcs = [
2031 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002032 "src/runtime.c",
2033 "src/subgraph.c",
2034 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002035 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002036 hdrs = ["include/xnnpack.h"],
2037 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002038 "-Isrc",
2039 "-Iinclude",
2040 ] + select({
2041 ":debug_build": [],
2042 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002043 }) + select({
2044 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2045 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002046 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002047 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002048 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002049 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002050 visibility = xnnpack_visibility(),
2051 deps = [
2052 ":enable_assembly",
2053 ":ukernels",
2054 ":operator_run",
2055 ":operators",
2056 "@clog",
2057 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002058 ] + select({
2059 ":emscripten": [],
2060 "//conditions:default": ["@cpuinfo"],
2061 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002062)
2063
Marat Dukhan10a38082020-04-17 03:58:35 -07002064xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002065 name = "xnnpack_operators_nhwc_f32",
2066 srcs = [
2067 "src/init.c",
2068 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002069 hdrs = ["include/xnnpack.h"],
2070 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002071 "-Isrc",
2072 "-Iinclude",
2073 ] + select({
2074 ":debug_build": [],
2075 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002076 }) + select({
2077 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2078 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002079 }),
2080 defines = [
2081 "XNN_NO_Q8_OPERATORS",
2082 "XNN_NO_U8_OPERATORS",
2083 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002084 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002085 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002086 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002087 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002088 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002089 visibility = xnnpack_visibility(),
2090 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002091 ":enable_assembly",
2092 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002093 ":operator_run",
2094 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002095 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002096 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002097 ] + select({
2098 ":emscripten": [],
2099 "//conditions:default": ["@cpuinfo"],
2100 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002101)
2102
Marat Dukhancf056b22019-10-07 10:26:29 -07002103xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002104 name = "bench_utils",
2105 srcs = ["bench/utils.cc"],
2106 hdrs = ["bench/utils.h"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002107 gcc_copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002108 deps = [
2109 "@com_google_benchmark//:benchmark",
2110 "@cpuinfo",
2111 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002112)
2113
Frank Barchard7e955972019-10-11 10:34:25 -07002114######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002115
2116xnnpack_benchmark(
2117 name = "q8_gemm_bench",
2118 srcs = [
2119 "bench/gemm.h",
2120 "bench/q8-gemm.cc",
2121 "src/xnnpack/AlignedAllocator.h",
2122 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002123 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
2124 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002125 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2126)
2127
2128xnnpack_benchmark(
2129 name = "f16_gemm_bench",
2130 srcs = [
2131 "bench/f16-gemm.cc",
2132 "bench/gemm.h",
2133 "src/xnnpack/AlignedAllocator.h",
2134 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002135 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002136 deps = MICROKERNEL_BENCHMARK_DEPS,
2137)
2138
2139xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002140 name = "f16_spmm_bench",
2141 srcs = [
2142 "bench/f16-spmm.cc",
2143 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002144 "src/xnnpack/AlignedAllocator.h",
2145 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002146 gcc_copts = ["-Wno-unused-function"],
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002147 deps = MICROKERNEL_BENCHMARK_DEPS,
2148)
2149
2150xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002151 name = "f32_igemm_bench",
2152 srcs = [
2153 "bench/f32-igemm.cc",
2154 "bench/conv.h",
2155 "src/xnnpack/AlignedAllocator.h",
2156 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002157 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002158)
2159
2160xnnpack_benchmark(
2161 name = "f32_conv_hwc_bench",
2162 srcs = [
2163 "bench/f32-conv-hwc.cc",
2164 "bench/dconv.h",
2165 "src/xnnpack/AlignedAllocator.h",
2166 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002167 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002168 deps = MICROKERNEL_BENCHMARK_DEPS,
2169)
2170
2171xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002172 name = "f32_conv_hwc2spchw_bench",
2173 srcs = [
2174 "bench/f32-conv-hwc2spchw.cc",
2175 "bench/dconv.h",
2176 "src/xnnpack/AlignedAllocator.h",
2177 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002178 gcc_copts = ["-Wno-unused-function"],
Erich Elsen563df5f2019-10-23 08:02:21 -07002179 deps = MICROKERNEL_BENCHMARK_DEPS,
2180)
2181
2182xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183 name = "f32_dwconv_bench",
2184 srcs = [
2185 "bench/f32-dwconv.cc",
2186 "bench/dwconv.h",
2187 "src/xnnpack/AlignedAllocator.h",
2188 ] + MICROKERNEL_BENCHMARK_HDRS,
2189 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2190)
2191
2192xnnpack_benchmark(
2193 name = "f32_dwconv_spchw_bench",
2194 srcs = [
2195 "bench/f32-dwconv-spchw.cc",
2196 "bench/dwconv.h",
2197 "src/xnnpack/AlignedAllocator.h",
2198 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2199 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2200)
2201
2202xnnpack_benchmark(
2203 name = "f32_gemm_bench",
2204 srcs = [
2205 "bench/f32-gemm.cc",
2206 "bench/gemm.h",
2207 "src/xnnpack/AlignedAllocator.h",
2208 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002209 copts = xnnpack_optional_ruy_copts(),
2210 gcc_copts = ["-Wno-unused-function"],
Frank Barchard7e955972019-10-11 10:34:25 -07002211 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002212)
2213
2214xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002215 name = "f32_raddexpminusmax_bench",
2216 srcs = [
2217 "bench/f32-raddexpminusmax.cc",
2218 "src/xnnpack/AlignedAllocator.h",
2219 ] + MICROKERNEL_BENCHMARK_HDRS,
2220 deps = MICROKERNEL_BENCHMARK_DEPS,
2221)
2222
2223xnnpack_benchmark(
2224 name = "f32_raddextexp_bench",
2225 srcs = [
2226 "bench/f32-raddextexp.cc",
2227 "src/xnnpack/AlignedAllocator.h",
2228 ] + MICROKERNEL_BENCHMARK_HDRS,
2229 deps = MICROKERNEL_BENCHMARK_DEPS,
2230)
2231
2232xnnpack_benchmark(
2233 name = "f32_raddstoreexpminusmax_bench",
2234 srcs = [
2235 "bench/f32-raddstoreexpminusmax.cc",
2236 "src/xnnpack/AlignedAllocator.h",
2237 ] + MICROKERNEL_BENCHMARK_HDRS,
2238 deps = MICROKERNEL_BENCHMARK_DEPS,
2239)
2240
2241xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002242 name = "f32_rmax_bench",
2243 srcs = [
2244 "bench/f32-rmax.cc",
2245 "src/xnnpack/AlignedAllocator.h",
2246 ] + MICROKERNEL_BENCHMARK_HDRS,
2247 deps = MICROKERNEL_BENCHMARK_DEPS,
2248)
2249
2250xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002251 name = "f32_sigmoid_bench",
2252 srcs = [
2253 "bench/f32-sigmoid.cc",
2254 "src/xnnpack/AlignedAllocator.h",
2255 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002256 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan14bec502019-11-18 11:35:31 -08002257 deps = MICROKERNEL_BENCHMARK_DEPS,
2258)
2259
2260xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002261 name = "f32_spmm_bench",
2262 srcs = [
2263 "bench/f32-spmm.cc",
2264 "bench/gemm.h",
2265 "src/xnnpack/AlignedAllocator.h",
2266 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002267 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268 deps = MICROKERNEL_BENCHMARK_DEPS,
2269)
2270
2271xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002272 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002273 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002274 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002275 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002276 copts = xnnpack_optional_dnnl_copts(),
2277 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002278 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002279)
2280
2281xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002282 name = "f32_vscaleexpminusmax_bench",
2283 srcs = [
2284 "bench/f32-vscaleexpminusmax.cc",
2285 "src/xnnpack/AlignedAllocator.h",
2286 ] + MICROKERNEL_BENCHMARK_HDRS,
2287 deps = MICROKERNEL_BENCHMARK_DEPS,
2288)
2289
2290xnnpack_benchmark(
2291 name = "f32_vscaleextexp_bench",
2292 srcs = [
2293 "bench/f32-vscaleextexp.cc",
2294 "src/xnnpack/AlignedAllocator.h",
2295 ] + MICROKERNEL_BENCHMARK_HDRS,
2296 deps = MICROKERNEL_BENCHMARK_DEPS,
2297)
2298
2299xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002300 name = "f32_im2col_gemm_bench",
2301 srcs = [
2302 "bench/f32-im2col-gemm.cc",
2303 "bench/conv.h",
2304 "src/xnnpack/AlignedAllocator.h",
2305 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2306 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2307)
2308
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002309xnnpack_benchmark(
2310 name = "requantization_bench",
2311 srcs = [
2312 "bench/requantization.cc",
2313 "src/xnnpack/requantization-stubs.h",
2314 "src/xnnpack/AlignedAllocator.h",
2315 ] + MICROKERNEL_BENCHMARK_HDRS,
2316 deps = MICROKERNEL_BENCHMARK_DEPS,
2317)
2318
Marat Dukhan08c4a432019-10-03 09:29:21 -07002319########################### Benchmarks for operators ###########################
2320
2321xnnpack_benchmark(
2322 name = "add_bench",
2323 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002324 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325)
2326
2327xnnpack_benchmark(
2328 name = "average_pooling_bench",
2329 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002330 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002331 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002332 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002333)
2334
2335xnnpack_benchmark(
2336 name = "channel_shuffle_bench",
2337 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002338 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002339)
2340
2341xnnpack_benchmark(
2342 name = "convolution_bench",
2343 srcs = ["bench/convolution.cc"],
2344 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002345 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002346 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002347)
2348
2349xnnpack_benchmark(
2350 name = "deconvolution_bench",
2351 srcs = ["bench/deconvolution.cc"],
2352 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002353 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002354 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002355)
2356
2357xnnpack_benchmark(
2358 name = "global_average_pooling_bench",
2359 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002360 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002361)
2362
2363xnnpack_benchmark(
2364 name = "max_pooling_bench",
2365 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002366 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002367)
2368
2369xnnpack_benchmark(
2370 name = "sigmoid_bench",
2371 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002372 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002373 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002374 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002375)
2376
2377xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002378 name = "prelu_bench",
2379 srcs = ["bench/prelu.cc"],
2380 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002381 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002382 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002383)
2384
2385xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002386 name = "softmax_bench",
2387 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002388 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002389 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002390 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391)
2392
Marat Dukhanc068bb62019-10-04 13:24:39 -07002393############################# End-to-end benchmarks ############################
2394
2395cc_library(
2396 name = "mobilenet_v1",
2397 srcs = ["models/mobilenet-v1.cc"],
2398 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002399 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002400 linkstatic = True,
2401 deps = [
2402 ":XNNPACK",
2403 "@pthreadpool",
2404 ],
2405)
2406
2407cc_library(
2408 name = "mobilenet_v2",
2409 srcs = ["models/mobilenet-v2.cc"],
2410 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002411 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002412 linkstatic = True,
2413 deps = [
2414 ":XNNPACK",
2415 "@pthreadpool",
2416 ],
2417)
2418
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002419cc_library(
2420 name = "mobilenet_v3_large",
2421 srcs = ["models/mobilenet-v3-large.cc"],
2422 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002423 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002424 linkstatic = True,
2425 deps = [
2426 ":XNNPACK",
2427 "@pthreadpool",
2428 ],
2429)
2430
2431cc_library(
2432 name = "mobilenet_v3_small",
2433 srcs = ["models/mobilenet-v3-small.cc"],
2434 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002435 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002436 linkstatic = True,
2437 deps = [
2438 ":XNNPACK",
2439 "@pthreadpool",
2440 ],
2441)
2442
Marat Dukhanc068bb62019-10-04 13:24:39 -07002443xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002444 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002445 srcs = [
2446 "bench/f32-dwconv-e2e.cc",
2447 "bench/end2end.h",
2448 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002449 gcc_copts = ["-Wno-unused-function"],
Marat Dukhanef4416e2019-10-31 13:44:40 -07002450 deps = MICROKERNEL_BENCHMARK_DEPS + [
2451 ":XNNPACK",
2452 ":mobilenet_v1",
2453 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002454 ":mobilenet_v3_large",
2455 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002456 ],
2457)
2458
2459xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002460 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002461 srcs = [
2462 "bench/f32-gemm-e2e.cc",
2463 "bench/end2end.h",
2464 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002465 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan5f18d262019-10-31 10:24:14 -07002466 deps = MICROKERNEL_BENCHMARK_DEPS + [
2467 ":XNNPACK",
2468 ":mobilenet_v1",
2469 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002470 ":mobilenet_v3_large",
2471 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002472 ],
2473)
2474
2475xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002476 name = "end2end_bench",
2477 srcs = ["bench/end2end.cc"],
2478 deps = [
2479 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002480 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002481 ":mobilenet_v1",
2482 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002483 ":mobilenet_v3_large",
2484 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002485 "@pthreadpool",
2486 ],
2487)
2488
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002489#################### Accuracy evaluation for math functions ####################
2490
2491xnnpack_benchmark(
2492 name = "f32_exp_eval",
2493 srcs = [
2494 "eval/f32-exp.cc",
2495 "src/xnnpack/AlignedAllocator.h",
2496 ] + ACCURACY_EVAL_HDRS,
2497 deps = ACCURACY_EVAL_DEPS,
2498)
2499
Marat Dukhan515c9772019-10-17 18:07:57 -07002500xnnpack_benchmark(
2501 name = "f32_expminus_eval",
2502 srcs = [
2503 "eval/f32-expminus.cc",
2504 "src/xnnpack/AlignedAllocator.h",
2505 ] + ACCURACY_EVAL_HDRS,
2506 deps = ACCURACY_EVAL_DEPS,
2507)
2508
Marat Dukhan98ba4412019-10-23 02:14:28 -07002509xnnpack_benchmark(
2510 name = "f32_extexp_eval",
2511 srcs = [
2512 "eval/f32-extexp.cc",
2513 "src/xnnpack/AlignedAllocator.h",
2514 ] + ACCURACY_EVAL_HDRS,
2515 deps = ACCURACY_EVAL_DEPS,
2516)
2517
Marat Dukhan346a9e52019-11-15 09:06:30 -08002518xnnpack_benchmark(
2519 name = "f32_sigmoid_eval",
2520 srcs = [
2521 "eval/f32-sigmoid.cc",
2522 "src/xnnpack/AlignedAllocator.h",
2523 ] + ACCURACY_EVAL_HDRS,
2524 deps = ACCURACY_EVAL_DEPS,
2525)
2526
Marat Dukhan08c4a432019-10-03 09:29:21 -07002527######################### Unit tests for micro-kernels #########################
2528
2529xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002530 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002531 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002532 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002533 "test/gemm-microkernel-tester.h",
2534 "src/xnnpack/AlignedAllocator.h",
2535 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2536 deps = MICROKERNEL_TEST_DEPS,
2537)
2538
2539xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002540 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002541 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002542 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002543 "test/spmm-microkernel-tester.h",
2544 "src/xnnpack/AlignedAllocator.h",
2545 ] + MICROKERNEL_TEST_HDRS,
2546 deps = MICROKERNEL_TEST_DEPS,
2547)
2548
2549xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002550 name = "f32_argmaxpool_test",
2551 srcs = [
2552 "test/f32-argmaxpool.cc",
2553 "test/argmaxpool-microkernel-tester.h",
2554 "src/xnnpack/AlignedAllocator.h",
2555 ] + MICROKERNEL_TEST_HDRS,
2556 deps = MICROKERNEL_TEST_DEPS,
2557)
2558
2559xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002560 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002561 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002562 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002563 "test/avgpool-microkernel-tester.h",
2564 "src/xnnpack/AlignedAllocator.h",
2565 ] + MICROKERNEL_TEST_HDRS,
2566 deps = MICROKERNEL_TEST_DEPS,
2567)
2568
2569xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002570 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002571 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002572 "test/f32-ibilinear.cc",
2573 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002574 "src/xnnpack/AlignedAllocator.h",
2575 ] + MICROKERNEL_TEST_HDRS,
2576 deps = MICROKERNEL_TEST_DEPS,
2577)
2578
2579xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002580 name = "f32_clamp_test",
2581 srcs = [
2582 "test/f32-clamp.cc",
2583 "test/clamp-microkernel-tester.h",
2584 ] + MICROKERNEL_TEST_HDRS,
2585 deps = MICROKERNEL_TEST_DEPS,
2586)
2587
2588xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002589 name = "f32_igemm_test",
2590 srcs = [
2591 "test/f32-igemm.cc",
2592 "test/gemm-microkernel-tester.h",
2593 "src/xnnpack/AlignedAllocator.h",
2594 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2595 deps = MICROKERNEL_TEST_DEPS,
2596)
2597
2598xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002599 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002600 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002601 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002602 "test/gemm-microkernel-tester.h",
2603 "src/xnnpack/AlignedAllocator.h",
2604 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002605 # Work-around for "too many sections" error
2606 mingw_copts = ["-Wa,-mbig-obj"],
2607 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002608 deps = MICROKERNEL_TEST_DEPS,
2609)
2610
2611xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002612 name = "f16_igemm_minmax_test",
2613 srcs = [
2614 "test/f16-igemm-minmax.cc",
2615 "test/gemm-microkernel-tester.h",
2616 "src/xnnpack/AlignedAllocator.h",
2617 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2618 deps = MICROKERNEL_TEST_DEPS,
2619)
2620
2621xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002622 name = "f32_conv_hwc_test",
2623 srcs = [
2624 "test/f32-conv-hwc.cc",
2625 "test/conv-hwc-microkernel-tester.h",
2626 "src/xnnpack/AlignedAllocator.h",
2627 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2628 deps = MICROKERNEL_TEST_DEPS,
2629)
2630
2631xnnpack_unit_test(
2632 name = "f32_conv_hwc2spchw_test",
2633 srcs = [
2634 "test/f32-conv-hwc2spchw.cc",
2635 "test/conv-hwc2spchw-microkernel-tester.h",
2636 "src/xnnpack/AlignedAllocator.h",
2637 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2638 deps = MICROKERNEL_TEST_DEPS,
2639)
2640
2641xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002642 name = "f32_dwconv_test",
2643 srcs = [
2644 "test/f32-dwconv.cc",
2645 "test/dwconv-microkernel-tester.h",
2646 "src/xnnpack/AlignedAllocator.h",
2647 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2648 deps = MICROKERNEL_TEST_DEPS,
2649)
2650
2651xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002652 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002653 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002654 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002655 "test/dwconv-microkernel-tester.h",
2656 "src/xnnpack/AlignedAllocator.h",
2657 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2658 deps = MICROKERNEL_TEST_DEPS,
2659)
2660
2661xnnpack_unit_test(
2662 name = "f32_dwconv_spchw_test",
2663 srcs = [
2664 "test/f32-dwconv-spchw.cc",
2665 "test/dwconv-spchw-microkernel-tester.h",
2666 "src/xnnpack/AlignedAllocator.h",
2667 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2668 deps = MICROKERNEL_TEST_DEPS,
2669)
2670
2671xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002672 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002673 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002674 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002675 "test/gavgpool-microkernel-tester.h",
2676 "src/xnnpack/AlignedAllocator.h",
2677 ] + MICROKERNEL_TEST_HDRS,
2678 deps = MICROKERNEL_TEST_DEPS,
2679)
2680
2681xnnpack_unit_test(
2682 name = "f32_gavgpool_spchw_test",
2683 srcs = [
2684 "test/f32-gavgpool-spchw.cc",
2685 "test/gavgpool-spchw-microkernel-tester.h",
2686 "src/xnnpack/AlignedAllocator.h",
2687 ] + MICROKERNEL_TEST_HDRS,
2688 deps = MICROKERNEL_TEST_DEPS,
2689)
2690
2691xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002692 name = "f32_gemm_test",
2693 srcs = [
2694 "test/f32-gemm.cc",
2695 "test/gemm-microkernel-tester.h",
2696 "src/xnnpack/AlignedAllocator.h",
2697 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2698 deps = MICROKERNEL_TEST_DEPS,
2699)
2700
2701xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002702 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002703 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002704 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002705 "test/gemm-microkernel-tester.h",
2706 "src/xnnpack/AlignedAllocator.h",
2707 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2708 deps = MICROKERNEL_TEST_DEPS,
2709)
2710
2711xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002712 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002713 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002714 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002715 "test/gemm-microkernel-tester.h",
2716 "src/xnnpack/AlignedAllocator.h",
2717 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2718 deps = MICROKERNEL_TEST_DEPS,
2719)
2720
2721xnnpack_unit_test(
2722 name = "f32_hswish_test",
2723 srcs = [
2724 "test/f32-hswish.cc",
2725 "test/hswish-microkernel-tester.h",
2726 ] + MICROKERNEL_TEST_HDRS,
2727 deps = MICROKERNEL_TEST_DEPS,
2728)
2729
2730xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002731 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002733 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002734 "test/maxpool-microkernel-tester.h",
2735 ] + MICROKERNEL_TEST_HDRS,
2736 deps = MICROKERNEL_TEST_DEPS,
2737)
2738
2739xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002740 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002741 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002742 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002743 "test/avgpool-microkernel-tester.h",
2744 "src/xnnpack/AlignedAllocator.h",
2745 ] + MICROKERNEL_TEST_HDRS,
2746 deps = MICROKERNEL_TEST_DEPS,
2747)
2748
2749xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002750 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002751 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002752 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002753 "test/gemm-microkernel-tester.h",
2754 "src/xnnpack/AlignedAllocator.h",
2755 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2756 deps = MICROKERNEL_TEST_DEPS,
2757)
2758
2759xnnpack_unit_test(
2760 name = "f32_prelu_test",
2761 srcs = [
2762 "test/f32-prelu.cc",
2763 "test/prelu-microkernel-tester.h",
2764 "src/xnnpack/AlignedAllocator.h",
2765 ] + MICROKERNEL_TEST_HDRS,
2766 deps = MICROKERNEL_TEST_DEPS,
2767)
2768
2769xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002770 name = "f32_raddexpminusmax_test",
2771 srcs = [
2772 "test/f32-raddexpminusmax.cc",
2773 "test/raddexpminusmax-microkernel-tester.h",
2774 ] + MICROKERNEL_TEST_HDRS,
2775 deps = MICROKERNEL_TEST_DEPS,
2776)
2777
2778xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002779 name = "f32_raddextexp_test",
2780 srcs = [
2781 "test/f32-raddextexp.cc",
2782 "test/raddextexp-microkernel-tester.h",
2783 ] + MICROKERNEL_TEST_HDRS,
2784 deps = MICROKERNEL_TEST_DEPS,
2785)
2786
2787xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002788 name = "f32_raddstoreexpminusmax_test",
2789 srcs = [
2790 "test/f32-raddstoreexpminusmax.cc",
2791 "test/raddstoreexpminusmax-microkernel-tester.h",
2792 ] + MICROKERNEL_TEST_HDRS,
2793 deps = MICROKERNEL_TEST_DEPS,
2794)
2795
2796xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002797 name = "f32_rmax_test",
2798 srcs = [
2799 "test/f32-rmax.cc",
2800 "test/rmax-microkernel-tester.h",
2801 ] + MICROKERNEL_TEST_HDRS,
2802 deps = MICROKERNEL_TEST_DEPS,
2803)
2804
2805xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002806 name = "f32_sigmoid_test",
2807 srcs = [
2808 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002809 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002810 ] + MICROKERNEL_TEST_HDRS,
2811 deps = MICROKERNEL_TEST_DEPS,
2812)
2813
2814xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002815 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002816 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002817 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002818 "test/spmm-microkernel-tester.h",
2819 "src/xnnpack/AlignedAllocator.h",
2820 ] + MICROKERNEL_TEST_HDRS,
2821 deps = MICROKERNEL_TEST_DEPS,
2822)
2823
2824xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002825 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002826 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002827 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002828 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002829 ] + MICROKERNEL_TEST_HDRS,
2830 deps = MICROKERNEL_TEST_DEPS,
2831)
2832
2833xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002834 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002835 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002836 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002837 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002838 ] + MICROKERNEL_TEST_HDRS,
2839 deps = MICROKERNEL_TEST_DEPS,
2840)
2841
2842xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002843 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002844 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002845 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002846 "test/vbinary-microkernel-tester.h",
2847 ] + MICROKERNEL_TEST_HDRS,
2848 deps = MICROKERNEL_TEST_DEPS,
2849)
2850
2851xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002852 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002853 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002854 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002855 "test/vbinaryc-microkernel-tester.h",
2856 ] + MICROKERNEL_TEST_HDRS,
2857 deps = MICROKERNEL_TEST_DEPS,
2858)
2859
2860xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002861 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002862 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002863 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002864 "test/vbinaryc-microkernel-tester.h",
2865 ] + MICROKERNEL_TEST_HDRS,
2866 deps = MICROKERNEL_TEST_DEPS,
2867)
2868
2869xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002870 name = "f32_vmax_test",
2871 srcs = [
2872 "test/f32-vmax.cc",
2873 "test/vbinary-microkernel-tester.h",
2874 ] + MICROKERNEL_TEST_HDRS,
2875 deps = MICROKERNEL_TEST_DEPS,
2876)
2877
2878xnnpack_unit_test(
2879 name = "f32_vmaxc_test",
2880 srcs = [
2881 "test/f32-vmaxc.cc",
2882 "test/vbinaryc-microkernel-tester.h",
2883 ] + MICROKERNEL_TEST_HDRS,
2884 deps = MICROKERNEL_TEST_DEPS,
2885)
2886
2887xnnpack_unit_test(
2888 name = "f32_vmin_test",
2889 srcs = [
2890 "test/f32-vmin.cc",
2891 "test/vbinary-microkernel-tester.h",
2892 ] + MICROKERNEL_TEST_HDRS,
2893 deps = MICROKERNEL_TEST_DEPS,
2894)
2895
2896xnnpack_unit_test(
2897 name = "f32_vminc_test",
2898 srcs = [
2899 "test/f32-vminc.cc",
2900 "test/vbinaryc-microkernel-tester.h",
2901 ] + MICROKERNEL_TEST_HDRS,
2902 deps = MICROKERNEL_TEST_DEPS,
2903)
2904
2905xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002906 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002907 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002908 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002909 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002910 ] + MICROKERNEL_TEST_HDRS,
2911 deps = MICROKERNEL_TEST_DEPS,
2912)
2913
2914xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002915 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002916 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002917 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002918 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919 ] + MICROKERNEL_TEST_HDRS,
2920 deps = MICROKERNEL_TEST_DEPS,
2921)
2922
2923xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002924 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002925 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002926 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002927 "test/vmulcaddc-microkernel-tester.h",
2928 "src/xnnpack/AlignedAllocator.h",
2929 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2930 deps = MICROKERNEL_TEST_DEPS,
2931)
2932
2933xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002934 name = "f32_vscale_test",
2935 srcs = [
2936 "test/f32-vscale.cc",
2937 "test/vscale-microkernel-tester.h",
2938 ] + MICROKERNEL_TEST_HDRS,
2939 deps = MICROKERNEL_TEST_DEPS,
2940)
2941
2942xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002943 name = "f32_vscaleexpminusmax_test",
2944 srcs = [
2945 "test/f32-vscaleexpminusmax.cc",
2946 "test/vscaleexpminusmax-microkernel-tester.h",
2947 ] + MICROKERNEL_TEST_HDRS,
2948 deps = MICROKERNEL_TEST_DEPS,
2949)
2950
2951xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002952 name = "f32_vscaleextexp_test",
2953 srcs = [
2954 "test/f32-vscaleextexp.cc",
2955 "test/vscaleextexp-microkernel-tester.h",
2956 ] + MICROKERNEL_TEST_HDRS,
2957 deps = MICROKERNEL_TEST_DEPS,
2958)
2959
2960xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002961 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07002962 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002963 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002964 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002965 ] + MICROKERNEL_TEST_HDRS,
2966 deps = MICROKERNEL_TEST_DEPS,
2967)
2968
2969xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002970 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002971 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002972 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002973 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002974 ] + MICROKERNEL_TEST_HDRS,
2975 deps = MICROKERNEL_TEST_DEPS,
2976)
2977
2978xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002979 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002980 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002981 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002982 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002983 ] + MICROKERNEL_TEST_HDRS,
2984 deps = MICROKERNEL_TEST_DEPS,
2985)
2986
2987xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002988 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002989 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002990 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 "test/avgpool-microkernel-tester.h",
2992 "src/xnnpack/AlignedAllocator.h",
2993 ] + MICROKERNEL_TEST_HDRS,
2994 deps = MICROKERNEL_TEST_DEPS,
2995)
2996
2997xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002998 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003000 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003001 "test/gemm-microkernel-tester.h",
3002 "src/xnnpack/AlignedAllocator.h",
3003 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3004 deps = MICROKERNEL_TEST_DEPS,
3005)
3006
3007xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003008 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003010 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011 "test/dwconv-microkernel-tester.h",
3012 "src/xnnpack/AlignedAllocator.h",
3013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3014 deps = MICROKERNEL_TEST_DEPS,
3015)
3016
3017xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003018 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003019 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003020 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021 "test/gavgpool-microkernel-tester.h",
3022 "src/xnnpack/AlignedAllocator.h",
3023 ] + MICROKERNEL_TEST_HDRS,
3024 deps = MICROKERNEL_TEST_DEPS,
3025)
3026
3027xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003028 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003029 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003030 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003031 "test/gemm-microkernel-tester.h",
3032 "src/xnnpack/AlignedAllocator.h",
3033 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3034 deps = MICROKERNEL_TEST_DEPS,
3035)
3036
3037xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003038 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003040 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003041 "test/vadd-microkernel-tester.h",
3042 ] + MICROKERNEL_TEST_HDRS,
3043 deps = MICROKERNEL_TEST_DEPS,
3044)
3045
3046xnnpack_unit_test(
3047 name = "u8_clamp_test",
3048 srcs = [
3049 "test/u8-clamp.cc",
3050 "test/clamp-microkernel-tester.h",
3051 ] + MICROKERNEL_TEST_HDRS,
3052 deps = MICROKERNEL_TEST_DEPS,
3053)
3054
3055xnnpack_unit_test(
3056 name = "u8_lut32norm_test",
3057 srcs = [
3058 "test/u8-lut32norm.cc",
3059 "test/lut-norm-microkernel-tester.h",
3060 ] + MICROKERNEL_TEST_HDRS,
3061 deps = MICROKERNEL_TEST_DEPS,
3062)
3063
3064xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003065 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003066 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003067 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003068 "test/maxpool-microkernel-tester.h",
3069 ] + MICROKERNEL_TEST_HDRS,
3070 deps = MICROKERNEL_TEST_DEPS,
3071)
3072
3073xnnpack_unit_test(
3074 name = "u8_rmax_test",
3075 srcs = [
3076 "test/u8-rmax.cc",
3077 "test/rmax-microkernel-tester.h",
3078 ] + MICROKERNEL_TEST_HDRS,
3079 deps = MICROKERNEL_TEST_DEPS,
3080)
3081
3082xnnpack_unit_test(
3083 name = "x32_packx_test",
3084 srcs = [
3085 "test/x32-packx.cc",
3086 "test/pack-microkernel-tester.h",
3087 "src/xnnpack/AlignedAllocator.h",
3088 ] + MICROKERNEL_TEST_HDRS,
3089 deps = MICROKERNEL_TEST_DEPS,
3090)
3091
3092xnnpack_unit_test(
3093 name = "x32_pad_test",
3094 srcs = [
3095 "test/x32-pad.cc",
3096 "test/pad-microkernel-tester.h",
3097 ] + MICROKERNEL_TEST_HDRS,
3098 deps = MICROKERNEL_TEST_DEPS,
3099)
3100
3101xnnpack_unit_test(
3102 name = "x32_unpool_test",
3103 srcs = [
3104 "test/x32-unpool.cc",
3105 "test/unpool-microkernel-tester.h",
3106 ] + MICROKERNEL_TEST_HDRS,
3107 deps = MICROKERNEL_TEST_DEPS,
3108)
3109
3110xnnpack_unit_test(
3111 name = "x32_zip_test",
3112 srcs = [
3113 "test/x32-zip.cc",
3114 "test/zip-microkernel-tester.h",
3115 ] + MICROKERNEL_TEST_HDRS,
3116 deps = MICROKERNEL_TEST_DEPS,
3117)
3118
3119xnnpack_unit_test(
3120 name = "x8_lut_test",
3121 srcs = [
3122 "test/x8-lut.cc",
3123 "test/lut-microkernel-tester.h",
3124 ] + MICROKERNEL_TEST_HDRS,
3125 deps = MICROKERNEL_TEST_DEPS,
3126)
3127
3128xnnpack_unit_test(
3129 name = "x8_zip_test",
3130 srcs = [
3131 "test/x8-zip.cc",
3132 "test/zip-microkernel-tester.h",
3133 ] + MICROKERNEL_TEST_HDRS,
3134 deps = MICROKERNEL_TEST_DEPS,
3135)
3136
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003137xnnpack_unit_test(
3138 name = "requantization_test",
3139 srcs = [
3140 "src/xnnpack/requantization-stubs.h",
3141 "test/requantization.cc",
3142 "test/requantization-tester.h",
3143 ] + MICROKERNEL_TEST_HDRS,
3144 deps = MICROKERNEL_TEST_DEPS,
3145)
3146
Marat Dukhan20c3b922020-03-10 03:45:06 -07003147########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003148
3149xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003150 name = "operator_size_test",
3151 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003152 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003153)
3154
Marat Dukhan20c3b922020-03-10 03:45:06 -07003155xnnpack_binary(
3156 name = "subgraph_size_test",
3157 srcs = ["test/subgraph-size.c"],
3158 deps = [":XNNPACK"],
3159)
3160
3161########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003162
3163xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003164 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003166 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 "test/add-operator-tester.h",
3168 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003169 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170)
3171
3172xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003173 name = "add_nd_test",
3174 srcs = [
3175 "test/add-nd.cc",
3176 "test/binary-elementwise-operator-tester.h",
3177 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003178 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003179)
3180
3181xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003182 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003183 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003184 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003185 "test/argmax-pooling-operator-tester.h",
3186 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003187 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003188)
3189
3190xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003191 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003192 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003193 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003194 "test/average-pooling-operator-tester.h",
3195 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003196 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003197)
3198
3199xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003200 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003201 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003202 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 "test/channel-pad-operator-tester.h",
3204 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003205 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003206)
3207
3208xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003209 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003211 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212 "test/channel-shuffle-operator-tester.h",
3213 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003214 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003215)
3216
3217xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003218 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003220 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003221 "test/clamp-operator-tester.h",
3222 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003223 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003224)
3225
3226xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003227 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003228 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003229 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003230 "test/convolution-operator-tester.h",
3231 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003232 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003233)
3234
3235xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003236 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003238 "test/convolution-nchw.cc",
3239 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003241 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003242)
3243
3244xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003245 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003246 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003247 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003248 "test/deconvolution-operator-tester.h",
3249 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003250 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003251)
3252
3253xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003254 name = "divide_nd_test",
3255 srcs = [
3256 "test/binary-elementwise-operator-tester.h",
3257 "test/divide-nd.cc",
3258 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003259 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003260)
3261
3262xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003263 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003264 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003265 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003266 "test/fully-connected-operator-tester.h",
3267 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003268 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003269)
3270
3271xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003272 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003273 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003274 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003275 "test/global-average-pooling-operator-tester.h",
3276 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003277 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278)
3279
3280xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003281 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003282 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003283 "test/global-average-pooling-ncw.cc",
3284 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003285 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003286 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003287)
3288
3289xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003290 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003291 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003292 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003293 "test/hardswish-operator-tester.h",
3294 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003295 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003296)
3297
3298xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003299 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003300 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003301 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003302 "test/leaky-relu-operator-tester.h",
3303 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003304 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003305)
3306
3307xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003308 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003309 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003310 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003311 "test/max-pooling-operator-tester.h",
3312 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003313 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003314)
3315
3316xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003317 name = "maximum_nd_test",
3318 srcs = [
3319 "test/binary-elementwise-operator-tester.h",
3320 "test/maximum-nd.cc",
3321 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003322 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003323)
3324
3325xnnpack_unit_test(
3326 name = "minimum_nd_test",
3327 srcs = [
3328 "test/binary-elementwise-operator-tester.h",
3329 "test/minimum-nd.cc",
3330 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003331 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003332)
3333
3334xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003335 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003336 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003337 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003338 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003339 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003340 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003341)
3342
3343xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003344 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003345 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003346 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003347 "test/prelu-operator-tester.h",
3348 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003349 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350)
3351
3352xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003353 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003354 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003355 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003356 "test/resize-bilinear-operator-tester.h",
3357 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003358 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003359)
3360
3361xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003362 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003363 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003364 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003365 "test/sigmoid-operator-tester.h",
3366 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003367 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003368)
3369
3370xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003371 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003373 "test/softmax-nc.cc",
3374 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003376 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377)
3378
3379xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003380 name = "subtract_nd_test",
3381 srcs = [
3382 "test/binary-elementwise-operator-tester.h",
3383 "test/subtract-nd.cc",
3384 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003385 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003386)
3387
3388xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003389 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003390 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003391 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003392 "test/unpooling-operator-tester.h",
3393 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003394 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003395)
3396
3397############################# Build configurations #############################
3398
Marat Dukhanb8642352019-10-30 15:43:02 -07003399# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003400config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003401 name = "xnn_enable_assembly_explicit_true",
3402 define_values = {"xnn_enable_assembly": "true"},
3403)
3404
3405# Disables usage of assembly kernels.
3406config_setting(
3407 name = "xnn_enable_assembly_explicit_false",
3408 define_values = {"xnn_enable_assembly": "false"},
3409)
3410
Marat Dukhan05702cf2020-03-26 15:41:33 -07003411# Disables usage of HMP-aware optimizations.
3412config_setting(
3413 name = "xnn_enable_hmp_explicit_false",
3414 define_values = {"xnn_enable_hmp": "false"},
3415)
3416
Marat Dukhanb8642352019-10-30 15:43:02 -07003417# Builds with -c dbg
3418config_setting(
3419 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003420 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003421 "compilation_mode": "dbg",
3422 },
3423)
3424
3425# Builds with -c opt
3426config_setting(
3427 name = "optimized_build",
3428 values = {
3429 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430 },
3431)
3432
3433config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003434 name = "linux_k8",
3435 values = {"cpu": "k8"},
3436)
3437
3438config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003439 name = "linux_arm",
3440 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003441)
3442
3443config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003444 name = "linux_armhf",
3445 values = {"cpu": "armhf"},
3446)
3447
3448config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003449 name = "linux_aarch64",
3450 values = {"cpu": "aarch64"},
3451)
3452
3453config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003454 name = "android",
3455 values = {"crosstool_top": "//external:android/crosstool"},
3456)
3457
3458config_setting(
3459 name = "android_armv7",
3460 values = {
3461 "crosstool_top": "//external:android/crosstool",
3462 "cpu": "armeabi-v7a",
3463 },
3464)
3465
3466config_setting(
3467 name = "android_arm64",
3468 values = {
3469 "crosstool_top": "//external:android/crosstool",
3470 "cpu": "arm64-v8a",
3471 },
3472)
3473
3474config_setting(
3475 name = "android_x86",
3476 values = {
3477 "crosstool_top": "//external:android/crosstool",
3478 "cpu": "x86",
3479 },
3480)
3481
3482config_setting(
3483 name = "android_x86_64",
3484 values = {
3485 "crosstool_top": "//external:android/crosstool",
3486 "cpu": "x86_64",
3487 },
3488)
3489
3490config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003491 name = "windows_x86_64",
3492 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003493)
3494
3495config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003496 name = "windows_x86_64_clang",
3497 values = {
3498 "compiler": "clang-cl",
3499 "cpu": "x64_windows",
3500 },
3501)
3502
3503config_setting(
3504 name = "windows_x86_64_mingw",
3505 values = {
3506 "compiler": "mingw-gcc",
3507 "cpu": "x64_windows",
3508 },
3509)
3510
3511config_setting(
3512 name = "windows_x86_64_msys",
3513 values = {
3514 "compiler": "msys-gcc",
3515 "cpu": "x64_windows",
3516 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003517)
3518
3519config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003520 name = "macos_x86_64",
3521 values = {
3522 "apple_platform_type": "macos",
3523 "cpu": "darwin",
3524 },
3525)
3526
3527config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003528 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003529 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530)
3531
3532config_setting(
3533 name = "emscripten_wasm",
3534 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003535 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003536 "cpu": "wasm",
3537 },
3538)
3539
3540config_setting(
3541 name = "emscripten_wasmsimd",
3542 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003543 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003544 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003545 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003546 },
3547)
3548
3549config_setting(
3550 name = "emscripten_asmjs",
3551 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003552 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553 "cpu": "asmjs",
3554 },
3555)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003556
3557config_setting(
3558 name = "ios_armv7",
3559 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003560 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003561 "cpu": "ios_armv7",
3562 },
3563)
3564
3565config_setting(
3566 name = "ios_arm64",
3567 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003568 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003569 "cpu": "ios_arm64",
3570 },
3571)
3572
3573config_setting(
3574 name = "ios_arm64e",
3575 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003576 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003577 "cpu": "ios_arm64e",
3578 },
3579)
3580
3581config_setting(
3582 name = "ios_x86",
3583 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003584 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003585 "cpu": "ios_i386",
3586 },
3587)
3588
3589config_setting(
3590 name = "ios_x86_64",
3591 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003592 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003593 "cpu": "ios_x86_64",
3594 },
3595)
3596
3597config_setting(
3598 name = "watchos_armv7k",
3599 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003600 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003601 "cpu": "watchos_armv7k",
3602 },
3603)
3604
3605config_setting(
3606 name = "watchos_arm64_32",
3607 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003608 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003609 "cpu": "watchos_arm64_32",
3610 },
3611)
3612
3613config_setting(
3614 name = "watchos_x86",
3615 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003616 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003617 "cpu": "watchos_i386",
3618 },
3619)
3620
3621config_setting(
3622 name = "watchos_x86_64",
3623 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003624 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003625 "cpu": "watchos_x86_64",
3626 },
3627)
3628
3629config_setting(
3630 name = "tvos_arm64",
3631 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003632 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003633 "cpu": "tvos_arm64",
3634 },
3635)
3636
3637config_setting(
3638 name = "tvos_x86_64",
3639 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003640 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003641 "cpu": "tvos_x86_64",
3642 },
3643)