blob: 2e0b28c511e568d32a410228799f631aba7d50d2 [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",
Marat Dukhan8853b822020-05-07 12:19:01 -0700238 "src/math/roundne-scalar.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800239 "src/math/sigmoid-scalar-lut2048-p1-div.c",
240 "src/math/sigmoid-scalar-lut64-p2-div.c",
241 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700242 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
243 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700244 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700245 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
246 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700247 "src/q8-gemm/2x2-minmax-scalar.c",
248 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700249 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700250 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700251 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700252 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700253 "src/u8-rmax/scalar.c",
254 "src/x32-packx/x2-scalar.c",
255 "src/x32-packx/x3-scalar.c",
256 "src/x32-packx/x4-scalar.c",
257 "src/x32-pad/x2-scalar.c",
258 "src/x32-unpool/scalar.c",
259 "src/x32-zip/x2-scalar.c",
260 "src/x32-zip/x3-scalar.c",
261 "src/x32-zip/x4-scalar.c",
262 "src/x32-zip/xm-scalar.c",
263 "src/x8-lut/scalar.c",
264 "src/x8-zip/x2-scalar.c",
265 "src/x8-zip/x3-scalar.c",
266 "src/x8-zip/x4-scalar.c",
267 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700268 "src/requantization/precise-scalar.c",
269 "src/requantization/fp32-scalar.c",
270 "src/requantization/q31-scalar.c",
271 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700272]
273
Marat Dukhan436ebe62019-12-04 15:10:12 -0800274WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700275 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
276 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700277 "src/f32-clamp/gen/wasm-x1.c",
278 "src/f32-clamp/gen/wasm-x2.c",
279 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700280 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
281 "src/f32-dwconv/gen/up1x4-wasm.c",
282 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
283 "src/f32-dwconv/gen/up1x9-wasm.c",
284 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
285 "src/f32-dwconv/gen/up1x25-wasm.c",
286 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
287 "src/f32-dwconv/gen/up2x4-wasm.c",
288 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
289 "src/f32-dwconv/gen/up2x9-wasm.c",
290 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
291 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700292 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
293 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
294 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
295 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700296 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
297 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700298 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
299 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
300 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
301 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700302 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
303 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700304 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
305 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700306 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
307 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
308 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700309 "src/f32-gemm/gen/1x4-wasm.c",
310 "src/f32-gemm/gen/2x4-wasm.c",
311 "src/f32-gemm/gen/4x2-wasm.c",
312 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700313 "src/f32-gemm/gen/1x4-minmax-wasm.c",
314 "src/f32-gemm/gen/2x4-minmax-wasm.c",
315 "src/f32-gemm/gen/4x2-minmax-wasm.c",
316 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800317 "src/f32-hswish/gen/wasm-x1.c",
318 "src/f32-hswish/gen/wasm-x2.c",
319 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700320 "src/f32-igemm/gen/1x4-wasm.c",
321 "src/f32-igemm/gen/2x4-wasm.c",
322 "src/f32-igemm/gen/4x2-wasm.c",
323 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700324 "src/f32-igemm/gen/1x4-minmax-wasm.c",
325 "src/f32-igemm/gen/2x4-minmax-wasm.c",
326 "src/f32-igemm/gen/4x2-minmax-wasm.c",
327 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700328 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
329 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
330 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700331 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
332 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
333 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
334 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
335 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
336 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
337 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
338 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
339 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
340 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
341 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
342 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vmax-wasm-x1.c",
344 "src/f32-vbinary/gen/vmax-wasm-x2.c",
345 "src/f32-vbinary/gen/vmax-wasm-x4.c",
346 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
347 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
348 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
349 "src/f32-vbinary/gen/vmin-wasm-x1.c",
350 "src/f32-vbinary/gen/vmin-wasm-x2.c",
351 "src/f32-vbinary/gen/vmin-wasm-x4.c",
352 "src/f32-vbinary/gen/vminc-wasm-x1.c",
353 "src/f32-vbinary/gen/vminc-wasm-x2.c",
354 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700355 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
356 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
357 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
358 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
359 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
360 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
361 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
362 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
363 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
364 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
365 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
366 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
367 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
368 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
369 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
370 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
371 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
372 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700373 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
374 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
375 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800376]
377
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800378PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800379 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800380 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800381 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700382 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
383 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700384 "src/f32-clamp/gen/psimd-x4.c",
385 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700386 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
387 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
388 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
389 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
390 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
391 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
392 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
393 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
394 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
395 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
396 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
397 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700398 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
399 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700400 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
401 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
402 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
403 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
404 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
405 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
406 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
407 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
408 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
409 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
410 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
411 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
412 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
413 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
414 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
415 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
416 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
417 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
418 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800419 "src/f32-hswish/gen/psimd-x4.c",
420 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700421 "src/f32-ibilinear/gen/psimd-c4.c",
422 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700423 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
424 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
425 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
426 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
427 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
428 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
429 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
430 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
431 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
432 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700433 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
434 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
435 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700436 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800437 "src/f32-prelu/gen/psimd-2x4.c",
438 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800439 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700440 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
441 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
442 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
443 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
444 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
445 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
446 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
447 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800448 "src/f32-vbinary/gen/vmax-psimd-x4.c",
449 "src/f32-vbinary/gen/vmax-psimd-x8.c",
450 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
451 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
452 "src/f32-vbinary/gen/vmin-psimd-x4.c",
453 "src/f32-vbinary/gen/vmin-psimd-x8.c",
454 "src/f32-vbinary/gen/vminc-psimd-x4.c",
455 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700456 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
457 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
458 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
459 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
460 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
461 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
462 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
463 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
464 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
465 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
466 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
467 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700468 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
469 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700470 "src/x32-packx/x4-psimd.c",
471 "src/x32-pad/x2-psimd.c",
472 "src/x32-unpool/psimd.c",
473 "src/x32-zip/x2-psimd.c",
474 "src/x32-zip/x3-psimd.c",
475 "src/x32-zip/x4-psimd.c",
476 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700477 "src/requantization/precise-psimd.c",
478 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700479]
480
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800481PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800482 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
483 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
484 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
485 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
486 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
487 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
488 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
489 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
490 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
491 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
492 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
493 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800494 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
495 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
496 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
497 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
498 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
499 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700500 "src/math/roundne-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800501 "src/math/sigmoid-psimd-p5-div.c",
502]
503
Marat Dukhan08c4a432019-10-03 09:29:21 -0700504# ISA-specific micro-kernels
505NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700506 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
507 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700508 "src/f32-clamp/gen/neon-x4.c",
509 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700510 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
511 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
512 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
513 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700514 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
515 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
516 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
517 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700518 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
519 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
520 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
521 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700522 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700523 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
524 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700525 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
526 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
527 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
528 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
529 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
530 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
531 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
532 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
533 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
534 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
535 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
536 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
537 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
538 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
539 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
540 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
541 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
542 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
543 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
544 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
545 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
546 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
547 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
548 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
549 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
550 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
551 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
552 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
553 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
554 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
555 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800556 "src/f32-hswish/gen/neon-x4.c",
557 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700558 "src/f32-ibilinear/gen/neon-c4.c",
559 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700560 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
561 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
562 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
563 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
564 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
565 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
566 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
567 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
568 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
569 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
570 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
571 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
572 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
573 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
574 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
575 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700576 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
577 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
578 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700579 "src/f32-ppmm/gen/4x8-minmax-neon.c",
580 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800581 "src/f32-prelu/gen/neon-2x4.c",
582 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800583 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
584 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
585 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
586 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
587 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
588 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
589 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
590 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
591 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
592 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
593 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
594 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
595 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
596 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
597 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
598 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
599 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
600 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
601 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
602 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
603 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
604 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
605 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
606 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700607 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800608 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800609 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
610 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
611 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
612 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
613 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
614 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
615 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
616 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
617 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
618 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
619 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
620 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
621 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
622 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
623 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
624 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
625 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
626 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700627 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
628 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
629 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
630 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800631 "src/f32-vbinary/gen/vmax-neon-x4.c",
632 "src/f32-vbinary/gen/vmax-neon-x8.c",
633 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
634 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
635 "src/f32-vbinary/gen/vmin-neon-x4.c",
636 "src/f32-vbinary/gen/vmin-neon-x8.c",
637 "src/f32-vbinary/gen/vminc-neon-x4.c",
638 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700639 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
640 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
641 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
642 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
643 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
644 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
645 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
646 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
647 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
648 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700649 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
650 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
651 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
652 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700653 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700654 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
655 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700656 "src/q8-gemm/4x8-minmax-neon.c",
657 "src/q8-gemm/8x8-minmax-neon.c",
658 "src/q8-igemm/4x8-minmax-neon.c",
659 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700660 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700661 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700662 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700663 "src/u8-rmax/neon.c",
664 "src/x32-packx/x4-neon-st4.c",
665 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700666 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700667 "src/x32-zip/x2-neon.c",
668 "src/x32-zip/x3-neon.c",
669 "src/x32-zip/x4-neon.c",
670 "src/x32-zip/xm-neon.c",
671 "src/x8-zip/x2-neon.c",
672 "src/x8-zip/x3-neon.c",
673 "src/x8-zip/x4-neon.c",
674 "src/x8-zip/xm-neon.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700675 "src/math/roundne-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800676 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800677 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
678 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
679 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
681 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
682 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700683 "src/requantization/precise-neon.c",
684 "src/requantization/fp32-neon.c",
685 "src/requantization/q31-neon.c",
686 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700687]
688
689NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700690 "src/f32-ibilinear/gen/neonfma-c4.c",
691 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700692 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
693 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
694 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
695 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
696 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
697 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
698 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
699 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
700 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700701 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
702 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
703 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
704 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700705 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
706 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
707 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
708 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700709 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
710 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
711 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
712 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700713 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
714 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
715 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
716 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
717 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
718 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
719 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
720 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
721 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
722 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
723 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
724 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
725 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
726 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
727 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
728 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
729 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
730 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800731 "src/f32-hswish/gen/neonfma-x4.c",
732 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700733 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
734 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800735 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
736 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
737 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
738 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
739 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
740 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
741 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
742 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
743 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
744 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
745 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
746 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
747 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
748 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
749 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
750 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
751 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
752 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
753 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
754 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
755 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
756 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
757 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
758 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800759 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
768 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
769 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
770 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
771 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
772 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
773 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
774 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
775 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
776 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
777 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
778 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
779 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700813 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
814 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800815 "src/math/exp-neonfma-lut64-p2.c",
816 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800817 "src/math/expminus-neonfma-lut2048-p1.c",
818 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800819 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800820 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
821 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
822 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
823 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
824 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
825 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
826 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
827 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
828 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
829 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
830 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
831 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
832 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
833 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
834 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
835 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
836 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
837 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700838]
839
840AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700841 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
842 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
843 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
844 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
845 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
846 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700847 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
848 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
849 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
850 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
851 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
852 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
853 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
854 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
855 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
856 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
857 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
858 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
859 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
860 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
861 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
862 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
863 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
864 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
865 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
866 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700867 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
868 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
869 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
870 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
871 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
872 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
873 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800874 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
875 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
876 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
877 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
878 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
879 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
880 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
881 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
882 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
883 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
884 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
885 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
886 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
887 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
888 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
889 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
890 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
891 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700892 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
893 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
894 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
895 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
896 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
897 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
898 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
899 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
900 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
901 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
902 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
903 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
904 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
905 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
906 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
907 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
908 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
909 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800910 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
911 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
912 "src/math/sigmoid-neonfma-rr1-p5-div.c",
913 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
914 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
915 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700916]
917
Marat Dukhan8853b822020-05-07 12:19:01 -0700918NEONV8_UKERNELS = [
919 "src/math/roundne-neonv8.c",
920]
921
Marat Dukhan08c4a432019-10-03 09:29:21 -0700922AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardd793f6c2020-05-08 13:37:43 -0700923 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
924 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
925 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
926 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
927 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
928 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
929 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
930 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
931 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
932 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
933 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
934 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
935 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
936 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
937 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
938 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
939 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
940 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
941 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
942 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
943 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
944 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
945 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
946 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
947 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
948 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
949 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
950 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard83ea9392020-04-20 20:15:14 -0700951 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
952 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
953 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
954 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
955 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
956 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
957 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
958 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -0700959 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
960 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
961 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
962 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -0700963 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
964 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
965 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
966 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
967 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
968 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
969 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
970 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
971 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
972 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
973 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
974 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700975 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
976 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
977 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
978 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
979 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
980 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
981 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
982 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700983]
984
985SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700986 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
987 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700988 "src/f32-clamp/gen/sse-x4.c",
989 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800990 "src/f32-dwconv-spchw/3x3p1-sse.c",
991 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700992 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
993 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
994 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
995 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
996 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
997 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
998 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
999 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1000 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1001 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1002 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1003 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001004 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001005 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1006 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001007 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1008 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1009 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1010 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1011 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1012 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1013 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1014 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1015 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1016 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1017 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1018 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1019 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001020 "src/f32-hswish/gen/sse-x4.c",
1021 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001022 "src/f32-ibilinear/gen/sse-c4.c",
1023 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001024 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1025 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1026 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1027 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1028 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1029 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1030 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001031 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1032 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1033 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001034 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001035 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001036 "src/f32-spmm/gen/4x1-minmax-sse.c",
1037 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001038 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1039 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1040 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1041 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1042 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1043 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1044 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1045 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001046 "src/f32-vbinary/gen/vmax-sse-x4.c",
1047 "src/f32-vbinary/gen/vmax-sse-x8.c",
1048 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1049 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1050 "src/f32-vbinary/gen/vmin-sse-x4.c",
1051 "src/f32-vbinary/gen/vmin-sse-x8.c",
1052 "src/f32-vbinary/gen/vminc-sse-x4.c",
1053 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001054 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1055 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1056 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1057 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1058 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1059 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1060 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1061 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1062 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1063 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1064 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1065 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001066 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1067 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001068 "src/x32-packx/x4-sse.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001069 "src/math/roundne-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001070]
1071
1072SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001073 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1074 "src/f32-argmaxpool/4x-sse2-c4.c",
1075 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001076 "src/f32-prelu/gen/sse2-2x4.c",
1077 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001078 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1079 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1080 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1081 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1082 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1083 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1084 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1085 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1086 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1087 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1088 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001090 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001091 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001092 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001093 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001094 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1095 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001096 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1097 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001098 "src/q8-igemm/4x4c2-minmax-sse2.c",
1099 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001100 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1101 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001102 "src/q8-gemm/2x4c8-minmax-sse2.c",
1103 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001104 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001105 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001106 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001107 "src/u8-rmax/sse2.c",
1108 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001109 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001110 "src/x32-zip/x2-sse2.c",
1111 "src/x32-zip/x3-sse2.c",
1112 "src/x32-zip/x4-sse2.c",
1113 "src/x32-zip/xm-sse2.c",
1114 "src/x8-zip/x2-sse2.c",
1115 "src/x8-zip/x3-sse2.c",
1116 "src/x8-zip/x4-sse2.c",
1117 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001118 "src/math/exp-sse2-p5.c",
1119 "src/math/expminus-sse2-p5.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001120 "src/math/roundne-sse2.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001121 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001122 "src/requantization/precise-sse2.c",
1123 "src/requantization/fp32-sse2.c",
1124 "src/requantization/q31-sse2.c",
1125 "src/requantization/gemmlowp-sse2.c",
1126]
1127
1128SSSE3_UKERNELS = [
1129 "src/requantization/precise-ssse3.c",
1130 "src/requantization/q31-ssse3.c",
1131 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001132]
1133
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001134SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001135 "src/f32-prelu/gen/sse41-2x4.c",
1136 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001137 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001138 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001139 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001140 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001141 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1142 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001143 "src/math/roundne-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001144 "src/requantization/precise-sse4.c",
1145 "src/requantization/q31-sse4.c",
1146 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001147]
1148
Marat Dukhan08c4a432019-10-03 09:29:21 -07001149AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001150 "src/f32-clamp/gen/avx-x8.c",
1151 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001152 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1153 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1154 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1155 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1156 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1157 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1158 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1159 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1160 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1161 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1162 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1163 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1164 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1165 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1166 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1167 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1168 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1169 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1170 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1171 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1172 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1173 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1174 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1175 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1176 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1177 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1178 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1179 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1180 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1181 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001182 "src/f32-hswish/gen/avx-x8.c",
1183 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001184 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1185 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1186 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1187 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1188 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1189 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1190 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1191 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1192 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001193 "src/f32-prelu/gen/avx-2x8.c",
1194 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001195 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001196 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1197 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1198 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1199 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1200 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1201 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1202 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1203 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001204 "src/f32-vbinary/gen/vmax-avx-x8.c",
1205 "src/f32-vbinary/gen/vmax-avx-x16.c",
1206 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1207 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1208 "src/f32-vbinary/gen/vmin-avx-x8.c",
1209 "src/f32-vbinary/gen/vmin-avx-x16.c",
1210 "src/f32-vbinary/gen/vminc-avx-x8.c",
1211 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001212 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1213 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1214 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1215 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1216 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1217 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1218 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1219 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1220 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1221 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1222 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1223 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001224 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001225]
1226
Marat Dukhanfda12b82019-11-21 12:27:59 -08001227FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001228 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1229 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1230 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1231 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1232 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1233 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1234 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1235 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1236 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1237 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1238 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1239 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1240 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1241 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1242 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1243 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1244 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1245 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1246 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1247 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1248 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1249 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1250 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1251 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1252 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1253 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1254 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1255 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1256 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1257 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1258 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1259 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1260 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1261 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1262 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1263 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1264 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1265 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1266 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1267 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001268 "src/f32-hswish/gen/fma3-x8.c",
1269 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001270 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1271 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1272 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1273 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1274 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1275 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1276 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1277 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1278 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1279 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1280 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1281 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1282 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1283 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001284]
1285
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001286AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001287 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1288 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1289 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1290 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1291 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1292 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1293 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1294 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1295 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1296 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1297 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1298 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1299 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1300 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1301 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1302 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1303 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1304 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1305 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1306 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1307 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1308 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1309 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1310 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1311 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1312 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1313 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1314 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1315 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1316 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1317 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1318 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1319 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1320 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1321 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1322 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001323 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1324 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1325 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1326 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1327 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1328 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1329 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1330 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1331 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1332 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1333 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1334 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1335 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1336 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1337 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1338 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1339 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1340 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1341 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1342 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1343 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1344 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1345 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1346 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1347 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1348 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1349 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1350 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1351 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1352 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001353 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1354 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1355 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1356 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1357 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1358 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1359 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1360 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1361 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1362 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1363 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1364 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1365 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1366 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1367 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1368 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1369 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1370 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1371 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1372 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1373 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1374 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1375 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1376 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001377 "src/math/exp-avx2-p5.c",
1378 "src/math/exp-avx2-perm-p3.c",
1379 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001380 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001381 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001382 "src/math/sigmoid-avx2-rr2-p5-div.c",
1383 "src/math/sigmoid-avx2-rr1-p5-div.c",
1384 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1385 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1386 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1387 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001388]
1389
Marat Dukhan08c4a432019-10-03 09:29:21 -07001390AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001391 "src/f32-clamp/gen/avx512f-x16.c",
1392 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001393 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1394 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1395 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1396 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1397 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1398 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1399 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1400 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1401 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1402 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1403 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1404 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1405 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1406 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1407 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1408 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1409 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1410 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1411 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1412 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1413 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1414 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1415 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1416 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001417 "src/f32-hswish/gen/avx512f-x16.c",
1418 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001419 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1420 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1421 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1422 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1423 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1424 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001425 "src/f32-prelu/gen/avx512f-2x16.c",
1426 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001427 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1428 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1429 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1430 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1431 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1432 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1433 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1434 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1435 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1436 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1437 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1438 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1439 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1440 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1441 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1442 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1443 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1444 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1445 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1446 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1447 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1448 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1449 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1450 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1451 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1452 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1453 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1454 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1455 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1456 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1457 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1458 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1459 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1460 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1461 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1462 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001463 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001464 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1465 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1466 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1467 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1468 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1469 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1470 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1471 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001472 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1473 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1474 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1475 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1476 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1477 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1478 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1479 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001480 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1481 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1482 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1483 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1484 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1485 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1486 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1487 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1488 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1489 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1490 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1491 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001492 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001493 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1494 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1495 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1496 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1497 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1498 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1499 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1500 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1501 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1502 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1503 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1504 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1505 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1506 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1507 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1508 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1509 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1510 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1511 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1512 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1513 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1514 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1515 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1516 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001517 "src/math/exp-avx512f-p5-scalef.c",
1518 "src/math/exp-avx512f-p5.c",
1519 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001520 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001521 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001522]
1523
1524AARCH32_ASM_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001525 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1526 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1527 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1528 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1529 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1530 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1531 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1532 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1533 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1534 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001535]
1536
1537AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001538 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001539 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1540 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001541 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001542 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1543 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001544 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1545 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1546 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001547 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001548 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1549 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1550 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001551 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001552 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1553 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001554 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001555 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1556 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1557 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1558 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1559 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1560 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1561 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1562 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1563 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1564 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1565 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1566 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1567 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1568 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1569 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1570 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1571 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1572 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1573 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1574 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1575 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001576 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001577 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1578 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1579 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1580 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1581 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1582 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1583 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1584 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1585 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1586 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1587 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1588 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1589 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1590 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1591 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1592 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1593 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1594 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1595 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1596 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1597 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1598 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1599 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1600 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1601 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1602 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1603 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1604 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1605 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1606 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1607 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1608 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1609 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1610 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1611 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1612 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1613 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1614 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001615]
1616
Marat Dukhan1b354632020-03-23 12:50:22 -07001617INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001618 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001619 "src/xnnpack/argmaxpool.h",
1620 "src/xnnpack/avgpool.h",
1621 "src/xnnpack/clamp.h",
1622 "src/xnnpack/common.h",
1623 "src/xnnpack/conv.h",
1624 "src/xnnpack/dwconv.h",
1625 "src/xnnpack/gavgpool.h",
1626 "src/xnnpack/gemm.h",
1627 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001628 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001629 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001630 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001631 "src/xnnpack/lut.h",
1632 "src/xnnpack/math.h",
1633 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001634 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001635 "src/xnnpack/packx.h",
1636 "src/xnnpack/pad.h",
1637 "src/xnnpack/params.h",
1638 "src/xnnpack/pavgpool.h",
1639 "src/xnnpack/ppmm.h",
1640 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001641 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001642 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001643 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001644 "src/xnnpack/rmax.h",
1645 "src/xnnpack/scalar-utils.h",
1646 "src/xnnpack/spmm.h",
1647 "src/xnnpack/unpool.h",
1648 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001649 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001650 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001651 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001652 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001653 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001654 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001655 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001656]
1657
1658INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001659 "include/xnnpack.h",
1660 "src/xnnpack/allocator.h",
1661 "src/xnnpack/compute.h",
1662 "src/xnnpack/im2col.h",
1663 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001664 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001665 "src/xnnpack/operator.h",
1666 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001667 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001668 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001669 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001670 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001671]
1672
Marat Dukhan1b354632020-03-23 12:50:22 -07001673ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001674 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001675]
1676
Marat Dukhan1b354632020-03-23 12:50:22 -07001677MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001678 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001679 "include/xnnpack.h",
1680]
1681
Marat Dukhan1b354632020-03-23 12:50:22 -07001682MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001683 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001684 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001685 "src/xnnpack/requantization.h",
1686 "include/xnnpack.h",
1687]
1688
1689OPERATOR_TEST_PARAMS_HDRS = [
1690 "src/xnnpack/params.h",
1691 "src/xnnpack/common.h",
1692]
1693
1694WEIGHTS_PACK_HDRS = [
1695 "src/xnnpack/pack.h",
1696 "src/xnnpack/operator.h",
1697 "src/xnnpack/compute.h",
1698]
1699
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001700LOGGING_COPTS = select({
1701 # No logging in optimized mode
1702 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1703 # Full logging in debug mode
1704 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1705 # Error-only logging in default (fastbuild) mode
1706 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1707})
1708
1709LOGGING_HDRS = [
1710 "src/xnnpack/log.h",
1711]
1712
Marat Dukhan08c4a432019-10-03 09:29:21 -07001713xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001714 name = "tables",
1715 srcs = TABLE_SRCS,
1716 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001717 gcc_copts = xnnpack_gcc_std_copts(),
1718 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001719)
1720
1721xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001722 name = "scalar_ukernels",
1723 srcs = SCALAR_UKERNELS,
1724 hdrs = INTERNAL_HDRS,
1725 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001726 gcc_copts = xnnpack_gcc_std_copts(),
1727 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001728 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001729 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730 "@FP16",
1731 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001732 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001733 ],
1734)
1735
1736xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001737 name = "wasm_ukernels",
1738 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001739 gcc_copts = xnnpack_gcc_std_copts(),
1740 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001741 wasm_srcs = WASM_UKERNELS,
1742 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001743 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001744 "@FP16",
1745 "@FXdiv",
1746 "@pthreadpool",
1747 ],
1748)
1749
1750xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001751 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001752 hdrs = INTERNAL_HDRS,
1753 aarch32_copts = [
1754 "-marm",
1755 "-mfpu=neon",
1756 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001757 gcc_copts = xnnpack_gcc_std_copts(),
1758 gcc_x86_copts = ["-msse2"],
1759 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001760 optimized_copts = [
1761 "-O3",
1762 "-ffast-math",
1763 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001764 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001765 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001766 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001767 "@FP16",
1768 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001769 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001770 ],
1771)
1772
1773xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001774 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001775 hdrs = INTERNAL_HDRS,
1776 aarch32_copts = [
1777 "-marm",
1778 "-mfpu=neon",
1779 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001780 gcc_copts = xnnpack_gcc_std_copts(),
1781 gcc_x86_copts = ["-msse2"],
1782 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001783 optimized_copts = [
1784 "-O3",
1785 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001786 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001787 deps = [
1788 ":tables",
1789 "@FP16",
1790 "@psimd",
1791 "@pthreadpool",
1792 ],
1793)
1794
1795xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001796 name = "neon_ukernels",
1797 hdrs = INTERNAL_HDRS,
1798 aarch32_copts = [
1799 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001800 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001801 "-mfpu=neon",
1802 ],
1803 aarch32_srcs = NEON_UKERNELS,
1804 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001805 gcc_copts = xnnpack_gcc_std_copts(),
1806 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001807 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001808 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001809 "@FP16",
1810 "@pthreadpool",
1811 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001812)
1813
1814xnnpack_cc_library(
1815 name = "neonfma_ukernels",
1816 hdrs = INTERNAL_HDRS,
1817 aarch32_copts = [
1818 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001819 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820 "-mfpu=neon-vfpv4",
1821 ],
1822 aarch32_srcs = NEONFMA_UKERNELS,
1823 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001824 gcc_copts = xnnpack_gcc_std_copts(),
1825 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001826 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001827 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001828 "@FP16",
1829 "@pthreadpool",
1830 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001831)
1832
1833xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001834 name = "neonv8_ukernels",
1835 hdrs = INTERNAL_HDRS,
1836 aarch32_copts = [
1837 "-marm",
1838 "-march=armv8-a",
1839 "-mfpu=neon-fp-armv8",
1840 ],
1841 aarch32_srcs = NEONV8_UKERNELS,
1842 aarch64_srcs = NEONV8_UKERNELS,
1843 gcc_copts = xnnpack_gcc_std_copts(),
1844 msvc_copts = xnnpack_msvc_std_copts(),
1845 deps = [
1846 ":tables",
1847 "@FP16",
1848 "@pthreadpool",
1849 ],
1850)
1851
1852xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001853 name = "neonfp16arith_ukernels",
1854 hdrs = INTERNAL_HDRS,
1855 aarch64_copts = ["-march=armv8.2-a+fp16"],
1856 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001857 gcc_copts = xnnpack_gcc_std_copts(),
1858 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001859 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001860 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001861 "@FP16",
1862 "@pthreadpool",
1863 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001864)
1865
1866xnnpack_cc_library(
1867 name = "sse2_ukernels",
1868 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001869 gcc_copts = xnnpack_gcc_std_copts(),
1870 gcc_x86_copts = ["-msse2"],
1871 msvc_copts = xnnpack_msvc_std_copts(),
1872 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001873 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001874 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001875 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001876 "@FP16",
1877 "@pthreadpool",
1878 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001879)
1880
1881xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001882 name = "ssse3_ukernels",
1883 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001884 gcc_copts = xnnpack_gcc_std_copts(),
1885 gcc_x86_copts = ["-mssse3"],
1886 msvc_copts = xnnpack_msvc_std_copts(),
1887 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001888 x86_srcs = SSSE3_UKERNELS,
1889 deps = [
1890 ":tables",
1891 "@FP16",
1892 "@pthreadpool",
1893 ],
1894)
1895
1896xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001897 name = "sse41_ukernels",
1898 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001899 gcc_copts = xnnpack_gcc_std_copts(),
1900 gcc_x86_copts = ["-msse4.1"],
1901 msvc_copts = xnnpack_msvc_std_copts(),
1902 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001903 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001904 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001905 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001906 "@FP16",
1907 "@pthreadpool",
1908 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001909)
1910
1911xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001912 name = "avx_ukernels",
1913 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001914 gcc_copts = xnnpack_gcc_std_copts(),
1915 gcc_x86_copts = ["-mavx"],
1916 msvc_copts = xnnpack_msvc_std_copts(),
1917 msvc_x86_32_copts = ["/arch:AVX"],
1918 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001919 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001920 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001921 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001922 "@FP16",
1923 "@pthreadpool",
1924 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001925)
1926
1927xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001928 name = "fma3_ukernels",
1929 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001930 gcc_copts = xnnpack_gcc_std_copts(),
1931 gcc_x86_copts = ["-mfma"],
1932 msvc_copts = xnnpack_msvc_std_copts(),
1933 msvc_x86_32_copts = ["/arch:AVX"],
1934 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001935 x86_srcs = FMA3_UKERNELS,
1936 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001937 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001938 "@FP16",
1939 "@pthreadpool",
1940 ],
1941)
1942
1943xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001944 name = "avx2_ukernels",
1945 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001946 gcc_copts = xnnpack_gcc_std_copts(),
1947 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001948 "-mfma",
1949 "-mavx2",
1950 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001951 msvc_copts = xnnpack_msvc_std_copts(),
1952 msvc_x86_32_copts = ["/arch:AVX2"],
1953 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001954 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001955 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001956 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001957 "@FP16",
1958 "@pthreadpool",
1959 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001960)
1961
1962xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001963 name = "avx512f_ukernels",
1964 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001965 gcc_copts = xnnpack_gcc_std_copts(),
1966 gcc_x86_copts = ["-mavx512f"],
1967 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1968 msvc_copts = xnnpack_msvc_std_copts(),
1969 msvc_x86_32_copts = ["/arch:AVX512"],
1970 msvc_x86_64_copts = ["/arch:AVX512"],
1971 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001972 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001973 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001974 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001975 "@FP16",
1976 "@pthreadpool",
1977 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001978)
1979
1980xnnpack_cc_library(
1981 name = "asm_ukernels",
1982 hdrs = ["src/xnnpack/assembly.h"],
1983 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001984 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001985 aarch64_srcs = AARCH64_ASM_UKERNELS,
1986)
1987
1988xnnpack_aggregate_library(
1989 name = "ukernels",
1990 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001991 ":neon_ukernels",
1992 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07001993 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001994 ":asm_ukernels",
1995 ],
1996 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997 ":neon_ukernels",
1998 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07001999 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002000 ":neonfp16arith_ukernels",
2001 ":asm_ukernels",
2002 ],
2003 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07002004 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002005 ":psimd_fastmath_ukernels",
2006 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002007 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002008 wasm_deps = [":wasm_ukernels"],
2009 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002010 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002011 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002012 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002013 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002014 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002015 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002016 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002017 ":avx512f_ukernels",
2018 ],
2019)
2020
2021xnnpack_cc_library(
2022 name = "im2col",
2023 srcs = ["src/im2col.c"],
2024 hdrs = [
2025 "src/xnnpack/common.h",
2026 "src/xnnpack/im2col.h",
2027 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002028 gcc_copts = xnnpack_gcc_std_copts(),
2029 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002030)
2031
2032xnnpack_cc_library(
2033 name = "indirection",
2034 srcs = ["src/indirection.c"],
2035 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002036 gcc_copts = xnnpack_gcc_std_copts(),
2037 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002038 deps = [
2039 "@FP16",
2040 "@FXdiv",
2041 "@pthreadpool",
2042 ],
2043)
2044
2045xnnpack_cc_library(
2046 name = "operator_run",
2047 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002048 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002049 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002050 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2051 "//conditions:default": [],
2052 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002053 gcc_copts = xnnpack_gcc_std_copts(),
2054 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002055 deps = [
2056 "@FP16",
2057 "@FXdiv",
2058 "@clog",
2059 "@pthreadpool",
2060 ],
2061)
2062
2063cc_library(
2064 name = "enable_assembly",
2065 defines = select({
2066 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2067 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002068 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002069 }),
2070)
2071
Marat Dukhancf056b22019-10-07 10:26:29 -07002072xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002073 name = "operators",
2074 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002075 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002076 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002077 ],
2078 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002079 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002080 "-Isrc",
2081 "-Iinclude",
2082 ] + select({
2083 ":debug_build": [],
2084 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002085 }) + select({
2086 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2087 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002088 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002089 gcc_copts = xnnpack_gcc_std_copts(),
2090 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002091 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002092 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002093 "@FP16",
2094 "@FXdiv",
2095 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002096 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002097 ],
2098)
2099
Marat Dukhan10a38082020-04-17 03:58:35 -07002100xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002101 name = "XNNPACK",
2102 srcs = [
2103 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002104 "src/runtime.c",
2105 "src/subgraph.c",
2106 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002107 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002108 hdrs = ["include/xnnpack.h"],
2109 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002110 "-Isrc",
2111 "-Iinclude",
2112 ] + select({
2113 ":debug_build": [],
2114 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002115 }) + select({
2116 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2117 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002118 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002119 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002120 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002121 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002122 visibility = xnnpack_visibility(),
2123 deps = [
2124 ":enable_assembly",
2125 ":ukernels",
2126 ":operator_run",
2127 ":operators",
2128 "@clog",
2129 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002130 ] + select({
2131 ":emscripten": [],
2132 "//conditions:default": ["@cpuinfo"],
2133 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002134)
2135
Marat Dukhan10a38082020-04-17 03:58:35 -07002136xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002137 name = "xnnpack_operators_nhwc_f32",
2138 srcs = [
2139 "src/init.c",
2140 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002141 hdrs = ["include/xnnpack.h"],
2142 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002143 "-Isrc",
2144 "-Iinclude",
2145 ] + select({
2146 ":debug_build": [],
2147 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002148 }) + select({
2149 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2150 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002151 }),
2152 defines = [
2153 "XNN_NO_Q8_OPERATORS",
2154 "XNN_NO_U8_OPERATORS",
2155 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002156 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002157 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002158 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002159 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002160 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002161 visibility = xnnpack_visibility(),
2162 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002163 ":enable_assembly",
2164 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002165 ":operator_run",
2166 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002167 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002168 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002169 ] + select({
2170 ":emscripten": [],
2171 "//conditions:default": ["@cpuinfo"],
2172 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002173)
2174
Marat Dukhancf056b22019-10-07 10:26:29 -07002175xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002176 name = "bench_utils",
2177 srcs = ["bench/utils.cc"],
2178 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002179 deps = [
2180 "@com_google_benchmark//:benchmark",
2181 "@cpuinfo",
2182 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183)
2184
Frank Barchard7e955972019-10-11 10:34:25 -07002185######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002186
2187xnnpack_benchmark(
2188 name = "q8_gemm_bench",
2189 srcs = [
2190 "bench/gemm.h",
2191 "bench/q8-gemm.cc",
2192 "src/xnnpack/AlignedAllocator.h",
2193 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002194 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002195 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2196)
2197
2198xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002199 name = "f16_igemm_bench",
2200 srcs = [
2201 "bench/f16-igemm.cc",
2202 "bench/conv.h",
2203 "bench/google/conv.h",
2204 "src/xnnpack/AlignedAllocator.h",
2205 ] + MICROKERNEL_BENCHMARK_HDRS,
2206 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2207)
2208
2209xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002210 name = "f16_gemm_bench",
2211 srcs = [
2212 "bench/f16-gemm.cc",
2213 "bench/gemm.h",
2214 "src/xnnpack/AlignedAllocator.h",
2215 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002216 deps = MICROKERNEL_BENCHMARK_DEPS,
2217)
2218
2219xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002220 name = "f16_spmm_bench",
2221 srcs = [
2222 "bench/f16-spmm.cc",
2223 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002224 "src/xnnpack/AlignedAllocator.h",
2225 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002226 deps = MICROKERNEL_BENCHMARK_DEPS,
2227)
2228
2229xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002230 name = "f32_igemm_bench",
2231 srcs = [
2232 "bench/f32-igemm.cc",
2233 "bench/conv.h",
2234 "src/xnnpack/AlignedAllocator.h",
2235 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002236 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002237)
2238
2239xnnpack_benchmark(
2240 name = "f32_conv_hwc_bench",
2241 srcs = [
2242 "bench/f32-conv-hwc.cc",
2243 "bench/dconv.h",
2244 "src/xnnpack/AlignedAllocator.h",
2245 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002246 deps = MICROKERNEL_BENCHMARK_DEPS,
2247)
2248
2249xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002250 name = "f32_conv_hwc2spchw_bench",
2251 srcs = [
2252 "bench/f32-conv-hwc2spchw.cc",
2253 "bench/dconv.h",
2254 "src/xnnpack/AlignedAllocator.h",
2255 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002256 deps = MICROKERNEL_BENCHMARK_DEPS,
2257)
2258
2259xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002260 name = "f32_dwconv_bench",
2261 srcs = [
2262 "bench/f32-dwconv.cc",
2263 "bench/dwconv.h",
2264 "src/xnnpack/AlignedAllocator.h",
2265 ] + MICROKERNEL_BENCHMARK_HDRS,
2266 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2267)
2268
2269xnnpack_benchmark(
2270 name = "f32_dwconv_spchw_bench",
2271 srcs = [
2272 "bench/f32-dwconv-spchw.cc",
2273 "bench/dwconv.h",
2274 "src/xnnpack/AlignedAllocator.h",
2275 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2276 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2277)
2278
2279xnnpack_benchmark(
2280 name = "f32_gemm_bench",
2281 srcs = [
2282 "bench/f32-gemm.cc",
2283 "bench/gemm.h",
2284 "src/xnnpack/AlignedAllocator.h",
2285 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002286 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002287 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002288)
2289
2290xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002291 name = "f32_raddexpminusmax_bench",
2292 srcs = [
2293 "bench/f32-raddexpminusmax.cc",
2294 "src/xnnpack/AlignedAllocator.h",
2295 ] + MICROKERNEL_BENCHMARK_HDRS,
2296 deps = MICROKERNEL_BENCHMARK_DEPS,
2297)
2298
2299xnnpack_benchmark(
2300 name = "f32_raddextexp_bench",
2301 srcs = [
2302 "bench/f32-raddextexp.cc",
2303 "src/xnnpack/AlignedAllocator.h",
2304 ] + MICROKERNEL_BENCHMARK_HDRS,
2305 deps = MICROKERNEL_BENCHMARK_DEPS,
2306)
2307
2308xnnpack_benchmark(
2309 name = "f32_raddstoreexpminusmax_bench",
2310 srcs = [
2311 "bench/f32-raddstoreexpminusmax.cc",
2312 "src/xnnpack/AlignedAllocator.h",
2313 ] + MICROKERNEL_BENCHMARK_HDRS,
2314 deps = MICROKERNEL_BENCHMARK_DEPS,
2315)
2316
2317xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002318 name = "f32_rmax_bench",
2319 srcs = [
2320 "bench/f32-rmax.cc",
2321 "src/xnnpack/AlignedAllocator.h",
2322 ] + MICROKERNEL_BENCHMARK_HDRS,
2323 deps = MICROKERNEL_BENCHMARK_DEPS,
2324)
2325
2326xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002327 name = "f32_sigmoid_bench",
2328 srcs = [
2329 "bench/f32-sigmoid.cc",
2330 "src/xnnpack/AlignedAllocator.h",
2331 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002332 deps = MICROKERNEL_BENCHMARK_DEPS,
2333)
2334
2335xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002336 name = "f32_spmm_bench",
2337 srcs = [
2338 "bench/f32-spmm.cc",
2339 "bench/gemm.h",
2340 "src/xnnpack/AlignedAllocator.h",
2341 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002342 deps = MICROKERNEL_BENCHMARK_DEPS,
2343)
2344
2345xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002346 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002347 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002348 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002349 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002350 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002351 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002352)
2353
2354xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002355 name = "f32_vscaleexpminusmax_bench",
2356 srcs = [
2357 "bench/f32-vscaleexpminusmax.cc",
2358 "src/xnnpack/AlignedAllocator.h",
2359 ] + MICROKERNEL_BENCHMARK_HDRS,
2360 deps = MICROKERNEL_BENCHMARK_DEPS,
2361)
2362
2363xnnpack_benchmark(
2364 name = "f32_vscaleextexp_bench",
2365 srcs = [
2366 "bench/f32-vscaleextexp.cc",
2367 "src/xnnpack/AlignedAllocator.h",
2368 ] + MICROKERNEL_BENCHMARK_HDRS,
2369 deps = MICROKERNEL_BENCHMARK_DEPS,
2370)
2371
2372xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002373 name = "f32_im2col_gemm_bench",
2374 srcs = [
2375 "bench/f32-im2col-gemm.cc",
2376 "bench/conv.h",
2377 "src/xnnpack/AlignedAllocator.h",
2378 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2379 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2380)
2381
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002382xnnpack_benchmark(
2383 name = "requantization_bench",
2384 srcs = [
2385 "bench/requantization.cc",
2386 "src/xnnpack/requantization-stubs.h",
2387 "src/xnnpack/AlignedAllocator.h",
2388 ] + MICROKERNEL_BENCHMARK_HDRS,
2389 deps = MICROKERNEL_BENCHMARK_DEPS,
2390)
2391
Marat Dukhan08c4a432019-10-03 09:29:21 -07002392########################### Benchmarks for operators ###########################
2393
2394xnnpack_benchmark(
2395 name = "add_bench",
2396 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002397 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002398)
2399
2400xnnpack_benchmark(
2401 name = "average_pooling_bench",
2402 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002403 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002404 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002405 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002406)
2407
2408xnnpack_benchmark(
2409 name = "channel_shuffle_bench",
2410 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002411 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002412)
2413
2414xnnpack_benchmark(
2415 name = "convolution_bench",
2416 srcs = ["bench/convolution.cc"],
2417 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002418 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002419 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002420)
2421
2422xnnpack_benchmark(
2423 name = "deconvolution_bench",
2424 srcs = ["bench/deconvolution.cc"],
2425 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002426 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002427 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002428)
2429
2430xnnpack_benchmark(
2431 name = "global_average_pooling_bench",
2432 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002433 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002434)
2435
2436xnnpack_benchmark(
2437 name = "max_pooling_bench",
2438 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002439 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002440)
2441
2442xnnpack_benchmark(
2443 name = "sigmoid_bench",
2444 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002445 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002446 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002447 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002448)
2449
2450xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002451 name = "prelu_bench",
2452 srcs = ["bench/prelu.cc"],
2453 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002454 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002455 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002456)
2457
2458xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002459 name = "softmax_bench",
2460 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002461 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002462 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002463 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002464)
2465
Marat Dukhanc068bb62019-10-04 13:24:39 -07002466############################# End-to-end benchmarks ############################
2467
2468cc_library(
2469 name = "mobilenet_v1",
2470 srcs = ["models/mobilenet-v1.cc"],
2471 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002472 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002473 linkstatic = True,
2474 deps = [
2475 ":XNNPACK",
2476 "@pthreadpool",
2477 ],
2478)
2479
2480cc_library(
2481 name = "mobilenet_v2",
2482 srcs = ["models/mobilenet-v2.cc"],
2483 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002484 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002485 linkstatic = True,
2486 deps = [
2487 ":XNNPACK",
2488 "@pthreadpool",
2489 ],
2490)
2491
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002492cc_library(
2493 name = "mobilenet_v3_large",
2494 srcs = ["models/mobilenet-v3-large.cc"],
2495 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002496 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002497 linkstatic = True,
2498 deps = [
2499 ":XNNPACK",
2500 "@pthreadpool",
2501 ],
2502)
2503
2504cc_library(
2505 name = "mobilenet_v3_small",
2506 srcs = ["models/mobilenet-v3-small.cc"],
2507 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002508 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002509 linkstatic = True,
2510 deps = [
2511 ":XNNPACK",
2512 "@pthreadpool",
2513 ],
2514)
2515
Marat Dukhanc068bb62019-10-04 13:24:39 -07002516xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002517 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002518 srcs = [
2519 "bench/f32-dwconv-e2e.cc",
2520 "bench/end2end.h",
2521 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002522 deps = MICROKERNEL_BENCHMARK_DEPS + [
2523 ":XNNPACK",
2524 ":mobilenet_v1",
2525 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002526 ":mobilenet_v3_large",
2527 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002528 ],
2529)
2530
2531xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002532 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002533 srcs = [
2534 "bench/f32-gemm-e2e.cc",
2535 "bench/end2end.h",
2536 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002537 deps = MICROKERNEL_BENCHMARK_DEPS + [
2538 ":XNNPACK",
2539 ":mobilenet_v1",
2540 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002541 ":mobilenet_v3_large",
2542 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002543 ],
2544)
2545
2546xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002547 name = "end2end_bench",
2548 srcs = ["bench/end2end.cc"],
2549 deps = [
2550 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002551 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002552 ":mobilenet_v1",
2553 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002554 ":mobilenet_v3_large",
2555 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002556 "@pthreadpool",
2557 ],
2558)
2559
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002560#################### Accuracy evaluation for math functions ####################
2561
2562xnnpack_benchmark(
2563 name = "f32_exp_eval",
2564 srcs = [
2565 "eval/f32-exp.cc",
2566 "src/xnnpack/AlignedAllocator.h",
2567 ] + ACCURACY_EVAL_HDRS,
2568 deps = ACCURACY_EVAL_DEPS,
2569)
2570
Marat Dukhan515c9772019-10-17 18:07:57 -07002571xnnpack_benchmark(
2572 name = "f32_expminus_eval",
2573 srcs = [
2574 "eval/f32-expminus.cc",
2575 "src/xnnpack/AlignedAllocator.h",
2576 ] + ACCURACY_EVAL_HDRS,
2577 deps = ACCURACY_EVAL_DEPS,
2578)
2579
Marat Dukhan98ba4412019-10-23 02:14:28 -07002580xnnpack_benchmark(
2581 name = "f32_extexp_eval",
2582 srcs = [
2583 "eval/f32-extexp.cc",
2584 "src/xnnpack/AlignedAllocator.h",
2585 ] + ACCURACY_EVAL_HDRS,
2586 deps = ACCURACY_EVAL_DEPS,
2587)
2588
Marat Dukhan8853b822020-05-07 12:19:01 -07002589xnnpack_unit_test(
2590 name = "f32_roundne_eval",
2591 srcs = [
2592 "eval/f32-roundne.cc",
2593 "src/xnnpack/AlignedAllocator.h",
2594 "src/xnnpack/math-stubs.h",
2595 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07002596 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07002597 deps = MICROKERNEL_TEST_DEPS,
2598)
2599
Marat Dukhan346a9e52019-11-15 09:06:30 -08002600xnnpack_benchmark(
2601 name = "f32_sigmoid_eval",
2602 srcs = [
2603 "eval/f32-sigmoid.cc",
2604 "src/xnnpack/AlignedAllocator.h",
2605 ] + ACCURACY_EVAL_HDRS,
2606 deps = ACCURACY_EVAL_DEPS,
2607)
2608
Marat Dukhan08c4a432019-10-03 09:29:21 -07002609######################### Unit tests for micro-kernels #########################
2610
2611xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002612 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002613 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002614 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002615 "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 Dukhan355ab432020-04-09 19:01:52 -07002622 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002623 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002624 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002625 "test/spmm-microkernel-tester.h",
2626 "src/xnnpack/AlignedAllocator.h",
2627 ] + MICROKERNEL_TEST_HDRS,
2628 deps = MICROKERNEL_TEST_DEPS,
2629)
2630
2631xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002632 name = "f32_argmaxpool_test",
2633 srcs = [
2634 "test/f32-argmaxpool.cc",
2635 "test/argmaxpool-microkernel-tester.h",
2636 "src/xnnpack/AlignedAllocator.h",
2637 ] + MICROKERNEL_TEST_HDRS,
2638 deps = MICROKERNEL_TEST_DEPS,
2639)
2640
2641xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002642 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002643 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002644 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002645 "test/avgpool-microkernel-tester.h",
2646 "src/xnnpack/AlignedAllocator.h",
2647 ] + MICROKERNEL_TEST_HDRS,
2648 deps = MICROKERNEL_TEST_DEPS,
2649)
2650
2651xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002652 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002653 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002654 "test/f32-ibilinear.cc",
2655 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002656 "src/xnnpack/AlignedAllocator.h",
2657 ] + MICROKERNEL_TEST_HDRS,
2658 deps = MICROKERNEL_TEST_DEPS,
2659)
2660
2661xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002662 name = "f32_clamp_test",
2663 srcs = [
2664 "test/f32-clamp.cc",
2665 "test/clamp-microkernel-tester.h",
2666 ] + MICROKERNEL_TEST_HDRS,
2667 deps = MICROKERNEL_TEST_DEPS,
2668)
2669
2670xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002671 name = "f32_igemm_test",
2672 srcs = [
2673 "test/f32-igemm.cc",
2674 "test/gemm-microkernel-tester.h",
2675 "src/xnnpack/AlignedAllocator.h",
2676 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2677 deps = MICROKERNEL_TEST_DEPS,
2678)
2679
2680xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002681 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002682 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002683 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002684 "test/gemm-microkernel-tester.h",
2685 "src/xnnpack/AlignedAllocator.h",
2686 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002687 # Work-around for "too many sections" error
2688 mingw_copts = ["-Wa,-mbig-obj"],
2689 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002690 deps = MICROKERNEL_TEST_DEPS,
2691)
2692
2693xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002694 name = "f16_igemm_minmax_test",
2695 srcs = [
2696 "test/f16-igemm-minmax.cc",
2697 "test/gemm-microkernel-tester.h",
2698 "src/xnnpack/AlignedAllocator.h",
2699 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2700 deps = MICROKERNEL_TEST_DEPS,
2701)
2702
2703xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002704 name = "f32_conv_hwc_test",
2705 srcs = [
2706 "test/f32-conv-hwc.cc",
2707 "test/conv-hwc-microkernel-tester.h",
2708 "src/xnnpack/AlignedAllocator.h",
2709 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2710 deps = MICROKERNEL_TEST_DEPS,
2711)
2712
2713xnnpack_unit_test(
2714 name = "f32_conv_hwc2spchw_test",
2715 srcs = [
2716 "test/f32-conv-hwc2spchw.cc",
2717 "test/conv-hwc2spchw-microkernel-tester.h",
2718 "src/xnnpack/AlignedAllocator.h",
2719 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2720 deps = MICROKERNEL_TEST_DEPS,
2721)
2722
2723xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002724 name = "f32_dwconv_test",
2725 srcs = [
2726 "test/f32-dwconv.cc",
2727 "test/dwconv-microkernel-tester.h",
2728 "src/xnnpack/AlignedAllocator.h",
2729 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2730 deps = MICROKERNEL_TEST_DEPS,
2731)
2732
2733xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002734 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002735 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002736 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737 "test/dwconv-microkernel-tester.h",
2738 "src/xnnpack/AlignedAllocator.h",
2739 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2740 deps = MICROKERNEL_TEST_DEPS,
2741)
2742
2743xnnpack_unit_test(
2744 name = "f32_dwconv_spchw_test",
2745 srcs = [
2746 "test/f32-dwconv-spchw.cc",
2747 "test/dwconv-spchw-microkernel-tester.h",
2748 "src/xnnpack/AlignedAllocator.h",
2749 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2750 deps = MICROKERNEL_TEST_DEPS,
2751)
2752
2753xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002754 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002755 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002756 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002757 "test/gavgpool-microkernel-tester.h",
2758 "src/xnnpack/AlignedAllocator.h",
2759 ] + MICROKERNEL_TEST_HDRS,
2760 deps = MICROKERNEL_TEST_DEPS,
2761)
2762
2763xnnpack_unit_test(
2764 name = "f32_gavgpool_spchw_test",
2765 srcs = [
2766 "test/f32-gavgpool-spchw.cc",
2767 "test/gavgpool-spchw-microkernel-tester.h",
2768 "src/xnnpack/AlignedAllocator.h",
2769 ] + MICROKERNEL_TEST_HDRS,
2770 deps = MICROKERNEL_TEST_DEPS,
2771)
2772
2773xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002774 name = "f32_gemm_test",
2775 srcs = [
2776 "test/f32-gemm.cc",
2777 "test/gemm-microkernel-tester.h",
2778 "src/xnnpack/AlignedAllocator.h",
2779 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2780 deps = MICROKERNEL_TEST_DEPS,
2781)
2782
2783xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002784 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002785 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002786 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002787 "test/gemm-microkernel-tester.h",
2788 "src/xnnpack/AlignedAllocator.h",
2789 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2790 deps = MICROKERNEL_TEST_DEPS,
2791)
2792
2793xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002794 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002795 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002796 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002797 "test/gemm-microkernel-tester.h",
2798 "src/xnnpack/AlignedAllocator.h",
2799 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2800 deps = MICROKERNEL_TEST_DEPS,
2801)
2802
2803xnnpack_unit_test(
2804 name = "f32_hswish_test",
2805 srcs = [
2806 "test/f32-hswish.cc",
2807 "test/hswish-microkernel-tester.h",
2808 ] + MICROKERNEL_TEST_HDRS,
2809 deps = MICROKERNEL_TEST_DEPS,
2810)
2811
2812xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002813 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002814 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002815 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002816 "test/maxpool-microkernel-tester.h",
2817 ] + MICROKERNEL_TEST_HDRS,
2818 deps = MICROKERNEL_TEST_DEPS,
2819)
2820
2821xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002822 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002823 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002824 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002825 "test/avgpool-microkernel-tester.h",
2826 "src/xnnpack/AlignedAllocator.h",
2827 ] + MICROKERNEL_TEST_HDRS,
2828 deps = MICROKERNEL_TEST_DEPS,
2829)
2830
2831xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002832 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002833 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002834 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002835 "test/gemm-microkernel-tester.h",
2836 "src/xnnpack/AlignedAllocator.h",
2837 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2838 deps = MICROKERNEL_TEST_DEPS,
2839)
2840
2841xnnpack_unit_test(
2842 name = "f32_prelu_test",
2843 srcs = [
2844 "test/f32-prelu.cc",
2845 "test/prelu-microkernel-tester.h",
2846 "src/xnnpack/AlignedAllocator.h",
2847 ] + MICROKERNEL_TEST_HDRS,
2848 deps = MICROKERNEL_TEST_DEPS,
2849)
2850
2851xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002852 name = "f32_raddexpminusmax_test",
2853 srcs = [
2854 "test/f32-raddexpminusmax.cc",
2855 "test/raddexpminusmax-microkernel-tester.h",
2856 ] + MICROKERNEL_TEST_HDRS,
2857 deps = MICROKERNEL_TEST_DEPS,
2858)
2859
2860xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002861 name = "f32_raddextexp_test",
2862 srcs = [
2863 "test/f32-raddextexp.cc",
2864 "test/raddextexp-microkernel-tester.h",
2865 ] + MICROKERNEL_TEST_HDRS,
2866 deps = MICROKERNEL_TEST_DEPS,
2867)
2868
2869xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002870 name = "f32_raddstoreexpminusmax_test",
2871 srcs = [
2872 "test/f32-raddstoreexpminusmax.cc",
2873 "test/raddstoreexpminusmax-microkernel-tester.h",
2874 ] + MICROKERNEL_TEST_HDRS,
2875 deps = MICROKERNEL_TEST_DEPS,
2876)
2877
2878xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002879 name = "f32_rmax_test",
2880 srcs = [
2881 "test/f32-rmax.cc",
2882 "test/rmax-microkernel-tester.h",
2883 ] + MICROKERNEL_TEST_HDRS,
2884 deps = MICROKERNEL_TEST_DEPS,
2885)
2886
2887xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002888 name = "f32_sigmoid_test",
2889 srcs = [
2890 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002891 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002892 ] + MICROKERNEL_TEST_HDRS,
2893 deps = MICROKERNEL_TEST_DEPS,
2894)
2895
2896xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002897 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002899 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002900 "test/spmm-microkernel-tester.h",
2901 "src/xnnpack/AlignedAllocator.h",
2902 ] + MICROKERNEL_TEST_HDRS,
2903 deps = MICROKERNEL_TEST_DEPS,
2904)
2905
2906xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07002907 name = "f16_vadd_minmax_test",
2908 srcs = [
2909 "test/f16-vadd-minmax.cc",
2910 "test/vbinary-microkernel-tester.h",
2911 ] + MICROKERNEL_TEST_HDRS,
2912 deps = MICROKERNEL_TEST_DEPS,
2913)
2914
2915xnnpack_unit_test(
2916 name = "f16_vaddc_minmax_test",
2917 srcs = [
2918 "test/f16-vaddc-minmax.cc",
2919 "test/vbinaryc-microkernel-tester.h",
2920 ] + MICROKERNEL_TEST_HDRS,
2921 deps = MICROKERNEL_TEST_DEPS,
2922)
2923
2924xnnpack_unit_test(
2925 name = "f16_vdiv_minmax_test",
2926 srcs = [
2927 "test/f16-vdiv-minmax.cc",
2928 "test/vbinary-microkernel-tester.h",
2929 ] + MICROKERNEL_TEST_HDRS,
2930 deps = MICROKERNEL_TEST_DEPS,
2931)
2932
2933xnnpack_unit_test(
2934 name = "f16_vdivc_minmax_test",
2935 srcs = [
2936 "test/f16-vdivc-minmax.cc",
2937 "test/vbinaryc-microkernel-tester.h",
2938 ] + MICROKERNEL_TEST_HDRS,
2939 deps = MICROKERNEL_TEST_DEPS,
2940)
2941
2942xnnpack_unit_test(
2943 name = "f16_vrdivc_minmax_test",
2944 srcs = [
2945 "test/f16-vrdivc-minmax.cc",
2946 "test/vbinaryc-microkernel-tester.h",
2947 ] + MICROKERNEL_TEST_HDRS,
2948 deps = MICROKERNEL_TEST_DEPS,
2949)
2950
2951xnnpack_unit_test(
2952 name = "f16_vmax_test",
2953 srcs = [
2954 "test/f16-vmax.cc",
2955 "test/vbinary-microkernel-tester.h",
2956 ] + MICROKERNEL_TEST_HDRS,
2957 deps = MICROKERNEL_TEST_DEPS,
2958)
2959
2960xnnpack_unit_test(
2961 name = "f16_vmaxc_test",
2962 srcs = [
2963 "test/f16-vmaxc.cc",
2964 "test/vbinaryc-microkernel-tester.h",
2965 ] + MICROKERNEL_TEST_HDRS,
2966 deps = MICROKERNEL_TEST_DEPS,
2967)
2968
2969xnnpack_unit_test(
2970 name = "f16_vmin_test",
2971 srcs = [
2972 "test/f16-vmin.cc",
2973 "test/vbinary-microkernel-tester.h",
2974 ] + MICROKERNEL_TEST_HDRS,
2975 deps = MICROKERNEL_TEST_DEPS,
2976)
2977
2978xnnpack_unit_test(
2979 name = "f16_vminc_test",
2980 srcs = [
2981 "test/f16-vminc.cc",
2982 "test/vbinaryc-microkernel-tester.h",
2983 ] + MICROKERNEL_TEST_HDRS,
2984 deps = MICROKERNEL_TEST_DEPS,
2985)
2986
2987xnnpack_unit_test(
2988 name = "f16_vmul_minmax_test",
2989 srcs = [
2990 "test/f16-vmul-minmax.cc",
2991 "test/vbinary-microkernel-tester.h",
2992 ] + MICROKERNEL_TEST_HDRS,
2993 deps = MICROKERNEL_TEST_DEPS,
2994)
2995
2996xnnpack_unit_test(
2997 name = "f16_vmulc_minmax_test",
2998 srcs = [
2999 "test/f16-vmulc-minmax.cc",
3000 "test/vbinaryc-microkernel-tester.h",
3001 ] + MICROKERNEL_TEST_HDRS,
3002 deps = MICROKERNEL_TEST_DEPS,
3003)
3004
3005xnnpack_unit_test(
3006 name = "f16_vsub_minmax_test",
3007 srcs = [
3008 "test/f16-vsub-minmax.cc",
3009 "test/vbinary-microkernel-tester.h",
3010 ] + MICROKERNEL_TEST_HDRS,
3011 deps = MICROKERNEL_TEST_DEPS,
3012)
3013
3014xnnpack_unit_test(
3015 name = "f16_vsubc_minmax_test",
3016 srcs = [
3017 "test/f16-vsubc-minmax.cc",
3018 "test/vbinaryc-microkernel-tester.h",
3019 ] + MICROKERNEL_TEST_HDRS,
3020 deps = MICROKERNEL_TEST_DEPS,
3021)
3022
3023xnnpack_unit_test(
3024 name = "f16_vrsubc_minmax_test",
3025 srcs = [
3026 "test/f16-vrsubc-minmax.cc",
3027 "test/vbinaryc-microkernel-tester.h",
3028 ] + MICROKERNEL_TEST_HDRS,
3029 deps = MICROKERNEL_TEST_DEPS,
3030)
3031
3032xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003033 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003034 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003035 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003036 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003037 ] + MICROKERNEL_TEST_HDRS,
3038 deps = MICROKERNEL_TEST_DEPS,
3039)
3040
3041xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003042 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003043 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003044 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003045 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003046 ] + MICROKERNEL_TEST_HDRS,
3047 deps = MICROKERNEL_TEST_DEPS,
3048)
3049
3050xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003051 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003052 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003053 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003054 "test/vbinary-microkernel-tester.h",
3055 ] + MICROKERNEL_TEST_HDRS,
3056 deps = MICROKERNEL_TEST_DEPS,
3057)
3058
3059xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003060 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003061 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003062 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003063 "test/vbinaryc-microkernel-tester.h",
3064 ] + MICROKERNEL_TEST_HDRS,
3065 deps = MICROKERNEL_TEST_DEPS,
3066)
3067
3068xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003069 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003070 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003071 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003072 "test/vbinaryc-microkernel-tester.h",
3073 ] + MICROKERNEL_TEST_HDRS,
3074 deps = MICROKERNEL_TEST_DEPS,
3075)
3076
3077xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003078 name = "f32_vmax_test",
3079 srcs = [
3080 "test/f32-vmax.cc",
3081 "test/vbinary-microkernel-tester.h",
3082 ] + MICROKERNEL_TEST_HDRS,
3083 deps = MICROKERNEL_TEST_DEPS,
3084)
3085
3086xnnpack_unit_test(
3087 name = "f32_vmaxc_test",
3088 srcs = [
3089 "test/f32-vmaxc.cc",
3090 "test/vbinaryc-microkernel-tester.h",
3091 ] + MICROKERNEL_TEST_HDRS,
3092 deps = MICROKERNEL_TEST_DEPS,
3093)
3094
3095xnnpack_unit_test(
3096 name = "f32_vmin_test",
3097 srcs = [
3098 "test/f32-vmin.cc",
3099 "test/vbinary-microkernel-tester.h",
3100 ] + MICROKERNEL_TEST_HDRS,
3101 deps = MICROKERNEL_TEST_DEPS,
3102)
3103
3104xnnpack_unit_test(
3105 name = "f32_vminc_test",
3106 srcs = [
3107 "test/f32-vminc.cc",
3108 "test/vbinaryc-microkernel-tester.h",
3109 ] + MICROKERNEL_TEST_HDRS,
3110 deps = MICROKERNEL_TEST_DEPS,
3111)
3112
3113xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003114 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003115 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003116 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003117 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003118 ] + MICROKERNEL_TEST_HDRS,
3119 deps = MICROKERNEL_TEST_DEPS,
3120)
3121
3122xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003123 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003124 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003125 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003126 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003127 ] + MICROKERNEL_TEST_HDRS,
3128 deps = MICROKERNEL_TEST_DEPS,
3129)
3130
3131xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003132 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003133 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003134 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003135 "test/vmulcaddc-microkernel-tester.h",
3136 "src/xnnpack/AlignedAllocator.h",
3137 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3138 deps = MICROKERNEL_TEST_DEPS,
3139)
3140
3141xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003142 name = "f32_vscale_test",
3143 srcs = [
3144 "test/f32-vscale.cc",
3145 "test/vscale-microkernel-tester.h",
3146 ] + MICROKERNEL_TEST_HDRS,
3147 deps = MICROKERNEL_TEST_DEPS,
3148)
3149
3150xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003151 name = "f32_vscaleexpminusmax_test",
3152 srcs = [
3153 "test/f32-vscaleexpminusmax.cc",
3154 "test/vscaleexpminusmax-microkernel-tester.h",
3155 ] + MICROKERNEL_TEST_HDRS,
3156 deps = MICROKERNEL_TEST_DEPS,
3157)
3158
3159xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003160 name = "f32_vscaleextexp_test",
3161 srcs = [
3162 "test/f32-vscaleextexp.cc",
3163 "test/vscaleextexp-microkernel-tester.h",
3164 ] + MICROKERNEL_TEST_HDRS,
3165 deps = MICROKERNEL_TEST_DEPS,
3166)
3167
3168xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003169 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003170 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003171 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003172 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003173 ] + MICROKERNEL_TEST_HDRS,
3174 deps = MICROKERNEL_TEST_DEPS,
3175)
3176
3177xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003178 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003179 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003180 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003181 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003182 ] + MICROKERNEL_TEST_HDRS,
3183 deps = MICROKERNEL_TEST_DEPS,
3184)
3185
3186xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003187 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003188 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003189 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003190 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003191 ] + MICROKERNEL_TEST_HDRS,
3192 deps = MICROKERNEL_TEST_DEPS,
3193)
3194
3195xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003196 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003197 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003198 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003199 "test/avgpool-microkernel-tester.h",
3200 "src/xnnpack/AlignedAllocator.h",
3201 ] + MICROKERNEL_TEST_HDRS,
3202 deps = MICROKERNEL_TEST_DEPS,
3203)
3204
3205xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003206 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003208 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003209 "test/gemm-microkernel-tester.h",
3210 "src/xnnpack/AlignedAllocator.h",
3211 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3212 deps = MICROKERNEL_TEST_DEPS,
3213)
3214
3215xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003216 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003217 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003218 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219 "test/dwconv-microkernel-tester.h",
3220 "src/xnnpack/AlignedAllocator.h",
3221 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3222 deps = MICROKERNEL_TEST_DEPS,
3223)
3224
3225xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003226 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003228 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003229 "test/gavgpool-microkernel-tester.h",
3230 "src/xnnpack/AlignedAllocator.h",
3231 ] + MICROKERNEL_TEST_HDRS,
3232 deps = MICROKERNEL_TEST_DEPS,
3233)
3234
3235xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003236 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003238 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003239 "test/gemm-microkernel-tester.h",
3240 "src/xnnpack/AlignedAllocator.h",
3241 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3242 deps = MICROKERNEL_TEST_DEPS,
3243)
3244
3245xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003246 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003247 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003248 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003249 "test/vadd-microkernel-tester.h",
3250 ] + MICROKERNEL_TEST_HDRS,
3251 deps = MICROKERNEL_TEST_DEPS,
3252)
3253
3254xnnpack_unit_test(
3255 name = "u8_clamp_test",
3256 srcs = [
3257 "test/u8-clamp.cc",
3258 "test/clamp-microkernel-tester.h",
3259 ] + MICROKERNEL_TEST_HDRS,
3260 deps = MICROKERNEL_TEST_DEPS,
3261)
3262
3263xnnpack_unit_test(
3264 name = "u8_lut32norm_test",
3265 srcs = [
3266 "test/u8-lut32norm.cc",
3267 "test/lut-norm-microkernel-tester.h",
3268 ] + MICROKERNEL_TEST_HDRS,
3269 deps = MICROKERNEL_TEST_DEPS,
3270)
3271
3272xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003273 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003274 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003275 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003276 "test/maxpool-microkernel-tester.h",
3277 ] + MICROKERNEL_TEST_HDRS,
3278 deps = MICROKERNEL_TEST_DEPS,
3279)
3280
3281xnnpack_unit_test(
3282 name = "u8_rmax_test",
3283 srcs = [
3284 "test/u8-rmax.cc",
3285 "test/rmax-microkernel-tester.h",
3286 ] + MICROKERNEL_TEST_HDRS,
3287 deps = MICROKERNEL_TEST_DEPS,
3288)
3289
3290xnnpack_unit_test(
3291 name = "x32_packx_test",
3292 srcs = [
3293 "test/x32-packx.cc",
3294 "test/pack-microkernel-tester.h",
3295 "src/xnnpack/AlignedAllocator.h",
3296 ] + MICROKERNEL_TEST_HDRS,
3297 deps = MICROKERNEL_TEST_DEPS,
3298)
3299
3300xnnpack_unit_test(
3301 name = "x32_pad_test",
3302 srcs = [
3303 "test/x32-pad.cc",
3304 "test/pad-microkernel-tester.h",
3305 ] + MICROKERNEL_TEST_HDRS,
3306 deps = MICROKERNEL_TEST_DEPS,
3307)
3308
3309xnnpack_unit_test(
3310 name = "x32_unpool_test",
3311 srcs = [
3312 "test/x32-unpool.cc",
3313 "test/unpool-microkernel-tester.h",
3314 ] + MICROKERNEL_TEST_HDRS,
3315 deps = MICROKERNEL_TEST_DEPS,
3316)
3317
3318xnnpack_unit_test(
3319 name = "x32_zip_test",
3320 srcs = [
3321 "test/x32-zip.cc",
3322 "test/zip-microkernel-tester.h",
3323 ] + MICROKERNEL_TEST_HDRS,
3324 deps = MICROKERNEL_TEST_DEPS,
3325)
3326
3327xnnpack_unit_test(
3328 name = "x8_lut_test",
3329 srcs = [
3330 "test/x8-lut.cc",
3331 "test/lut-microkernel-tester.h",
3332 ] + MICROKERNEL_TEST_HDRS,
3333 deps = MICROKERNEL_TEST_DEPS,
3334)
3335
3336xnnpack_unit_test(
3337 name = "x8_zip_test",
3338 srcs = [
3339 "test/x8-zip.cc",
3340 "test/zip-microkernel-tester.h",
3341 ] + MICROKERNEL_TEST_HDRS,
3342 deps = MICROKERNEL_TEST_DEPS,
3343)
3344
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003345xnnpack_unit_test(
3346 name = "requantization_test",
3347 srcs = [
3348 "src/xnnpack/requantization-stubs.h",
3349 "test/requantization.cc",
3350 "test/requantization-tester.h",
3351 ] + MICROKERNEL_TEST_HDRS,
3352 deps = MICROKERNEL_TEST_DEPS,
3353)
3354
Marat Dukhan20c3b922020-03-10 03:45:06 -07003355########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003356
3357xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003358 name = "operator_size_test",
3359 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003360 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003361)
3362
Marat Dukhan20c3b922020-03-10 03:45:06 -07003363xnnpack_binary(
3364 name = "subgraph_size_test",
3365 srcs = ["test/subgraph-size.c"],
3366 deps = [":XNNPACK"],
3367)
3368
3369########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003370
3371xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003372 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003373 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003374 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375 "test/add-operator-tester.h",
3376 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003377 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003378)
3379
3380xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003381 name = "add_nd_test",
3382 srcs = [
3383 "test/add-nd.cc",
3384 "test/binary-elementwise-operator-tester.h",
3385 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003386 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003387)
3388
3389xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003390 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003391 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003392 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003393 "test/argmax-pooling-operator-tester.h",
3394 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003395 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003396)
3397
3398xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003399 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003400 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003401 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 "test/average-pooling-operator-tester.h",
3403 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003404 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003405)
3406
3407xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003408 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003409 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003410 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003411 "test/channel-pad-operator-tester.h",
3412 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003413 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003414)
3415
3416xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003417 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003418 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003419 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003420 "test/channel-shuffle-operator-tester.h",
3421 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003422 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003423)
3424
3425xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003426 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003428 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003429 "test/clamp-operator-tester.h",
3430 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003431 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003432)
3433
3434xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003435 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003437 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003438 "test/convolution-operator-tester.h",
3439 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003440 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003441)
3442
3443xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003444 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003445 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003446 "test/convolution-nchw.cc",
3447 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003449 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003450)
3451
3452xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003453 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003454 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003455 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003456 "test/deconvolution-operator-tester.h",
3457 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003458 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003459)
3460
3461xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003462 name = "divide_nd_test",
3463 srcs = [
3464 "test/binary-elementwise-operator-tester.h",
3465 "test/divide-nd.cc",
3466 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003467 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003468)
3469
3470xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003471 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003472 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003473 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003474 "test/fully-connected-operator-tester.h",
3475 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003476 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003477)
3478
3479xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003480 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003481 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003482 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003483 "test/global-average-pooling-operator-tester.h",
3484 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003485 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003486)
3487
3488xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003489 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003491 "test/global-average-pooling-ncw.cc",
3492 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003494 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003495)
3496
3497xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003498 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003499 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003500 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003501 "test/hardswish-operator-tester.h",
3502 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003503 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003504)
3505
3506xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003507 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003509 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003510 "test/leaky-relu-operator-tester.h",
3511 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003512 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513)
3514
3515xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003516 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003518 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003519 "test/max-pooling-operator-tester.h",
3520 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003521 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003522)
3523
3524xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003525 name = "maximum_nd_test",
3526 srcs = [
3527 "test/binary-elementwise-operator-tester.h",
3528 "test/maximum-nd.cc",
3529 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003530 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003531)
3532
3533xnnpack_unit_test(
3534 name = "minimum_nd_test",
3535 srcs = [
3536 "test/binary-elementwise-operator-tester.h",
3537 "test/minimum-nd.cc",
3538 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003539 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003540)
3541
3542xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003543 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003544 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003545 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003546 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003547 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003548 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003549)
3550
3551xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003552 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003554 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555 "test/prelu-operator-tester.h",
3556 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003557 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003558)
3559
3560xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003561 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003562 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003563 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003564 "test/resize-bilinear-operator-tester.h",
3565 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003566 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003567)
3568
3569xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003570 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003571 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003572 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003573 "test/sigmoid-operator-tester.h",
3574 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003575 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576)
3577
3578xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003579 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003580 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003581 "test/softmax-nc.cc",
3582 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003584 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585)
3586
3587xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003588 name = "subtract_nd_test",
3589 srcs = [
3590 "test/binary-elementwise-operator-tester.h",
3591 "test/subtract-nd.cc",
3592 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003593 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003594)
3595
3596xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003597 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003598 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003599 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600 "test/unpooling-operator-tester.h",
3601 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003602 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603)
3604
3605############################# Build configurations #############################
3606
Marat Dukhanb8642352019-10-30 15:43:02 -07003607# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003608config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003609 name = "xnn_enable_assembly_explicit_true",
3610 define_values = {"xnn_enable_assembly": "true"},
3611)
3612
3613# Disables usage of assembly kernels.
3614config_setting(
3615 name = "xnn_enable_assembly_explicit_false",
3616 define_values = {"xnn_enable_assembly": "false"},
3617)
3618
Marat Dukhan05702cf2020-03-26 15:41:33 -07003619# Disables usage of HMP-aware optimizations.
3620config_setting(
3621 name = "xnn_enable_hmp_explicit_false",
3622 define_values = {"xnn_enable_hmp": "false"},
3623)
3624
Marat Dukhanb8642352019-10-30 15:43:02 -07003625# Builds with -c dbg
3626config_setting(
3627 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003628 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003629 "compilation_mode": "dbg",
3630 },
3631)
3632
3633# Builds with -c opt
3634config_setting(
3635 name = "optimized_build",
3636 values = {
3637 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003638 },
3639)
3640
3641config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003642 name = "linux_k8",
3643 values = {"cpu": "k8"},
3644)
3645
3646config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003647 name = "linux_arm",
3648 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003649)
3650
3651config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003652 name = "linux_armhf",
3653 values = {"cpu": "armhf"},
3654)
3655
3656config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003657 name = "linux_aarch64",
3658 values = {"cpu": "aarch64"},
3659)
3660
3661config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 name = "android",
3663 values = {"crosstool_top": "//external:android/crosstool"},
3664)
3665
3666config_setting(
3667 name = "android_armv7",
3668 values = {
3669 "crosstool_top": "//external:android/crosstool",
3670 "cpu": "armeabi-v7a",
3671 },
3672)
3673
3674config_setting(
3675 name = "android_arm64",
3676 values = {
3677 "crosstool_top": "//external:android/crosstool",
3678 "cpu": "arm64-v8a",
3679 },
3680)
3681
3682config_setting(
3683 name = "android_x86",
3684 values = {
3685 "crosstool_top": "//external:android/crosstool",
3686 "cpu": "x86",
3687 },
3688)
3689
3690config_setting(
3691 name = "android_x86_64",
3692 values = {
3693 "crosstool_top": "//external:android/crosstool",
3694 "cpu": "x86_64",
3695 },
3696)
3697
3698config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003699 name = "windows_x86_64",
3700 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003701)
3702
3703config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003704 name = "windows_x86_64_clang",
3705 values = {
3706 "compiler": "clang-cl",
3707 "cpu": "x64_windows",
3708 },
3709)
3710
3711config_setting(
3712 name = "windows_x86_64_mingw",
3713 values = {
3714 "compiler": "mingw-gcc",
3715 "cpu": "x64_windows",
3716 },
3717)
3718
3719config_setting(
3720 name = "windows_x86_64_msys",
3721 values = {
3722 "compiler": "msys-gcc",
3723 "cpu": "x64_windows",
3724 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003725)
3726
3727config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003728 name = "macos_x86_64",
3729 values = {
3730 "apple_platform_type": "macos",
3731 "cpu": "darwin",
3732 },
3733)
3734
3735config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003736 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003737 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003738)
3739
3740config_setting(
3741 name = "emscripten_wasm",
3742 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003743 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003744 "cpu": "wasm",
3745 },
3746)
3747
3748config_setting(
3749 name = "emscripten_wasmsimd",
3750 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003751 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003752 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003753 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003754 },
3755)
3756
3757config_setting(
3758 name = "emscripten_asmjs",
3759 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003760 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003761 "cpu": "asmjs",
3762 },
3763)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003764
3765config_setting(
3766 name = "ios_armv7",
3767 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003768 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003769 "cpu": "ios_armv7",
3770 },
3771)
3772
3773config_setting(
3774 name = "ios_arm64",
3775 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003776 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003777 "cpu": "ios_arm64",
3778 },
3779)
3780
3781config_setting(
3782 name = "ios_arm64e",
3783 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003784 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003785 "cpu": "ios_arm64e",
3786 },
3787)
3788
3789config_setting(
3790 name = "ios_x86",
3791 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003792 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003793 "cpu": "ios_i386",
3794 },
3795)
3796
3797config_setting(
3798 name = "ios_x86_64",
3799 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003800 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003801 "cpu": "ios_x86_64",
3802 },
3803)
3804
3805config_setting(
3806 name = "watchos_armv7k",
3807 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003808 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003809 "cpu": "watchos_armv7k",
3810 },
3811)
3812
3813config_setting(
3814 name = "watchos_arm64_32",
3815 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003816 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003817 "cpu": "watchos_arm64_32",
3818 },
3819)
3820
3821config_setting(
3822 name = "watchos_x86",
3823 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003824 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003825 "cpu": "watchos_i386",
3826 },
3827)
3828
3829config_setting(
3830 name = "watchos_x86_64",
3831 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003832 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003833 "cpu": "watchos_x86_64",
3834 },
3835)
3836
3837config_setting(
3838 name = "tvos_arm64",
3839 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003840 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003841 "cpu": "tvos_arm64",
3842 },
3843)
3844
3845config_setting(
3846 name = "tvos_x86_64",
3847 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003848 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003849 "cpu": "tvos_x86_64",
3850 },
3851)