blob: 009d8a8f1a68606733b32cdf62c49249388875e5 [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 Dukhande06f492020-04-09 00:19:31 -07001525 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001526 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1527 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1528 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1529 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1530 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1531 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1532 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1533 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1534 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1535 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001536]
1537
1538AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001539 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001540 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1541 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001542 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001543 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1544 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001545 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1546 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1547 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001548 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001549 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1550 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1551 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001552 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001553 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1554 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001555 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001556 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1557 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1558 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1559 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1560 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1561 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1562 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1563 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1564 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1565 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1566 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1567 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1568 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1569 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1570 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1571 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1572 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1573 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1574 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1575 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1576 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001577 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001578 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1579 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1580 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1581 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1582 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1583 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1584 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1585 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1586 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1587 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1588 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1589 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1590 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1591 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1592 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1593 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1594 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1595 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1596 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1597 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1598 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1599 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1600 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1601 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1602 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1603 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1604 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1605 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1606 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1607 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1608 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1609 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1610 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1611 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1612 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1613 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1614 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1615 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001616]
1617
Marat Dukhan1b354632020-03-23 12:50:22 -07001618INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001619 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001620 "src/xnnpack/argmaxpool.h",
1621 "src/xnnpack/avgpool.h",
1622 "src/xnnpack/clamp.h",
1623 "src/xnnpack/common.h",
1624 "src/xnnpack/conv.h",
1625 "src/xnnpack/dwconv.h",
1626 "src/xnnpack/gavgpool.h",
1627 "src/xnnpack/gemm.h",
1628 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001629 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001630 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001631 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001632 "src/xnnpack/lut.h",
1633 "src/xnnpack/math.h",
1634 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001635 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001636 "src/xnnpack/packx.h",
1637 "src/xnnpack/pad.h",
1638 "src/xnnpack/params.h",
1639 "src/xnnpack/pavgpool.h",
1640 "src/xnnpack/ppmm.h",
1641 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001642 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001643 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001644 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001645 "src/xnnpack/rmax.h",
1646 "src/xnnpack/scalar-utils.h",
1647 "src/xnnpack/spmm.h",
1648 "src/xnnpack/unpool.h",
1649 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001650 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001651 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001652 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001653 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001654 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001655 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001656 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001657]
1658
1659INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001660 "include/xnnpack.h",
1661 "src/xnnpack/allocator.h",
1662 "src/xnnpack/compute.h",
1663 "src/xnnpack/im2col.h",
1664 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001665 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001666 "src/xnnpack/operator.h",
1667 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001668 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001669 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001670 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001671 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001672]
1673
Marat Dukhan1b354632020-03-23 12:50:22 -07001674ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001675 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001676]
1677
Marat Dukhan1b354632020-03-23 12:50:22 -07001678MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001679 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001680 "include/xnnpack.h",
1681]
1682
Marat Dukhan1b354632020-03-23 12:50:22 -07001683MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001684 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001685 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001686 "src/xnnpack/requantization.h",
1687 "include/xnnpack.h",
1688]
1689
1690OPERATOR_TEST_PARAMS_HDRS = [
1691 "src/xnnpack/params.h",
1692 "src/xnnpack/common.h",
1693]
1694
1695WEIGHTS_PACK_HDRS = [
1696 "src/xnnpack/pack.h",
1697 "src/xnnpack/operator.h",
1698 "src/xnnpack/compute.h",
1699]
1700
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001701LOGGING_COPTS = select({
1702 # No logging in optimized mode
1703 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1704 # Full logging in debug mode
1705 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1706 # Error-only logging in default (fastbuild) mode
1707 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1708})
1709
1710LOGGING_HDRS = [
1711 "src/xnnpack/log.h",
1712]
1713
Marat Dukhan08c4a432019-10-03 09:29:21 -07001714xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001715 name = "tables",
1716 srcs = TABLE_SRCS,
1717 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001718 gcc_copts = xnnpack_gcc_std_copts(),
1719 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001720)
1721
1722xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723 name = "scalar_ukernels",
1724 srcs = SCALAR_UKERNELS,
1725 hdrs = INTERNAL_HDRS,
1726 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001727 gcc_copts = xnnpack_gcc_std_copts(),
1728 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001729 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001730 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001731 "@FP16",
1732 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001733 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001734 ],
1735)
1736
1737xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001738 name = "wasm_ukernels",
1739 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001740 gcc_copts = xnnpack_gcc_std_copts(),
1741 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001742 wasm_srcs = WASM_UKERNELS,
1743 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001744 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001745 "@FP16",
1746 "@FXdiv",
1747 "@pthreadpool",
1748 ],
1749)
1750
1751xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001752 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001753 hdrs = INTERNAL_HDRS,
1754 aarch32_copts = [
1755 "-marm",
1756 "-mfpu=neon",
1757 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001758 gcc_copts = xnnpack_gcc_std_copts(),
1759 gcc_x86_copts = ["-msse2"],
1760 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001761 optimized_copts = [
1762 "-O3",
1763 "-ffast-math",
1764 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001765 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001766 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001767 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001768 "@FP16",
1769 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001770 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001771 ],
1772)
1773
1774xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001775 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001776 hdrs = INTERNAL_HDRS,
1777 aarch32_copts = [
1778 "-marm",
1779 "-mfpu=neon",
1780 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001781 gcc_copts = xnnpack_gcc_std_copts(),
1782 gcc_x86_copts = ["-msse2"],
1783 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001784 optimized_copts = [
1785 "-O3",
1786 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001787 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001788 deps = [
1789 ":tables",
1790 "@FP16",
1791 "@psimd",
1792 "@pthreadpool",
1793 ],
1794)
1795
1796xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001797 name = "neon_ukernels",
1798 hdrs = INTERNAL_HDRS,
1799 aarch32_copts = [
1800 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001801 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001802 "-mfpu=neon",
1803 ],
1804 aarch32_srcs = NEON_UKERNELS,
1805 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001806 gcc_copts = xnnpack_gcc_std_copts(),
1807 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001808 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001809 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001810 "@FP16",
1811 "@pthreadpool",
1812 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001813)
1814
1815xnnpack_cc_library(
1816 name = "neonfma_ukernels",
1817 hdrs = INTERNAL_HDRS,
1818 aarch32_copts = [
1819 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001820 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001821 "-mfpu=neon-vfpv4",
1822 ],
1823 aarch32_srcs = NEONFMA_UKERNELS,
1824 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001825 gcc_copts = xnnpack_gcc_std_copts(),
1826 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001827 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001828 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001829 "@FP16",
1830 "@pthreadpool",
1831 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001832)
1833
1834xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001835 name = "neonv8_ukernels",
1836 hdrs = INTERNAL_HDRS,
1837 aarch32_copts = [
1838 "-marm",
1839 "-march=armv8-a",
1840 "-mfpu=neon-fp-armv8",
1841 ],
1842 aarch32_srcs = NEONV8_UKERNELS,
1843 aarch64_srcs = NEONV8_UKERNELS,
1844 gcc_copts = xnnpack_gcc_std_copts(),
1845 msvc_copts = xnnpack_msvc_std_copts(),
1846 deps = [
1847 ":tables",
1848 "@FP16",
1849 "@pthreadpool",
1850 ],
1851)
1852
1853xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001854 name = "neonfp16arith_ukernels",
1855 hdrs = INTERNAL_HDRS,
1856 aarch64_copts = ["-march=armv8.2-a+fp16"],
1857 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001858 gcc_copts = xnnpack_gcc_std_copts(),
1859 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001860 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001861 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001862 "@FP16",
1863 "@pthreadpool",
1864 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001865)
1866
1867xnnpack_cc_library(
1868 name = "sse2_ukernels",
1869 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001870 gcc_copts = xnnpack_gcc_std_copts(),
1871 gcc_x86_copts = ["-msse2"],
1872 msvc_copts = xnnpack_msvc_std_copts(),
1873 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001874 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001875 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001876 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001877 "@FP16",
1878 "@pthreadpool",
1879 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001880)
1881
1882xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001883 name = "ssse3_ukernels",
1884 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001885 gcc_copts = xnnpack_gcc_std_copts(),
1886 gcc_x86_copts = ["-mssse3"],
1887 msvc_copts = xnnpack_msvc_std_copts(),
1888 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001889 x86_srcs = SSSE3_UKERNELS,
1890 deps = [
1891 ":tables",
1892 "@FP16",
1893 "@pthreadpool",
1894 ],
1895)
1896
1897xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001898 name = "sse41_ukernels",
1899 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001900 gcc_copts = xnnpack_gcc_std_copts(),
1901 gcc_x86_copts = ["-msse4.1"],
1902 msvc_copts = xnnpack_msvc_std_copts(),
1903 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001904 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001905 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001906 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001907 "@FP16",
1908 "@pthreadpool",
1909 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001910)
1911
1912xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001913 name = "avx_ukernels",
1914 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001915 gcc_copts = xnnpack_gcc_std_copts(),
1916 gcc_x86_copts = ["-mavx"],
1917 msvc_copts = xnnpack_msvc_std_copts(),
1918 msvc_x86_32_copts = ["/arch:AVX"],
1919 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001920 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001921 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001922 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001923 "@FP16",
1924 "@pthreadpool",
1925 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001926)
1927
1928xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001929 name = "fma3_ukernels",
1930 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001931 gcc_copts = xnnpack_gcc_std_copts(),
1932 gcc_x86_copts = ["-mfma"],
1933 msvc_copts = xnnpack_msvc_std_copts(),
1934 msvc_x86_32_copts = ["/arch:AVX"],
1935 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001936 x86_srcs = FMA3_UKERNELS,
1937 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001938 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001939 "@FP16",
1940 "@pthreadpool",
1941 ],
1942)
1943
1944xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001945 name = "avx2_ukernels",
1946 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001947 gcc_copts = xnnpack_gcc_std_copts(),
1948 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001949 "-mfma",
1950 "-mavx2",
1951 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001952 msvc_copts = xnnpack_msvc_std_copts(),
1953 msvc_x86_32_copts = ["/arch:AVX2"],
1954 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001955 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001956 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001957 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001958 "@FP16",
1959 "@pthreadpool",
1960 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001961)
1962
1963xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001964 name = "avx512f_ukernels",
1965 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001966 gcc_copts = xnnpack_gcc_std_copts(),
1967 gcc_x86_copts = ["-mavx512f"],
1968 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1969 msvc_copts = xnnpack_msvc_std_copts(),
1970 msvc_x86_32_copts = ["/arch:AVX512"],
1971 msvc_x86_64_copts = ["/arch:AVX512"],
1972 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001973 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001974 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001975 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001976 "@FP16",
1977 "@pthreadpool",
1978 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001979)
1980
1981xnnpack_cc_library(
1982 name = "asm_ukernels",
1983 hdrs = ["src/xnnpack/assembly.h"],
1984 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001985 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001986 aarch64_srcs = AARCH64_ASM_UKERNELS,
1987)
1988
1989xnnpack_aggregate_library(
1990 name = "ukernels",
1991 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001992 ":neon_ukernels",
1993 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07001994 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001995 ":asm_ukernels",
1996 ],
1997 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001998 ":neon_ukernels",
1999 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002000 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002001 ":neonfp16arith_ukernels",
2002 ":asm_ukernels",
2003 ],
2004 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07002005 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002006 ":psimd_fastmath_ukernels",
2007 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002008 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002009 wasm_deps = [":wasm_ukernels"],
2010 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002011 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002012 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002013 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002014 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002015 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002016 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002017 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002018 ":avx512f_ukernels",
2019 ],
2020)
2021
2022xnnpack_cc_library(
2023 name = "im2col",
2024 srcs = ["src/im2col.c"],
2025 hdrs = [
2026 "src/xnnpack/common.h",
2027 "src/xnnpack/im2col.h",
2028 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002029 gcc_copts = xnnpack_gcc_std_copts(),
2030 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002031)
2032
2033xnnpack_cc_library(
2034 name = "indirection",
2035 srcs = ["src/indirection.c"],
2036 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002037 gcc_copts = xnnpack_gcc_std_copts(),
2038 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002039 deps = [
2040 "@FP16",
2041 "@FXdiv",
2042 "@pthreadpool",
2043 ],
2044)
2045
2046xnnpack_cc_library(
2047 name = "operator_run",
2048 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002049 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002050 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002051 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2052 "//conditions:default": [],
2053 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002054 gcc_copts = xnnpack_gcc_std_copts(),
2055 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002056 deps = [
2057 "@FP16",
2058 "@FXdiv",
2059 "@clog",
2060 "@pthreadpool",
2061 ],
2062)
2063
2064cc_library(
2065 name = "enable_assembly",
2066 defines = select({
2067 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2068 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002069 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070 }),
2071)
2072
Marat Dukhancf056b22019-10-07 10:26:29 -07002073xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002074 name = "operators",
2075 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002076 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002077 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002078 ],
2079 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002080 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002081 "-Isrc",
2082 "-Iinclude",
2083 ] + select({
2084 ":debug_build": [],
2085 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002086 }) + select({
2087 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2088 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002089 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002090 gcc_copts = xnnpack_gcc_std_copts(),
2091 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002092 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002093 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002094 "@FP16",
2095 "@FXdiv",
2096 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002097 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002098 ],
2099)
2100
Marat Dukhan10a38082020-04-17 03:58:35 -07002101xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002102 name = "XNNPACK",
2103 srcs = [
2104 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002105 "src/runtime.c",
2106 "src/subgraph.c",
2107 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002108 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002109 hdrs = ["include/xnnpack.h"],
2110 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002111 "-Isrc",
2112 "-Iinclude",
2113 ] + select({
2114 ":debug_build": [],
2115 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002116 }) + select({
2117 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2118 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002119 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002120 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002121 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002122 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002123 visibility = xnnpack_visibility(),
2124 deps = [
2125 ":enable_assembly",
2126 ":ukernels",
2127 ":operator_run",
2128 ":operators",
2129 "@clog",
2130 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002131 ] + select({
2132 ":emscripten": [],
2133 "//conditions:default": ["@cpuinfo"],
2134 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002135)
2136
Marat Dukhan10a38082020-04-17 03:58:35 -07002137xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002138 name = "xnnpack_operators_nhwc_f32",
2139 srcs = [
2140 "src/init.c",
2141 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002142 hdrs = ["include/xnnpack.h"],
2143 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002144 "-Isrc",
2145 "-Iinclude",
2146 ] + select({
2147 ":debug_build": [],
2148 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002149 }) + select({
2150 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2151 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002152 }),
2153 defines = [
2154 "XNN_NO_Q8_OPERATORS",
2155 "XNN_NO_U8_OPERATORS",
2156 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002157 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002158 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002159 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002160 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002161 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002162 visibility = xnnpack_visibility(),
2163 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002164 ":enable_assembly",
2165 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166 ":operator_run",
2167 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002168 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002169 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002170 ] + select({
2171 ":emscripten": [],
2172 "//conditions:default": ["@cpuinfo"],
2173 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002174)
2175
Marat Dukhancf056b22019-10-07 10:26:29 -07002176xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002177 name = "bench_utils",
2178 srcs = ["bench/utils.cc"],
2179 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002180 deps = [
2181 "@com_google_benchmark//:benchmark",
2182 "@cpuinfo",
2183 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002184)
2185
Frank Barchard7e955972019-10-11 10:34:25 -07002186######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002187
2188xnnpack_benchmark(
2189 name = "q8_gemm_bench",
2190 srcs = [
2191 "bench/gemm.h",
2192 "bench/q8-gemm.cc",
2193 "src/xnnpack/AlignedAllocator.h",
2194 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002195 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002196 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2197)
2198
2199xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002200 name = "f16_igemm_bench",
2201 srcs = [
2202 "bench/f16-igemm.cc",
2203 "bench/conv.h",
2204 "bench/google/conv.h",
2205 "src/xnnpack/AlignedAllocator.h",
2206 ] + MICROKERNEL_BENCHMARK_HDRS,
2207 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2208)
2209
2210xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002211 name = "f16_gemm_bench",
2212 srcs = [
2213 "bench/f16-gemm.cc",
2214 "bench/gemm.h",
2215 "src/xnnpack/AlignedAllocator.h",
2216 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002217 deps = MICROKERNEL_BENCHMARK_DEPS,
2218)
2219
2220xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002221 name = "f16_spmm_bench",
2222 srcs = [
2223 "bench/f16-spmm.cc",
2224 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002225 "src/xnnpack/AlignedAllocator.h",
2226 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002227 deps = MICROKERNEL_BENCHMARK_DEPS,
2228)
2229
2230xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002231 name = "f32_igemm_bench",
2232 srcs = [
2233 "bench/f32-igemm.cc",
2234 "bench/conv.h",
2235 "src/xnnpack/AlignedAllocator.h",
2236 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002237 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238)
2239
2240xnnpack_benchmark(
2241 name = "f32_conv_hwc_bench",
2242 srcs = [
2243 "bench/f32-conv-hwc.cc",
2244 "bench/dconv.h",
2245 "src/xnnpack/AlignedAllocator.h",
2246 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002247 deps = MICROKERNEL_BENCHMARK_DEPS,
2248)
2249
2250xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002251 name = "f32_conv_hwc2spchw_bench",
2252 srcs = [
2253 "bench/f32-conv-hwc2spchw.cc",
2254 "bench/dconv.h",
2255 "src/xnnpack/AlignedAllocator.h",
2256 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002257 deps = MICROKERNEL_BENCHMARK_DEPS,
2258)
2259
2260xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002261 name = "f32_dwconv_bench",
2262 srcs = [
2263 "bench/f32-dwconv.cc",
2264 "bench/dwconv.h",
2265 "src/xnnpack/AlignedAllocator.h",
2266 ] + MICROKERNEL_BENCHMARK_HDRS,
2267 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2268)
2269
2270xnnpack_benchmark(
2271 name = "f32_dwconv_spchw_bench",
2272 srcs = [
2273 "bench/f32-dwconv-spchw.cc",
2274 "bench/dwconv.h",
2275 "src/xnnpack/AlignedAllocator.h",
2276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2277 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2278)
2279
2280xnnpack_benchmark(
2281 name = "f32_gemm_bench",
2282 srcs = [
2283 "bench/f32-gemm.cc",
2284 "bench/gemm.h",
2285 "src/xnnpack/AlignedAllocator.h",
2286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002287 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002288 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289)
2290
2291xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002292 name = "f32_raddexpminusmax_bench",
2293 srcs = [
2294 "bench/f32-raddexpminusmax.cc",
2295 "src/xnnpack/AlignedAllocator.h",
2296 ] + MICROKERNEL_BENCHMARK_HDRS,
2297 deps = MICROKERNEL_BENCHMARK_DEPS,
2298)
2299
2300xnnpack_benchmark(
2301 name = "f32_raddextexp_bench",
2302 srcs = [
2303 "bench/f32-raddextexp.cc",
2304 "src/xnnpack/AlignedAllocator.h",
2305 ] + MICROKERNEL_BENCHMARK_HDRS,
2306 deps = MICROKERNEL_BENCHMARK_DEPS,
2307)
2308
2309xnnpack_benchmark(
2310 name = "f32_raddstoreexpminusmax_bench",
2311 srcs = [
2312 "bench/f32-raddstoreexpminusmax.cc",
2313 "src/xnnpack/AlignedAllocator.h",
2314 ] + MICROKERNEL_BENCHMARK_HDRS,
2315 deps = MICROKERNEL_BENCHMARK_DEPS,
2316)
2317
2318xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002319 name = "f32_rmax_bench",
2320 srcs = [
2321 "bench/f32-rmax.cc",
2322 "src/xnnpack/AlignedAllocator.h",
2323 ] + MICROKERNEL_BENCHMARK_HDRS,
2324 deps = MICROKERNEL_BENCHMARK_DEPS,
2325)
2326
2327xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002328 name = "f32_sigmoid_bench",
2329 srcs = [
2330 "bench/f32-sigmoid.cc",
2331 "src/xnnpack/AlignedAllocator.h",
2332 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002333 deps = MICROKERNEL_BENCHMARK_DEPS,
2334)
2335
2336xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002337 name = "f32_spmm_bench",
2338 srcs = [
2339 "bench/f32-spmm.cc",
2340 "bench/gemm.h",
2341 "src/xnnpack/AlignedAllocator.h",
2342 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002343 deps = MICROKERNEL_BENCHMARK_DEPS,
2344)
2345
2346xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002347 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002348 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002349 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002350 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002351 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002352 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002353)
2354
2355xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002356 name = "f32_vscaleexpminusmax_bench",
2357 srcs = [
2358 "bench/f32-vscaleexpminusmax.cc",
2359 "src/xnnpack/AlignedAllocator.h",
2360 ] + MICROKERNEL_BENCHMARK_HDRS,
2361 deps = MICROKERNEL_BENCHMARK_DEPS,
2362)
2363
2364xnnpack_benchmark(
2365 name = "f32_vscaleextexp_bench",
2366 srcs = [
2367 "bench/f32-vscaleextexp.cc",
2368 "src/xnnpack/AlignedAllocator.h",
2369 ] + MICROKERNEL_BENCHMARK_HDRS,
2370 deps = MICROKERNEL_BENCHMARK_DEPS,
2371)
2372
2373xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002374 name = "f32_im2col_gemm_bench",
2375 srcs = [
2376 "bench/f32-im2col-gemm.cc",
2377 "bench/conv.h",
2378 "src/xnnpack/AlignedAllocator.h",
2379 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2380 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2381)
2382
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002383xnnpack_benchmark(
2384 name = "requantization_bench",
2385 srcs = [
2386 "bench/requantization.cc",
2387 "src/xnnpack/requantization-stubs.h",
2388 "src/xnnpack/AlignedAllocator.h",
2389 ] + MICROKERNEL_BENCHMARK_HDRS,
2390 deps = MICROKERNEL_BENCHMARK_DEPS,
2391)
2392
Marat Dukhan08c4a432019-10-03 09:29:21 -07002393########################### Benchmarks for operators ###########################
2394
2395xnnpack_benchmark(
2396 name = "add_bench",
2397 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002398 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002399)
2400
2401xnnpack_benchmark(
2402 name = "average_pooling_bench",
2403 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002404 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002405 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002406 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002407)
2408
2409xnnpack_benchmark(
2410 name = "channel_shuffle_bench",
2411 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002412 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002413)
2414
2415xnnpack_benchmark(
2416 name = "convolution_bench",
2417 srcs = ["bench/convolution.cc"],
2418 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002419 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002420 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002421)
2422
2423xnnpack_benchmark(
2424 name = "deconvolution_bench",
2425 srcs = ["bench/deconvolution.cc"],
2426 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002427 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002428 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002429)
2430
2431xnnpack_benchmark(
2432 name = "global_average_pooling_bench",
2433 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002434 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002435)
2436
2437xnnpack_benchmark(
2438 name = "max_pooling_bench",
2439 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002440 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002441)
2442
2443xnnpack_benchmark(
2444 name = "sigmoid_bench",
2445 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002446 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002447 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002448 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002449)
2450
2451xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002452 name = "prelu_bench",
2453 srcs = ["bench/prelu.cc"],
2454 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002455 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002456 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002457)
2458
2459xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002460 name = "softmax_bench",
2461 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002462 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002463 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002464 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465)
2466
Marat Dukhanc068bb62019-10-04 13:24:39 -07002467############################# End-to-end benchmarks ############################
2468
2469cc_library(
2470 name = "mobilenet_v1",
2471 srcs = ["models/mobilenet-v1.cc"],
2472 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002473 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002474 linkstatic = True,
2475 deps = [
2476 ":XNNPACK",
2477 "@pthreadpool",
2478 ],
2479)
2480
2481cc_library(
2482 name = "mobilenet_v2",
2483 srcs = ["models/mobilenet-v2.cc"],
2484 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002485 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002486 linkstatic = True,
2487 deps = [
2488 ":XNNPACK",
2489 "@pthreadpool",
2490 ],
2491)
2492
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002493cc_library(
2494 name = "mobilenet_v3_large",
2495 srcs = ["models/mobilenet-v3-large.cc"],
2496 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002497 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002498 linkstatic = True,
2499 deps = [
2500 ":XNNPACK",
2501 "@pthreadpool",
2502 ],
2503)
2504
2505cc_library(
2506 name = "mobilenet_v3_small",
2507 srcs = ["models/mobilenet-v3-small.cc"],
2508 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002509 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002510 linkstatic = True,
2511 deps = [
2512 ":XNNPACK",
2513 "@pthreadpool",
2514 ],
2515)
2516
Marat Dukhanc068bb62019-10-04 13:24:39 -07002517xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002518 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002519 srcs = [
2520 "bench/f32-dwconv-e2e.cc",
2521 "bench/end2end.h",
2522 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002523 deps = MICROKERNEL_BENCHMARK_DEPS + [
2524 ":XNNPACK",
2525 ":mobilenet_v1",
2526 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002527 ":mobilenet_v3_large",
2528 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002529 ],
2530)
2531
2532xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002533 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002534 srcs = [
2535 "bench/f32-gemm-e2e.cc",
2536 "bench/end2end.h",
2537 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002538 deps = MICROKERNEL_BENCHMARK_DEPS + [
2539 ":XNNPACK",
2540 ":mobilenet_v1",
2541 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002542 ":mobilenet_v3_large",
2543 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002544 ],
2545)
2546
2547xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002548 name = "end2end_bench",
2549 srcs = ["bench/end2end.cc"],
2550 deps = [
2551 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002552 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002553 ":mobilenet_v1",
2554 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002555 ":mobilenet_v3_large",
2556 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002557 "@pthreadpool",
2558 ],
2559)
2560
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002561#################### Accuracy evaluation for math functions ####################
2562
2563xnnpack_benchmark(
2564 name = "f32_exp_eval",
2565 srcs = [
2566 "eval/f32-exp.cc",
2567 "src/xnnpack/AlignedAllocator.h",
2568 ] + ACCURACY_EVAL_HDRS,
2569 deps = ACCURACY_EVAL_DEPS,
2570)
2571
Marat Dukhan515c9772019-10-17 18:07:57 -07002572xnnpack_benchmark(
2573 name = "f32_expminus_eval",
2574 srcs = [
2575 "eval/f32-expminus.cc",
2576 "src/xnnpack/AlignedAllocator.h",
2577 ] + ACCURACY_EVAL_HDRS,
2578 deps = ACCURACY_EVAL_DEPS,
2579)
2580
Marat Dukhan98ba4412019-10-23 02:14:28 -07002581xnnpack_benchmark(
2582 name = "f32_extexp_eval",
2583 srcs = [
2584 "eval/f32-extexp.cc",
2585 "src/xnnpack/AlignedAllocator.h",
2586 ] + ACCURACY_EVAL_HDRS,
2587 deps = ACCURACY_EVAL_DEPS,
2588)
2589
Marat Dukhan8853b822020-05-07 12:19:01 -07002590xnnpack_unit_test(
2591 name = "f32_roundne_eval",
2592 srcs = [
2593 "eval/f32-roundne.cc",
2594 "src/xnnpack/AlignedAllocator.h",
2595 "src/xnnpack/math-stubs.h",
2596 ] + MICROKERNEL_TEST_HDRS,
2597 tags = [
2598 "notap",
2599 ],
2600 deps = MICROKERNEL_TEST_DEPS,
2601)
2602
Marat Dukhan346a9e52019-11-15 09:06:30 -08002603xnnpack_benchmark(
2604 name = "f32_sigmoid_eval",
2605 srcs = [
2606 "eval/f32-sigmoid.cc",
2607 "src/xnnpack/AlignedAllocator.h",
2608 ] + ACCURACY_EVAL_HDRS,
2609 deps = ACCURACY_EVAL_DEPS,
2610)
2611
Marat Dukhan08c4a432019-10-03 09:29:21 -07002612######################### Unit tests for micro-kernels #########################
2613
2614xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002615 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002616 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002617 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002618 "test/gemm-microkernel-tester.h",
2619 "src/xnnpack/AlignedAllocator.h",
2620 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2621 deps = MICROKERNEL_TEST_DEPS,
2622)
2623
2624xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002625 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002626 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002627 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002628 "test/spmm-microkernel-tester.h",
2629 "src/xnnpack/AlignedAllocator.h",
2630 ] + MICROKERNEL_TEST_HDRS,
2631 deps = MICROKERNEL_TEST_DEPS,
2632)
2633
2634xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002635 name = "f32_argmaxpool_test",
2636 srcs = [
2637 "test/f32-argmaxpool.cc",
2638 "test/argmaxpool-microkernel-tester.h",
2639 "src/xnnpack/AlignedAllocator.h",
2640 ] + MICROKERNEL_TEST_HDRS,
2641 deps = MICROKERNEL_TEST_DEPS,
2642)
2643
2644xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002645 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002646 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002647 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002648 "test/avgpool-microkernel-tester.h",
2649 "src/xnnpack/AlignedAllocator.h",
2650 ] + MICROKERNEL_TEST_HDRS,
2651 deps = MICROKERNEL_TEST_DEPS,
2652)
2653
2654xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002655 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002656 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002657 "test/f32-ibilinear.cc",
2658 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002659 "src/xnnpack/AlignedAllocator.h",
2660 ] + MICROKERNEL_TEST_HDRS,
2661 deps = MICROKERNEL_TEST_DEPS,
2662)
2663
2664xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002665 name = "f32_clamp_test",
2666 srcs = [
2667 "test/f32-clamp.cc",
2668 "test/clamp-microkernel-tester.h",
2669 ] + MICROKERNEL_TEST_HDRS,
2670 deps = MICROKERNEL_TEST_DEPS,
2671)
2672
2673xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002674 name = "f32_igemm_test",
2675 srcs = [
2676 "test/f32-igemm.cc",
2677 "test/gemm-microkernel-tester.h",
2678 "src/xnnpack/AlignedAllocator.h",
2679 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2680 deps = MICROKERNEL_TEST_DEPS,
2681)
2682
2683xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002684 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002685 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002686 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002687 "test/gemm-microkernel-tester.h",
2688 "src/xnnpack/AlignedAllocator.h",
2689 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002690 # Work-around for "too many sections" error
2691 mingw_copts = ["-Wa,-mbig-obj"],
2692 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002693 deps = MICROKERNEL_TEST_DEPS,
2694)
2695
2696xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002697 name = "f16_igemm_minmax_test",
2698 srcs = [
2699 "test/f16-igemm-minmax.cc",
2700 "test/gemm-microkernel-tester.h",
2701 "src/xnnpack/AlignedAllocator.h",
2702 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2703 deps = MICROKERNEL_TEST_DEPS,
2704)
2705
2706xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002707 name = "f32_conv_hwc_test",
2708 srcs = [
2709 "test/f32-conv-hwc.cc",
2710 "test/conv-hwc-microkernel-tester.h",
2711 "src/xnnpack/AlignedAllocator.h",
2712 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2713 deps = MICROKERNEL_TEST_DEPS,
2714)
2715
2716xnnpack_unit_test(
2717 name = "f32_conv_hwc2spchw_test",
2718 srcs = [
2719 "test/f32-conv-hwc2spchw.cc",
2720 "test/conv-hwc2spchw-microkernel-tester.h",
2721 "src/xnnpack/AlignedAllocator.h",
2722 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2723 deps = MICROKERNEL_TEST_DEPS,
2724)
2725
2726xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002727 name = "f32_dwconv_test",
2728 srcs = [
2729 "test/f32-dwconv.cc",
2730 "test/dwconv-microkernel-tester.h",
2731 "src/xnnpack/AlignedAllocator.h",
2732 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2733 deps = MICROKERNEL_TEST_DEPS,
2734)
2735
2736xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002737 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002738 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002739 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 "test/dwconv-microkernel-tester.h",
2741 "src/xnnpack/AlignedAllocator.h",
2742 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2743 deps = MICROKERNEL_TEST_DEPS,
2744)
2745
2746xnnpack_unit_test(
2747 name = "f32_dwconv_spchw_test",
2748 srcs = [
2749 "test/f32-dwconv-spchw.cc",
2750 "test/dwconv-spchw-microkernel-tester.h",
2751 "src/xnnpack/AlignedAllocator.h",
2752 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2753 deps = MICROKERNEL_TEST_DEPS,
2754)
2755
2756xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002757 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002758 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002759 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002760 "test/gavgpool-microkernel-tester.h",
2761 "src/xnnpack/AlignedAllocator.h",
2762 ] + MICROKERNEL_TEST_HDRS,
2763 deps = MICROKERNEL_TEST_DEPS,
2764)
2765
2766xnnpack_unit_test(
2767 name = "f32_gavgpool_spchw_test",
2768 srcs = [
2769 "test/f32-gavgpool-spchw.cc",
2770 "test/gavgpool-spchw-microkernel-tester.h",
2771 "src/xnnpack/AlignedAllocator.h",
2772 ] + MICROKERNEL_TEST_HDRS,
2773 deps = MICROKERNEL_TEST_DEPS,
2774)
2775
2776xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002777 name = "f32_gemm_test",
2778 srcs = [
2779 "test/f32-gemm.cc",
2780 "test/gemm-microkernel-tester.h",
2781 "src/xnnpack/AlignedAllocator.h",
2782 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2783 deps = MICROKERNEL_TEST_DEPS,
2784)
2785
2786xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002787 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002788 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002789 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002790 "test/gemm-microkernel-tester.h",
2791 "src/xnnpack/AlignedAllocator.h",
2792 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2793 deps = MICROKERNEL_TEST_DEPS,
2794)
2795
2796xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002797 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002798 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002799 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002800 "test/gemm-microkernel-tester.h",
2801 "src/xnnpack/AlignedAllocator.h",
2802 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2803 deps = MICROKERNEL_TEST_DEPS,
2804)
2805
2806xnnpack_unit_test(
2807 name = "f32_hswish_test",
2808 srcs = [
2809 "test/f32-hswish.cc",
2810 "test/hswish-microkernel-tester.h",
2811 ] + MICROKERNEL_TEST_HDRS,
2812 deps = MICROKERNEL_TEST_DEPS,
2813)
2814
2815xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002816 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002817 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002818 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002819 "test/maxpool-microkernel-tester.h",
2820 ] + MICROKERNEL_TEST_HDRS,
2821 deps = MICROKERNEL_TEST_DEPS,
2822)
2823
2824xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002825 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002826 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002827 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002828 "test/avgpool-microkernel-tester.h",
2829 "src/xnnpack/AlignedAllocator.h",
2830 ] + MICROKERNEL_TEST_HDRS,
2831 deps = MICROKERNEL_TEST_DEPS,
2832)
2833
2834xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002835 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002836 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002837 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002838 "test/gemm-microkernel-tester.h",
2839 "src/xnnpack/AlignedAllocator.h",
2840 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2841 deps = MICROKERNEL_TEST_DEPS,
2842)
2843
2844xnnpack_unit_test(
2845 name = "f32_prelu_test",
2846 srcs = [
2847 "test/f32-prelu.cc",
2848 "test/prelu-microkernel-tester.h",
2849 "src/xnnpack/AlignedAllocator.h",
2850 ] + MICROKERNEL_TEST_HDRS,
2851 deps = MICROKERNEL_TEST_DEPS,
2852)
2853
2854xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002855 name = "f32_raddexpminusmax_test",
2856 srcs = [
2857 "test/f32-raddexpminusmax.cc",
2858 "test/raddexpminusmax-microkernel-tester.h",
2859 ] + MICROKERNEL_TEST_HDRS,
2860 deps = MICROKERNEL_TEST_DEPS,
2861)
2862
2863xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002864 name = "f32_raddextexp_test",
2865 srcs = [
2866 "test/f32-raddextexp.cc",
2867 "test/raddextexp-microkernel-tester.h",
2868 ] + MICROKERNEL_TEST_HDRS,
2869 deps = MICROKERNEL_TEST_DEPS,
2870)
2871
2872xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002873 name = "f32_raddstoreexpminusmax_test",
2874 srcs = [
2875 "test/f32-raddstoreexpminusmax.cc",
2876 "test/raddstoreexpminusmax-microkernel-tester.h",
2877 ] + MICROKERNEL_TEST_HDRS,
2878 deps = MICROKERNEL_TEST_DEPS,
2879)
2880
2881xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 name = "f32_rmax_test",
2883 srcs = [
2884 "test/f32-rmax.cc",
2885 "test/rmax-microkernel-tester.h",
2886 ] + MICROKERNEL_TEST_HDRS,
2887 deps = MICROKERNEL_TEST_DEPS,
2888)
2889
2890xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002891 name = "f32_sigmoid_test",
2892 srcs = [
2893 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002894 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002895 ] + MICROKERNEL_TEST_HDRS,
2896 deps = MICROKERNEL_TEST_DEPS,
2897)
2898
2899xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002900 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002902 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903 "test/spmm-microkernel-tester.h",
2904 "src/xnnpack/AlignedAllocator.h",
2905 ] + MICROKERNEL_TEST_HDRS,
2906 deps = MICROKERNEL_TEST_DEPS,
2907)
2908
2909xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07002910 name = "f16_vadd_minmax_test",
2911 srcs = [
2912 "test/f16-vadd-minmax.cc",
2913 "test/vbinary-microkernel-tester.h",
2914 ] + MICROKERNEL_TEST_HDRS,
2915 deps = MICROKERNEL_TEST_DEPS,
2916)
2917
2918xnnpack_unit_test(
2919 name = "f16_vaddc_minmax_test",
2920 srcs = [
2921 "test/f16-vaddc-minmax.cc",
2922 "test/vbinaryc-microkernel-tester.h",
2923 ] + MICROKERNEL_TEST_HDRS,
2924 deps = MICROKERNEL_TEST_DEPS,
2925)
2926
2927xnnpack_unit_test(
2928 name = "f16_vdiv_minmax_test",
2929 srcs = [
2930 "test/f16-vdiv-minmax.cc",
2931 "test/vbinary-microkernel-tester.h",
2932 ] + MICROKERNEL_TEST_HDRS,
2933 deps = MICROKERNEL_TEST_DEPS,
2934)
2935
2936xnnpack_unit_test(
2937 name = "f16_vdivc_minmax_test",
2938 srcs = [
2939 "test/f16-vdivc-minmax.cc",
2940 "test/vbinaryc-microkernel-tester.h",
2941 ] + MICROKERNEL_TEST_HDRS,
2942 deps = MICROKERNEL_TEST_DEPS,
2943)
2944
2945xnnpack_unit_test(
2946 name = "f16_vrdivc_minmax_test",
2947 srcs = [
2948 "test/f16-vrdivc-minmax.cc",
2949 "test/vbinaryc-microkernel-tester.h",
2950 ] + MICROKERNEL_TEST_HDRS,
2951 deps = MICROKERNEL_TEST_DEPS,
2952)
2953
2954xnnpack_unit_test(
2955 name = "f16_vmax_test",
2956 srcs = [
2957 "test/f16-vmax.cc",
2958 "test/vbinary-microkernel-tester.h",
2959 ] + MICROKERNEL_TEST_HDRS,
2960 deps = MICROKERNEL_TEST_DEPS,
2961)
2962
2963xnnpack_unit_test(
2964 name = "f16_vmaxc_test",
2965 srcs = [
2966 "test/f16-vmaxc.cc",
2967 "test/vbinaryc-microkernel-tester.h",
2968 ] + MICROKERNEL_TEST_HDRS,
2969 deps = MICROKERNEL_TEST_DEPS,
2970)
2971
2972xnnpack_unit_test(
2973 name = "f16_vmin_test",
2974 srcs = [
2975 "test/f16-vmin.cc",
2976 "test/vbinary-microkernel-tester.h",
2977 ] + MICROKERNEL_TEST_HDRS,
2978 deps = MICROKERNEL_TEST_DEPS,
2979)
2980
2981xnnpack_unit_test(
2982 name = "f16_vminc_test",
2983 srcs = [
2984 "test/f16-vminc.cc",
2985 "test/vbinaryc-microkernel-tester.h",
2986 ] + MICROKERNEL_TEST_HDRS,
2987 deps = MICROKERNEL_TEST_DEPS,
2988)
2989
2990xnnpack_unit_test(
2991 name = "f16_vmul_minmax_test",
2992 srcs = [
2993 "test/f16-vmul-minmax.cc",
2994 "test/vbinary-microkernel-tester.h",
2995 ] + MICROKERNEL_TEST_HDRS,
2996 deps = MICROKERNEL_TEST_DEPS,
2997)
2998
2999xnnpack_unit_test(
3000 name = "f16_vmulc_minmax_test",
3001 srcs = [
3002 "test/f16-vmulc-minmax.cc",
3003 "test/vbinaryc-microkernel-tester.h",
3004 ] + MICROKERNEL_TEST_HDRS,
3005 deps = MICROKERNEL_TEST_DEPS,
3006)
3007
3008xnnpack_unit_test(
3009 name = "f16_vsub_minmax_test",
3010 srcs = [
3011 "test/f16-vsub-minmax.cc",
3012 "test/vbinary-microkernel-tester.h",
3013 ] + MICROKERNEL_TEST_HDRS,
3014 deps = MICROKERNEL_TEST_DEPS,
3015)
3016
3017xnnpack_unit_test(
3018 name = "f16_vsubc_minmax_test",
3019 srcs = [
3020 "test/f16-vsubc-minmax.cc",
3021 "test/vbinaryc-microkernel-tester.h",
3022 ] + MICROKERNEL_TEST_HDRS,
3023 deps = MICROKERNEL_TEST_DEPS,
3024)
3025
3026xnnpack_unit_test(
3027 name = "f16_vrsubc_minmax_test",
3028 srcs = [
3029 "test/f16-vrsubc-minmax.cc",
3030 "test/vbinaryc-microkernel-tester.h",
3031 ] + MICROKERNEL_TEST_HDRS,
3032 deps = MICROKERNEL_TEST_DEPS,
3033)
3034
3035xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003036 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003037 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003038 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003039 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003040 ] + MICROKERNEL_TEST_HDRS,
3041 deps = MICROKERNEL_TEST_DEPS,
3042)
3043
3044xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003045 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003046 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003047 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003048 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003049 ] + MICROKERNEL_TEST_HDRS,
3050 deps = MICROKERNEL_TEST_DEPS,
3051)
3052
3053xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003054 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003055 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003056 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003057 "test/vbinary-microkernel-tester.h",
3058 ] + MICROKERNEL_TEST_HDRS,
3059 deps = MICROKERNEL_TEST_DEPS,
3060)
3061
3062xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003063 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003064 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003065 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003066 "test/vbinaryc-microkernel-tester.h",
3067 ] + MICROKERNEL_TEST_HDRS,
3068 deps = MICROKERNEL_TEST_DEPS,
3069)
3070
3071xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003072 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003073 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003074 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003075 "test/vbinaryc-microkernel-tester.h",
3076 ] + MICROKERNEL_TEST_HDRS,
3077 deps = MICROKERNEL_TEST_DEPS,
3078)
3079
3080xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003081 name = "f32_vmax_test",
3082 srcs = [
3083 "test/f32-vmax.cc",
3084 "test/vbinary-microkernel-tester.h",
3085 ] + MICROKERNEL_TEST_HDRS,
3086 deps = MICROKERNEL_TEST_DEPS,
3087)
3088
3089xnnpack_unit_test(
3090 name = "f32_vmaxc_test",
3091 srcs = [
3092 "test/f32-vmaxc.cc",
3093 "test/vbinaryc-microkernel-tester.h",
3094 ] + MICROKERNEL_TEST_HDRS,
3095 deps = MICROKERNEL_TEST_DEPS,
3096)
3097
3098xnnpack_unit_test(
3099 name = "f32_vmin_test",
3100 srcs = [
3101 "test/f32-vmin.cc",
3102 "test/vbinary-microkernel-tester.h",
3103 ] + MICROKERNEL_TEST_HDRS,
3104 deps = MICROKERNEL_TEST_DEPS,
3105)
3106
3107xnnpack_unit_test(
3108 name = "f32_vminc_test",
3109 srcs = [
3110 "test/f32-vminc.cc",
3111 "test/vbinaryc-microkernel-tester.h",
3112 ] + MICROKERNEL_TEST_HDRS,
3113 deps = MICROKERNEL_TEST_DEPS,
3114)
3115
3116xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003117 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003118 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003119 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003120 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003121 ] + MICROKERNEL_TEST_HDRS,
3122 deps = MICROKERNEL_TEST_DEPS,
3123)
3124
3125xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003126 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003127 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003128 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003129 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130 ] + MICROKERNEL_TEST_HDRS,
3131 deps = MICROKERNEL_TEST_DEPS,
3132)
3133
3134xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003135 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003136 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003137 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003138 "test/vmulcaddc-microkernel-tester.h",
3139 "src/xnnpack/AlignedAllocator.h",
3140 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3141 deps = MICROKERNEL_TEST_DEPS,
3142)
3143
3144xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003145 name = "f32_vscale_test",
3146 srcs = [
3147 "test/f32-vscale.cc",
3148 "test/vscale-microkernel-tester.h",
3149 ] + MICROKERNEL_TEST_HDRS,
3150 deps = MICROKERNEL_TEST_DEPS,
3151)
3152
3153xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003154 name = "f32_vscaleexpminusmax_test",
3155 srcs = [
3156 "test/f32-vscaleexpminusmax.cc",
3157 "test/vscaleexpminusmax-microkernel-tester.h",
3158 ] + MICROKERNEL_TEST_HDRS,
3159 deps = MICROKERNEL_TEST_DEPS,
3160)
3161
3162xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003163 name = "f32_vscaleextexp_test",
3164 srcs = [
3165 "test/f32-vscaleextexp.cc",
3166 "test/vscaleextexp-microkernel-tester.h",
3167 ] + MICROKERNEL_TEST_HDRS,
3168 deps = MICROKERNEL_TEST_DEPS,
3169)
3170
3171xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003172 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003173 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003174 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003175 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003176 ] + MICROKERNEL_TEST_HDRS,
3177 deps = MICROKERNEL_TEST_DEPS,
3178)
3179
3180xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003181 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003182 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003183 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003184 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003185 ] + MICROKERNEL_TEST_HDRS,
3186 deps = MICROKERNEL_TEST_DEPS,
3187)
3188
3189xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003190 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003191 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003192 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003193 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003194 ] + MICROKERNEL_TEST_HDRS,
3195 deps = MICROKERNEL_TEST_DEPS,
3196)
3197
3198xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003199 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003200 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003201 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003202 "test/avgpool-microkernel-tester.h",
3203 "src/xnnpack/AlignedAllocator.h",
3204 ] + MICROKERNEL_TEST_HDRS,
3205 deps = MICROKERNEL_TEST_DEPS,
3206)
3207
3208xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003209 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003211 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212 "test/gemm-microkernel-tester.h",
3213 "src/xnnpack/AlignedAllocator.h",
3214 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3215 deps = MICROKERNEL_TEST_DEPS,
3216)
3217
3218xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003219 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003220 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003221 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003222 "test/dwconv-microkernel-tester.h",
3223 "src/xnnpack/AlignedAllocator.h",
3224 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3225 deps = MICROKERNEL_TEST_DEPS,
3226)
3227
3228xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003229 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003230 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003231 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003232 "test/gavgpool-microkernel-tester.h",
3233 "src/xnnpack/AlignedAllocator.h",
3234 ] + MICROKERNEL_TEST_HDRS,
3235 deps = MICROKERNEL_TEST_DEPS,
3236)
3237
3238xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003239 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003241 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003242 "test/gemm-microkernel-tester.h",
3243 "src/xnnpack/AlignedAllocator.h",
3244 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3245 deps = MICROKERNEL_TEST_DEPS,
3246)
3247
3248xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003249 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003251 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003252 "test/vadd-microkernel-tester.h",
3253 ] + MICROKERNEL_TEST_HDRS,
3254 deps = MICROKERNEL_TEST_DEPS,
3255)
3256
3257xnnpack_unit_test(
3258 name = "u8_clamp_test",
3259 srcs = [
3260 "test/u8-clamp.cc",
3261 "test/clamp-microkernel-tester.h",
3262 ] + MICROKERNEL_TEST_HDRS,
3263 deps = MICROKERNEL_TEST_DEPS,
3264)
3265
3266xnnpack_unit_test(
3267 name = "u8_lut32norm_test",
3268 srcs = [
3269 "test/u8-lut32norm.cc",
3270 "test/lut-norm-microkernel-tester.h",
3271 ] + MICROKERNEL_TEST_HDRS,
3272 deps = MICROKERNEL_TEST_DEPS,
3273)
3274
3275xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003276 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003277 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003278 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003279 "test/maxpool-microkernel-tester.h",
3280 ] + MICROKERNEL_TEST_HDRS,
3281 deps = MICROKERNEL_TEST_DEPS,
3282)
3283
3284xnnpack_unit_test(
3285 name = "u8_rmax_test",
3286 srcs = [
3287 "test/u8-rmax.cc",
3288 "test/rmax-microkernel-tester.h",
3289 ] + MICROKERNEL_TEST_HDRS,
3290 deps = MICROKERNEL_TEST_DEPS,
3291)
3292
3293xnnpack_unit_test(
3294 name = "x32_packx_test",
3295 srcs = [
3296 "test/x32-packx.cc",
3297 "test/pack-microkernel-tester.h",
3298 "src/xnnpack/AlignedAllocator.h",
3299 ] + MICROKERNEL_TEST_HDRS,
3300 deps = MICROKERNEL_TEST_DEPS,
3301)
3302
3303xnnpack_unit_test(
3304 name = "x32_pad_test",
3305 srcs = [
3306 "test/x32-pad.cc",
3307 "test/pad-microkernel-tester.h",
3308 ] + MICROKERNEL_TEST_HDRS,
3309 deps = MICROKERNEL_TEST_DEPS,
3310)
3311
3312xnnpack_unit_test(
3313 name = "x32_unpool_test",
3314 srcs = [
3315 "test/x32-unpool.cc",
3316 "test/unpool-microkernel-tester.h",
3317 ] + MICROKERNEL_TEST_HDRS,
3318 deps = MICROKERNEL_TEST_DEPS,
3319)
3320
3321xnnpack_unit_test(
3322 name = "x32_zip_test",
3323 srcs = [
3324 "test/x32-zip.cc",
3325 "test/zip-microkernel-tester.h",
3326 ] + MICROKERNEL_TEST_HDRS,
3327 deps = MICROKERNEL_TEST_DEPS,
3328)
3329
3330xnnpack_unit_test(
3331 name = "x8_lut_test",
3332 srcs = [
3333 "test/x8-lut.cc",
3334 "test/lut-microkernel-tester.h",
3335 ] + MICROKERNEL_TEST_HDRS,
3336 deps = MICROKERNEL_TEST_DEPS,
3337)
3338
3339xnnpack_unit_test(
3340 name = "x8_zip_test",
3341 srcs = [
3342 "test/x8-zip.cc",
3343 "test/zip-microkernel-tester.h",
3344 ] + MICROKERNEL_TEST_HDRS,
3345 deps = MICROKERNEL_TEST_DEPS,
3346)
3347
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003348xnnpack_unit_test(
3349 name = "requantization_test",
3350 srcs = [
3351 "src/xnnpack/requantization-stubs.h",
3352 "test/requantization.cc",
3353 "test/requantization-tester.h",
3354 ] + MICROKERNEL_TEST_HDRS,
3355 deps = MICROKERNEL_TEST_DEPS,
3356)
3357
Marat Dukhan20c3b922020-03-10 03:45:06 -07003358########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003359
3360xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003361 name = "operator_size_test",
3362 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003363 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364)
3365
Marat Dukhan20c3b922020-03-10 03:45:06 -07003366xnnpack_binary(
3367 name = "subgraph_size_test",
3368 srcs = ["test/subgraph-size.c"],
3369 deps = [":XNNPACK"],
3370)
3371
3372########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003373
3374xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003375 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003376 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003377 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003378 "test/add-operator-tester.h",
3379 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003380 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003381)
3382
3383xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003384 name = "add_nd_test",
3385 srcs = [
3386 "test/add-nd.cc",
3387 "test/binary-elementwise-operator-tester.h",
3388 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003389 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003390)
3391
3392xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003393 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003394 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003395 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003396 "test/argmax-pooling-operator-tester.h",
3397 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003398 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003399)
3400
3401xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003402 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003403 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003404 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003405 "test/average-pooling-operator-tester.h",
3406 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003407 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003408)
3409
3410xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003411 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003412 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003413 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003414 "test/channel-pad-operator-tester.h",
3415 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003416 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003417)
3418
3419xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003420 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003421 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003422 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003423 "test/channel-shuffle-operator-tester.h",
3424 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003425 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003426)
3427
3428xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003429 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003431 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003432 "test/clamp-operator-tester.h",
3433 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003434 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003435)
3436
3437xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003438 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003439 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003440 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003441 "test/convolution-operator-tester.h",
3442 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003443 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003444)
3445
3446xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003447 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003449 "test/convolution-nchw.cc",
3450 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003451 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003452 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003453)
3454
3455xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003456 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003457 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003458 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003459 "test/deconvolution-operator-tester.h",
3460 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003461 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003462)
3463
3464xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003465 name = "divide_nd_test",
3466 srcs = [
3467 "test/binary-elementwise-operator-tester.h",
3468 "test/divide-nd.cc",
3469 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003470 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003471)
3472
3473xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003474 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003475 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003476 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003477 "test/fully-connected-operator-tester.h",
3478 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003479 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003480)
3481
3482xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003483 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003485 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003486 "test/global-average-pooling-operator-tester.h",
3487 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003488 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003489)
3490
3491xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003492 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003494 "test/global-average-pooling-ncw.cc",
3495 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003496 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003497 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003498)
3499
3500xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003501 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003502 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003503 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003504 "test/hardswish-operator-tester.h",
3505 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003506 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003507)
3508
3509xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003510 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003512 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513 "test/leaky-relu-operator-tester.h",
3514 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003515 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003516)
3517
3518xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003519 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003520 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003521 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003522 "test/max-pooling-operator-tester.h",
3523 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003524 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003525)
3526
3527xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003528 name = "maximum_nd_test",
3529 srcs = [
3530 "test/binary-elementwise-operator-tester.h",
3531 "test/maximum-nd.cc",
3532 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003533 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003534)
3535
3536xnnpack_unit_test(
3537 name = "minimum_nd_test",
3538 srcs = [
3539 "test/binary-elementwise-operator-tester.h",
3540 "test/minimum-nd.cc",
3541 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003542 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003543)
3544
3545xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003546 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003547 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003548 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003549 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003550 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003551 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003552)
3553
3554xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003555 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003556 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003557 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003558 "test/prelu-operator-tester.h",
3559 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003560 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561)
3562
3563xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003564 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003565 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003566 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003567 "test/resize-bilinear-operator-tester.h",
3568 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003569 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003570)
3571
3572xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003573 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003575 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576 "test/sigmoid-operator-tester.h",
3577 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003578 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579)
3580
3581xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003582 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003584 "test/softmax-nc.cc",
3585 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003587 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588)
3589
3590xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003591 name = "subtract_nd_test",
3592 srcs = [
3593 "test/binary-elementwise-operator-tester.h",
3594 "test/subtract-nd.cc",
3595 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003596 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003597)
3598
3599xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003600 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003602 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603 "test/unpooling-operator-tester.h",
3604 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003605 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003606)
3607
3608############################# Build configurations #############################
3609
Marat Dukhanb8642352019-10-30 15:43:02 -07003610# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003612 name = "xnn_enable_assembly_explicit_true",
3613 define_values = {"xnn_enable_assembly": "true"},
3614)
3615
3616# Disables usage of assembly kernels.
3617config_setting(
3618 name = "xnn_enable_assembly_explicit_false",
3619 define_values = {"xnn_enable_assembly": "false"},
3620)
3621
Marat Dukhan05702cf2020-03-26 15:41:33 -07003622# Disables usage of HMP-aware optimizations.
3623config_setting(
3624 name = "xnn_enable_hmp_explicit_false",
3625 define_values = {"xnn_enable_hmp": "false"},
3626)
3627
Marat Dukhanb8642352019-10-30 15:43:02 -07003628# Builds with -c dbg
3629config_setting(
3630 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003631 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003632 "compilation_mode": "dbg",
3633 },
3634)
3635
3636# Builds with -c opt
3637config_setting(
3638 name = "optimized_build",
3639 values = {
3640 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003641 },
3642)
3643
3644config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003645 name = "linux_k8",
3646 values = {"cpu": "k8"},
3647)
3648
3649config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003650 name = "linux_arm",
3651 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003652)
3653
3654config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003655 name = "linux_armhf",
3656 values = {"cpu": "armhf"},
3657)
3658
3659config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003660 name = "linux_aarch64",
3661 values = {"cpu": "aarch64"},
3662)
3663
3664config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003665 name = "android",
3666 values = {"crosstool_top": "//external:android/crosstool"},
3667)
3668
3669config_setting(
3670 name = "android_armv7",
3671 values = {
3672 "crosstool_top": "//external:android/crosstool",
3673 "cpu": "armeabi-v7a",
3674 },
3675)
3676
3677config_setting(
3678 name = "android_arm64",
3679 values = {
3680 "crosstool_top": "//external:android/crosstool",
3681 "cpu": "arm64-v8a",
3682 },
3683)
3684
3685config_setting(
3686 name = "android_x86",
3687 values = {
3688 "crosstool_top": "//external:android/crosstool",
3689 "cpu": "x86",
3690 },
3691)
3692
3693config_setting(
3694 name = "android_x86_64",
3695 values = {
3696 "crosstool_top": "//external:android/crosstool",
3697 "cpu": "x86_64",
3698 },
3699)
3700
3701config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003702 name = "windows_x86_64",
3703 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003704)
3705
3706config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003707 name = "windows_x86_64_clang",
3708 values = {
3709 "compiler": "clang-cl",
3710 "cpu": "x64_windows",
3711 },
3712)
3713
3714config_setting(
3715 name = "windows_x86_64_mingw",
3716 values = {
3717 "compiler": "mingw-gcc",
3718 "cpu": "x64_windows",
3719 },
3720)
3721
3722config_setting(
3723 name = "windows_x86_64_msys",
3724 values = {
3725 "compiler": "msys-gcc",
3726 "cpu": "x64_windows",
3727 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003728)
3729
3730config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003731 name = "macos_x86_64",
3732 values = {
3733 "apple_platform_type": "macos",
3734 "cpu": "darwin",
3735 },
3736)
3737
3738config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003739 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003740 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003741)
3742
3743config_setting(
3744 name = "emscripten_wasm",
3745 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003746 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003747 "cpu": "wasm",
3748 },
3749)
3750
3751config_setting(
3752 name = "emscripten_wasmsimd",
3753 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003754 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003755 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003756 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003757 },
3758)
3759
3760config_setting(
3761 name = "emscripten_asmjs",
3762 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003763 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003764 "cpu": "asmjs",
3765 },
3766)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003767
3768config_setting(
3769 name = "ios_armv7",
3770 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003771 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003772 "cpu": "ios_armv7",
3773 },
3774)
3775
3776config_setting(
3777 name = "ios_arm64",
3778 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003779 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003780 "cpu": "ios_arm64",
3781 },
3782)
3783
3784config_setting(
3785 name = "ios_arm64e",
3786 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003787 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003788 "cpu": "ios_arm64e",
3789 },
3790)
3791
3792config_setting(
3793 name = "ios_x86",
3794 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003795 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003796 "cpu": "ios_i386",
3797 },
3798)
3799
3800config_setting(
3801 name = "ios_x86_64",
3802 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003803 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003804 "cpu": "ios_x86_64",
3805 },
3806)
3807
3808config_setting(
3809 name = "watchos_armv7k",
3810 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003811 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003812 "cpu": "watchos_armv7k",
3813 },
3814)
3815
3816config_setting(
3817 name = "watchos_arm64_32",
3818 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003819 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003820 "cpu": "watchos_arm64_32",
3821 },
3822)
3823
3824config_setting(
3825 name = "watchos_x86",
3826 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003827 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003828 "cpu": "watchos_i386",
3829 },
3830)
3831
3832config_setting(
3833 name = "watchos_x86_64",
3834 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003835 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003836 "cpu": "watchos_x86_64",
3837 },
3838)
3839
3840config_setting(
3841 name = "tvos_arm64",
3842 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003843 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003844 "cpu": "tvos_arm64",
3845 },
3846)
3847
3848config_setting(
3849 name = "tvos_x86_64",
3850 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003851 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003852 "cpu": "tvos_x86_64",
3853 },
3854)