blob: 9b0287b212b77a8c82848ca622037999334ed4df [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 Barchardb1966592020-05-12 13:47:06 -0700923 "src/f16-clamp/gen/neonfp16arith-x16.c",
924 "src/f16-clamp/gen/neonfp16arith-x8.c",
925 "src/f16-hswish/gen/neonfp16arith-x16.c",
926 "src/f16-hswish/gen/neonfp16arith-x8.c",
927 "src/f16-prelu/gen/neonfp16arith-2x16.c",
928 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -0700929 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
930 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
931 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
932 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
933 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
934 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
935 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
936 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
937 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
938 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
939 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
940 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
941 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
942 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
943 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
944 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
945 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
946 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
947 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
948 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
949 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
950 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
951 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
952 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
953 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
954 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
955 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
956 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard83ea9392020-04-20 20:15:14 -0700957 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
958 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
959 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
960 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
961 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
962 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
963 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
964 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -0700965 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
966 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
967 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
968 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -0700969 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
970 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
971 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
972 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
973 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
974 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
975 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
976 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
977 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
978 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
979 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
980 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700981 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
982 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
983 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
984 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
985 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
986 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
987 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
988 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700989]
990
991SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700992 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
993 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700994 "src/f32-clamp/gen/sse-x4.c",
995 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800996 "src/f32-dwconv-spchw/3x3p1-sse.c",
997 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700998 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
999 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1000 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1001 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1002 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1003 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1004 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1005 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1006 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1007 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1008 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1009 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001010 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001011 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1012 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001013 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1014 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1015 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1016 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1017 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1018 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1019 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1020 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1021 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1022 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1023 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1024 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1025 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001026 "src/f32-hswish/gen/sse-x4.c",
1027 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001028 "src/f32-ibilinear/gen/sse-c4.c",
1029 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001030 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1031 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1032 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1033 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1034 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1035 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1036 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001037 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1038 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1039 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001040 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001041 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001042 "src/f32-spmm/gen/4x1-minmax-sse.c",
1043 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001044 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1045 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1046 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1047 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1048 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1049 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1050 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1051 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001052 "src/f32-vbinary/gen/vmax-sse-x4.c",
1053 "src/f32-vbinary/gen/vmax-sse-x8.c",
1054 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1055 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1056 "src/f32-vbinary/gen/vmin-sse-x4.c",
1057 "src/f32-vbinary/gen/vmin-sse-x8.c",
1058 "src/f32-vbinary/gen/vminc-sse-x4.c",
1059 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001060 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1061 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1062 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1063 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1064 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1065 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1066 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1067 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1068 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1069 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1070 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1071 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001072 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1073 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001074 "src/x32-packx/x4-sse.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001075 "src/math/roundne-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001076]
1077
1078SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001079 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1080 "src/f32-argmaxpool/4x-sse2-c4.c",
1081 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001082 "src/f32-prelu/gen/sse2-2x4.c",
1083 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001084 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1085 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1086 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1087 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1088 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1090 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1091 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1092 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1093 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1094 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1095 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001096 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001097 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001098 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001099 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001100 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1101 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001102 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1103 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001104 "src/q8-igemm/4x4c2-minmax-sse2.c",
1105 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001106 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1107 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001108 "src/q8-gemm/2x4c8-minmax-sse2.c",
1109 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001110 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001111 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001112 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001113 "src/u8-rmax/sse2.c",
1114 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001115 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001116 "src/x32-zip/x2-sse2.c",
1117 "src/x32-zip/x3-sse2.c",
1118 "src/x32-zip/x4-sse2.c",
1119 "src/x32-zip/xm-sse2.c",
1120 "src/x8-zip/x2-sse2.c",
1121 "src/x8-zip/x3-sse2.c",
1122 "src/x8-zip/x4-sse2.c",
1123 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001124 "src/math/exp-sse2-p5.c",
1125 "src/math/expminus-sse2-p5.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001126 "src/math/roundne-sse2.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001127 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001128 "src/requantization/precise-sse2.c",
1129 "src/requantization/fp32-sse2.c",
1130 "src/requantization/q31-sse2.c",
1131 "src/requantization/gemmlowp-sse2.c",
1132]
1133
1134SSSE3_UKERNELS = [
1135 "src/requantization/precise-ssse3.c",
1136 "src/requantization/q31-ssse3.c",
1137 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001138]
1139
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001140SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001141 "src/f32-prelu/gen/sse41-2x4.c",
1142 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001143 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001144 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001145 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001146 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001147 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1148 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001149 "src/math/roundne-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001150 "src/requantization/precise-sse4.c",
1151 "src/requantization/q31-sse4.c",
1152 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001153]
1154
Marat Dukhan08c4a432019-10-03 09:29:21 -07001155AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001156 "src/f32-clamp/gen/avx-x8.c",
1157 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001158 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1159 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1160 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1161 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1162 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1163 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1164 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1165 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1166 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1167 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1168 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1169 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1170 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1171 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1172 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1173 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1174 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1175 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1176 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1177 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1178 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1179 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1180 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1181 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1182 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1183 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1184 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1185 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1186 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1187 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001188 "src/f32-hswish/gen/avx-x8.c",
1189 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001190 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1191 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1192 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1193 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1194 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1195 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1196 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1197 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1198 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001199 "src/f32-prelu/gen/avx-2x8.c",
1200 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001201 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001202 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1203 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1204 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1205 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1206 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1207 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1208 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1209 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001210 "src/f32-vbinary/gen/vmax-avx-x8.c",
1211 "src/f32-vbinary/gen/vmax-avx-x16.c",
1212 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1213 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1214 "src/f32-vbinary/gen/vmin-avx-x8.c",
1215 "src/f32-vbinary/gen/vmin-avx-x16.c",
1216 "src/f32-vbinary/gen/vminc-avx-x8.c",
1217 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001218 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1219 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1220 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1221 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1222 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1223 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1224 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1225 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1226 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1227 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1228 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1229 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001230 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001231]
1232
Marat Dukhanfda12b82019-11-21 12:27:59 -08001233FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001234 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1235 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1236 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1237 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1238 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1239 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1240 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1241 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1242 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1243 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1244 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1245 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1246 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1247 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1248 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1249 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1250 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1251 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1252 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1253 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1254 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1255 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1256 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1257 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1258 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1259 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1260 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1261 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1262 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1263 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1264 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1265 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1266 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1267 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1268 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1269 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1270 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1271 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1272 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1273 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001274 "src/f32-hswish/gen/fma3-x8.c",
1275 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001276 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1277 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1278 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1279 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1280 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1281 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1282 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1283 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1284 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1285 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1286 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1287 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1288 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1289 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001290]
1291
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001292AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001293 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1294 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1295 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1296 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1297 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1298 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1299 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1300 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1301 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1302 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1303 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1304 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1305 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1306 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1307 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1308 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1309 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1310 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1311 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1312 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1313 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1314 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1315 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1316 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1317 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1318 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1319 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1320 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1321 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1322 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1323 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1324 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1325 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1326 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1327 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1328 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001329 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1330 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1331 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1332 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1333 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1334 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1335 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1336 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1337 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1338 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1339 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1340 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1341 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1342 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1343 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1344 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1345 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1346 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1347 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1348 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1349 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1350 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1351 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1352 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1353 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1354 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1355 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1356 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1357 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1358 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001359 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1360 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1361 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1362 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1363 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1364 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1365 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1366 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1367 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1368 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1369 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1370 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1371 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1372 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1373 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1374 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1375 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1376 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1377 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1378 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1379 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1380 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1381 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1382 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001383 "src/math/exp-avx2-p5.c",
1384 "src/math/exp-avx2-perm-p3.c",
1385 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001386 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001387 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001388 "src/math/sigmoid-avx2-rr2-p5-div.c",
1389 "src/math/sigmoid-avx2-rr1-p5-div.c",
1390 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1391 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1392 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1393 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001394]
1395
Marat Dukhan08c4a432019-10-03 09:29:21 -07001396AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001397 "src/f32-clamp/gen/avx512f-x16.c",
1398 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001399 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1400 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1401 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1402 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1403 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1404 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1405 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1406 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1407 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1408 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1409 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1410 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1411 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1412 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1413 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1414 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1415 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1416 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1417 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1418 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1419 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1420 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1421 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1422 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001423 "src/f32-hswish/gen/avx512f-x16.c",
1424 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001425 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1426 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1427 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1428 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1429 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1430 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001431 "src/f32-prelu/gen/avx512f-2x16.c",
1432 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001433 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1434 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1435 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1436 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1437 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1438 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1439 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1440 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1441 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1442 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1443 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1444 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1445 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1446 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1447 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1448 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1449 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1450 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1451 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1452 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1453 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1454 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1455 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1456 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1457 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1458 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1459 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1460 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1461 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1462 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1463 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1464 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1465 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1466 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1467 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1468 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001469 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001470 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1471 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1472 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1473 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1474 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1475 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1476 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1477 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001478 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1479 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1480 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1481 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1482 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1483 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1484 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1485 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001486 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1487 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1488 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1489 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1490 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1491 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1492 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1493 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1494 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1495 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1496 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1497 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001498 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001499 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1500 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1501 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1502 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1503 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1504 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1505 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1506 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1507 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1508 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1509 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1510 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1511 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1512 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1513 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1514 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1515 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1516 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1517 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1518 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1519 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1520 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1521 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1522 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001523 "src/math/exp-avx512f-p5-scalef.c",
1524 "src/math/exp-avx512f-p5.c",
1525 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001526 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001527 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001528]
1529
1530AARCH32_ASM_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1532 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1533 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1534 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1535 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1536 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1537 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1538 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1539 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1540 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001541]
1542
1543AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001544 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001545 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1546 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001547 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001548 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1549 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001550 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1551 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1552 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001553 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001554 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1555 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1556 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001557 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001558 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1559 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001560 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001561 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1562 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1563 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1564 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1565 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1566 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1567 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1568 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1569 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1570 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1571 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1572 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1573 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1574 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1575 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1576 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1577 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1578 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1579 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1580 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1581 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001582 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001583 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1584 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1585 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1586 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1587 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1588 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1589 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1590 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1591 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1592 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1593 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1594 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1595 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1596 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1597 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1598 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1599 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1600 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1601 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1602 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1603 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1604 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1605 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1606 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1607 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1608 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1609 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1610 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1611 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1612 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1613 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1614 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1615 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1616 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1617 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1618 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1619 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1620 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001621]
1622
Marat Dukhan1b354632020-03-23 12:50:22 -07001623INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001624 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001625 "src/xnnpack/argmaxpool.h",
1626 "src/xnnpack/avgpool.h",
1627 "src/xnnpack/clamp.h",
1628 "src/xnnpack/common.h",
1629 "src/xnnpack/conv.h",
1630 "src/xnnpack/dwconv.h",
1631 "src/xnnpack/gavgpool.h",
1632 "src/xnnpack/gemm.h",
1633 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001634 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001635 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001636 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001637 "src/xnnpack/lut.h",
1638 "src/xnnpack/math.h",
1639 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001640 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001641 "src/xnnpack/packx.h",
1642 "src/xnnpack/pad.h",
1643 "src/xnnpack/params.h",
1644 "src/xnnpack/pavgpool.h",
1645 "src/xnnpack/ppmm.h",
1646 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001647 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001648 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001649 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001650 "src/xnnpack/rmax.h",
1651 "src/xnnpack/scalar-utils.h",
1652 "src/xnnpack/spmm.h",
1653 "src/xnnpack/unpool.h",
1654 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001655 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001656 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001657 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001658 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001659 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001660 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001661 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001662]
1663
1664INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001665 "include/xnnpack.h",
1666 "src/xnnpack/allocator.h",
1667 "src/xnnpack/compute.h",
1668 "src/xnnpack/im2col.h",
1669 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001670 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001671 "src/xnnpack/operator.h",
1672 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001673 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001674 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001675 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001676 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001677]
1678
Marat Dukhan1b354632020-03-23 12:50:22 -07001679ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001680 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001681]
1682
Marat Dukhan1b354632020-03-23 12:50:22 -07001683MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001684 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001685 "include/xnnpack.h",
1686]
1687
Marat Dukhan1b354632020-03-23 12:50:22 -07001688MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001689 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001690 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001691 "src/xnnpack/requantization.h",
1692 "include/xnnpack.h",
1693]
1694
1695OPERATOR_TEST_PARAMS_HDRS = [
1696 "src/xnnpack/params.h",
1697 "src/xnnpack/common.h",
1698]
1699
1700WEIGHTS_PACK_HDRS = [
1701 "src/xnnpack/pack.h",
1702 "src/xnnpack/operator.h",
1703 "src/xnnpack/compute.h",
1704]
1705
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001706LOGGING_COPTS = select({
1707 # No logging in optimized mode
1708 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1709 # Full logging in debug mode
1710 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1711 # Error-only logging in default (fastbuild) mode
1712 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1713})
1714
1715LOGGING_HDRS = [
1716 "src/xnnpack/log.h",
1717]
1718
Marat Dukhan08c4a432019-10-03 09:29:21 -07001719xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001720 name = "tables",
1721 srcs = TABLE_SRCS,
1722 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001723 gcc_copts = xnnpack_gcc_std_copts(),
1724 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001725)
1726
1727xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001728 name = "scalar_ukernels",
1729 srcs = SCALAR_UKERNELS,
1730 hdrs = INTERNAL_HDRS,
1731 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001732 gcc_copts = xnnpack_gcc_std_copts(),
1733 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001734 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001735 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001736 "@FP16",
1737 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001738 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001739 ],
1740)
1741
1742xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001743 name = "wasm_ukernels",
1744 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001745 gcc_copts = xnnpack_gcc_std_copts(),
1746 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001747 wasm_srcs = WASM_UKERNELS,
1748 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001749 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001750 "@FP16",
1751 "@FXdiv",
1752 "@pthreadpool",
1753 ],
1754)
1755
1756xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001757 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001758 hdrs = INTERNAL_HDRS,
1759 aarch32_copts = [
1760 "-marm",
1761 "-mfpu=neon",
1762 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001763 gcc_copts = xnnpack_gcc_std_copts(),
1764 gcc_x86_copts = ["-msse2"],
1765 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001766 optimized_copts = [
1767 "-O3",
1768 "-ffast-math",
1769 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001770 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001771 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001772 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001773 "@FP16",
1774 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001775 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001776 ],
1777)
1778
1779xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001780 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001781 hdrs = INTERNAL_HDRS,
1782 aarch32_copts = [
1783 "-marm",
1784 "-mfpu=neon",
1785 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001786 gcc_copts = xnnpack_gcc_std_copts(),
1787 gcc_x86_copts = ["-msse2"],
1788 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001789 optimized_copts = [
1790 "-O3",
1791 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001792 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001793 deps = [
1794 ":tables",
1795 "@FP16",
1796 "@psimd",
1797 "@pthreadpool",
1798 ],
1799)
1800
1801xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001802 name = "neon_ukernels",
1803 hdrs = INTERNAL_HDRS,
1804 aarch32_copts = [
1805 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001806 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001807 "-mfpu=neon",
1808 ],
1809 aarch32_srcs = NEON_UKERNELS,
1810 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001811 gcc_copts = xnnpack_gcc_std_copts(),
1812 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001813 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001814 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001815 "@FP16",
1816 "@pthreadpool",
1817 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001818)
1819
1820xnnpack_cc_library(
1821 name = "neonfma_ukernels",
1822 hdrs = INTERNAL_HDRS,
1823 aarch32_copts = [
1824 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001825 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001826 "-mfpu=neon-vfpv4",
1827 ],
1828 aarch32_srcs = NEONFMA_UKERNELS,
1829 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001830 gcc_copts = xnnpack_gcc_std_copts(),
1831 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001832 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001833 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001834 "@FP16",
1835 "@pthreadpool",
1836 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001837)
1838
1839xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001840 name = "neonv8_ukernels",
1841 hdrs = INTERNAL_HDRS,
1842 aarch32_copts = [
1843 "-marm",
1844 "-march=armv8-a",
1845 "-mfpu=neon-fp-armv8",
1846 ],
1847 aarch32_srcs = NEONV8_UKERNELS,
1848 aarch64_srcs = NEONV8_UKERNELS,
1849 gcc_copts = xnnpack_gcc_std_copts(),
1850 msvc_copts = xnnpack_msvc_std_copts(),
1851 deps = [
1852 ":tables",
1853 "@FP16",
1854 "@pthreadpool",
1855 ],
1856)
1857
1858xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001859 name = "neonfp16arith_ukernels",
1860 hdrs = INTERNAL_HDRS,
1861 aarch64_copts = ["-march=armv8.2-a+fp16"],
1862 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001863 gcc_copts = xnnpack_gcc_std_copts(),
1864 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001865 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001866 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001867 "@FP16",
1868 "@pthreadpool",
1869 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001870)
1871
1872xnnpack_cc_library(
1873 name = "sse2_ukernels",
1874 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001875 gcc_copts = xnnpack_gcc_std_copts(),
1876 gcc_x86_copts = ["-msse2"],
1877 msvc_copts = xnnpack_msvc_std_copts(),
1878 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001879 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001880 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001881 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001882 "@FP16",
1883 "@pthreadpool",
1884 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001885)
1886
1887xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001888 name = "ssse3_ukernels",
1889 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001890 gcc_copts = xnnpack_gcc_std_copts(),
1891 gcc_x86_copts = ["-mssse3"],
1892 msvc_copts = xnnpack_msvc_std_copts(),
1893 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001894 x86_srcs = SSSE3_UKERNELS,
1895 deps = [
1896 ":tables",
1897 "@FP16",
1898 "@pthreadpool",
1899 ],
1900)
1901
1902xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001903 name = "sse41_ukernels",
1904 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001905 gcc_copts = xnnpack_gcc_std_copts(),
1906 gcc_x86_copts = ["-msse4.1"],
1907 msvc_copts = xnnpack_msvc_std_copts(),
1908 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001909 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001910 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001911 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001912 "@FP16",
1913 "@pthreadpool",
1914 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001915)
1916
1917xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001918 name = "avx_ukernels",
1919 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001920 gcc_copts = xnnpack_gcc_std_copts(),
1921 gcc_x86_copts = ["-mavx"],
1922 msvc_copts = xnnpack_msvc_std_copts(),
1923 msvc_x86_32_copts = ["/arch:AVX"],
1924 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001925 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001926 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001927 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001928 "@FP16",
1929 "@pthreadpool",
1930 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001931)
1932
1933xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001934 name = "fma3_ukernels",
1935 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001936 gcc_copts = xnnpack_gcc_std_copts(),
1937 gcc_x86_copts = ["-mfma"],
1938 msvc_copts = xnnpack_msvc_std_copts(),
1939 msvc_x86_32_copts = ["/arch:AVX"],
1940 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001941 x86_srcs = FMA3_UKERNELS,
1942 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001943 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001944 "@FP16",
1945 "@pthreadpool",
1946 ],
1947)
1948
1949xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001950 name = "avx2_ukernels",
1951 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001952 gcc_copts = xnnpack_gcc_std_copts(),
1953 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001954 "-mfma",
1955 "-mavx2",
1956 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001957 msvc_copts = xnnpack_msvc_std_copts(),
1958 msvc_x86_32_copts = ["/arch:AVX2"],
1959 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001960 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001961 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001962 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001963 "@FP16",
1964 "@pthreadpool",
1965 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001966)
1967
1968xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001969 name = "avx512f_ukernels",
1970 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001971 gcc_copts = xnnpack_gcc_std_copts(),
1972 gcc_x86_copts = ["-mavx512f"],
1973 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1974 msvc_copts = xnnpack_msvc_std_copts(),
1975 msvc_x86_32_copts = ["/arch:AVX512"],
1976 msvc_x86_64_copts = ["/arch:AVX512"],
1977 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001978 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001979 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001980 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001981 "@FP16",
1982 "@pthreadpool",
1983 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001984)
1985
1986xnnpack_cc_library(
1987 name = "asm_ukernels",
1988 hdrs = ["src/xnnpack/assembly.h"],
1989 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001990 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001991 aarch64_srcs = AARCH64_ASM_UKERNELS,
1992)
1993
1994xnnpack_aggregate_library(
1995 name = "ukernels",
1996 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997 ":neon_ukernels",
1998 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07001999 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002000 ":asm_ukernels",
2001 ],
2002 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002003 ":neon_ukernels",
2004 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002005 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002006 ":neonfp16arith_ukernels",
2007 ":asm_ukernels",
2008 ],
2009 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07002010 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002011 ":psimd_fastmath_ukernels",
2012 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002013 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002014 wasm_deps = [":wasm_ukernels"],
2015 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002016 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002017 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002018 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002019 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002020 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002021 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002022 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002023 ":avx512f_ukernels",
2024 ],
2025)
2026
2027xnnpack_cc_library(
2028 name = "im2col",
2029 srcs = ["src/im2col.c"],
2030 hdrs = [
2031 "src/xnnpack/common.h",
2032 "src/xnnpack/im2col.h",
2033 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002034 gcc_copts = xnnpack_gcc_std_copts(),
2035 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002036)
2037
2038xnnpack_cc_library(
2039 name = "indirection",
2040 srcs = ["src/indirection.c"],
2041 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002042 gcc_copts = xnnpack_gcc_std_copts(),
2043 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002044 deps = [
2045 "@FP16",
2046 "@FXdiv",
2047 "@pthreadpool",
2048 ],
2049)
2050
2051xnnpack_cc_library(
2052 name = "operator_run",
2053 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002054 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002055 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002056 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2057 "//conditions:default": [],
2058 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002059 gcc_copts = xnnpack_gcc_std_copts(),
2060 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002061 deps = [
2062 "@FP16",
2063 "@FXdiv",
2064 "@clog",
2065 "@pthreadpool",
2066 ],
2067)
2068
2069cc_library(
2070 name = "enable_assembly",
2071 defines = select({
2072 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2073 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002074 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002075 }),
2076)
2077
Marat Dukhancf056b22019-10-07 10:26:29 -07002078xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002079 name = "operators",
2080 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002081 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002082 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002083 ],
2084 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002085 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002086 "-Isrc",
2087 "-Iinclude",
2088 ] + select({
2089 ":debug_build": [],
2090 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002091 }) + select({
2092 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2093 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002094 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002095 gcc_copts = xnnpack_gcc_std_copts(),
2096 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002097 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002098 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002099 "@FP16",
2100 "@FXdiv",
2101 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002102 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002103 ],
2104)
2105
Marat Dukhan10a38082020-04-17 03:58:35 -07002106xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002107 name = "XNNPACK",
2108 srcs = [
2109 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002110 "src/runtime.c",
2111 "src/subgraph.c",
2112 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002113 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002114 hdrs = ["include/xnnpack.h"],
2115 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002116 "-Isrc",
2117 "-Iinclude",
2118 ] + select({
2119 ":debug_build": [],
2120 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002121 }) + select({
2122 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2123 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002124 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002125 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002126 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002127 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002128 visibility = xnnpack_visibility(),
2129 deps = [
2130 ":enable_assembly",
2131 ":ukernels",
2132 ":operator_run",
2133 ":operators",
2134 "@clog",
2135 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002136 ] + select({
2137 ":emscripten": [],
2138 "//conditions:default": ["@cpuinfo"],
2139 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002140)
2141
Marat Dukhan10a38082020-04-17 03:58:35 -07002142xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002143 name = "xnnpack_operators_nhwc_f32",
2144 srcs = [
2145 "src/init.c",
2146 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002147 hdrs = ["include/xnnpack.h"],
2148 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002149 "-Isrc",
2150 "-Iinclude",
2151 ] + select({
2152 ":debug_build": [],
2153 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002154 }) + select({
2155 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2156 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002157 }),
2158 defines = [
2159 "XNN_NO_Q8_OPERATORS",
2160 "XNN_NO_U8_OPERATORS",
2161 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002162 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002163 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002164 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002165 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002166 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002167 visibility = xnnpack_visibility(),
2168 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002169 ":enable_assembly",
2170 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002171 ":operator_run",
2172 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002173 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002174 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002175 ] + select({
2176 ":emscripten": [],
2177 "//conditions:default": ["@cpuinfo"],
2178 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002179)
2180
Marat Dukhancf056b22019-10-07 10:26:29 -07002181xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002182 name = "bench_utils",
2183 srcs = ["bench/utils.cc"],
2184 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002185 deps = [
2186 "@com_google_benchmark//:benchmark",
2187 "@cpuinfo",
2188 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002189)
2190
Frank Barchard7e955972019-10-11 10:34:25 -07002191######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002192
2193xnnpack_benchmark(
2194 name = "q8_gemm_bench",
2195 srcs = [
2196 "bench/gemm.h",
2197 "bench/q8-gemm.cc",
2198 "src/xnnpack/AlignedAllocator.h",
2199 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002200 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002201 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2202)
2203
2204xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002205 name = "f16_igemm_bench",
2206 srcs = [
2207 "bench/f16-igemm.cc",
2208 "bench/conv.h",
2209 "bench/google/conv.h",
2210 "src/xnnpack/AlignedAllocator.h",
2211 ] + MICROKERNEL_BENCHMARK_HDRS,
2212 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2213)
2214
2215xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002216 name = "f16_gemm_bench",
2217 srcs = [
2218 "bench/f16-gemm.cc",
2219 "bench/gemm.h",
2220 "src/xnnpack/AlignedAllocator.h",
2221 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002222 deps = MICROKERNEL_BENCHMARK_DEPS,
2223)
2224
2225xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002226 name = "f16_spmm_bench",
2227 srcs = [
2228 "bench/f16-spmm.cc",
2229 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002230 "src/xnnpack/AlignedAllocator.h",
2231 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002232 deps = MICROKERNEL_BENCHMARK_DEPS,
2233)
2234
2235xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002236 name = "f32_igemm_bench",
2237 srcs = [
2238 "bench/f32-igemm.cc",
2239 "bench/conv.h",
2240 "src/xnnpack/AlignedAllocator.h",
2241 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002242 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243)
2244
2245xnnpack_benchmark(
2246 name = "f32_conv_hwc_bench",
2247 srcs = [
2248 "bench/f32-conv-hwc.cc",
2249 "bench/dconv.h",
2250 "src/xnnpack/AlignedAllocator.h",
2251 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002252 deps = MICROKERNEL_BENCHMARK_DEPS,
2253)
2254
2255xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002256 name = "f32_conv_hwc2spchw_bench",
2257 srcs = [
2258 "bench/f32-conv-hwc2spchw.cc",
2259 "bench/dconv.h",
2260 "src/xnnpack/AlignedAllocator.h",
2261 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002262 deps = MICROKERNEL_BENCHMARK_DEPS,
2263)
2264
2265xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002266 name = "f32_dwconv_bench",
2267 srcs = [
2268 "bench/f32-dwconv.cc",
2269 "bench/dwconv.h",
2270 "src/xnnpack/AlignedAllocator.h",
2271 ] + MICROKERNEL_BENCHMARK_HDRS,
2272 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2273)
2274
2275xnnpack_benchmark(
2276 name = "f32_dwconv_spchw_bench",
2277 srcs = [
2278 "bench/f32-dwconv-spchw.cc",
2279 "bench/dwconv.h",
2280 "src/xnnpack/AlignedAllocator.h",
2281 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2282 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2283)
2284
2285xnnpack_benchmark(
2286 name = "f32_gemm_bench",
2287 srcs = [
2288 "bench/f32-gemm.cc",
2289 "bench/gemm.h",
2290 "src/xnnpack/AlignedAllocator.h",
2291 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002292 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002293 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002294)
2295
2296xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002297 name = "f32_raddexpminusmax_bench",
2298 srcs = [
2299 "bench/f32-raddexpminusmax.cc",
2300 "src/xnnpack/AlignedAllocator.h",
2301 ] + MICROKERNEL_BENCHMARK_HDRS,
2302 deps = MICROKERNEL_BENCHMARK_DEPS,
2303)
2304
2305xnnpack_benchmark(
2306 name = "f32_raddextexp_bench",
2307 srcs = [
2308 "bench/f32-raddextexp.cc",
2309 "src/xnnpack/AlignedAllocator.h",
2310 ] + MICROKERNEL_BENCHMARK_HDRS,
2311 deps = MICROKERNEL_BENCHMARK_DEPS,
2312)
2313
2314xnnpack_benchmark(
2315 name = "f32_raddstoreexpminusmax_bench",
2316 srcs = [
2317 "bench/f32-raddstoreexpminusmax.cc",
2318 "src/xnnpack/AlignedAllocator.h",
2319 ] + MICROKERNEL_BENCHMARK_HDRS,
2320 deps = MICROKERNEL_BENCHMARK_DEPS,
2321)
2322
2323xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002324 name = "f32_rmax_bench",
2325 srcs = [
2326 "bench/f32-rmax.cc",
2327 "src/xnnpack/AlignedAllocator.h",
2328 ] + MICROKERNEL_BENCHMARK_HDRS,
2329 deps = MICROKERNEL_BENCHMARK_DEPS,
2330)
2331
2332xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002333 name = "f32_sigmoid_bench",
2334 srcs = [
2335 "bench/f32-sigmoid.cc",
2336 "src/xnnpack/AlignedAllocator.h",
2337 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002338 deps = MICROKERNEL_BENCHMARK_DEPS,
2339)
2340
2341xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002342 name = "f32_spmm_bench",
2343 srcs = [
2344 "bench/f32-spmm.cc",
2345 "bench/gemm.h",
2346 "src/xnnpack/AlignedAllocator.h",
2347 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002348 deps = MICROKERNEL_BENCHMARK_DEPS,
2349)
2350
2351xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002352 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002353 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002354 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002355 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002356 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002357 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002358)
2359
2360xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002361 name = "f32_vscaleexpminusmax_bench",
2362 srcs = [
2363 "bench/f32-vscaleexpminusmax.cc",
2364 "src/xnnpack/AlignedAllocator.h",
2365 ] + MICROKERNEL_BENCHMARK_HDRS,
2366 deps = MICROKERNEL_BENCHMARK_DEPS,
2367)
2368
2369xnnpack_benchmark(
2370 name = "f32_vscaleextexp_bench",
2371 srcs = [
2372 "bench/f32-vscaleextexp.cc",
2373 "src/xnnpack/AlignedAllocator.h",
2374 ] + MICROKERNEL_BENCHMARK_HDRS,
2375 deps = MICROKERNEL_BENCHMARK_DEPS,
2376)
2377
2378xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002379 name = "f32_im2col_gemm_bench",
2380 srcs = [
2381 "bench/f32-im2col-gemm.cc",
2382 "bench/conv.h",
2383 "src/xnnpack/AlignedAllocator.h",
2384 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2385 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2386)
2387
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002388xnnpack_benchmark(
2389 name = "requantization_bench",
2390 srcs = [
2391 "bench/requantization.cc",
2392 "src/xnnpack/requantization-stubs.h",
2393 "src/xnnpack/AlignedAllocator.h",
2394 ] + MICROKERNEL_BENCHMARK_HDRS,
2395 deps = MICROKERNEL_BENCHMARK_DEPS,
2396)
2397
Marat Dukhan08c4a432019-10-03 09:29:21 -07002398########################### Benchmarks for operators ###########################
2399
2400xnnpack_benchmark(
2401 name = "add_bench",
2402 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002403 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404)
2405
2406xnnpack_benchmark(
2407 name = "average_pooling_bench",
2408 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002409 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002410 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002411 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002412)
2413
2414xnnpack_benchmark(
2415 name = "channel_shuffle_bench",
2416 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002417 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002418)
2419
2420xnnpack_benchmark(
2421 name = "convolution_bench",
2422 srcs = ["bench/convolution.cc"],
2423 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002424 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002425 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002426)
2427
2428xnnpack_benchmark(
2429 name = "deconvolution_bench",
2430 srcs = ["bench/deconvolution.cc"],
2431 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002432 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002433 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002434)
2435
2436xnnpack_benchmark(
2437 name = "global_average_pooling_bench",
2438 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002439 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002440)
2441
2442xnnpack_benchmark(
2443 name = "max_pooling_bench",
2444 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002445 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002446)
2447
2448xnnpack_benchmark(
2449 name = "sigmoid_bench",
2450 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002451 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002452 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002453 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002454)
2455
2456xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002457 name = "prelu_bench",
2458 srcs = ["bench/prelu.cc"],
2459 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002460 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002461 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002462)
2463
2464xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002465 name = "softmax_bench",
2466 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002467 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002468 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002469 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002470)
2471
Marat Dukhanc068bb62019-10-04 13:24:39 -07002472############################# End-to-end benchmarks ############################
2473
2474cc_library(
2475 name = "mobilenet_v1",
2476 srcs = ["models/mobilenet-v1.cc"],
2477 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002478 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002479 linkstatic = True,
2480 deps = [
2481 ":XNNPACK",
2482 "@pthreadpool",
2483 ],
2484)
2485
2486cc_library(
2487 name = "mobilenet_v2",
2488 srcs = ["models/mobilenet-v2.cc"],
2489 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002490 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002491 linkstatic = True,
2492 deps = [
2493 ":XNNPACK",
2494 "@pthreadpool",
2495 ],
2496)
2497
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002498cc_library(
2499 name = "mobilenet_v3_large",
2500 srcs = ["models/mobilenet-v3-large.cc"],
2501 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002502 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002503 linkstatic = True,
2504 deps = [
2505 ":XNNPACK",
2506 "@pthreadpool",
2507 ],
2508)
2509
2510cc_library(
2511 name = "mobilenet_v3_small",
2512 srcs = ["models/mobilenet-v3-small.cc"],
2513 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002514 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002515 linkstatic = True,
2516 deps = [
2517 ":XNNPACK",
2518 "@pthreadpool",
2519 ],
2520)
2521
Marat Dukhanc068bb62019-10-04 13:24:39 -07002522xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002523 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002524 srcs = [
2525 "bench/f32-dwconv-e2e.cc",
2526 "bench/end2end.h",
2527 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002528 deps = MICROKERNEL_BENCHMARK_DEPS + [
2529 ":XNNPACK",
2530 ":mobilenet_v1",
2531 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002532 ":mobilenet_v3_large",
2533 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002534 ],
2535)
2536
2537xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002538 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002539 srcs = [
2540 "bench/f32-gemm-e2e.cc",
2541 "bench/end2end.h",
2542 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002543 deps = MICROKERNEL_BENCHMARK_DEPS + [
2544 ":XNNPACK",
2545 ":mobilenet_v1",
2546 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002547 ":mobilenet_v3_large",
2548 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002549 ],
2550)
2551
2552xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002553 name = "end2end_bench",
2554 srcs = ["bench/end2end.cc"],
2555 deps = [
2556 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002557 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002558 ":mobilenet_v1",
2559 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002560 ":mobilenet_v3_large",
2561 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002562 "@pthreadpool",
2563 ],
2564)
2565
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002566#################### Accuracy evaluation for math functions ####################
2567
2568xnnpack_benchmark(
2569 name = "f32_exp_eval",
2570 srcs = [
2571 "eval/f32-exp.cc",
2572 "src/xnnpack/AlignedAllocator.h",
2573 ] + ACCURACY_EVAL_HDRS,
2574 deps = ACCURACY_EVAL_DEPS,
2575)
2576
Marat Dukhan515c9772019-10-17 18:07:57 -07002577xnnpack_benchmark(
2578 name = "f32_expminus_eval",
2579 srcs = [
2580 "eval/f32-expminus.cc",
2581 "src/xnnpack/AlignedAllocator.h",
2582 ] + ACCURACY_EVAL_HDRS,
2583 deps = ACCURACY_EVAL_DEPS,
2584)
2585
Marat Dukhan98ba4412019-10-23 02:14:28 -07002586xnnpack_benchmark(
2587 name = "f32_extexp_eval",
2588 srcs = [
2589 "eval/f32-extexp.cc",
2590 "src/xnnpack/AlignedAllocator.h",
2591 ] + ACCURACY_EVAL_HDRS,
2592 deps = ACCURACY_EVAL_DEPS,
2593)
2594
Marat Dukhan8853b822020-05-07 12:19:01 -07002595xnnpack_unit_test(
2596 name = "f32_roundne_eval",
2597 srcs = [
2598 "eval/f32-roundne.cc",
2599 "src/xnnpack/AlignedAllocator.h",
2600 "src/xnnpack/math-stubs.h",
2601 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07002602 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07002603 deps = MICROKERNEL_TEST_DEPS,
2604)
2605
Marat Dukhan346a9e52019-11-15 09:06:30 -08002606xnnpack_benchmark(
2607 name = "f32_sigmoid_eval",
2608 srcs = [
2609 "eval/f32-sigmoid.cc",
2610 "src/xnnpack/AlignedAllocator.h",
2611 ] + ACCURACY_EVAL_HDRS,
2612 deps = ACCURACY_EVAL_DEPS,
2613)
2614
Marat Dukhan08c4a432019-10-03 09:29:21 -07002615######################### Unit tests for micro-kernels #########################
2616
2617xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002618 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002619 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002620 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002621 "test/gemm-microkernel-tester.h",
2622 "src/xnnpack/AlignedAllocator.h",
2623 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2624 deps = MICROKERNEL_TEST_DEPS,
2625)
2626
2627xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002628 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002629 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002630 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002631 "test/spmm-microkernel-tester.h",
2632 "src/xnnpack/AlignedAllocator.h",
2633 ] + MICROKERNEL_TEST_HDRS,
2634 deps = MICROKERNEL_TEST_DEPS,
2635)
2636
2637xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002638 name = "f32_argmaxpool_test",
2639 srcs = [
2640 "test/f32-argmaxpool.cc",
2641 "test/argmaxpool-microkernel-tester.h",
2642 "src/xnnpack/AlignedAllocator.h",
2643 ] + MICROKERNEL_TEST_HDRS,
2644 deps = MICROKERNEL_TEST_DEPS,
2645)
2646
2647xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002648 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002649 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002650 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002651 "test/avgpool-microkernel-tester.h",
2652 "src/xnnpack/AlignedAllocator.h",
2653 ] + MICROKERNEL_TEST_HDRS,
2654 deps = MICROKERNEL_TEST_DEPS,
2655)
2656
2657xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002658 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002659 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002660 "test/f32-ibilinear.cc",
2661 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002662 "src/xnnpack/AlignedAllocator.h",
2663 ] + MICROKERNEL_TEST_HDRS,
2664 deps = MICROKERNEL_TEST_DEPS,
2665)
2666
2667xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002668 name = "f16_clamp_test",
2669 srcs = [
2670 "test/f16-clamp.cc",
2671 "test/clamp-microkernel-tester.h",
2672 ] + MICROKERNEL_TEST_HDRS,
2673 deps = MICROKERNEL_TEST_DEPS,
2674)
2675
2676xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002677 name = "f32_clamp_test",
2678 srcs = [
2679 "test/f32-clamp.cc",
2680 "test/clamp-microkernel-tester.h",
2681 ] + MICROKERNEL_TEST_HDRS,
2682 deps = MICROKERNEL_TEST_DEPS,
2683)
2684
2685xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002686 name = "f32_igemm_test",
2687 srcs = [
2688 "test/f32-igemm.cc",
2689 "test/gemm-microkernel-tester.h",
2690 "src/xnnpack/AlignedAllocator.h",
2691 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2692 deps = MICROKERNEL_TEST_DEPS,
2693)
2694
2695xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002696 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002697 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002698 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002699 "test/gemm-microkernel-tester.h",
2700 "src/xnnpack/AlignedAllocator.h",
2701 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002702 # Work-around for "too many sections" error
2703 mingw_copts = ["-Wa,-mbig-obj"],
2704 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002705 deps = MICROKERNEL_TEST_DEPS,
2706)
2707
2708xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002709 name = "f16_igemm_minmax_test",
2710 srcs = [
2711 "test/f16-igemm-minmax.cc",
2712 "test/gemm-microkernel-tester.h",
2713 "src/xnnpack/AlignedAllocator.h",
2714 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2715 deps = MICROKERNEL_TEST_DEPS,
2716)
2717
2718xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002719 name = "f32_conv_hwc_test",
2720 srcs = [
2721 "test/f32-conv-hwc.cc",
2722 "test/conv-hwc-microkernel-tester.h",
2723 "src/xnnpack/AlignedAllocator.h",
2724 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2725 deps = MICROKERNEL_TEST_DEPS,
2726)
2727
2728xnnpack_unit_test(
2729 name = "f32_conv_hwc2spchw_test",
2730 srcs = [
2731 "test/f32-conv-hwc2spchw.cc",
2732 "test/conv-hwc2spchw-microkernel-tester.h",
2733 "src/xnnpack/AlignedAllocator.h",
2734 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2735 deps = MICROKERNEL_TEST_DEPS,
2736)
2737
2738xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002739 name = "f32_dwconv_test",
2740 srcs = [
2741 "test/f32-dwconv.cc",
2742 "test/dwconv-microkernel-tester.h",
2743 "src/xnnpack/AlignedAllocator.h",
2744 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2745 deps = MICROKERNEL_TEST_DEPS,
2746)
2747
2748xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002749 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002750 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002751 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002752 "test/dwconv-microkernel-tester.h",
2753 "src/xnnpack/AlignedAllocator.h",
2754 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2755 deps = MICROKERNEL_TEST_DEPS,
2756)
2757
2758xnnpack_unit_test(
2759 name = "f32_dwconv_spchw_test",
2760 srcs = [
2761 "test/f32-dwconv-spchw.cc",
2762 "test/dwconv-spchw-microkernel-tester.h",
2763 "src/xnnpack/AlignedAllocator.h",
2764 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2765 deps = MICROKERNEL_TEST_DEPS,
2766)
2767
2768xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002769 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002770 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002771 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002772 "test/gavgpool-microkernel-tester.h",
2773 "src/xnnpack/AlignedAllocator.h",
2774 ] + MICROKERNEL_TEST_HDRS,
2775 deps = MICROKERNEL_TEST_DEPS,
2776)
2777
2778xnnpack_unit_test(
2779 name = "f32_gavgpool_spchw_test",
2780 srcs = [
2781 "test/f32-gavgpool-spchw.cc",
2782 "test/gavgpool-spchw-microkernel-tester.h",
2783 "src/xnnpack/AlignedAllocator.h",
2784 ] + MICROKERNEL_TEST_HDRS,
2785 deps = MICROKERNEL_TEST_DEPS,
2786)
2787
2788xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002789 name = "f32_gemm_test",
2790 srcs = [
2791 "test/f32-gemm.cc",
2792 "test/gemm-microkernel-tester.h",
2793 "src/xnnpack/AlignedAllocator.h",
2794 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2795 deps = MICROKERNEL_TEST_DEPS,
2796)
2797
2798xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002799 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002800 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002801 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002802 "test/gemm-microkernel-tester.h",
2803 "src/xnnpack/AlignedAllocator.h",
2804 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2805 deps = MICROKERNEL_TEST_DEPS,
2806)
2807
2808xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002809 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002810 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002811 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002812 "test/gemm-microkernel-tester.h",
2813 "src/xnnpack/AlignedAllocator.h",
2814 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2815 deps = MICROKERNEL_TEST_DEPS,
2816)
2817
2818xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002819 name = "f16_hswish_test",
2820 srcs = [
2821 "test/f16-hswish.cc",
2822 "test/hswish-microkernel-tester.h",
2823 ] + MICROKERNEL_TEST_HDRS,
2824 deps = MICROKERNEL_TEST_DEPS,
2825)
2826
2827xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002828 name = "f32_hswish_test",
2829 srcs = [
2830 "test/f32-hswish.cc",
2831 "test/hswish-microkernel-tester.h",
2832 ] + MICROKERNEL_TEST_HDRS,
2833 deps = MICROKERNEL_TEST_DEPS,
2834)
2835
2836xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002837 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002838 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002839 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840 "test/maxpool-microkernel-tester.h",
2841 ] + MICROKERNEL_TEST_HDRS,
2842 deps = MICROKERNEL_TEST_DEPS,
2843)
2844
2845xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002846 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002847 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002848 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002849 "test/avgpool-microkernel-tester.h",
2850 "src/xnnpack/AlignedAllocator.h",
2851 ] + MICROKERNEL_TEST_HDRS,
2852 deps = MICROKERNEL_TEST_DEPS,
2853)
2854
2855xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002856 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002857 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002858 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002859 "test/gemm-microkernel-tester.h",
2860 "src/xnnpack/AlignedAllocator.h",
2861 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2862 deps = MICROKERNEL_TEST_DEPS,
2863)
2864
2865xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002866 name = "f16_prelu_test",
2867 srcs = [
2868 "test/f16-prelu.cc",
2869 "test/prelu-microkernel-tester.h",
2870 "src/xnnpack/AlignedAllocator.h",
2871 ] + MICROKERNEL_TEST_HDRS,
2872 deps = MICROKERNEL_TEST_DEPS,
2873)
2874
2875xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002876 name = "f32_prelu_test",
2877 srcs = [
2878 "test/f32-prelu.cc",
2879 "test/prelu-microkernel-tester.h",
2880 "src/xnnpack/AlignedAllocator.h",
2881 ] + MICROKERNEL_TEST_HDRS,
2882 deps = MICROKERNEL_TEST_DEPS,
2883)
2884
2885xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002886 name = "f32_raddexpminusmax_test",
2887 srcs = [
2888 "test/f32-raddexpminusmax.cc",
2889 "test/raddexpminusmax-microkernel-tester.h",
2890 ] + MICROKERNEL_TEST_HDRS,
2891 deps = MICROKERNEL_TEST_DEPS,
2892)
2893
2894xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002895 name = "f32_raddextexp_test",
2896 srcs = [
2897 "test/f32-raddextexp.cc",
2898 "test/raddextexp-microkernel-tester.h",
2899 ] + MICROKERNEL_TEST_HDRS,
2900 deps = MICROKERNEL_TEST_DEPS,
2901)
2902
2903xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002904 name = "f32_raddstoreexpminusmax_test",
2905 srcs = [
2906 "test/f32-raddstoreexpminusmax.cc",
2907 "test/raddstoreexpminusmax-microkernel-tester.h",
2908 ] + MICROKERNEL_TEST_HDRS,
2909 deps = MICROKERNEL_TEST_DEPS,
2910)
2911
2912xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002913 name = "f32_rmax_test",
2914 srcs = [
2915 "test/f32-rmax.cc",
2916 "test/rmax-microkernel-tester.h",
2917 ] + MICROKERNEL_TEST_HDRS,
2918 deps = MICROKERNEL_TEST_DEPS,
2919)
2920
2921xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002922 name = "f32_sigmoid_test",
2923 srcs = [
2924 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002925 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002926 ] + MICROKERNEL_TEST_HDRS,
2927 deps = MICROKERNEL_TEST_DEPS,
2928)
2929
2930xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002931 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002932 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002933 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 "test/spmm-microkernel-tester.h",
2935 "src/xnnpack/AlignedAllocator.h",
2936 ] + MICROKERNEL_TEST_HDRS,
2937 deps = MICROKERNEL_TEST_DEPS,
2938)
2939
2940xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07002941 name = "f16_vadd_minmax_test",
2942 srcs = [
2943 "test/f16-vadd-minmax.cc",
2944 "test/vbinary-microkernel-tester.h",
2945 ] + MICROKERNEL_TEST_HDRS,
2946 deps = MICROKERNEL_TEST_DEPS,
2947)
2948
2949xnnpack_unit_test(
2950 name = "f16_vaddc_minmax_test",
2951 srcs = [
2952 "test/f16-vaddc-minmax.cc",
2953 "test/vbinaryc-microkernel-tester.h",
2954 ] + MICROKERNEL_TEST_HDRS,
2955 deps = MICROKERNEL_TEST_DEPS,
2956)
2957
2958xnnpack_unit_test(
2959 name = "f16_vdiv_minmax_test",
2960 srcs = [
2961 "test/f16-vdiv-minmax.cc",
2962 "test/vbinary-microkernel-tester.h",
2963 ] + MICROKERNEL_TEST_HDRS,
2964 deps = MICROKERNEL_TEST_DEPS,
2965)
2966
2967xnnpack_unit_test(
2968 name = "f16_vdivc_minmax_test",
2969 srcs = [
2970 "test/f16-vdivc-minmax.cc",
2971 "test/vbinaryc-microkernel-tester.h",
2972 ] + MICROKERNEL_TEST_HDRS,
2973 deps = MICROKERNEL_TEST_DEPS,
2974)
2975
2976xnnpack_unit_test(
2977 name = "f16_vrdivc_minmax_test",
2978 srcs = [
2979 "test/f16-vrdivc-minmax.cc",
2980 "test/vbinaryc-microkernel-tester.h",
2981 ] + MICROKERNEL_TEST_HDRS,
2982 deps = MICROKERNEL_TEST_DEPS,
2983)
2984
2985xnnpack_unit_test(
2986 name = "f16_vmax_test",
2987 srcs = [
2988 "test/f16-vmax.cc",
2989 "test/vbinary-microkernel-tester.h",
2990 ] + MICROKERNEL_TEST_HDRS,
2991 deps = MICROKERNEL_TEST_DEPS,
2992)
2993
2994xnnpack_unit_test(
2995 name = "f16_vmaxc_test",
2996 srcs = [
2997 "test/f16-vmaxc.cc",
2998 "test/vbinaryc-microkernel-tester.h",
2999 ] + MICROKERNEL_TEST_HDRS,
3000 deps = MICROKERNEL_TEST_DEPS,
3001)
3002
3003xnnpack_unit_test(
3004 name = "f16_vmin_test",
3005 srcs = [
3006 "test/f16-vmin.cc",
3007 "test/vbinary-microkernel-tester.h",
3008 ] + MICROKERNEL_TEST_HDRS,
3009 deps = MICROKERNEL_TEST_DEPS,
3010)
3011
3012xnnpack_unit_test(
3013 name = "f16_vminc_test",
3014 srcs = [
3015 "test/f16-vminc.cc",
3016 "test/vbinaryc-microkernel-tester.h",
3017 ] + MICROKERNEL_TEST_HDRS,
3018 deps = MICROKERNEL_TEST_DEPS,
3019)
3020
3021xnnpack_unit_test(
3022 name = "f16_vmul_minmax_test",
3023 srcs = [
3024 "test/f16-vmul-minmax.cc",
3025 "test/vbinary-microkernel-tester.h",
3026 ] + MICROKERNEL_TEST_HDRS,
3027 deps = MICROKERNEL_TEST_DEPS,
3028)
3029
3030xnnpack_unit_test(
3031 name = "f16_vmulc_minmax_test",
3032 srcs = [
3033 "test/f16-vmulc-minmax.cc",
3034 "test/vbinaryc-microkernel-tester.h",
3035 ] + MICROKERNEL_TEST_HDRS,
3036 deps = MICROKERNEL_TEST_DEPS,
3037)
3038
3039xnnpack_unit_test(
3040 name = "f16_vsub_minmax_test",
3041 srcs = [
3042 "test/f16-vsub-minmax.cc",
3043 "test/vbinary-microkernel-tester.h",
3044 ] + MICROKERNEL_TEST_HDRS,
3045 deps = MICROKERNEL_TEST_DEPS,
3046)
3047
3048xnnpack_unit_test(
3049 name = "f16_vsubc_minmax_test",
3050 srcs = [
3051 "test/f16-vsubc-minmax.cc",
3052 "test/vbinaryc-microkernel-tester.h",
3053 ] + MICROKERNEL_TEST_HDRS,
3054 deps = MICROKERNEL_TEST_DEPS,
3055)
3056
3057xnnpack_unit_test(
3058 name = "f16_vrsubc_minmax_test",
3059 srcs = [
3060 "test/f16-vrsubc-minmax.cc",
3061 "test/vbinaryc-microkernel-tester.h",
3062 ] + MICROKERNEL_TEST_HDRS,
3063 deps = MICROKERNEL_TEST_DEPS,
3064)
3065
3066xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003067 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003068 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003069 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003070 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003071 ] + MICROKERNEL_TEST_HDRS,
3072 deps = MICROKERNEL_TEST_DEPS,
3073)
3074
3075xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003076 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003077 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003078 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003079 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003080 ] + MICROKERNEL_TEST_HDRS,
3081 deps = MICROKERNEL_TEST_DEPS,
3082)
3083
3084xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003085 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003086 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003087 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003088 "test/vbinary-microkernel-tester.h",
3089 ] + MICROKERNEL_TEST_HDRS,
3090 deps = MICROKERNEL_TEST_DEPS,
3091)
3092
3093xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003094 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003095 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003096 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003097 "test/vbinaryc-microkernel-tester.h",
3098 ] + MICROKERNEL_TEST_HDRS,
3099 deps = MICROKERNEL_TEST_DEPS,
3100)
3101
3102xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003103 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003104 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003105 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003106 "test/vbinaryc-microkernel-tester.h",
3107 ] + MICROKERNEL_TEST_HDRS,
3108 deps = MICROKERNEL_TEST_DEPS,
3109)
3110
3111xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003112 name = "f32_vmax_test",
3113 srcs = [
3114 "test/f32-vmax.cc",
3115 "test/vbinary-microkernel-tester.h",
3116 ] + MICROKERNEL_TEST_HDRS,
3117 deps = MICROKERNEL_TEST_DEPS,
3118)
3119
3120xnnpack_unit_test(
3121 name = "f32_vmaxc_test",
3122 srcs = [
3123 "test/f32-vmaxc.cc",
3124 "test/vbinaryc-microkernel-tester.h",
3125 ] + MICROKERNEL_TEST_HDRS,
3126 deps = MICROKERNEL_TEST_DEPS,
3127)
3128
3129xnnpack_unit_test(
3130 name = "f32_vmin_test",
3131 srcs = [
3132 "test/f32-vmin.cc",
3133 "test/vbinary-microkernel-tester.h",
3134 ] + MICROKERNEL_TEST_HDRS,
3135 deps = MICROKERNEL_TEST_DEPS,
3136)
3137
3138xnnpack_unit_test(
3139 name = "f32_vminc_test",
3140 srcs = [
3141 "test/f32-vminc.cc",
3142 "test/vbinaryc-microkernel-tester.h",
3143 ] + MICROKERNEL_TEST_HDRS,
3144 deps = MICROKERNEL_TEST_DEPS,
3145)
3146
3147xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003148 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003149 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003150 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003151 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003152 ] + MICROKERNEL_TEST_HDRS,
3153 deps = MICROKERNEL_TEST_DEPS,
3154)
3155
3156xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003157 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003158 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003159 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003160 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161 ] + MICROKERNEL_TEST_HDRS,
3162 deps = MICROKERNEL_TEST_DEPS,
3163)
3164
3165xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003166 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003168 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003169 "test/vmulcaddc-microkernel-tester.h",
3170 "src/xnnpack/AlignedAllocator.h",
3171 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3172 deps = MICROKERNEL_TEST_DEPS,
3173)
3174
3175xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003176 name = "f32_vscale_test",
3177 srcs = [
3178 "test/f32-vscale.cc",
3179 "test/vscale-microkernel-tester.h",
3180 ] + MICROKERNEL_TEST_HDRS,
3181 deps = MICROKERNEL_TEST_DEPS,
3182)
3183
3184xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003185 name = "f32_vscaleexpminusmax_test",
3186 srcs = [
3187 "test/f32-vscaleexpminusmax.cc",
3188 "test/vscaleexpminusmax-microkernel-tester.h",
3189 ] + MICROKERNEL_TEST_HDRS,
3190 deps = MICROKERNEL_TEST_DEPS,
3191)
3192
3193xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003194 name = "f32_vscaleextexp_test",
3195 srcs = [
3196 "test/f32-vscaleextexp.cc",
3197 "test/vscaleextexp-microkernel-tester.h",
3198 ] + MICROKERNEL_TEST_HDRS,
3199 deps = MICROKERNEL_TEST_DEPS,
3200)
3201
3202xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003203 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003204 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003205 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003206 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003207 ] + MICROKERNEL_TEST_HDRS,
3208 deps = MICROKERNEL_TEST_DEPS,
3209)
3210
3211xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003212 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003213 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003214 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003215 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003216 ] + MICROKERNEL_TEST_HDRS,
3217 deps = MICROKERNEL_TEST_DEPS,
3218)
3219
3220xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003221 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003222 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003223 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003224 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003225 ] + MICROKERNEL_TEST_HDRS,
3226 deps = MICROKERNEL_TEST_DEPS,
3227)
3228
3229xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003230 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003231 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003232 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003233 "test/avgpool-microkernel-tester.h",
3234 "src/xnnpack/AlignedAllocator.h",
3235 ] + MICROKERNEL_TEST_HDRS,
3236 deps = MICROKERNEL_TEST_DEPS,
3237)
3238
3239xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003240 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003241 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003242 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003243 "test/gemm-microkernel-tester.h",
3244 "src/xnnpack/AlignedAllocator.h",
3245 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3246 deps = MICROKERNEL_TEST_DEPS,
3247)
3248
3249xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003250 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003251 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003252 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 "test/dwconv-microkernel-tester.h",
3254 "src/xnnpack/AlignedAllocator.h",
3255 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3256 deps = MICROKERNEL_TEST_DEPS,
3257)
3258
3259xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003260 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003261 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003262 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003263 "test/gavgpool-microkernel-tester.h",
3264 "src/xnnpack/AlignedAllocator.h",
3265 ] + MICROKERNEL_TEST_HDRS,
3266 deps = MICROKERNEL_TEST_DEPS,
3267)
3268
3269xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003270 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003272 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003273 "test/gemm-microkernel-tester.h",
3274 "src/xnnpack/AlignedAllocator.h",
3275 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3276 deps = MICROKERNEL_TEST_DEPS,
3277)
3278
3279xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003280 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003282 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003283 "test/vadd-microkernel-tester.h",
3284 ] + MICROKERNEL_TEST_HDRS,
3285 deps = MICROKERNEL_TEST_DEPS,
3286)
3287
3288xnnpack_unit_test(
3289 name = "u8_clamp_test",
3290 srcs = [
3291 "test/u8-clamp.cc",
3292 "test/clamp-microkernel-tester.h",
3293 ] + MICROKERNEL_TEST_HDRS,
3294 deps = MICROKERNEL_TEST_DEPS,
3295)
3296
3297xnnpack_unit_test(
3298 name = "u8_lut32norm_test",
3299 srcs = [
3300 "test/u8-lut32norm.cc",
3301 "test/lut-norm-microkernel-tester.h",
3302 ] + MICROKERNEL_TEST_HDRS,
3303 deps = MICROKERNEL_TEST_DEPS,
3304)
3305
3306xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003307 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003308 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003309 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003310 "test/maxpool-microkernel-tester.h",
3311 ] + MICROKERNEL_TEST_HDRS,
3312 deps = MICROKERNEL_TEST_DEPS,
3313)
3314
3315xnnpack_unit_test(
3316 name = "u8_rmax_test",
3317 srcs = [
3318 "test/u8-rmax.cc",
3319 "test/rmax-microkernel-tester.h",
3320 ] + MICROKERNEL_TEST_HDRS,
3321 deps = MICROKERNEL_TEST_DEPS,
3322)
3323
3324xnnpack_unit_test(
3325 name = "x32_packx_test",
3326 srcs = [
3327 "test/x32-packx.cc",
3328 "test/pack-microkernel-tester.h",
3329 "src/xnnpack/AlignedAllocator.h",
3330 ] + MICROKERNEL_TEST_HDRS,
3331 deps = MICROKERNEL_TEST_DEPS,
3332)
3333
3334xnnpack_unit_test(
3335 name = "x32_pad_test",
3336 srcs = [
3337 "test/x32-pad.cc",
3338 "test/pad-microkernel-tester.h",
3339 ] + MICROKERNEL_TEST_HDRS,
3340 deps = MICROKERNEL_TEST_DEPS,
3341)
3342
3343xnnpack_unit_test(
3344 name = "x32_unpool_test",
3345 srcs = [
3346 "test/x32-unpool.cc",
3347 "test/unpool-microkernel-tester.h",
3348 ] + MICROKERNEL_TEST_HDRS,
3349 deps = MICROKERNEL_TEST_DEPS,
3350)
3351
3352xnnpack_unit_test(
3353 name = "x32_zip_test",
3354 srcs = [
3355 "test/x32-zip.cc",
3356 "test/zip-microkernel-tester.h",
3357 ] + MICROKERNEL_TEST_HDRS,
3358 deps = MICROKERNEL_TEST_DEPS,
3359)
3360
3361xnnpack_unit_test(
3362 name = "x8_lut_test",
3363 srcs = [
3364 "test/x8-lut.cc",
3365 "test/lut-microkernel-tester.h",
3366 ] + MICROKERNEL_TEST_HDRS,
3367 deps = MICROKERNEL_TEST_DEPS,
3368)
3369
3370xnnpack_unit_test(
3371 name = "x8_zip_test",
3372 srcs = [
3373 "test/x8-zip.cc",
3374 "test/zip-microkernel-tester.h",
3375 ] + MICROKERNEL_TEST_HDRS,
3376 deps = MICROKERNEL_TEST_DEPS,
3377)
3378
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003379xnnpack_unit_test(
3380 name = "requantization_test",
3381 srcs = [
3382 "src/xnnpack/requantization-stubs.h",
3383 "test/requantization.cc",
3384 "test/requantization-tester.h",
3385 ] + MICROKERNEL_TEST_HDRS,
3386 deps = MICROKERNEL_TEST_DEPS,
3387)
3388
Marat Dukhan20c3b922020-03-10 03:45:06 -07003389########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003390
3391xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003392 name = "operator_size_test",
3393 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003394 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003395)
3396
Marat Dukhan20c3b922020-03-10 03:45:06 -07003397xnnpack_binary(
3398 name = "subgraph_size_test",
3399 srcs = ["test/subgraph-size.c"],
3400 deps = [":XNNPACK"],
3401)
3402
3403########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003404
3405xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003406 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003407 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003408 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003409 "test/add-operator-tester.h",
3410 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003411 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003412)
3413
3414xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003415 name = "add_nd_test",
3416 srcs = [
3417 "test/add-nd.cc",
3418 "test/binary-elementwise-operator-tester.h",
3419 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003420 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003421)
3422
3423xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003424 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003425 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003426 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427 "test/argmax-pooling-operator-tester.h",
3428 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003429 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430)
3431
3432xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003433 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003434 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003435 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436 "test/average-pooling-operator-tester.h",
3437 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003438 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003439)
3440
3441xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003442 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003443 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003444 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003445 "test/channel-pad-operator-tester.h",
3446 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003447 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448)
3449
3450xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003451 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003452 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003453 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003454 "test/channel-shuffle-operator-tester.h",
3455 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003456 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003457)
3458
3459xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003460 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003461 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003462 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003463 "test/clamp-operator-tester.h",
3464 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003465 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003466)
3467
3468xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003469 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003470 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003471 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003472 "test/convolution-operator-tester.h",
3473 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003474 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003475)
3476
3477xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003478 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003480 "test/convolution-nchw.cc",
3481 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003482 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003483 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484)
3485
3486xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003487 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003488 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003489 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490 "test/deconvolution-operator-tester.h",
3491 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003492 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493)
3494
3495xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003496 name = "divide_nd_test",
3497 srcs = [
3498 "test/binary-elementwise-operator-tester.h",
3499 "test/divide-nd.cc",
3500 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003501 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003502)
3503
3504xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003505 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003506 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003507 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 "test/fully-connected-operator-tester.h",
3509 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003510 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511)
3512
3513xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003514 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003516 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517 "test/global-average-pooling-operator-tester.h",
3518 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003519 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003520)
3521
3522xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003523 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003524 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003525 "test/global-average-pooling-ncw.cc",
3526 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003527 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003528 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003529)
3530
3531xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003532 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003533 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003534 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003535 "test/hardswish-operator-tester.h",
3536 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003537 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003538)
3539
3540xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003541 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003542 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003543 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003544 "test/leaky-relu-operator-tester.h",
3545 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003546 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003547)
3548
3549xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003550 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003551 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003552 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553 "test/max-pooling-operator-tester.h",
3554 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003555 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003556)
3557
3558xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003559 name = "maximum_nd_test",
3560 srcs = [
3561 "test/binary-elementwise-operator-tester.h",
3562 "test/maximum-nd.cc",
3563 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003564 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003565)
3566
3567xnnpack_unit_test(
3568 name = "minimum_nd_test",
3569 srcs = [
3570 "test/binary-elementwise-operator-tester.h",
3571 "test/minimum-nd.cc",
3572 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003573 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003574)
3575
3576xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003577 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003578 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003579 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003580 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003581 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003582 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003583)
3584
3585xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003586 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003587 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003588 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589 "test/prelu-operator-tester.h",
3590 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003591 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003592)
3593
3594xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003595 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003596 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003597 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003598 "test/resize-bilinear-operator-tester.h",
3599 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003600 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003601)
3602
3603xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003604 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003605 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003606 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607 "test/sigmoid-operator-tester.h",
3608 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003609 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610)
3611
3612xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003613 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003614 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003615 "test/softmax-nc.cc",
3616 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003618 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003619)
3620
3621xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003622 name = "subtract_nd_test",
3623 srcs = [
3624 "test/binary-elementwise-operator-tester.h",
3625 "test/subtract-nd.cc",
3626 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003627 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003628)
3629
3630xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003631 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003633 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 "test/unpooling-operator-tester.h",
3635 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003636 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637)
3638
3639############################# Build configurations #############################
3640
Marat Dukhanb8642352019-10-30 15:43:02 -07003641# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003643 name = "xnn_enable_assembly_explicit_true",
3644 define_values = {"xnn_enable_assembly": "true"},
3645)
3646
3647# Disables usage of assembly kernels.
3648config_setting(
3649 name = "xnn_enable_assembly_explicit_false",
3650 define_values = {"xnn_enable_assembly": "false"},
3651)
3652
Marat Dukhan05702cf2020-03-26 15:41:33 -07003653# Disables usage of HMP-aware optimizations.
3654config_setting(
3655 name = "xnn_enable_hmp_explicit_false",
3656 define_values = {"xnn_enable_hmp": "false"},
3657)
3658
Marat Dukhanb8642352019-10-30 15:43:02 -07003659# Builds with -c dbg
3660config_setting(
3661 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003663 "compilation_mode": "dbg",
3664 },
3665)
3666
3667# Builds with -c opt
3668config_setting(
3669 name = "optimized_build",
3670 values = {
3671 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003672 },
3673)
3674
3675config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003676 name = "linux_k8",
3677 values = {"cpu": "k8"},
3678)
3679
3680config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003681 name = "linux_arm",
3682 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003683)
3684
3685config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003686 name = "linux_armhf",
3687 values = {"cpu": "armhf"},
3688)
3689
3690config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003691 name = "linux_aarch64",
3692 values = {"cpu": "aarch64"},
3693)
3694
3695config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 name = "android",
3697 values = {"crosstool_top": "//external:android/crosstool"},
3698)
3699
3700config_setting(
3701 name = "android_armv7",
3702 values = {
3703 "crosstool_top": "//external:android/crosstool",
3704 "cpu": "armeabi-v7a",
3705 },
3706)
3707
3708config_setting(
3709 name = "android_arm64",
3710 values = {
3711 "crosstool_top": "//external:android/crosstool",
3712 "cpu": "arm64-v8a",
3713 },
3714)
3715
3716config_setting(
3717 name = "android_x86",
3718 values = {
3719 "crosstool_top": "//external:android/crosstool",
3720 "cpu": "x86",
3721 },
3722)
3723
3724config_setting(
3725 name = "android_x86_64",
3726 values = {
3727 "crosstool_top": "//external:android/crosstool",
3728 "cpu": "x86_64",
3729 },
3730)
3731
3732config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003733 name = "windows_x86_64",
3734 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003735)
3736
3737config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003738 name = "windows_x86_64_clang",
3739 values = {
3740 "compiler": "clang-cl",
3741 "cpu": "x64_windows",
3742 },
3743)
3744
3745config_setting(
3746 name = "windows_x86_64_mingw",
3747 values = {
3748 "compiler": "mingw-gcc",
3749 "cpu": "x64_windows",
3750 },
3751)
3752
3753config_setting(
3754 name = "windows_x86_64_msys",
3755 values = {
3756 "compiler": "msys-gcc",
3757 "cpu": "x64_windows",
3758 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003759)
3760
3761config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003762 name = "macos_x86_64",
3763 values = {
3764 "apple_platform_type": "macos",
3765 "cpu": "darwin",
3766 },
3767)
3768
3769config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003770 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003771 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772)
3773
3774config_setting(
3775 name = "emscripten_wasm",
3776 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003777 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003778 "cpu": "wasm",
3779 },
3780)
3781
3782config_setting(
3783 name = "emscripten_wasmsimd",
3784 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003785 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003786 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003787 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003788 },
3789)
3790
3791config_setting(
3792 name = "emscripten_asmjs",
3793 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003794 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003795 "cpu": "asmjs",
3796 },
3797)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003798
3799config_setting(
3800 name = "ios_armv7",
3801 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003802 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003803 "cpu": "ios_armv7",
3804 },
3805)
3806
3807config_setting(
3808 name = "ios_arm64",
3809 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003810 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003811 "cpu": "ios_arm64",
3812 },
3813)
3814
3815config_setting(
3816 name = "ios_arm64e",
3817 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003818 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003819 "cpu": "ios_arm64e",
3820 },
3821)
3822
3823config_setting(
3824 name = "ios_x86",
3825 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003826 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003827 "cpu": "ios_i386",
3828 },
3829)
3830
3831config_setting(
3832 name = "ios_x86_64",
3833 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003834 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003835 "cpu": "ios_x86_64",
3836 },
3837)
3838
3839config_setting(
3840 name = "watchos_armv7k",
3841 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003842 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003843 "cpu": "watchos_armv7k",
3844 },
3845)
3846
3847config_setting(
3848 name = "watchos_arm64_32",
3849 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003850 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003851 "cpu": "watchos_arm64_32",
3852 },
3853)
3854
3855config_setting(
3856 name = "watchos_x86",
3857 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003858 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003859 "cpu": "watchos_i386",
3860 },
3861)
3862
3863config_setting(
3864 name = "watchos_x86_64",
3865 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003866 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003867 "cpu": "watchos_x86_64",
3868 },
3869)
3870
3871config_setting(
3872 name = "tvos_arm64",
3873 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003874 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003875 "cpu": "tvos_arm64",
3876 },
3877)
3878
3879config_setting(
3880 name = "tvos_x86_64",
3881 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003882 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003883 "cpu": "tvos_x86_64",
3884 },
3885)