blob: de80a64e2c53280864521e4e5219cc7a019a80db [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 Dukhan8d3c6932020-03-06 20:27:27 -08009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_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_copts", "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 Dukhan08c4a432019-10-03 09:29:21 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
22MICROKERNEL_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
46OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
52OPERATOR_SRCS = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
Marat Dukhanfd8e6892020-01-27 15:25:25 -080072 "src/softmax-nc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080073 "src/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 Dukhan6ee435a2020-02-26 22:33:38 -080085 "src/f32-avgpool/9p8x-scalar-c1.c",
86 "src/f32-avgpool/9x-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 Dukhan40a672f2019-11-25 03:08:22 -080097 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x25-scalar.c",
99 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
100 "src/f32-dwconv/gen/up1x4-scalar.c",
101 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
102 "src/f32-dwconv/gen/up1x9-scalar.c",
103 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x25-scalar.c",
105 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
106 "src/f32-dwconv/gen/up2x4-scalar.c",
107 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
108 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700109 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700110 "src/f32-gavgpool/7p7x-scalar-c1.c",
111 "src/f32-gavgpool/7x-scalar-c1.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800112 "src/f32-gemm/gen-inc/1x4-scalar.c",
113 "src/f32-gemm/gen-inc/2x4-scalar.c",
114 "src/f32-gemm/gen-inc/4x4-scalar.c",
115 "src/f32-gemm/gen/1x4-scalar.c",
116 "src/f32-gemm/gen/2x4-scalar.c",
117 "src/f32-gemm/gen/4x2-scalar.c",
118 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800119 "src/f32-hswish/gen/scalar-x1.c",
120 "src/f32-hswish/gen/scalar-x2.c",
121 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700122 "src/f32-ibilinear/gen/scalar-c1.c",
123 "src/f32-ibilinear/gen/scalar-c2.c",
124 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800125 "src/f32-igemm/gen/1x4-scalar.c",
126 "src/f32-igemm/gen/2x4-scalar.c",
127 "src/f32-igemm/gen/4x2-scalar.c",
128 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800129 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800130 "src/f32-pavgpool/9p8x-scalar-c1.c",
131 "src/f32-pavgpool/9x-scalar-c1.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800132 "src/f32-ppmm/gen/2x4-scalar.c",
133 "src/f32-ppmm/gen/3x3-scalar.c",
134 "src/f32-ppmm/gen/4x2-scalar.c",
135 "src/f32-ppmm/gen/4x4-scalar.c",
136 "src/f32-prelu/gen/scalar-2x1.c",
137 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800138 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
139 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
140 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
141 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
142 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
143 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
144 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
145 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
146 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
147 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
148 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
149 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700150 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800151 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
152 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
153 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
154 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
155 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
156 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
157 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
158 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
159 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800160 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
161 "src/f32-spmm/gen/1x1-scalar.c",
162 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
163 "src/f32-spmm/gen/2x1-scalar.c",
164 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
165 "src/f32-spmm/gen/4x1-scalar.c",
166 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
167 "src/f32-spmm/gen/8x1-scalar.c",
168 "src/f32-spmm/gen/8x2-scalar.c",
169 "src/f32-spmm/gen/8x4-scalar.c",
170 "src/f32-vbinary/gen/vadd-scalar-x1.c",
171 "src/f32-vbinary/gen/vadd-scalar-x2.c",
172 "src/f32-vbinary/gen/vadd-scalar-x4.c",
173 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
174 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
175 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800176 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
177 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
178 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
179 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
180 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
181 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800182 "src/f32-vbinary/gen/vmax-scalar-x1.c",
183 "src/f32-vbinary/gen/vmax-scalar-x2.c",
184 "src/f32-vbinary/gen/vmax-scalar-x4.c",
185 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
186 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
187 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
188 "src/f32-vbinary/gen/vmin-scalar-x1.c",
189 "src/f32-vbinary/gen/vmin-scalar-x2.c",
190 "src/f32-vbinary/gen/vmin-scalar-x4.c",
191 "src/f32-vbinary/gen/vminc-scalar-x1.c",
192 "src/f32-vbinary/gen/vminc-scalar-x2.c",
193 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800194 "src/f32-vbinary/gen/vmul-scalar-x1.c",
195 "src/f32-vbinary/gen/vmul-scalar-x2.c",
196 "src/f32-vbinary/gen/vmul-scalar-x4.c",
197 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
198 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
199 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800200 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
201 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
202 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800203 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
204 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
205 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
206 "src/f32-vbinary/gen/vsub-scalar-x1.c",
207 "src/f32-vbinary/gen/vsub-scalar-x2.c",
208 "src/f32-vbinary/gen/vsub-scalar-x4.c",
209 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
210 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
211 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
212 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
213 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
214 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800215 "src/math/expminus-scalar-lut2048-p1.c",
216 "src/math/expminus-scalar-lut64-p2.c",
217 "src/math/expminus-scalar-p5.c",
218 "src/math/sigmoid-scalar-lut2048-p1-div.c",
219 "src/math/sigmoid-scalar-lut64-p2-div.c",
220 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800221 "src/q8-avgpool/9p8x-scalar-c1.c",
222 "src/q8-avgpool/9x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700223 "src/q8-dwconv/up1x9-scalar.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700224 "src/q8-gavgpool/7p7x-scalar-c1.c",
225 "src/q8-gavgpool/7x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700226 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800227 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700228 "src/q8-vadd/scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700229 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700230 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800231 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700232 "src/u8-rmax/scalar.c",
233 "src/x32-packx/x2-scalar.c",
234 "src/x32-packx/x3-scalar.c",
235 "src/x32-packx/x4-scalar.c",
236 "src/x32-pad/x2-scalar.c",
237 "src/x32-unpool/scalar.c",
238 "src/x32-zip/x2-scalar.c",
239 "src/x32-zip/x3-scalar.c",
240 "src/x32-zip/x4-scalar.c",
241 "src/x32-zip/xm-scalar.c",
242 "src/x8-lut/scalar.c",
243 "src/x8-zip/x2-scalar.c",
244 "src/x8-zip/x3-scalar.c",
245 "src/x8-zip/x4-scalar.c",
246 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700247 "src/requantization/precise-scalar.c",
248 "src/requantization/fp32-scalar.c",
249 "src/requantization/q31-scalar.c",
250 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700251]
252
Marat Dukhan436ebe62019-12-04 15:10:12 -0800253WASM_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800254 "src/f32-avgpool/9p8x-wasm-c1.c",
255 "src/f32-avgpool/9x-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700256 "src/f32-clamp/gen/wasm-x1.c",
257 "src/f32-clamp/gen/wasm-x2.c",
258 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800259 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
260 "src/f32-dwconv/gen/up1x25-wasm.c",
261 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
262 "src/f32-dwconv/gen/up1x4-wasm.c",
263 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
264 "src/f32-dwconv/gen/up1x9-wasm.c",
265 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
266 "src/f32-dwconv/gen/up2x25-wasm.c",
267 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
268 "src/f32-dwconv/gen/up2x4-wasm.c",
269 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
270 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700271 "src/f32-gavgpool/7p7x-wasm-c1.c",
272 "src/f32-gavgpool/7x-wasm-c1.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800273 "src/f32-gemm/gen-inc/1x4-wasm.c",
274 "src/f32-gemm/gen-inc/2x4-wasm.c",
275 "src/f32-gemm/gen-inc/4x4-wasm.c",
276 "src/f32-gemm/gen/1x4-wasm.c",
277 "src/f32-gemm/gen/2x4-wasm.c",
278 "src/f32-gemm/gen/4x2-wasm.c",
279 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800280 "src/f32-hswish/gen/wasm-x1.c",
281 "src/f32-hswish/gen/wasm-x2.c",
282 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800283 "src/f32-igemm/gen/1x4-wasm.c",
284 "src/f32-igemm/gen/2x4-wasm.c",
285 "src/f32-igemm/gen/4x2-wasm.c",
286 "src/f32-igemm/gen/4x4-wasm.c",
287 "src/f32-maxpool/9p8x-wasm-c1.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800288 "src/f32-pavgpool/9p8x-wasm-c1.c",
289 "src/f32-pavgpool/9x-wasm-c1.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800290 "src/f32-vbinary/gen/vadd-wasm-x1.c",
291 "src/f32-vbinary/gen/vadd-wasm-x2.c",
292 "src/f32-vbinary/gen/vadd-wasm-x4.c",
293 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
294 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
295 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800296 "src/f32-vbinary/gen/vdiv-wasm-x1.c",
297 "src/f32-vbinary/gen/vdiv-wasm-x2.c",
298 "src/f32-vbinary/gen/vdiv-wasm-x4.c",
299 "src/f32-vbinary/gen/vdivc-wasm-x1.c",
300 "src/f32-vbinary/gen/vdivc-wasm-x2.c",
301 "src/f32-vbinary/gen/vdivc-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800302 "src/f32-vbinary/gen/vmax-wasm-x1.c",
303 "src/f32-vbinary/gen/vmax-wasm-x2.c",
304 "src/f32-vbinary/gen/vmax-wasm-x4.c",
305 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
306 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
307 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
308 "src/f32-vbinary/gen/vmin-wasm-x1.c",
309 "src/f32-vbinary/gen/vmin-wasm-x2.c",
310 "src/f32-vbinary/gen/vmin-wasm-x4.c",
311 "src/f32-vbinary/gen/vminc-wasm-x1.c",
312 "src/f32-vbinary/gen/vminc-wasm-x2.c",
313 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800314 "src/f32-vbinary/gen/vmul-wasm-x1.c",
315 "src/f32-vbinary/gen/vmul-wasm-x2.c",
316 "src/f32-vbinary/gen/vmul-wasm-x4.c",
317 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
318 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
319 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800320 "src/f32-vbinary/gen/vrdivc-wasm-x1.c",
321 "src/f32-vbinary/gen/vrdivc-wasm-x2.c",
322 "src/f32-vbinary/gen/vrdivc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800323 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
324 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
325 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
326 "src/f32-vbinary/gen/vsub-wasm-x1.c",
327 "src/f32-vbinary/gen/vsub-wasm-x2.c",
328 "src/f32-vbinary/gen/vsub-wasm-x4.c",
329 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
330 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
331 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
332 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
333 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
334 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
335]
336
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800337PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800338 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800339 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800340 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800341 "src/f32-avgpool/9p8x-psimd-c4.c",
342 "src/f32-avgpool/9x-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700343 "src/f32-clamp/gen/psimd-x4.c",
344 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800345 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
346 "src/f32-dwconv/gen/up4x25-psimd.c",
347 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
348 "src/f32-dwconv/gen/up4x4-psimd.c",
349 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
350 "src/f32-dwconv/gen/up4x9-psimd.c",
351 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
352 "src/f32-dwconv/gen/up8x25-psimd.c",
353 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
354 "src/f32-dwconv/gen/up8x4-psimd.c",
355 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
356 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700357 "src/f32-gavgpool/7p7x-psimd-c4.c",
358 "src/f32-gavgpool/7x-psimd-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800359 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
360 "src/f32-gemm/gen/1x8-psimd-splat.c",
361 "src/f32-gemm/gen/1x8s4-psimd.c",
Marat Dukhanb00004d2020-02-13 09:28:35 -0800362 "src/f32-gemm/gen/4x2c4-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800363 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
364 "src/f32-gemm/gen/4x8-psimd-splat.c",
365 "src/f32-gemm/gen/4x8s4-psimd.c",
366 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
367 "src/f32-gemm/gen/6x8-psimd-splat.c",
368 "src/f32-gemm/gen/6x8s4-psimd.c",
369 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
370 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
371 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
372 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
373 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
374 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
375 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
376 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
377 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800378 "src/f32-hswish/gen/psimd-x4.c",
379 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700380 "src/f32-ibilinear/gen/psimd-c4.c",
381 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800382 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
383 "src/f32-igemm/gen/1x8-psimd-splat.c",
384 "src/f32-igemm/gen/1x8s4-psimd.c",
385 "src/f32-igemm/gen/4x2c4-psimd.c",
386 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
387 "src/f32-igemm/gen/4x8-psimd-splat.c",
388 "src/f32-igemm/gen/4x8s4-psimd.c",
389 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
390 "src/f32-igemm/gen/6x8-psimd-splat.c",
391 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800392 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800393 "src/f32-pavgpool/9p8x-psimd-c4.c",
394 "src/f32-pavgpool/9x-psimd-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800395 "src/f32-ppmm/gen/4x8-psimd.c",
396 "src/f32-prelu/gen/psimd-2x4.c",
397 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800398 "src/f32-rmax/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800399 "src/f32-vbinary/gen/vadd-psimd-x4.c",
400 "src/f32-vbinary/gen/vadd-psimd-x8.c",
401 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
402 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800403 "src/f32-vbinary/gen/vdiv-psimd-x4.c",
404 "src/f32-vbinary/gen/vdiv-psimd-x8.c",
405 "src/f32-vbinary/gen/vdivc-psimd-x4.c",
406 "src/f32-vbinary/gen/vdivc-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800407 "src/f32-vbinary/gen/vmax-psimd-x4.c",
408 "src/f32-vbinary/gen/vmax-psimd-x8.c",
409 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
410 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
411 "src/f32-vbinary/gen/vmin-psimd-x4.c",
412 "src/f32-vbinary/gen/vmin-psimd-x8.c",
413 "src/f32-vbinary/gen/vminc-psimd-x4.c",
414 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800415 "src/f32-vbinary/gen/vmul-psimd-x4.c",
416 "src/f32-vbinary/gen/vmul-psimd-x8.c",
417 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
418 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800419 "src/f32-vbinary/gen/vrdivc-psimd-x4.c",
420 "src/f32-vbinary/gen/vrdivc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800421 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
422 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
423 "src/f32-vbinary/gen/vsub-psimd-x4.c",
424 "src/f32-vbinary/gen/vsub-psimd-x8.c",
425 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
426 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
427 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
428 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700429 "src/x32-packx/x4-psimd.c",
430 "src/x32-pad/x2-psimd.c",
431 "src/x32-unpool/psimd.c",
432 "src/x32-zip/x2-psimd.c",
433 "src/x32-zip/x3-psimd.c",
434 "src/x32-zip/x4-psimd.c",
435 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700436 "src/requantization/precise-psimd.c",
437 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700438]
439
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800440PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800441 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
442 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
443 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
444 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
445 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
446 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
447 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
448 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
449 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
450 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
451 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
452 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800453 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
454 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
455 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
456 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
457 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
458 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
459 "src/math/sigmoid-psimd-p5-div.c",
460]
461
Marat Dukhan08c4a432019-10-03 09:29:21 -0700462# ISA-specific micro-kernels
463NEON_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800464 "src/f32-avgpool/9p8x-neon-c4.c",
465 "src/f32-avgpool/9x-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700466 "src/f32-clamp/gen/neon-x4.c",
467 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800468 "src/f32-dwconv/gen/up4x9-neon.c",
469 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
470 "src/f32-dwconv/gen/up8x9-neon.c",
471 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700472 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700473 "src/f32-gavgpool/7p7x-neon-c4.c",
474 "src/f32-gavgpool/7x-neon-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800475 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
476 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
477 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
478 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
479 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
480 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800481 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800482 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
483 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
484 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
485 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800486 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800487 "src/f32-gemm/gen/1x8s4-neon.c",
488 "src/f32-gemm/gen/4x8s4-neon.c",
489 "src/f32-gemm/gen/6x8s4-neon.c",
490 "src/f32-gemm/gen/8x8s4-neon.c",
491 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
492 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
493 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
494 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
495 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800496 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800497 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
498 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
499 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
500 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800501 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800502 "src/f32-gemm/gen-inc/1x8s4-neon.c",
503 "src/f32-gemm/gen-inc/4x8s4-neon.c",
504 "src/f32-gemm/gen-inc/6x8s4-neon.c",
505 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800506 "src/f32-hswish/gen/neon-x4.c",
507 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700508 "src/f32-ibilinear/gen/neon-c4.c",
509 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800510 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
511 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
512 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
513 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
514 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
515 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800516 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800517 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
518 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
519 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
520 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800521 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800522 "src/f32-igemm/gen/1x8s4-neon.c",
523 "src/f32-igemm/gen/4x8s4-neon.c",
524 "src/f32-igemm/gen/6x8s4-neon.c",
525 "src/f32-igemm/gen/8x8s4-neon.c",
Frank Barchardf092a4a2020-03-03 14:22:46 -0800526 "src/f32-maxpool/9p8x-neon-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800527 "src/f32-pavgpool/9p8x-neon-c4.c",
528 "src/f32-pavgpool/9x-neon-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800529 "src/f32-ppmm/gen/4x8-neon.c",
530 "src/f32-ppmm/gen/8x8-neon.c",
531 "src/f32-prelu/gen/neon-2x4.c",
532 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800533 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
534 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
535 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
536 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
537 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
538 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
539 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
540 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
541 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
542 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
543 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
544 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
545 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
546 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
547 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
548 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
549 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
550 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
551 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
552 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
553 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
554 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
555 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700557 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800558 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800559 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
560 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
561 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
562 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
563 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
564 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
565 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
566 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
567 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
568 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
569 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
570 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
571 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
572 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
573 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
574 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
575 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
576 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800577 "src/f32-vbinary/gen/vadd-neon-x4.c",
578 "src/f32-vbinary/gen/vadd-neon-x8.c",
579 "src/f32-vbinary/gen/vaddc-neon-x4.c",
580 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800581 "src/f32-vbinary/gen/vmax-neon-x4.c",
582 "src/f32-vbinary/gen/vmax-neon-x8.c",
583 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
584 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
585 "src/f32-vbinary/gen/vmin-neon-x4.c",
586 "src/f32-vbinary/gen/vmin-neon-x8.c",
587 "src/f32-vbinary/gen/vminc-neon-x4.c",
588 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800589 "src/f32-vbinary/gen/vmul-neon-x4.c",
590 "src/f32-vbinary/gen/vmul-neon-x8.c",
591 "src/f32-vbinary/gen/vmulc-neon-x4.c",
592 "src/f32-vbinary/gen/vmulc-neon-x8.c",
593 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
594 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
595 "src/f32-vbinary/gen/vsub-neon-x4.c",
596 "src/f32-vbinary/gen/vsub-neon-x8.c",
597 "src/f32-vbinary/gen/vsubc-neon-x4.c",
598 "src/f32-vbinary/gen/vsubc-neon-x8.c",
599 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
600 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800601 "src/q8-avgpool/9p8x-neon-c8.c",
602 "src/q8-avgpool/9x-neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700603 "src/q8-dwconv/up8x9-neon.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700604 "src/q8-gavgpool/7p7x-neon-c8.c",
605 "src/q8-gavgpool/7x-neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700606 "src/q8-gemm/4x8-neon.c",
607 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800608 "src/q8-igemm/4x8-neon.c",
609 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700610 "src/q8-vadd/neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700611 "src/u8-clamp/neon-x64.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800612 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700613 "src/u8-rmax/neon.c",
614 "src/x32-packx/x4-neon-st4.c",
615 "src/x32-pad/x2-neon.c",
616 "src/x32-zip/x2-neon.c",
617 "src/x32-zip/x3-neon.c",
618 "src/x32-zip/x4-neon.c",
619 "src/x32-zip/xm-neon.c",
620 "src/x8-zip/x2-neon.c",
621 "src/x8-zip/x3-neon.c",
622 "src/x8-zip/x4-neon.c",
623 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800624 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800625 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
626 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
627 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
628 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
629 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
630 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700631 "src/requantization/precise-neon.c",
632 "src/requantization/fp32-neon.c",
633 "src/requantization/q31-neon.c",
634 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700635]
636
637NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700638 "src/f32-ibilinear/gen/neonfma-c4.c",
639 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800640 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
641 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
642 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
643 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800644 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800645 "src/f32-igemm/gen/1x8s4-neonfma.c",
646 "src/f32-igemm/gen/4x8s4-neonfma.c",
647 "src/f32-igemm/gen/6x8s4-neonfma.c",
648 "src/f32-igemm/gen/8x8s4-neonfma.c",
649 "src/f32-dwconv/gen/up4x9-neonfma.c",
650 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
651 "src/f32-dwconv/gen/up8x9-neonfma.c",
652 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
653 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
654 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
655 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
656 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800657 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800658 "src/f32-gemm/gen/1x8s4-neonfma.c",
659 "src/f32-gemm/gen/4x8s4-neonfma.c",
660 "src/f32-gemm/gen/6x8s4-neonfma.c",
661 "src/f32-gemm/gen/8x8s4-neonfma.c",
662 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
663 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
664 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
665 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800666 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800667 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
668 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
669 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
670 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800671 "src/f32-hswish/gen/neonfma-x4.c",
672 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800673 "src/f32-ppmm/gen/4x8-neonfma.c",
674 "src/f32-ppmm/gen/8x8-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800675 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
676 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
677 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
678 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
679 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
680 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
681 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
682 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
683 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
684 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
685 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
686 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
687 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
688 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
689 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
690 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
691 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
692 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
693 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
694 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
695 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
696 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
697 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
698 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800699 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
700 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
701 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
702 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
703 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
704 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
705 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
706 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
707 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
708 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
709 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
710 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
711 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
712 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
713 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
714 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
715 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
716 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
717 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
718 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
719 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
720 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
721 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
722 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
723 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
724 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
725 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
726 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
727 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
728 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
729 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
730 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
731 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
732 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
733 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
734 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
735 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
736 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
737 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
738 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
739 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
740 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
741 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
742 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
743 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
744 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
745 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
746 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
747 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
748 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
749 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
750 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
751 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
752 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800753 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
754 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800755 "src/math/exp-neonfma-lut64-p2.c",
756 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800757 "src/math/expminus-neonfma-lut2048-p1.c",
758 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800759 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800760 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
761 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
762 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
763 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
764 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
765 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
766 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
767 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
768 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
769 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
770 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
771 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
772 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
773 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
774 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
775 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
776 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
777 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700778]
779
780AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800781 "src/f32-vbinary/gen/vdiv-neon-x4.c",
782 "src/f32-vbinary/gen/vdiv-neon-x8.c",
783 "src/f32-vbinary/gen/vdivc-neon-x4.c",
784 "src/f32-vbinary/gen/vdivc-neon-x8.c",
785 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
786 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800787 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
788 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
789 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
790 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
791 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
792 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800793 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800794 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
795 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
796 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
797 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
798 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800799 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800800 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
801 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
802 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
803 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
804 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
805 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800806 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700807 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
808 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
809 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
810 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
811 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
812 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
813 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800814 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
816 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
817 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
818 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
819 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
820 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
821 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
822 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
823 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
824 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
825 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
826 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
827 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
828 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
829 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
830 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
831 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800832 "src/f32-spmm/gen/12x1-neonfma.c",
833 "src/f32-spmm/gen/12x2-neonfma.c",
834 "src/f32-spmm/gen/12x4-neonfma.c",
835 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
836 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
837 "src/f32-spmm/gen/16x1-neonfma.c",
838 "src/f32-spmm/gen/16x2-neonfma.c",
839 "src/f32-spmm/gen/16x4-neonfma.c",
840 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
841 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
842 "src/f32-spmm/gen/4x1-neonfma.c",
843 "src/f32-spmm/gen/4x2-neonfma.c",
844 "src/f32-spmm/gen/4x4-neonfma.c",
845 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
846 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
847 "src/f32-spmm/gen/8x1-neonfma.c",
848 "src/f32-spmm/gen/8x2-neonfma.c",
849 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800850 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
851 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
852 "src/math/sigmoid-neonfma-rr1-p5-div.c",
853 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
854 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
855 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700856]
857
858AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800859 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
860 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
861 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhanbdb56f52020-02-05 21:42:49 -0800862 "src/f16-spmm/gen/8x1-neonfp16arith.c",
863 "src/f16-spmm/gen/8x1-neonfp16arith-unroll2.c",
864 "src/f16-spmm/gen/16x1-neonfp16arith.c",
865 "src/f16-spmm/gen/16x1-neonfp16arith-unroll2.c",
866 "src/f16-spmm/gen/24x1-neonfp16arith.c",
867 "src/f16-spmm/gen/24x1-neonfp16arith-unroll2.c",
868 "src/f16-spmm/gen/32x1-neonfp16arith.c",
869 "src/f16-spmm/gen/32x1-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700870]
871
872SSE_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800873 "src/f32-avgpool/9p8x-sse-c4.c",
874 "src/f32-avgpool/9x-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700875 "src/f32-clamp/gen/sse-x4.c",
876 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800877 "src/f32-dwconv-spchw/3x3p1-sse.c",
878 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800879 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
880 "src/f32-dwconv/gen/up4x25-sse.c",
881 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
882 "src/f32-dwconv/gen/up4x4-sse.c",
883 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
884 "src/f32-dwconv/gen/up4x9-sse.c",
885 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
886 "src/f32-dwconv/gen/up8x25-sse.c",
887 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
888 "src/f32-dwconv/gen/up8x4-sse.c",
889 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
890 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700891 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700892 "src/f32-gavgpool/7p7x-sse-c4.c",
893 "src/f32-gavgpool/7x-sse-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800894 "src/f32-gemm/gen/1x8-sse-dup.c",
895 "src/f32-gemm/gen/1x8-sse-load1.c",
896 "src/f32-gemm/gen/1x8s4-sse.c",
Marat Dukhanb00004d2020-02-13 09:28:35 -0800897 "src/f32-gemm/gen/4x2c4-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800898 "src/f32-gemm/gen/4x8-sse-dup.c",
899 "src/f32-gemm/gen/4x8-sse-load1.c",
900 "src/f32-gemm/gen/4x8s4-sse.c",
901 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
902 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
903 "src/f32-gemm/gen-inc/1x8s4-sse.c",
904 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
905 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
906 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800907 "src/f32-hswish/gen/sse-x4.c",
908 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700909 "src/f32-ibilinear/gen/sse-c4.c",
910 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800911 "src/f32-igemm/gen/1x8-sse-dup.c",
912 "src/f32-igemm/gen/1x8-sse-load1.c",
913 "src/f32-igemm/gen/1x8s4-sse.c",
914 "src/f32-igemm/gen/4x2c4-sse.c",
915 "src/f32-igemm/gen/4x8-sse-dup.c",
916 "src/f32-igemm/gen/4x8-sse-load1.c",
917 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800918 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800919 "src/f32-pavgpool/9p8x-sse-c4.c",
920 "src/f32-pavgpool/9x-sse-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800921 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700922 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800923 "src/f32-spmm/gen/4x1-sse.c",
924 "src/f32-spmm/gen/8x1-sse.c",
925 "src/f32-vbinary/gen/vadd-sse-x4.c",
926 "src/f32-vbinary/gen/vadd-sse-x8.c",
927 "src/f32-vbinary/gen/vaddc-sse-x4.c",
928 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800929 "src/f32-vbinary/gen/vdiv-sse-x4.c",
930 "src/f32-vbinary/gen/vdiv-sse-x8.c",
931 "src/f32-vbinary/gen/vdivc-sse-x4.c",
932 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800933 "src/f32-vbinary/gen/vmax-sse-x4.c",
934 "src/f32-vbinary/gen/vmax-sse-x8.c",
935 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
936 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
937 "src/f32-vbinary/gen/vmin-sse-x4.c",
938 "src/f32-vbinary/gen/vmin-sse-x8.c",
939 "src/f32-vbinary/gen/vminc-sse-x4.c",
940 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800941 "src/f32-vbinary/gen/vmul-sse-x4.c",
942 "src/f32-vbinary/gen/vmul-sse-x8.c",
943 "src/f32-vbinary/gen/vmulc-sse-x4.c",
944 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800945 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
946 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800947 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
948 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
949 "src/f32-vbinary/gen/vsub-sse-x4.c",
950 "src/f32-vbinary/gen/vsub-sse-x8.c",
951 "src/f32-vbinary/gen/vsubc-sse-x4.c",
952 "src/f32-vbinary/gen/vsubc-sse-x8.c",
953 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
954 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700955 "src/x32-packx/x4-sse.c",
956]
957
958SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800959 "src/f32-argmaxpool/9p8x-sse2-c4.c",
960 "src/f32-argmaxpool/4x-sse2-c4.c",
961 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800962 "src/f32-prelu/gen/sse2-2x4.c",
963 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800964 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
965 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
966 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
967 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
968 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
969 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
970 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
971 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
972 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
973 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
974 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
975 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800976 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800977 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800978 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800979 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800980 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
981 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800982 "src/q8-avgpool/9p8x-sse2-c8.c",
983 "src/q8-avgpool/9x-sse2-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700984 "src/q8-igemm/4x4c2-sse2.c",
985 "src/q8-dwconv/up8x9-sse2.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700986 "src/q8-gavgpool/7p7x-sse2-c8.c",
987 "src/q8-gavgpool/7x-sse2-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700988 "src/q8-gemm/2x4c8-sse2.c",
989 "src/q8-gemm/4x4c2-sse2.c",
990 "src/q8-vadd/sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700991 "src/u8-clamp/sse2-x64.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800992 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700993 "src/u8-rmax/sse2.c",
994 "src/x32-pad/x2-sse2.c",
995 "src/x32-zip/x2-sse2.c",
996 "src/x32-zip/x3-sse2.c",
997 "src/x32-zip/x4-sse2.c",
998 "src/x32-zip/xm-sse2.c",
999 "src/x8-zip/x2-sse2.c",
1000 "src/x8-zip/x3-sse2.c",
1001 "src/x8-zip/x4-sse2.c",
1002 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001003 "src/math/exp-sse2-p5.c",
1004 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001005 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001006 "src/requantization/precise-sse2.c",
1007 "src/requantization/fp32-sse2.c",
1008 "src/requantization/q31-sse2.c",
1009 "src/requantization/gemmlowp-sse2.c",
1010]
1011
1012SSSE3_UKERNELS = [
1013 "src/requantization/precise-ssse3.c",
1014 "src/requantization/q31-ssse3.c",
1015 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001016]
1017
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001018SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001019 "src/f32-prelu/gen/sse41-2x4.c",
1020 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001021 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001022 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001023 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001024 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001025 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1026 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001027 "src/requantization/precise-sse4.c",
1028 "src/requantization/q31-sse4.c",
1029 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001030]
1031
Marat Dukhan08c4a432019-10-03 09:29:21 -07001032AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001033 "src/f32-clamp/gen/avx-x8.c",
1034 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001035 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
1036 "src/f32-dwconv/gen/up16x4-avx.c",
1037 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
1038 "src/f32-dwconv/gen/up8x4-avx.c",
1039 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
1040 "src/f32-dwconv/gen/up16x9-avx.c",
1041 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
1042 "src/f32-dwconv/gen/up8x9-avx.c",
1043 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
1044 "src/f32-dwconv/gen/up16x25-avx.c",
1045 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
1046 "src/f32-dwconv/gen/up8x25-avx.c",
1047 "src/f32-gemm/gen/1x8-avx-broadcast.c",
1048 "src/f32-gemm/gen/4x8-avx-broadcast.c",
1049 "src/f32-gemm/gen/5x8-avx-broadcast.c",
1050 "src/f32-gemm/gen/6x8-avx-broadcast.c",
1051 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001052 "src/f32-gemm/gen/1x16-avx-broadcast.c",
1053 "src/f32-gemm/gen/3x16-avx-broadcast.c",
1054 "src/f32-gemm/gen/4x16-avx-broadcast.c",
1055 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001056 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
1057 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
1058 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
1059 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
1060 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001061 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
1062 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
1063 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
1064 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001065 "src/f32-hswish/gen/avx-x8.c",
1066 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001067 "src/f32-igemm/gen/1x8-avx-broadcast.c",
1068 "src/f32-igemm/gen/4x8-avx-broadcast.c",
1069 "src/f32-igemm/gen/5x8-avx-broadcast.c",
1070 "src/f32-igemm/gen/6x8-avx-broadcast.c",
1071 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001072 "src/f32-igemm/gen/1x16-avx-broadcast.c",
1073 "src/f32-igemm/gen/3x16-avx-broadcast.c",
1074 "src/f32-igemm/gen/4x16-avx-broadcast.c",
1075 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001076 "src/f32-prelu/gen/avx-2x8.c",
1077 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001078 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001079 "src/f32-vbinary/gen/vadd-avx-x8.c",
1080 "src/f32-vbinary/gen/vadd-avx-x16.c",
1081 "src/f32-vbinary/gen/vaddc-avx-x8.c",
1082 "src/f32-vbinary/gen/vaddc-avx-x16.c",
1083 "src/f32-vbinary/gen/vdiv-avx-x8.c",
1084 "src/f32-vbinary/gen/vdiv-avx-x16.c",
1085 "src/f32-vbinary/gen/vdivc-avx-x8.c",
1086 "src/f32-vbinary/gen/vdivc-avx-x16.c",
1087 "src/f32-vbinary/gen/vmax-avx-x8.c",
1088 "src/f32-vbinary/gen/vmax-avx-x16.c",
1089 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1090 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1091 "src/f32-vbinary/gen/vmin-avx-x8.c",
1092 "src/f32-vbinary/gen/vmin-avx-x16.c",
1093 "src/f32-vbinary/gen/vminc-avx-x8.c",
1094 "src/f32-vbinary/gen/vminc-avx-x16.c",
1095 "src/f32-vbinary/gen/vmul-avx-x8.c",
1096 "src/f32-vbinary/gen/vmul-avx-x16.c",
1097 "src/f32-vbinary/gen/vmulc-avx-x8.c",
1098 "src/f32-vbinary/gen/vmulc-avx-x16.c",
1099 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
1100 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
1101 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
1102 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
1103 "src/f32-vbinary/gen/vsub-avx-x8.c",
1104 "src/f32-vbinary/gen/vsub-avx-x16.c",
1105 "src/f32-vbinary/gen/vsubc-avx-x8.c",
1106 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001107 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001108]
1109
Marat Dukhanfda12b82019-11-21 12:27:59 -08001110FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001111 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
1112 "src/f32-dwconv/gen/up16x4-fma3.c",
1113 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
1114 "src/f32-dwconv/gen/up8x4-fma3.c",
1115 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
1116 "src/f32-dwconv/gen/up16x9-fma3.c",
1117 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
1118 "src/f32-dwconv/gen/up8x9-fma3.c",
1119 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
1120 "src/f32-dwconv/gen/up16x25-fma3.c",
1121 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
1122 "src/f32-dwconv/gen/up8x25-fma3.c",
1123 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
1124 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
1125 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
1126 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
1127 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
1128 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001129 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
1130 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
1131 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
1132 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001133 "src/f32-gemm/gen/1x16s4-fma3-broadcast.c",
1134 "src/f32-gemm/gen/3x16s4-fma3-broadcast.c",
1135 "src/f32-gemm/gen/4x16s4-fma3-broadcast.c",
1136 "src/f32-gemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001137 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
1138 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
1139 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
1140 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
1141 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
1142 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001143 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
1144 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
1145 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
1146 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001147 "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c",
1148 "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c",
1149 "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c",
1150 "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001151 "src/f32-hswish/gen/fma3-x8.c",
1152 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001153 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
1154 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
1155 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
1156 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
1157 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
1158 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001159 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
1160 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
1161 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
1162 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001163 "src/f32-igemm/gen/1x16s4-fma3-broadcast.c",
1164 "src/f32-igemm/gen/3x16s4-fma3-broadcast.c",
1165 "src/f32-igemm/gen/4x16s4-fma3-broadcast.c",
1166 "src/f32-igemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001167]
1168
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001169AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001170 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1171 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1172 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1173 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1174 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1175 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1176 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1177 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1178 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1179 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1180 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1181 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1182 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1183 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1184 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1185 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1186 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1187 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1188 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1189 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1190 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1191 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1192 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1193 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1194 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1195 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1196 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1197 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1198 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1199 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1200 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1201 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1202 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1203 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1204 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1205 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001206 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1207 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1208 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1209 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1210 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1211 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1212 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1213 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1214 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1215 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1216 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1217 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1218 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1219 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1220 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1221 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1222 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1223 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1224 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1225 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1226 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1227 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1228 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1229 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1230 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1231 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1232 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1233 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1234 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1235 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001236 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1237 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1238 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1239 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1240 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1241 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1242 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1243 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1244 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1245 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1246 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1247 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1248 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1249 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1250 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1251 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1252 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1253 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1254 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1255 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1256 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1257 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1258 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1259 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001260 "src/math/exp-avx2-p5.c",
1261 "src/math/exp-avx2-perm-p3.c",
1262 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001263 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001264 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001265 "src/math/sigmoid-avx2-rr2-p5-div.c",
1266 "src/math/sigmoid-avx2-rr1-p5-div.c",
1267 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1268 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1269 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1270 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001271]
1272
Marat Dukhan08c4a432019-10-03 09:29:21 -07001273AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001274 "src/f32-clamp/gen/avx512f-x16.c",
1275 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -08001276 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
1277 "src/f32-dwconv/gen/up32x4-avx512f.c",
1278 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
1279 "src/f32-dwconv/gen/up16x4-avx512f.c",
1280 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
1281 "src/f32-dwconv/gen/up32x9-avx512f.c",
1282 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
1283 "src/f32-dwconv/gen/up16x9-avx512f.c",
1284 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
1285 "src/f32-dwconv/gen/up32x25-avx512f.c",
1286 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
1287 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001288 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
1289 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
1290 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
1291 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
1292 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
1293 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
1294 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
1295 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
1296 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
1297 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
1298 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
1299 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001300 "src/f32-hswish/gen/avx512f-x16.c",
1301 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001302 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
1303 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
1304 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
1305 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
1306 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
1307 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001308 "src/f32-prelu/gen/avx512f-2x16.c",
1309 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001310 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1311 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1312 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1313 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1314 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1315 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1316 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1317 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1318 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1319 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1320 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1321 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1322 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1323 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1324 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1325 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1326 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1327 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1328 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1329 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1330 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1331 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1332 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1333 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1334 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1335 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1336 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1337 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1338 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1339 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1340 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1341 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1342 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1343 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1344 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1345 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001346 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001347 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1348 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1349 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1350 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1351 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1352 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1353 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1354 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1355 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1356 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1357 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1358 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1359 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1360 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1361 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1362 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1363 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1364 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1365 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1366 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1367 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1368 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1369 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1370 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1371 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1372 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1373 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1374 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001375 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001376 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1377 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1378 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1379 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1380 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1381 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1382 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1383 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1384 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1385 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1386 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1387 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1388 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1389 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1390 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1391 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1392 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1393 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1394 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1395 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1396 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1397 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1398 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1399 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001400 "src/math/exp-avx512f-p5-scalef.c",
1401 "src/math/exp-avx512f-p5.c",
1402 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001403 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001404 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001405]
1406
1407AARCH32_ASM_UKERNELS = [
1408 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard13916042019-12-11 10:56:34 -08001409 "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S",
Frank Barchardabf81542019-12-13 16:18:30 -08001410 "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S",
1411 "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001412 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Frank Barcharddc38f072020-02-10 13:21:42 -08001413 "src/f32-igemm/4x8-aarch32-neon-ld64.S",
Frank Barchard90ce7892020-02-10 23:35:45 -08001414 "src/f32-igemm/gen/4x8-aarch32-neon-cortex-a75.S",
1415 "src/f32-igemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchardc87a8fd2020-02-12 13:02:52 -08001416 "src/f32-igemm/4x8-aarch32-neon-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001417]
1418
1419AARCH64_ASM_UKERNELS = [
1420 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1421 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001422 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1423 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1424 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1425 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1426 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1427 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1428 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1429 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1430 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1431 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001432 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001433 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1434 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard91e19992020-03-09 18:46:14 -07001435 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a55.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001436 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001437 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001438 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1439 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1440 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1441 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1442 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1443 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1444 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1445 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1446 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1447 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1448 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1449 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1450 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001451 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001452 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1453 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard91e19992020-03-09 18:46:14 -07001454 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a55.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001455 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001456 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001457 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1458 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1459 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001460 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001461 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001462 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1463 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001464 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001465 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001466 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1467 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1468 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
1469 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001470 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard91e19992020-03-09 18:46:14 -07001471 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001472 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001473 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1474 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001475]
1476
1477INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001478 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001479 "src/xnnpack/argmaxpool.h",
1480 "src/xnnpack/avgpool.h",
1481 "src/xnnpack/clamp.h",
1482 "src/xnnpack/common.h",
1483 "src/xnnpack/conv.h",
1484 "src/xnnpack/dwconv.h",
1485 "src/xnnpack/gavgpool.h",
1486 "src/xnnpack/gemm.h",
1487 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001488 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001489 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001490 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001491 "src/xnnpack/lut.h",
1492 "src/xnnpack/math.h",
1493 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001494 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001495 "src/xnnpack/packx.h",
1496 "src/xnnpack/pad.h",
1497 "src/xnnpack/params.h",
1498 "src/xnnpack/pavgpool.h",
1499 "src/xnnpack/ppmm.h",
1500 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001501 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001502 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001503 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001504 "src/xnnpack/rmax.h",
1505 "src/xnnpack/scalar-utils.h",
1506 "src/xnnpack/spmm.h",
1507 "src/xnnpack/unpool.h",
1508 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001509 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001510 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001511 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001512 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001513 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001514 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001515 "src/xnnpack/zip.h",
1516]
1517
1518INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1519 "include/xnnpack.h",
1520 "src/xnnpack/allocator.h",
1521 "src/xnnpack/compute.h",
1522 "src/xnnpack/im2col.h",
1523 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001524 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001525 "src/xnnpack/operator.h",
1526 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001527 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001528 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001529 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001530 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001531]
1532
1533ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1534 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001535]
1536
1537MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001538 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001539 "include/xnnpack.h",
1540]
1541
1542MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1543 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001544 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001545 "src/xnnpack/requantization.h",
1546 "include/xnnpack.h",
1547]
1548
1549OPERATOR_TEST_PARAMS_HDRS = [
1550 "src/xnnpack/params.h",
1551 "src/xnnpack/common.h",
1552]
1553
1554WEIGHTS_PACK_HDRS = [
1555 "src/xnnpack/pack.h",
1556 "src/xnnpack/operator.h",
1557 "src/xnnpack/compute.h",
1558]
1559
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001560LOGGING_COPTS = select({
1561 # No logging in optimized mode
1562 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1563 # Full logging in debug mode
1564 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1565 # Error-only logging in default (fastbuild) mode
1566 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1567})
1568
1569LOGGING_HDRS = [
1570 "src/xnnpack/log.h",
1571]
1572
Marat Dukhan08c4a432019-10-03 09:29:21 -07001573xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001574 name = "tables",
1575 srcs = TABLE_SRCS,
1576 hdrs = INTERNAL_HDRS,
1577 copts = xnnpack_std_copts(),
1578)
1579
1580xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581 name = "scalar_ukernels",
1582 srcs = SCALAR_UKERNELS,
1583 hdrs = INTERNAL_HDRS,
1584 aarch32_copts = ["-marm"],
1585 copts = xnnpack_std_copts(),
1586 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001587 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001588 "@FP16",
1589 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001590 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001591 ],
1592)
1593
1594xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001595 name = "wasm_ukernels",
1596 hdrs = INTERNAL_HDRS,
1597 copts = xnnpack_std_copts(),
1598 wasm_srcs = WASM_UKERNELS,
1599 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001600 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001601 "@FP16",
1602 "@FXdiv",
1603 "@pthreadpool",
1604 ],
1605)
1606
1607xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001608 name = "psimd_fastmath_ukernels",
1609 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001610 hdrs = INTERNAL_HDRS,
1611 aarch32_copts = [
1612 "-marm",
1613 "-mfpu=neon",
1614 ],
1615 copts = xnnpack_std_copts(),
1616 optimized_copts = [
1617 "-O3",
1618 "-ffast-math",
1619 ],
1620 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001621 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001622 "@FP16",
1623 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001624 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001625 ],
1626)
1627
1628xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001629 name = "psimd_accmath_ukernels",
1630 srcs = PSIMD_ACCMATH_UKERNELS,
1631 hdrs = INTERNAL_HDRS,
1632 aarch32_copts = [
1633 "-marm",
1634 "-mfpu=neon",
1635 ],
1636 copts = xnnpack_std_copts(),
1637 optimized_copts = [
1638 "-O3",
1639 ],
1640 deps = [
1641 ":tables",
1642 "@FP16",
1643 "@psimd",
1644 "@pthreadpool",
1645 ],
1646)
1647
1648xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001649 name = "neon_ukernels",
1650 hdrs = INTERNAL_HDRS,
1651 aarch32_copts = [
1652 "-marm",
1653 "-mfpu=neon",
1654 ],
1655 aarch32_srcs = NEON_UKERNELS,
1656 aarch64_srcs = NEON_UKERNELS,
1657 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001658 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001659 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001660 "@FP16",
1661 "@pthreadpool",
1662 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001663)
1664
1665xnnpack_cc_library(
1666 name = "neonfma_ukernels",
1667 hdrs = INTERNAL_HDRS,
1668 aarch32_copts = [
1669 "-marm",
1670 "-mfpu=neon-vfpv4",
1671 ],
1672 aarch32_srcs = NEONFMA_UKERNELS,
1673 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1674 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001675 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001676 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001677 "@FP16",
1678 "@pthreadpool",
1679 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001680)
1681
1682xnnpack_cc_library(
1683 name = "neonfp16arith_ukernels",
1684 hdrs = INTERNAL_HDRS,
1685 aarch64_copts = ["-march=armv8.2-a+fp16"],
1686 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1687 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001688 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001689 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001690 "@FP16",
1691 "@pthreadpool",
1692 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693)
1694
1695xnnpack_cc_library(
1696 name = "sse2_ukernels",
1697 hdrs = INTERNAL_HDRS,
1698 copts = xnnpack_std_copts(),
1699 x86_copts = ["-msse2"],
1700 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001701 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001702 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001703 "@FP16",
1704 "@pthreadpool",
1705 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001706)
1707
1708xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001709 name = "ssse3_ukernels",
1710 hdrs = INTERNAL_HDRS,
1711 copts = xnnpack_std_copts(),
1712 x86_copts = ["-mssse3"],
1713 x86_srcs = SSSE3_UKERNELS,
1714 deps = [
1715 ":tables",
1716 "@FP16",
1717 "@pthreadpool",
1718 ],
1719)
1720
1721xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001722 name = "sse41_ukernels",
1723 hdrs = INTERNAL_HDRS,
1724 copts = xnnpack_std_copts(),
1725 x86_copts = ["-msse4.1"],
1726 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001727 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001728 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001729 "@FP16",
1730 "@pthreadpool",
1731 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001732)
1733
1734xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001735 name = "avx_ukernels",
1736 hdrs = INTERNAL_HDRS,
1737 copts = xnnpack_std_copts(),
1738 x86_copts = ["-mavx"],
1739 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001740 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001741 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001742 "@FP16",
1743 "@pthreadpool",
1744 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001745)
1746
1747xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001748 name = "fma3_ukernels",
1749 hdrs = INTERNAL_HDRS,
1750 copts = xnnpack_std_copts(),
1751 x86_copts = [
1752 "-mfma",
1753 ],
1754 x86_srcs = FMA3_UKERNELS,
1755 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001756 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001757 "@FP16",
1758 "@pthreadpool",
1759 ],
1760)
1761
1762xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001763 name = "avx2_ukernels",
1764 hdrs = INTERNAL_HDRS,
1765 copts = xnnpack_std_copts(),
1766 x86_copts = [
1767 "-mfma",
1768 "-mavx2",
1769 ],
1770 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001771 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001772 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001773 "@FP16",
1774 "@pthreadpool",
1775 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001776)
1777
1778xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001779 name = "avx512f_ukernels",
1780 hdrs = INTERNAL_HDRS,
1781 copts = xnnpack_std_copts(),
1782 x86_copts = ["-mavx512f"],
1783 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001784 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001785 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001786 "@FP16",
1787 "@pthreadpool",
1788 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001789)
1790
1791xnnpack_cc_library(
1792 name = "asm_ukernels",
1793 hdrs = ["src/xnnpack/assembly.h"],
1794 aarch32_srcs = AARCH32_ASM_UKERNELS,
1795 aarch64_srcs = AARCH64_ASM_UKERNELS,
1796)
1797
1798xnnpack_aggregate_library(
1799 name = "ukernels",
1800 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001801 ":psimd_fastmath_ukernels",
1802 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001803 ":neon_ukernels",
1804 ":neonfma_ukernels",
1805 ":asm_ukernels",
1806 ],
1807 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001808 ":psimd_fastmath_ukernels",
1809 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001810 ":neon_ukernels",
1811 ":neonfma_ukernels",
1812 ":neonfp16arith_ukernels",
1813 ":asm_ukernels",
1814 ],
1815 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001816 wasm_deps = [
1817 ":wasm_ukernels",
1818 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001819 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001820 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001821 ":psimd_fastmath_ukernels",
1822 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001823 ],
1824 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001825 ":psimd_fastmath_ukernels",
1826 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001827 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001828 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001829 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001830 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001831 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001832 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001833 ":avx512f_ukernels",
1834 ],
1835)
1836
1837xnnpack_cc_library(
1838 name = "im2col",
1839 srcs = ["src/im2col.c"],
1840 hdrs = [
1841 "src/xnnpack/common.h",
1842 "src/xnnpack/im2col.h",
1843 ],
1844 copts = xnnpack_std_copts(),
1845)
1846
1847xnnpack_cc_library(
1848 name = "indirection",
1849 srcs = ["src/indirection.c"],
1850 hdrs = INTERNAL_HDRS,
1851 copts = xnnpack_std_copts(),
1852 deps = [
1853 "@FP16",
1854 "@FXdiv",
1855 "@pthreadpool",
1856 ],
1857)
1858
1859xnnpack_cc_library(
1860 name = "operator_run",
1861 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001862 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1863 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001864 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1865 "-Wno-vla",
1866 ],
1867 deps = [
1868 "@FP16",
1869 "@FXdiv",
1870 "@clog",
1871 "@pthreadpool",
1872 ],
1873)
1874
1875cc_library(
1876 name = "enable_assembly",
1877 defines = select({
1878 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1879 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001880 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001881 }),
1882)
1883
Marat Dukhancf056b22019-10-07 10:26:29 -07001884xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001885 name = "operators",
1886 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001887 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001888 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001889 ],
1890 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001891 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001892 "-Isrc",
1893 "-Iinclude",
1894 ] + select({
1895 ":debug_build": [],
1896 "//conditions:default": xnnpack_min_size_copts(),
1897 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001898 wasm_srcs = ["src/wasm-stubs.c"],
1899 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001900 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001901 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001902 "@FP16",
1903 "@FXdiv",
1904 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001905 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001906 ],
1907)
1908
1909cc_library(
1910 name = "XNNPACK",
1911 srcs = [
1912 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08001913 "src/runtime.c",
1914 "src/subgraph.c",
1915 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001916 ],
1917 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1918 "-Isrc",
1919 "-Iinclude",
1920 ] + select({
1921 ":debug_build": [],
1922 "//conditions:default": xnnpack_min_size_copts(),
1923 }),
1924 includes = ["include"],
1925 linkstatic = True,
1926 textual_hdrs = ["include/xnnpack.h"],
1927 visibility = xnnpack_visibility(),
1928 deps = [
1929 ":enable_assembly",
1930 ":ukernels",
1931 ":operator_run",
1932 ":operators",
1933 "@clog",
1934 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001935 ] + select({
1936 ":emscripten": [],
1937 "//conditions:default": ["@cpuinfo"],
1938 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001939)
1940
1941cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001942 name = "xnnpack_operators_nhwc_f32",
1943 srcs = [
1944 "src/init.c",
1945 ],
1946 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1947 "-Isrc",
1948 "-Iinclude",
1949 ] + select({
1950 ":debug_build": [],
1951 "//conditions:default": xnnpack_min_size_copts(),
1952 }),
1953 defines = [
1954 "XNN_NO_Q8_OPERATORS",
1955 "XNN_NO_U8_OPERATORS",
1956 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001957 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001958 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001959 includes = ["include"],
1960 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001961 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001962 visibility = xnnpack_visibility(),
1963 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001964 ":enable_assembly",
1965 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001966 ":operator_run",
1967 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001968 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001969 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001970 ] + select({
1971 ":emscripten": [],
1972 "//conditions:default": ["@cpuinfo"],
1973 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001974)
1975
Marat Dukhancf056b22019-10-07 10:26:29 -07001976xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001977 name = "bench_utils",
1978 srcs = ["bench/utils.cc"],
1979 hdrs = ["bench/utils.h"],
1980 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001981 deps = [
1982 "@com_google_benchmark//:benchmark",
1983 "@cpuinfo",
1984 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001985)
1986
Frank Barchard7e955972019-10-11 10:34:25 -07001987######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001988
1989xnnpack_benchmark(
1990 name = "q8_gemm_bench",
1991 srcs = [
1992 "bench/gemm.h",
1993 "bench/q8-gemm.cc",
1994 "src/xnnpack/AlignedAllocator.h",
1995 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1996 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1997 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1998)
1999
2000xnnpack_benchmark(
2001 name = "f16_gemm_bench",
2002 srcs = [
2003 "bench/f16-gemm.cc",
2004 "bench/gemm.h",
2005 "src/xnnpack/AlignedAllocator.h",
2006 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2007 copts = ["-Wno-unused-function"],
2008 deps = MICROKERNEL_BENCHMARK_DEPS,
2009)
2010
2011xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002012 name = "f16_spmm_bench",
2013 srcs = [
2014 "bench/f16-spmm.cc",
2015 "bench/gemm.h",
2016 "bench/google/gemm.h",
2017 "src/xnnpack/AlignedAllocator.h",
2018 ] + MICROKERNEL_BENCHMARK_HDRS,
2019 copts = ["-Wno-unused-function"],
2020 deps = MICROKERNEL_BENCHMARK_DEPS,
2021)
2022
2023xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002024 name = "f32_igemm_bench",
2025 srcs = [
2026 "bench/f32-igemm.cc",
2027 "bench/conv.h",
2028 "src/xnnpack/AlignedAllocator.h",
2029 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002030 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002031)
2032
2033xnnpack_benchmark(
2034 name = "f32_conv_hwc_bench",
2035 srcs = [
2036 "bench/f32-conv-hwc.cc",
2037 "bench/dconv.h",
2038 "src/xnnpack/AlignedAllocator.h",
2039 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2040 copts = ["-Wno-unused-function"],
2041 deps = MICROKERNEL_BENCHMARK_DEPS,
2042)
2043
2044xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002045 name = "f32_conv_hwc2spchw_bench",
2046 srcs = [
2047 "bench/f32-conv-hwc2spchw.cc",
2048 "bench/dconv.h",
2049 "src/xnnpack/AlignedAllocator.h",
2050 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2051 copts = ["-Wno-unused-function"],
2052 deps = MICROKERNEL_BENCHMARK_DEPS,
2053)
2054
2055xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002056 name = "f32_dwconv_bench",
2057 srcs = [
2058 "bench/f32-dwconv.cc",
2059 "bench/dwconv.h",
2060 "src/xnnpack/AlignedAllocator.h",
2061 ] + MICROKERNEL_BENCHMARK_HDRS,
2062 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2063)
2064
2065xnnpack_benchmark(
2066 name = "f32_dwconv_spchw_bench",
2067 srcs = [
2068 "bench/f32-dwconv-spchw.cc",
2069 "bench/dwconv.h",
2070 "src/xnnpack/AlignedAllocator.h",
2071 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2072 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2073)
2074
2075xnnpack_benchmark(
2076 name = "f32_gemm_bench",
2077 srcs = [
2078 "bench/f32-gemm.cc",
2079 "bench/gemm.h",
2080 "src/xnnpack/AlignedAllocator.h",
2081 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2082 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002083 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002084)
2085
2086xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002087 name = "f32_raddexpminusmax_bench",
2088 srcs = [
2089 "bench/f32-raddexpminusmax.cc",
2090 "src/xnnpack/AlignedAllocator.h",
2091 ] + MICROKERNEL_BENCHMARK_HDRS,
2092 deps = MICROKERNEL_BENCHMARK_DEPS,
2093)
2094
2095xnnpack_benchmark(
2096 name = "f32_raddextexp_bench",
2097 srcs = [
2098 "bench/f32-raddextexp.cc",
2099 "src/xnnpack/AlignedAllocator.h",
2100 ] + MICROKERNEL_BENCHMARK_HDRS,
2101 deps = MICROKERNEL_BENCHMARK_DEPS,
2102)
2103
2104xnnpack_benchmark(
2105 name = "f32_raddstoreexpminusmax_bench",
2106 srcs = [
2107 "bench/f32-raddstoreexpminusmax.cc",
2108 "src/xnnpack/AlignedAllocator.h",
2109 ] + MICROKERNEL_BENCHMARK_HDRS,
2110 deps = MICROKERNEL_BENCHMARK_DEPS,
2111)
2112
2113xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002114 name = "f32_rmax_bench",
2115 srcs = [
2116 "bench/f32-rmax.cc",
2117 "src/xnnpack/AlignedAllocator.h",
2118 ] + MICROKERNEL_BENCHMARK_HDRS,
2119 deps = MICROKERNEL_BENCHMARK_DEPS,
2120)
2121
2122xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002123 name = "f32_sigmoid_bench",
2124 srcs = [
2125 "bench/f32-sigmoid.cc",
2126 "src/xnnpack/AlignedAllocator.h",
2127 ] + MICROKERNEL_BENCHMARK_HDRS,
2128 copts = ["-Wno-unused-function"],
2129 deps = MICROKERNEL_BENCHMARK_DEPS,
2130)
2131
2132xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002133 name = "f32_spmm_bench",
2134 srcs = [
2135 "bench/f32-spmm.cc",
2136 "bench/gemm.h",
2137 "src/xnnpack/AlignedAllocator.h",
2138 ] + MICROKERNEL_BENCHMARK_HDRS,
2139 copts = ["-Wno-unused-function"],
2140 deps = MICROKERNEL_BENCHMARK_DEPS,
2141)
2142
2143xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002144 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002145 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002146 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002147 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002148 copts = ["-Wno-unused-function"] + xnnpack_optional_dnnl_copts(),
2149 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002150)
2151
2152xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002153 name = "f32_vscaleexpminusmax_bench",
2154 srcs = [
2155 "bench/f32-vscaleexpminusmax.cc",
2156 "src/xnnpack/AlignedAllocator.h",
2157 ] + MICROKERNEL_BENCHMARK_HDRS,
2158 deps = MICROKERNEL_BENCHMARK_DEPS,
2159)
2160
2161xnnpack_benchmark(
2162 name = "f32_vscaleextexp_bench",
2163 srcs = [
2164 "bench/f32-vscaleextexp.cc",
2165 "src/xnnpack/AlignedAllocator.h",
2166 ] + MICROKERNEL_BENCHMARK_HDRS,
2167 deps = MICROKERNEL_BENCHMARK_DEPS,
2168)
2169
2170xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002171 name = "f32_im2col_gemm_bench",
2172 srcs = [
2173 "bench/f32-im2col-gemm.cc",
2174 "bench/conv.h",
2175 "src/xnnpack/AlignedAllocator.h",
2176 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2177 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2178)
2179
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002180xnnpack_benchmark(
2181 name = "requantization_bench",
2182 srcs = [
2183 "bench/requantization.cc",
2184 "src/xnnpack/requantization-stubs.h",
2185 "src/xnnpack/AlignedAllocator.h",
2186 ] + MICROKERNEL_BENCHMARK_HDRS,
2187 deps = MICROKERNEL_BENCHMARK_DEPS,
2188)
2189
Marat Dukhan08c4a432019-10-03 09:29:21 -07002190########################### Benchmarks for operators ###########################
2191
2192xnnpack_benchmark(
2193 name = "add_bench",
2194 srcs = ["bench/add.cc"],
2195 deps = OPERATOR_BENCHMARK_DEPS,
2196)
2197
2198xnnpack_benchmark(
2199 name = "average_pooling_bench",
2200 srcs = ["bench/average-pooling.cc"],
2201 deps = OPERATOR_BENCHMARK_DEPS,
2202)
2203
2204xnnpack_benchmark(
2205 name = "channel_shuffle_bench",
2206 srcs = ["bench/channel-shuffle.cc"],
2207 deps = OPERATOR_BENCHMARK_DEPS,
2208)
2209
2210xnnpack_benchmark(
2211 name = "convolution_bench",
2212 srcs = ["bench/convolution.cc"],
2213 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
2214 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
2215)
2216
2217xnnpack_benchmark(
2218 name = "deconvolution_bench",
2219 srcs = ["bench/deconvolution.cc"],
2220 copts = xnnpack_optional_tflite_copts(),
2221 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2222)
2223
2224xnnpack_benchmark(
2225 name = "global_average_pooling_bench",
2226 srcs = ["bench/global-average-pooling.cc"],
2227 deps = OPERATOR_BENCHMARK_DEPS,
2228)
2229
2230xnnpack_benchmark(
2231 name = "max_pooling_bench",
2232 srcs = ["bench/max-pooling.cc"],
2233 deps = OPERATOR_BENCHMARK_DEPS,
2234)
2235
2236xnnpack_benchmark(
2237 name = "sigmoid_bench",
2238 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002239 copts = xnnpack_optional_tflite_copts(),
2240 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002241)
2242
2243xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002244 name = "prelu_bench",
2245 srcs = ["bench/prelu.cc"],
2246 copts = xnnpack_optional_tflite_copts(),
2247 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2248)
2249
2250xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002251 name = "softmax_bench",
2252 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002253 copts = xnnpack_optional_tflite_copts(),
2254 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002255)
2256
Marat Dukhanc068bb62019-10-04 13:24:39 -07002257############################# End-to-end benchmarks ############################
2258
2259cc_library(
2260 name = "mobilenet_v1",
2261 srcs = ["models/mobilenet-v1.cc"],
2262 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002263 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002264 linkstatic = True,
2265 deps = [
2266 ":XNNPACK",
2267 "@pthreadpool",
2268 ],
2269)
2270
2271cc_library(
2272 name = "mobilenet_v2",
2273 srcs = ["models/mobilenet-v2.cc"],
2274 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002275 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002276 linkstatic = True,
2277 deps = [
2278 ":XNNPACK",
2279 "@pthreadpool",
2280 ],
2281)
2282
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002283cc_library(
2284 name = "mobilenet_v3_large",
2285 srcs = ["models/mobilenet-v3-large.cc"],
2286 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002287 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002288 linkstatic = True,
2289 deps = [
2290 ":XNNPACK",
2291 "@pthreadpool",
2292 ],
2293)
2294
2295cc_library(
2296 name = "mobilenet_v3_small",
2297 srcs = ["models/mobilenet-v3-small.cc"],
2298 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002299 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002300 linkstatic = True,
2301 deps = [
2302 ":XNNPACK",
2303 "@pthreadpool",
2304 ],
2305)
2306
Marat Dukhanc068bb62019-10-04 13:24:39 -07002307xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002308 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002309 srcs = [
2310 "bench/f32-dwconv-e2e.cc",
2311 "bench/end2end.h",
2312 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002313 copts = ["-Wno-unused-function"],
2314 deps = MICROKERNEL_BENCHMARK_DEPS + [
2315 ":XNNPACK",
2316 ":mobilenet_v1",
2317 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002318 ":mobilenet_v3_large",
2319 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002320 ],
2321)
2322
2323xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002324 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002325 srcs = [
2326 "bench/f32-gemm-e2e.cc",
2327 "bench/end2end.h",
2328 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002329 copts = ["-Wno-unused-function"],
2330 deps = MICROKERNEL_BENCHMARK_DEPS + [
2331 ":XNNPACK",
2332 ":mobilenet_v1",
2333 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002334 ":mobilenet_v3_large",
2335 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002336 ],
2337)
2338
2339xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002340 name = "end2end_bench",
2341 srcs = ["bench/end2end.cc"],
2342 deps = [
2343 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002344 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002345 ":mobilenet_v1",
2346 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002347 ":mobilenet_v3_large",
2348 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002349 "@pthreadpool",
2350 ],
2351)
2352
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002353#################### Accuracy evaluation for math functions ####################
2354
2355xnnpack_benchmark(
2356 name = "f32_exp_eval",
2357 srcs = [
2358 "eval/f32-exp.cc",
2359 "src/xnnpack/AlignedAllocator.h",
2360 ] + ACCURACY_EVAL_HDRS,
2361 deps = ACCURACY_EVAL_DEPS,
2362)
2363
Marat Dukhan515c9772019-10-17 18:07:57 -07002364xnnpack_benchmark(
2365 name = "f32_expminus_eval",
2366 srcs = [
2367 "eval/f32-expminus.cc",
2368 "src/xnnpack/AlignedAllocator.h",
2369 ] + ACCURACY_EVAL_HDRS,
2370 deps = ACCURACY_EVAL_DEPS,
2371)
2372
Marat Dukhan98ba4412019-10-23 02:14:28 -07002373xnnpack_benchmark(
2374 name = "f32_extexp_eval",
2375 srcs = [
2376 "eval/f32-extexp.cc",
2377 "src/xnnpack/AlignedAllocator.h",
2378 ] + ACCURACY_EVAL_HDRS,
2379 deps = ACCURACY_EVAL_DEPS,
2380)
2381
Marat Dukhan346a9e52019-11-15 09:06:30 -08002382xnnpack_benchmark(
2383 name = "f32_sigmoid_eval",
2384 srcs = [
2385 "eval/f32-sigmoid.cc",
2386 "src/xnnpack/AlignedAllocator.h",
2387 ] + ACCURACY_EVAL_HDRS,
2388 deps = ACCURACY_EVAL_DEPS,
2389)
2390
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391######################### Unit tests for micro-kernels #########################
2392
2393xnnpack_unit_test(
2394 name = "f16_gemm_test",
2395 srcs = [
2396 "test/f16-gemm.cc",
2397 "test/gemm-microkernel-tester.h",
2398 "src/xnnpack/AlignedAllocator.h",
2399 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2400 deps = MICROKERNEL_TEST_DEPS,
2401)
2402
2403xnnpack_unit_test(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002404 name = "f16_spmm_test",
2405 srcs = [
2406 "test/f16-spmm.cc",
2407 "test/spmm-microkernel-tester.h",
2408 "src/xnnpack/AlignedAllocator.h",
2409 ] + MICROKERNEL_TEST_HDRS,
2410 deps = MICROKERNEL_TEST_DEPS,
2411)
2412
2413xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002414 name = "f32_argmaxpool_test",
2415 srcs = [
2416 "test/f32-argmaxpool.cc",
2417 "test/argmaxpool-microkernel-tester.h",
2418 "src/xnnpack/AlignedAllocator.h",
2419 ] + MICROKERNEL_TEST_HDRS,
2420 deps = MICROKERNEL_TEST_DEPS,
2421)
2422
2423xnnpack_unit_test(
2424 name = "f32_avgpool_test",
2425 srcs = [
2426 "test/f32-avgpool.cc",
2427 "test/avgpool-microkernel-tester.h",
2428 "src/xnnpack/AlignedAllocator.h",
2429 ] + MICROKERNEL_TEST_HDRS,
2430 deps = MICROKERNEL_TEST_DEPS,
2431)
2432
2433xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002434 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002435 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002436 "test/f32-ibilinear.cc",
2437 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002438 "src/xnnpack/AlignedAllocator.h",
2439 ] + MICROKERNEL_TEST_HDRS,
2440 deps = MICROKERNEL_TEST_DEPS,
2441)
2442
2443xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002444 name = "f32_clamp_test",
2445 srcs = [
2446 "test/f32-clamp.cc",
2447 "test/clamp-microkernel-tester.h",
2448 ] + MICROKERNEL_TEST_HDRS,
2449 deps = MICROKERNEL_TEST_DEPS,
2450)
2451
2452xnnpack_unit_test(
2453 name = "f32_igemm_test",
2454 srcs = [
2455 "test/f32-igemm.cc",
2456 "test/gemm-microkernel-tester.h",
2457 "src/xnnpack/AlignedAllocator.h",
2458 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2459 deps = MICROKERNEL_TEST_DEPS,
2460)
2461
2462xnnpack_unit_test(
2463 name = "f32_conv_hwc_test",
2464 srcs = [
2465 "test/f32-conv-hwc.cc",
2466 "test/conv-hwc-microkernel-tester.h",
2467 "src/xnnpack/AlignedAllocator.h",
2468 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2469 deps = MICROKERNEL_TEST_DEPS,
2470)
2471
2472xnnpack_unit_test(
2473 name = "f32_conv_hwc2spchw_test",
2474 srcs = [
2475 "test/f32-conv-hwc2spchw.cc",
2476 "test/conv-hwc2spchw-microkernel-tester.h",
2477 "src/xnnpack/AlignedAllocator.h",
2478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2479 deps = MICROKERNEL_TEST_DEPS,
2480)
2481
2482xnnpack_unit_test(
2483 name = "f32_dwconv_test",
2484 srcs = [
2485 "test/f32-dwconv.cc",
2486 "test/dwconv-microkernel-tester.h",
2487 "src/xnnpack/AlignedAllocator.h",
2488 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2489 deps = MICROKERNEL_TEST_DEPS,
2490)
2491
2492xnnpack_unit_test(
2493 name = "f32_dwconv_spchw_test",
2494 srcs = [
2495 "test/f32-dwconv-spchw.cc",
2496 "test/dwconv-spchw-microkernel-tester.h",
2497 "src/xnnpack/AlignedAllocator.h",
2498 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2499 deps = MICROKERNEL_TEST_DEPS,
2500)
2501
2502xnnpack_unit_test(
2503 name = "f32_gavgpool_test",
2504 srcs = [
2505 "test/f32-gavgpool.cc",
2506 "test/gavgpool-microkernel-tester.h",
2507 "src/xnnpack/AlignedAllocator.h",
2508 ] + MICROKERNEL_TEST_HDRS,
2509 deps = MICROKERNEL_TEST_DEPS,
2510)
2511
2512xnnpack_unit_test(
2513 name = "f32_gavgpool_spchw_test",
2514 srcs = [
2515 "test/f32-gavgpool-spchw.cc",
2516 "test/gavgpool-spchw-microkernel-tester.h",
2517 "src/xnnpack/AlignedAllocator.h",
2518 ] + MICROKERNEL_TEST_HDRS,
2519 deps = MICROKERNEL_TEST_DEPS,
2520)
2521
2522xnnpack_unit_test(
2523 name = "f32_gemm_test",
2524 srcs = [
2525 "test/f32-gemm.cc",
2526 "test/gemm-microkernel-tester.h",
2527 "src/xnnpack/AlignedAllocator.h",
2528 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2529 deps = MICROKERNEL_TEST_DEPS,
2530)
2531
2532xnnpack_unit_test(
2533 name = "f32_gemminc_test",
2534 srcs = [
2535 "test/f32-gemminc.cc",
2536 "test/gemm-microkernel-tester.h",
2537 "src/xnnpack/AlignedAllocator.h",
2538 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2539 deps = MICROKERNEL_TEST_DEPS,
2540)
2541
2542xnnpack_unit_test(
2543 name = "f32_hswish_test",
2544 srcs = [
2545 "test/f32-hswish.cc",
2546 "test/hswish-microkernel-tester.h",
2547 ] + MICROKERNEL_TEST_HDRS,
2548 deps = MICROKERNEL_TEST_DEPS,
2549)
2550
2551xnnpack_unit_test(
2552 name = "f32_maxpool_test",
2553 srcs = [
2554 "test/f32-maxpool.cc",
2555 "test/maxpool-microkernel-tester.h",
2556 ] + MICROKERNEL_TEST_HDRS,
2557 deps = MICROKERNEL_TEST_DEPS,
2558)
2559
2560xnnpack_unit_test(
2561 name = "f32_pavgpool_test",
2562 srcs = [
2563 "test/f32-pavgpool.cc",
2564 "test/avgpool-microkernel-tester.h",
2565 "src/xnnpack/AlignedAllocator.h",
2566 ] + MICROKERNEL_TEST_HDRS,
2567 deps = MICROKERNEL_TEST_DEPS,
2568)
2569
2570xnnpack_unit_test(
2571 name = "f32_ppmm_test",
2572 srcs = [
2573 "test/f32-ppmm.cc",
2574 "test/gemm-microkernel-tester.h",
2575 "src/xnnpack/AlignedAllocator.h",
2576 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2577 deps = MICROKERNEL_TEST_DEPS,
2578)
2579
2580xnnpack_unit_test(
2581 name = "f32_prelu_test",
2582 srcs = [
2583 "test/f32-prelu.cc",
2584 "test/prelu-microkernel-tester.h",
2585 "src/xnnpack/AlignedAllocator.h",
2586 ] + MICROKERNEL_TEST_HDRS,
2587 deps = MICROKERNEL_TEST_DEPS,
2588)
2589
2590xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002591 name = "f32_raddexpminusmax_test",
2592 srcs = [
2593 "test/f32-raddexpminusmax.cc",
2594 "test/raddexpminusmax-microkernel-tester.h",
2595 ] + MICROKERNEL_TEST_HDRS,
2596 deps = MICROKERNEL_TEST_DEPS,
2597)
2598
2599xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002600 name = "f32_raddextexp_test",
2601 srcs = [
2602 "test/f32-raddextexp.cc",
2603 "test/raddextexp-microkernel-tester.h",
2604 ] + MICROKERNEL_TEST_HDRS,
2605 deps = MICROKERNEL_TEST_DEPS,
2606)
2607
2608xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002609 name = "f32_raddstoreexpminusmax_test",
2610 srcs = [
2611 "test/f32-raddstoreexpminusmax.cc",
2612 "test/raddstoreexpminusmax-microkernel-tester.h",
2613 ] + MICROKERNEL_TEST_HDRS,
2614 deps = MICROKERNEL_TEST_DEPS,
2615)
2616
2617xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002618 name = "f32_rmax_test",
2619 srcs = [
2620 "test/f32-rmax.cc",
2621 "test/rmax-microkernel-tester.h",
2622 ] + MICROKERNEL_TEST_HDRS,
2623 deps = MICROKERNEL_TEST_DEPS,
2624)
2625
2626xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002627 name = "f32_sigmoid_test",
2628 srcs = [
2629 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002630 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002631 ] + MICROKERNEL_TEST_HDRS,
2632 deps = MICROKERNEL_TEST_DEPS,
2633)
2634
2635xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002636 name = "f32_spmm_test",
2637 srcs = [
2638 "test/f32-spmm.cc",
2639 "test/spmm-microkernel-tester.h",
2640 "src/xnnpack/AlignedAllocator.h",
2641 ] + MICROKERNEL_TEST_HDRS,
2642 deps = MICROKERNEL_TEST_DEPS,
2643)
2644
2645xnnpack_unit_test(
2646 name = "f32_vadd_test",
2647 srcs = [
2648 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002649 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002650 ] + MICROKERNEL_TEST_HDRS,
2651 deps = MICROKERNEL_TEST_DEPS,
2652)
2653
2654xnnpack_unit_test(
2655 name = "f32_vaddc_test",
2656 srcs = [
2657 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002658 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002659 ] + MICROKERNEL_TEST_HDRS,
2660 deps = MICROKERNEL_TEST_DEPS,
2661)
2662
2663xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002664 name = "f32_vdiv_test",
2665 srcs = [
2666 "test/f32-vdiv.cc",
2667 "test/vbinary-microkernel-tester.h",
2668 ] + MICROKERNEL_TEST_HDRS,
2669 deps = MICROKERNEL_TEST_DEPS,
2670)
2671
2672xnnpack_unit_test(
2673 name = "f32_vdivc_test",
2674 srcs = [
2675 "test/f32-vdivc.cc",
2676 "test/vbinaryc-microkernel-tester.h",
2677 ] + MICROKERNEL_TEST_HDRS,
2678 deps = MICROKERNEL_TEST_DEPS,
2679)
2680
2681xnnpack_unit_test(
2682 name = "f32_vrdivc_test",
2683 srcs = [
2684 "test/f32-vrdivc.cc",
2685 "test/vbinaryc-microkernel-tester.h",
2686 ] + MICROKERNEL_TEST_HDRS,
2687 deps = MICROKERNEL_TEST_DEPS,
2688)
2689
2690xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002691 name = "f32_vmax_test",
2692 srcs = [
2693 "test/f32-vmax.cc",
2694 "test/vbinary-microkernel-tester.h",
2695 ] + MICROKERNEL_TEST_HDRS,
2696 deps = MICROKERNEL_TEST_DEPS,
2697)
2698
2699xnnpack_unit_test(
2700 name = "f32_vmaxc_test",
2701 srcs = [
2702 "test/f32-vmaxc.cc",
2703 "test/vbinaryc-microkernel-tester.h",
2704 ] + MICROKERNEL_TEST_HDRS,
2705 deps = MICROKERNEL_TEST_DEPS,
2706)
2707
2708xnnpack_unit_test(
2709 name = "f32_vmin_test",
2710 srcs = [
2711 "test/f32-vmin.cc",
2712 "test/vbinary-microkernel-tester.h",
2713 ] + MICROKERNEL_TEST_HDRS,
2714 deps = MICROKERNEL_TEST_DEPS,
2715)
2716
2717xnnpack_unit_test(
2718 name = "f32_vminc_test",
2719 srcs = [
2720 "test/f32-vminc.cc",
2721 "test/vbinaryc-microkernel-tester.h",
2722 ] + MICROKERNEL_TEST_HDRS,
2723 deps = MICROKERNEL_TEST_DEPS,
2724)
2725
2726xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002727 name = "f32_vmul_test",
2728 srcs = [
2729 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002730 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002731 ] + MICROKERNEL_TEST_HDRS,
2732 deps = MICROKERNEL_TEST_DEPS,
2733)
2734
2735xnnpack_unit_test(
2736 name = "f32_vmulc_test",
2737 srcs = [
2738 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002739 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 ] + MICROKERNEL_TEST_HDRS,
2741 deps = MICROKERNEL_TEST_DEPS,
2742)
2743
2744xnnpack_unit_test(
2745 name = "f32_vmulcaddc_test",
2746 srcs = [
2747 "test/f32-vmulcaddc.cc",
2748 "test/vmulcaddc-microkernel-tester.h",
2749 "src/xnnpack/AlignedAllocator.h",
2750 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2751 deps = MICROKERNEL_TEST_DEPS,
2752)
2753
2754xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002755 name = "f32_vscale_test",
2756 srcs = [
2757 "test/f32-vscale.cc",
2758 "test/vscale-microkernel-tester.h",
2759 ] + MICROKERNEL_TEST_HDRS,
2760 deps = MICROKERNEL_TEST_DEPS,
2761)
2762
2763xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002764 name = "f32_vscaleexpminusmax_test",
2765 srcs = [
2766 "test/f32-vscaleexpminusmax.cc",
2767 "test/vscaleexpminusmax-microkernel-tester.h",
2768 ] + MICROKERNEL_TEST_HDRS,
2769 deps = MICROKERNEL_TEST_DEPS,
2770)
2771
2772xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002773 name = "f32_vscaleextexp_test",
2774 srcs = [
2775 "test/f32-vscaleextexp.cc",
2776 "test/vscaleextexp-microkernel-tester.h",
2777 ] + MICROKERNEL_TEST_HDRS,
2778 deps = MICROKERNEL_TEST_DEPS,
2779)
2780
2781xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002782 name = "f32_vsub_test",
2783 srcs = [
2784 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002785 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002786 ] + MICROKERNEL_TEST_HDRS,
2787 deps = MICROKERNEL_TEST_DEPS,
2788)
2789
2790xnnpack_unit_test(
2791 name = "f32_vsubc_test",
2792 srcs = [
2793 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002794 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002795 ] + MICROKERNEL_TEST_HDRS,
2796 deps = MICROKERNEL_TEST_DEPS,
2797)
2798
2799xnnpack_unit_test(
2800 name = "f32_vrsubc_test",
2801 srcs = [
2802 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002803 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002804 ] + MICROKERNEL_TEST_HDRS,
2805 deps = MICROKERNEL_TEST_DEPS,
2806)
2807
2808xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002809 name = "q8_avgpool_test",
2810 srcs = [
2811 "test/q8-avgpool.cc",
2812 "test/avgpool-microkernel-tester.h",
2813 "src/xnnpack/AlignedAllocator.h",
2814 ] + MICROKERNEL_TEST_HDRS,
2815 deps = MICROKERNEL_TEST_DEPS,
2816)
2817
2818xnnpack_unit_test(
2819 name = "q8_igemm_test",
2820 srcs = [
2821 "test/q8-igemm.cc",
2822 "test/gemm-microkernel-tester.h",
2823 "src/xnnpack/AlignedAllocator.h",
2824 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2825 deps = MICROKERNEL_TEST_DEPS,
2826)
2827
2828xnnpack_unit_test(
2829 name = "q8_dwconv_test",
2830 srcs = [
2831 "test/q8-dwconv.cc",
2832 "test/dwconv-microkernel-tester.h",
2833 "src/xnnpack/AlignedAllocator.h",
2834 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2835 deps = MICROKERNEL_TEST_DEPS,
2836)
2837
2838xnnpack_unit_test(
2839 name = "q8_gavgpool_test",
2840 srcs = [
2841 "test/q8-gavgpool.cc",
2842 "test/gavgpool-microkernel-tester.h",
2843 "src/xnnpack/AlignedAllocator.h",
2844 ] + MICROKERNEL_TEST_HDRS,
2845 deps = MICROKERNEL_TEST_DEPS,
2846)
2847
2848xnnpack_unit_test(
2849 name = "q8_gemm_test",
2850 srcs = [
2851 "test/q8-gemm.cc",
2852 "test/gemm-microkernel-tester.h",
2853 "src/xnnpack/AlignedAllocator.h",
2854 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2855 deps = MICROKERNEL_TEST_DEPS,
2856)
2857
2858xnnpack_unit_test(
2859 name = "q8_vadd_test",
2860 srcs = [
2861 "test/q8-vadd.cc",
2862 "test/vadd-microkernel-tester.h",
2863 ] + MICROKERNEL_TEST_HDRS,
2864 deps = MICROKERNEL_TEST_DEPS,
2865)
2866
2867xnnpack_unit_test(
2868 name = "u8_clamp_test",
2869 srcs = [
2870 "test/u8-clamp.cc",
2871 "test/clamp-microkernel-tester.h",
2872 ] + MICROKERNEL_TEST_HDRS,
2873 deps = MICROKERNEL_TEST_DEPS,
2874)
2875
2876xnnpack_unit_test(
2877 name = "u8_lut32norm_test",
2878 srcs = [
2879 "test/u8-lut32norm.cc",
2880 "test/lut-norm-microkernel-tester.h",
2881 ] + MICROKERNEL_TEST_HDRS,
2882 deps = MICROKERNEL_TEST_DEPS,
2883)
2884
2885xnnpack_unit_test(
2886 name = "u8_maxpool_test",
2887 srcs = [
2888 "test/u8-maxpool.cc",
2889 "test/maxpool-microkernel-tester.h",
2890 ] + MICROKERNEL_TEST_HDRS,
2891 deps = MICROKERNEL_TEST_DEPS,
2892)
2893
2894xnnpack_unit_test(
2895 name = "u8_rmax_test",
2896 srcs = [
2897 "test/u8-rmax.cc",
2898 "test/rmax-microkernel-tester.h",
2899 ] + MICROKERNEL_TEST_HDRS,
2900 deps = MICROKERNEL_TEST_DEPS,
2901)
2902
2903xnnpack_unit_test(
2904 name = "x32_packx_test",
2905 srcs = [
2906 "test/x32-packx.cc",
2907 "test/pack-microkernel-tester.h",
2908 "src/xnnpack/AlignedAllocator.h",
2909 ] + MICROKERNEL_TEST_HDRS,
2910 deps = MICROKERNEL_TEST_DEPS,
2911)
2912
2913xnnpack_unit_test(
2914 name = "x32_pad_test",
2915 srcs = [
2916 "test/x32-pad.cc",
2917 "test/pad-microkernel-tester.h",
2918 ] + MICROKERNEL_TEST_HDRS,
2919 deps = MICROKERNEL_TEST_DEPS,
2920)
2921
2922xnnpack_unit_test(
2923 name = "x32_unpool_test",
2924 srcs = [
2925 "test/x32-unpool.cc",
2926 "test/unpool-microkernel-tester.h",
2927 ] + MICROKERNEL_TEST_HDRS,
2928 deps = MICROKERNEL_TEST_DEPS,
2929)
2930
2931xnnpack_unit_test(
2932 name = "x32_zip_test",
2933 srcs = [
2934 "test/x32-zip.cc",
2935 "test/zip-microkernel-tester.h",
2936 ] + MICROKERNEL_TEST_HDRS,
2937 deps = MICROKERNEL_TEST_DEPS,
2938)
2939
2940xnnpack_unit_test(
2941 name = "x8_lut_test",
2942 srcs = [
2943 "test/x8-lut.cc",
2944 "test/lut-microkernel-tester.h",
2945 ] + MICROKERNEL_TEST_HDRS,
2946 deps = MICROKERNEL_TEST_DEPS,
2947)
2948
2949xnnpack_unit_test(
2950 name = "x8_zip_test",
2951 srcs = [
2952 "test/x8-zip.cc",
2953 "test/zip-microkernel-tester.h",
2954 ] + MICROKERNEL_TEST_HDRS,
2955 deps = MICROKERNEL_TEST_DEPS,
2956)
2957
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002958xnnpack_unit_test(
2959 name = "requantization_test",
2960 srcs = [
2961 "src/xnnpack/requantization-stubs.h",
2962 "test/requantization.cc",
2963 "test/requantization-tester.h",
2964 ] + MICROKERNEL_TEST_HDRS,
2965 deps = MICROKERNEL_TEST_DEPS,
2966)
2967
Marat Dukhan20c3b922020-03-10 03:45:06 -07002968########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002969
2970xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07002971 name = "operator_size_test",
2972 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002973 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002974)
2975
Marat Dukhan20c3b922020-03-10 03:45:06 -07002976xnnpack_binary(
2977 name = "subgraph_size_test",
2978 srcs = ["test/subgraph-size.c"],
2979 deps = [":XNNPACK"],
2980)
2981
2982########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002983
2984xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002985 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002986 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002987 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002988 "test/add-operator-tester.h",
2989 ],
2990 deps = OPERATOR_TEST_DEPS,
2991)
2992
2993xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002994 name = "add_nd_test",
2995 srcs = [
2996 "test/add-nd.cc",
2997 "test/binary-elementwise-operator-tester.h",
2998 ],
2999 deps = OPERATOR_TEST_DEPS,
3000)
3001
3002xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003003 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003004 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003005 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003006 "test/argmax-pooling-operator-tester.h",
3007 ] + OPERATOR_TEST_PARAMS_HDRS,
3008 deps = OPERATOR_TEST_DEPS,
3009)
3010
3011xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003012 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003013 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003014 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003015 "test/average-pooling-operator-tester.h",
3016 ] + OPERATOR_TEST_PARAMS_HDRS,
3017 deps = OPERATOR_TEST_DEPS,
3018)
3019
3020xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003021 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003023 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 "test/channel-pad-operator-tester.h",
3025 ] + OPERATOR_TEST_PARAMS_HDRS,
3026 deps = OPERATOR_TEST_DEPS,
3027)
3028
3029xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003030 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003031 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003032 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003033 "test/channel-shuffle-operator-tester.h",
3034 ],
3035 deps = OPERATOR_TEST_DEPS,
3036)
3037
3038xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003039 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003040 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003041 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003042 "test/clamp-operator-tester.h",
3043 ],
3044 deps = OPERATOR_TEST_DEPS,
3045)
3046
3047xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003048 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003049 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003050 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003051 "test/convolution-operator-tester.h",
3052 ],
3053 deps = OPERATOR_TEST_DEPS,
3054)
3055
3056xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003057 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003058 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003059 "test/convolution-nchw.cc",
3060 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003061 ],
3062 deps = OPERATOR_TEST_DEPS,
3063)
3064
3065xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003066 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003067 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003068 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069 "test/deconvolution-operator-tester.h",
3070 ] + OPERATOR_TEST_PARAMS_HDRS,
3071 deps = OPERATOR_TEST_DEPS,
3072)
3073
3074xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003075 name = "divide_nd_test",
3076 srcs = [
3077 "test/binary-elementwise-operator-tester.h",
3078 "test/divide-nd.cc",
3079 ],
3080 deps = OPERATOR_TEST_DEPS,
3081)
3082
3083xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003084 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003085 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003086 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003087 "test/fully-connected-operator-tester.h",
3088 ],
3089 deps = OPERATOR_TEST_DEPS,
3090)
3091
3092xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003093 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003094 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003095 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003096 "test/global-average-pooling-operator-tester.h",
3097 ] + OPERATOR_TEST_PARAMS_HDRS,
3098 deps = OPERATOR_TEST_DEPS,
3099)
3100
3101xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003102 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003103 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003104 "test/global-average-pooling-ncw.cc",
3105 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106 ],
3107 deps = OPERATOR_TEST_DEPS,
3108)
3109
3110xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003111 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003112 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003113 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114 "test/hardswish-operator-tester.h",
3115 ],
3116 deps = OPERATOR_TEST_DEPS,
3117)
3118
3119xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003120 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003121 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003122 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003123 "test/leaky-relu-operator-tester.h",
3124 ],
3125 deps = OPERATOR_TEST_DEPS,
3126)
3127
3128xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003129 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003131 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003132 "test/max-pooling-operator-tester.h",
3133 ] + OPERATOR_TEST_PARAMS_HDRS,
3134 deps = OPERATOR_TEST_DEPS,
3135)
3136
3137xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003138 name = "maximum_nd_test",
3139 srcs = [
3140 "test/binary-elementwise-operator-tester.h",
3141 "test/maximum-nd.cc",
3142 ],
3143 deps = OPERATOR_TEST_DEPS,
3144)
3145
3146xnnpack_unit_test(
3147 name = "minimum_nd_test",
3148 srcs = [
3149 "test/binary-elementwise-operator-tester.h",
3150 "test/minimum-nd.cc",
3151 ],
3152 deps = OPERATOR_TEST_DEPS,
3153)
3154
3155xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003156 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003157 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003158 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003159 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003160 ],
3161 deps = OPERATOR_TEST_DEPS,
3162)
3163
3164xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003165 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003167 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003168 "test/prelu-operator-tester.h",
3169 ] + OPERATOR_TEST_PARAMS_HDRS,
3170 deps = OPERATOR_TEST_DEPS,
3171)
3172
3173xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003174 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003175 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003176 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003177 "test/resize-bilinear-operator-tester.h",
3178 ] + OPERATOR_TEST_PARAMS_HDRS,
3179 deps = OPERATOR_TEST_DEPS,
3180)
3181
3182xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003183 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003185 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003186 "test/sigmoid-operator-tester.h",
3187 ],
3188 deps = OPERATOR_TEST_DEPS,
3189)
3190
3191xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003192 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003194 "test/softmax-nc.cc",
3195 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003196 ],
3197 deps = OPERATOR_TEST_DEPS,
3198)
3199
3200xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003201 name = "subtract_nd_test",
3202 srcs = [
3203 "test/binary-elementwise-operator-tester.h",
3204 "test/subtract-nd.cc",
3205 ],
3206 deps = OPERATOR_TEST_DEPS,
3207)
3208
3209xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003210 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003211 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003212 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003213 "test/unpooling-operator-tester.h",
3214 ],
3215 deps = OPERATOR_TEST_DEPS,
3216)
3217
3218############################# Build configurations #############################
3219
Marat Dukhanb8642352019-10-30 15:43:02 -07003220# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003221config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003222 name = "xnn_enable_assembly_explicit_true",
3223 define_values = {"xnn_enable_assembly": "true"},
3224)
3225
3226# Disables usage of assembly kernels.
3227config_setting(
3228 name = "xnn_enable_assembly_explicit_false",
3229 define_values = {"xnn_enable_assembly": "false"},
3230)
3231
3232# Builds with -c dbg
3233config_setting(
3234 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003235 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003236 "compilation_mode": "dbg",
3237 },
3238)
3239
3240# Builds with -c opt
3241config_setting(
3242 name = "optimized_build",
3243 values = {
3244 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245 },
3246)
3247
3248config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003249 name = "linux_k8",
3250 values = {"cpu": "k8"},
3251)
3252
3253config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003254 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003255 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003256)
3257
3258config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003259 name = "android",
3260 values = {"crosstool_top": "//external:android/crosstool"},
3261)
3262
3263config_setting(
3264 name = "android_armv7",
3265 values = {
3266 "crosstool_top": "//external:android/crosstool",
3267 "cpu": "armeabi-v7a",
3268 },
3269)
3270
3271config_setting(
3272 name = "android_arm64",
3273 values = {
3274 "crosstool_top": "//external:android/crosstool",
3275 "cpu": "arm64-v8a",
3276 },
3277)
3278
3279config_setting(
3280 name = "android_x86",
3281 values = {
3282 "crosstool_top": "//external:android/crosstool",
3283 "cpu": "x86",
3284 },
3285)
3286
3287config_setting(
3288 name = "android_x86_64",
3289 values = {
3290 "crosstool_top": "//external:android/crosstool",
3291 "cpu": "x86_64",
3292 },
3293)
3294
3295config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003296 name = "macos_x86_64",
3297 values = {
3298 "apple_platform_type": "macos",
3299 "cpu": "darwin",
3300 },
3301)
3302
3303config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003304 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003305 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003306)
3307
3308config_setting(
3309 name = "emscripten_wasm",
3310 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003311 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003312 "cpu": "wasm",
3313 },
3314)
3315
3316config_setting(
3317 name = "emscripten_wasmsimd",
3318 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003319 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003320 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003321 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003322 },
3323)
3324
3325config_setting(
3326 name = "emscripten_asmjs",
3327 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003328 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003329 "cpu": "asmjs",
3330 },
3331)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003332
3333config_setting(
3334 name = "ios_armv7",
3335 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003336 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003337 "cpu": "ios_armv7",
3338 },
3339)
3340
3341config_setting(
3342 name = "ios_arm64",
3343 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003344 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003345 "cpu": "ios_arm64",
3346 },
3347)
3348
3349config_setting(
3350 name = "ios_arm64e",
3351 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003352 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003353 "cpu": "ios_arm64e",
3354 },
3355)
3356
3357config_setting(
3358 name = "ios_x86",
3359 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003360 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003361 "cpu": "ios_i386",
3362 },
3363)
3364
3365config_setting(
3366 name = "ios_x86_64",
3367 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003368 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003369 "cpu": "ios_x86_64",
3370 },
3371)
3372
3373config_setting(
3374 name = "watchos_armv7k",
3375 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003376 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003377 "cpu": "watchos_armv7k",
3378 },
3379)
3380
3381config_setting(
3382 name = "watchos_arm64_32",
3383 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003384 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003385 "cpu": "watchos_arm64_32",
3386 },
3387)
3388
3389config_setting(
3390 name = "watchos_x86",
3391 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003392 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003393 "cpu": "watchos_i386",
3394 },
3395)
3396
3397config_setting(
3398 name = "watchos_x86_64",
3399 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003400 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003401 "cpu": "watchos_x86_64",
3402 },
3403)
3404
3405config_setting(
3406 name = "tvos_arm64",
3407 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003408 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003409 "cpu": "tvos_arm64",
3410 },
3411)
3412
3413config_setting(
3414 name = "tvos_x86_64",
3415 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003416 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003417 "cpu": "tvos_x86_64",
3418 },
3419)