blob: 9c52c295c1d2904d1f18512cb3d9031b5a0e32a3 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
Marat Dukhan08c4a432019-10-03 09:29:21 -070022MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
Marat Dukhan1b354632020-03-23 12:50:22 -070046OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070053 "src/operators/add-nc.c",
54 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
57 "src/operators/channel-pad-nc.c",
58 "src/operators/channel-shuffle-nc.c",
59 "src/operators/clamp-nc.c",
60 "src/operators/convolution-nchw.c",
61 "src/operators/convolution-nhwc.c",
62 "src/operators/deconvolution-nhwc.c",
63 "src/operators/fully-connected-nc.c",
64 "src/operators/global-average-pooling-ncw.c",
65 "src/operators/global-average-pooling-nwc.c",
66 "src/operators/hardswish-nc.c",
67 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
70 "src/operators/resize-bilinear-nhwc.c",
71 "src/operators/sigmoid-nc.c",
72 "src/operators/softmax-nc.c",
73 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -070085 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
86 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -070087 "src/f32-clamp/gen/scalar-x1.c",
88 "src/f32-clamp/gen/scalar-x2.c",
89 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080090 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080091 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070092 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070093 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070094 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080095 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080096 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -070097 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x4-scalar.c",
99 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
100 "src/f32-dwconv/gen/up1x9-scalar.c",
101 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
102 "src/f32-dwconv/gen/up1x25-scalar.c",
103 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x4-scalar.c",
105 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
106 "src/f32-dwconv/gen/up2x9-scalar.c",
107 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
108 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700109 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
110 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
111 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
112 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700113 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
114 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700115 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
116 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
117 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
118 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700119 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
120 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700121 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700122 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
123 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700124 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
125 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
126 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700127 "src/f32-gemm/gen/1x4-scalar.c",
128 "src/f32-gemm/gen/2x4-scalar.c",
129 "src/f32-gemm/gen/4x2-scalar.c",
130 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700131 "src/f32-gemm/gen/1x4-minmax-scalar.c",
132 "src/f32-gemm/gen/2x4-minmax-scalar.c",
133 "src/f32-gemm/gen/4x2-minmax-scalar.c",
134 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800135 "src/f32-hswish/gen/scalar-x1.c",
136 "src/f32-hswish/gen/scalar-x2.c",
137 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700138 "src/f32-ibilinear/gen/scalar-c1.c",
139 "src/f32-ibilinear/gen/scalar-c2.c",
140 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-igemm/gen/1x4-scalar.c",
142 "src/f32-igemm/gen/2x4-scalar.c",
143 "src/f32-igemm/gen/4x2-scalar.c",
144 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-igemm/gen/1x4-minmax-scalar.c",
146 "src/f32-igemm/gen/2x4-minmax-scalar.c",
147 "src/f32-igemm/gen/4x2-minmax-scalar.c",
148 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700149 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
150 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
151 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700152 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
153 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
154 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
155 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800156 "src/f32-prelu/gen/scalar-2x1.c",
157 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800158 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
159 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
160 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
161 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
162 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
163 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
164 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
165 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
166 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
167 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
168 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
169 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700170 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800171 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
172 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
173 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
174 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
175 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
176 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
177 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
178 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
179 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700180 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
181 "src/f32-spmm/gen/1x1-minmax-scalar.c",
182 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
183 "src/f32-spmm/gen/2x1-minmax-scalar.c",
184 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
185 "src/f32-spmm/gen/4x1-minmax-scalar.c",
186 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
187 "src/f32-spmm/gen/8x1-minmax-scalar.c",
188 "src/f32-spmm/gen/8x2-minmax-scalar.c",
189 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700190 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
191 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
192 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
193 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
194 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
195 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
196 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
197 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
198 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
199 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
200 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
201 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800202 "src/f32-vbinary/gen/vmax-scalar-x1.c",
203 "src/f32-vbinary/gen/vmax-scalar-x2.c",
204 "src/f32-vbinary/gen/vmax-scalar-x4.c",
205 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
206 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
207 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
208 "src/f32-vbinary/gen/vmin-scalar-x1.c",
209 "src/f32-vbinary/gen/vmin-scalar-x2.c",
210 "src/f32-vbinary/gen/vmin-scalar-x4.c",
211 "src/f32-vbinary/gen/vminc-scalar-x1.c",
212 "src/f32-vbinary/gen/vminc-scalar-x2.c",
213 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700214 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
215 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
216 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
217 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
218 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
219 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
220 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
221 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
222 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
223 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
224 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
225 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
226 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
227 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
228 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
229 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700232 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
233 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
234 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800235 "src/math/expminus-scalar-lut2048-p1.c",
236 "src/math/expminus-scalar-lut64-p2.c",
237 "src/math/expminus-scalar-p5.c",
238 "src/math/sigmoid-scalar-lut2048-p1-div.c",
239 "src/math/sigmoid-scalar-lut64-p2-div.c",
240 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700241 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
242 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700243 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700244 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
245 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700246 "src/q8-gemm/2x2-minmax-scalar.c",
247 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700248 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700249 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700250 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700251 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700252 "src/u8-rmax/scalar.c",
253 "src/x32-packx/x2-scalar.c",
254 "src/x32-packx/x3-scalar.c",
255 "src/x32-packx/x4-scalar.c",
256 "src/x32-pad/x2-scalar.c",
257 "src/x32-unpool/scalar.c",
258 "src/x32-zip/x2-scalar.c",
259 "src/x32-zip/x3-scalar.c",
260 "src/x32-zip/x4-scalar.c",
261 "src/x32-zip/xm-scalar.c",
262 "src/x8-lut/scalar.c",
263 "src/x8-zip/x2-scalar.c",
264 "src/x8-zip/x3-scalar.c",
265 "src/x8-zip/x4-scalar.c",
266 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700267 "src/requantization/precise-scalar.c",
268 "src/requantization/fp32-scalar.c",
269 "src/requantization/q31-scalar.c",
270 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700271]
272
Marat Dukhan436ebe62019-12-04 15:10:12 -0800273WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700274 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
275 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700276 "src/f32-clamp/gen/wasm-x1.c",
277 "src/f32-clamp/gen/wasm-x2.c",
278 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700279 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
280 "src/f32-dwconv/gen/up1x4-wasm.c",
281 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
282 "src/f32-dwconv/gen/up1x9-wasm.c",
283 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
284 "src/f32-dwconv/gen/up1x25-wasm.c",
285 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
286 "src/f32-dwconv/gen/up2x4-wasm.c",
287 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
288 "src/f32-dwconv/gen/up2x9-wasm.c",
289 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
290 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700291 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
292 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
293 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
294 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700295 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
296 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700297 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
298 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
299 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
300 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700301 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
302 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700303 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
304 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700305 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
306 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
307 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700308 "src/f32-gemm/gen/1x4-wasm.c",
309 "src/f32-gemm/gen/2x4-wasm.c",
310 "src/f32-gemm/gen/4x2-wasm.c",
311 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700312 "src/f32-gemm/gen/1x4-minmax-wasm.c",
313 "src/f32-gemm/gen/2x4-minmax-wasm.c",
314 "src/f32-gemm/gen/4x2-minmax-wasm.c",
315 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800316 "src/f32-hswish/gen/wasm-x1.c",
317 "src/f32-hswish/gen/wasm-x2.c",
318 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700319 "src/f32-igemm/gen/1x4-wasm.c",
320 "src/f32-igemm/gen/2x4-wasm.c",
321 "src/f32-igemm/gen/4x2-wasm.c",
322 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700323 "src/f32-igemm/gen/1x4-minmax-wasm.c",
324 "src/f32-igemm/gen/2x4-minmax-wasm.c",
325 "src/f32-igemm/gen/4x2-minmax-wasm.c",
326 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700327 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
328 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
329 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700330 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
331 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
332 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
333 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
334 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
335 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
336 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
337 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
338 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
339 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
340 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
341 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800342 "src/f32-vbinary/gen/vmax-wasm-x1.c",
343 "src/f32-vbinary/gen/vmax-wasm-x2.c",
344 "src/f32-vbinary/gen/vmax-wasm-x4.c",
345 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
346 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
347 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
348 "src/f32-vbinary/gen/vmin-wasm-x1.c",
349 "src/f32-vbinary/gen/vmin-wasm-x2.c",
350 "src/f32-vbinary/gen/vmin-wasm-x4.c",
351 "src/f32-vbinary/gen/vminc-wasm-x1.c",
352 "src/f32-vbinary/gen/vminc-wasm-x2.c",
353 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700354 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
355 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
356 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
357 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
358 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
359 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
360 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
361 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
362 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
363 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
364 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
365 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
366 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
367 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
368 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
369 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
370 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
371 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700372 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
373 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
374 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800375]
376
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800377PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800378 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800379 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800380 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700381 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
382 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700383 "src/f32-clamp/gen/psimd-x4.c",
384 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700385 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
386 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
387 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
388 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
389 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
390 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
391 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
392 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
393 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
394 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
395 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
396 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700397 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
398 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700399 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
400 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
401 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
402 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
403 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
404 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
405 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
406 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
407 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
408 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
409 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
410 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
411 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
412 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
413 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
414 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
415 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
416 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
417 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800418 "src/f32-hswish/gen/psimd-x4.c",
419 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700420 "src/f32-ibilinear/gen/psimd-c4.c",
421 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
423 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
424 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
425 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
426 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
427 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
428 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
429 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
430 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
431 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700432 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
433 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
434 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700435 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800436 "src/f32-prelu/gen/psimd-2x4.c",
437 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800438 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700439 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
440 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
441 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
442 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
443 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
444 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
445 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
446 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800447 "src/f32-vbinary/gen/vmax-psimd-x4.c",
448 "src/f32-vbinary/gen/vmax-psimd-x8.c",
449 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
450 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
451 "src/f32-vbinary/gen/vmin-psimd-x4.c",
452 "src/f32-vbinary/gen/vmin-psimd-x8.c",
453 "src/f32-vbinary/gen/vminc-psimd-x4.c",
454 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700455 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
456 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
457 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
458 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
459 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
460 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
461 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
462 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
463 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
464 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
465 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
466 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700467 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
468 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700469 "src/x32-packx/x4-psimd.c",
470 "src/x32-pad/x2-psimd.c",
471 "src/x32-unpool/psimd.c",
472 "src/x32-zip/x2-psimd.c",
473 "src/x32-zip/x3-psimd.c",
474 "src/x32-zip/x4-psimd.c",
475 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700476 "src/requantization/precise-psimd.c",
477 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700478]
479
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800480PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800481 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
482 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
483 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
484 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
485 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
486 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
487 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
488 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
489 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
490 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
491 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
492 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800493 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
494 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
495 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
496 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
497 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
498 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
499 "src/math/sigmoid-psimd-p5-div.c",
500]
501
Marat Dukhan08c4a432019-10-03 09:29:21 -0700502# ISA-specific micro-kernels
503NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700504 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
505 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700506 "src/f32-clamp/gen/neon-x4.c",
507 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700508 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
509 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
510 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
511 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700512 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
513 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
514 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
515 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700516 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
517 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
518 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
519 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700520 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700521 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
522 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700523 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
524 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
525 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
526 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
527 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
528 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
529 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
530 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
531 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
532 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
533 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
534 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
535 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
536 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
537 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
538 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
539 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
540 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
541 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
542 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
543 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
544 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
545 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
546 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
547 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
548 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
549 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
550 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
551 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
552 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
553 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800554 "src/f32-hswish/gen/neon-x4.c",
555 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700556 "src/f32-ibilinear/gen/neon-c4.c",
557 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700558 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
559 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
560 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
561 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
562 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
563 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
564 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
565 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
566 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
567 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
568 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
569 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
570 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
571 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
572 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
573 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700574 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
575 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
576 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-ppmm/gen/4x8-minmax-neon.c",
578 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800579 "src/f32-prelu/gen/neon-2x4.c",
580 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800581 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
582 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
583 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
584 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
585 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
586 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
587 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
588 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
589 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
590 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
591 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
592 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
593 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
594 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
595 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
596 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
597 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
598 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
599 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
600 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
601 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
602 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
603 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
604 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700605 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800606 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800607 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
608 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
609 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
610 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
611 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
612 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
613 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
614 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
615 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
616 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
617 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
618 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
619 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
620 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
621 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
622 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
623 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
624 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700625 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
626 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
627 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
628 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800629 "src/f32-vbinary/gen/vmax-neon-x4.c",
630 "src/f32-vbinary/gen/vmax-neon-x8.c",
631 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
632 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
633 "src/f32-vbinary/gen/vmin-neon-x4.c",
634 "src/f32-vbinary/gen/vmin-neon-x8.c",
635 "src/f32-vbinary/gen/vminc-neon-x4.c",
636 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700637 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
638 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
639 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
640 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
641 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
642 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
643 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
644 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
645 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
646 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700647 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
648 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
649 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
650 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700651 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700652 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
653 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700654 "src/q8-gemm/4x8-minmax-neon.c",
655 "src/q8-gemm/8x8-minmax-neon.c",
656 "src/q8-igemm/4x8-minmax-neon.c",
657 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700658 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700659 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700660 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700661 "src/u8-rmax/neon.c",
662 "src/x32-packx/x4-neon-st4.c",
663 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700664 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700665 "src/x32-zip/x2-neon.c",
666 "src/x32-zip/x3-neon.c",
667 "src/x32-zip/x4-neon.c",
668 "src/x32-zip/xm-neon.c",
669 "src/x8-zip/x2-neon.c",
670 "src/x8-zip/x3-neon.c",
671 "src/x8-zip/x4-neon.c",
672 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800673 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800674 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
675 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
676 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
677 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
678 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
679 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700680 "src/requantization/precise-neon.c",
681 "src/requantization/fp32-neon.c",
682 "src/requantization/q31-neon.c",
683 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700684]
685
686NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700687 "src/f32-ibilinear/gen/neonfma-c4.c",
688 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700689 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
690 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
691 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
692 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
693 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
694 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
695 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
696 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
697 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700698 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
699 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
700 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
701 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700702 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
703 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
704 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
705 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700706 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
707 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
708 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
709 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700710 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
711 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
712 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
713 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
714 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
715 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
716 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
717 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
718 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
719 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
720 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
721 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
722 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
723 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
724 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
725 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
726 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
727 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800728 "src/f32-hswish/gen/neonfma-x4.c",
729 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700730 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
731 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800732 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
733 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
734 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
735 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
736 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
737 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
738 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
739 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
740 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
741 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
742 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
743 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
744 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
745 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
746 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
747 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
748 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
749 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
750 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
751 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
752 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
753 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
754 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
755 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800756 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
757 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
758 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
759 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
768 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
769 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
770 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
771 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
772 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
773 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
774 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
775 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
776 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
777 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
778 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
779 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700810 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
811 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800812 "src/math/exp-neonfma-lut64-p2.c",
813 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800814 "src/math/expminus-neonfma-lut2048-p1.c",
815 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800816 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800817 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
818 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
819 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
820 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
821 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
822 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
823 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
824 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
825 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
826 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
827 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
828 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
829 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
830 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
831 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
832 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
833 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
834 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700835]
836
837AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700838 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
839 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
840 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
841 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
842 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
843 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700844 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
845 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
846 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
847 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
848 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
849 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
850 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
851 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
852 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
853 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
854 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
855 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
856 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
857 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
858 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
859 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
860 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
861 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
862 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
863 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700864 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
865 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
866 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
867 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
868 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
869 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
870 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800871 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
872 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
873 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
874 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
875 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
876 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
877 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
878 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
879 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
880 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
881 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
882 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
883 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
884 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
885 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
886 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
887 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
888 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700889 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
890 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
891 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
892 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
893 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
894 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
895 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
896 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
897 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
898 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
899 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
900 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
901 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
902 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
903 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
904 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
905 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
906 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800907 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
908 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
909 "src/math/sigmoid-neonfma-rr1-p5-div.c",
910 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
911 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
912 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700913]
914
915AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard83ea9392020-04-20 20:15:14 -0700916 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
917 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
918 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
919 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
920 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
921 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
922 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
923 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -0700924 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
925 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
926 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
927 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -0700928 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
929 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
930 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
931 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
932 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
933 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
934 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
935 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
936 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
937 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
938 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
939 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700940 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
941 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
942 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
943 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
944 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
945 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
946 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
947 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700948]
949
950SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700951 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
952 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700953 "src/f32-clamp/gen/sse-x4.c",
954 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800955 "src/f32-dwconv-spchw/3x3p1-sse.c",
956 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700957 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
958 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
959 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
960 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
961 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
962 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
963 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
964 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
965 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
966 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
967 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
968 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700969 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700970 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
971 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700972 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
973 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
974 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
975 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
976 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
977 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
978 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
979 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
980 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
981 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
982 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
983 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
984 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800985 "src/f32-hswish/gen/sse-x4.c",
986 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700987 "src/f32-ibilinear/gen/sse-c4.c",
988 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700989 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
990 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
991 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
992 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
993 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
994 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
995 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700996 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
997 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
998 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700999 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001000 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001001 "src/f32-spmm/gen/4x1-minmax-sse.c",
1002 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001003 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1004 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1005 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1006 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1007 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1008 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1009 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1010 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001011 "src/f32-vbinary/gen/vmax-sse-x4.c",
1012 "src/f32-vbinary/gen/vmax-sse-x8.c",
1013 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1014 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1015 "src/f32-vbinary/gen/vmin-sse-x4.c",
1016 "src/f32-vbinary/gen/vmin-sse-x8.c",
1017 "src/f32-vbinary/gen/vminc-sse-x4.c",
1018 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001019 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1020 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1021 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1022 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1023 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1024 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1025 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1026 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1027 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1028 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1029 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1030 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001031 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1032 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001033 "src/x32-packx/x4-sse.c",
1034]
1035
1036SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001037 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1038 "src/f32-argmaxpool/4x-sse2-c4.c",
1039 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001040 "src/f32-prelu/gen/sse2-2x4.c",
1041 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001042 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1043 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1044 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1045 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1046 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1047 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1048 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1049 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1050 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1051 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1052 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1053 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001054 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001055 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001056 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001057 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001058 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1059 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001060 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1061 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001062 "src/q8-igemm/4x4c2-minmax-sse2.c",
1063 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001064 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1065 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001066 "src/q8-gemm/2x4c8-minmax-sse2.c",
1067 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001068 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001069 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001070 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001071 "src/u8-rmax/sse2.c",
1072 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001073 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001074 "src/x32-zip/x2-sse2.c",
1075 "src/x32-zip/x3-sse2.c",
1076 "src/x32-zip/x4-sse2.c",
1077 "src/x32-zip/xm-sse2.c",
1078 "src/x8-zip/x2-sse2.c",
1079 "src/x8-zip/x3-sse2.c",
1080 "src/x8-zip/x4-sse2.c",
1081 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001082 "src/math/exp-sse2-p5.c",
1083 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001084 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001085 "src/requantization/precise-sse2.c",
1086 "src/requantization/fp32-sse2.c",
1087 "src/requantization/q31-sse2.c",
1088 "src/requantization/gemmlowp-sse2.c",
1089]
1090
1091SSSE3_UKERNELS = [
1092 "src/requantization/precise-ssse3.c",
1093 "src/requantization/q31-ssse3.c",
1094 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001095]
1096
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001097SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001098 "src/f32-prelu/gen/sse41-2x4.c",
1099 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001100 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001101 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001102 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001103 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001104 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1105 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001106 "src/requantization/precise-sse4.c",
1107 "src/requantization/q31-sse4.c",
1108 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001109]
1110
Marat Dukhan08c4a432019-10-03 09:29:21 -07001111AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001112 "src/f32-clamp/gen/avx-x8.c",
1113 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001114 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1115 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1116 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1117 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1118 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1119 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1120 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1121 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1122 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1123 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1124 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1125 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1126 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1127 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1128 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1129 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1130 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1131 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1132 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1133 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1134 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1135 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1136 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1137 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1138 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1139 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1140 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1141 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1142 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1143 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001144 "src/f32-hswish/gen/avx-x8.c",
1145 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001146 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1147 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1148 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1149 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1150 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1151 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1152 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1153 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1154 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001155 "src/f32-prelu/gen/avx-2x8.c",
1156 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001157 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001158 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1159 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1160 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1161 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1162 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1163 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1164 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1165 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001166 "src/f32-vbinary/gen/vmax-avx-x8.c",
1167 "src/f32-vbinary/gen/vmax-avx-x16.c",
1168 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1169 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1170 "src/f32-vbinary/gen/vmin-avx-x8.c",
1171 "src/f32-vbinary/gen/vmin-avx-x16.c",
1172 "src/f32-vbinary/gen/vminc-avx-x8.c",
1173 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001174 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1175 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1176 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1177 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1178 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1179 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1180 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1181 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1182 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1183 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1184 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1185 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001186 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001187]
1188
Marat Dukhanfda12b82019-11-21 12:27:59 -08001189FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001190 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1191 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1192 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1193 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1194 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1195 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1196 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1197 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1198 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1199 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1200 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1201 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1202 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1203 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1204 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1205 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1206 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1207 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1208 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1209 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1210 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1211 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1212 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1213 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1214 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1215 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1216 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1217 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1218 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1219 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1220 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1221 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1222 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1223 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1224 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1225 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1226 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1227 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1228 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1229 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001230 "src/f32-hswish/gen/fma3-x8.c",
1231 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001232 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1233 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1234 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1235 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1236 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1237 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1238 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1239 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1240 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1241 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1242 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1243 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1244 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1245 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001246]
1247
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001248AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001249 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1250 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1251 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1252 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1253 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1254 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1255 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1256 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1257 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1258 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1259 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1260 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1261 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1262 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1263 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1264 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1265 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1266 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1267 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1268 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1269 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1270 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1271 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1272 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1273 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1274 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1275 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1276 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1277 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1278 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1279 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1280 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1281 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1282 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1283 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1284 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001285 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1286 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1287 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1288 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1289 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1290 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1291 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1292 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1293 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1294 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1295 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1296 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1297 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1298 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1299 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1300 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1301 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1302 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1303 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1304 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1305 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1306 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1307 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1308 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1309 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1310 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1311 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1312 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1313 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1314 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001315 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1316 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1317 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1318 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1319 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1320 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1321 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1322 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1323 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1324 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1325 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1326 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1327 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1328 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1329 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1330 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1331 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1332 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1333 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1334 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1335 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1336 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1337 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1338 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001339 "src/math/exp-avx2-p5.c",
1340 "src/math/exp-avx2-perm-p3.c",
1341 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001342 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001343 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001344 "src/math/sigmoid-avx2-rr2-p5-div.c",
1345 "src/math/sigmoid-avx2-rr1-p5-div.c",
1346 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1347 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1348 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1349 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001350]
1351
Marat Dukhan08c4a432019-10-03 09:29:21 -07001352AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001353 "src/f32-clamp/gen/avx512f-x16.c",
1354 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001355 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1356 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1357 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1358 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1359 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1360 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1361 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1362 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1363 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1364 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1365 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1366 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1367 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1368 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1369 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1370 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1371 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1372 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1373 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1374 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1375 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1376 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1377 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1378 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001379 "src/f32-hswish/gen/avx512f-x16.c",
1380 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001381 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1382 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1383 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1384 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1385 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1386 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001387 "src/f32-prelu/gen/avx512f-2x16.c",
1388 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001389 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1390 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1391 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1392 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1393 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1394 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1395 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1396 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1397 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1398 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1399 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1400 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1401 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1402 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1403 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1404 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1405 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1406 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1407 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1408 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1409 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1410 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1411 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1412 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1413 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1414 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1415 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1416 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1417 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1418 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1419 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1420 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1421 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1422 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1423 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1424 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001425 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001426 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1427 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1428 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1429 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1430 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1431 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1432 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1433 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001434 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1435 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1436 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1437 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1438 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1439 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1440 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1441 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001442 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1443 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1444 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1445 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1446 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1447 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1448 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1449 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1450 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1451 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1452 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1453 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001454 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001455 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1456 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1457 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1458 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1459 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1460 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1461 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1462 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1463 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1464 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1465 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1466 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1467 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1468 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1469 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1470 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1471 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1472 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1473 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1474 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1475 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1476 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1477 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1478 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001479 "src/math/exp-avx512f-p5-scalef.c",
1480 "src/math/exp-avx512f-p5.c",
1481 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001482 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001483 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001484]
1485
1486AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001487 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1489 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1490 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1491 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1492 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1493 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1494 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1495 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1496 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1497 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001498]
1499
1500AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001501 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001502 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1503 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001504 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001505 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1506 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001507 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1508 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1509 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001510 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001511 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1512 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1513 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001514 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001515 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1516 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001517 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001518 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1519 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1520 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1521 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1522 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1523 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1524 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1525 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1526 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1527 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1528 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1529 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1530 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1531 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1532 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1533 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1534 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1535 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1536 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1537 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1538 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001539 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001540 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1541 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1542 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1543 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1544 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1545 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1546 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1547 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1548 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1549 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1550 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1551 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1552 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1553 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1554 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1555 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1556 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1557 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1558 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1559 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1560 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1561 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1562 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1563 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1564 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1565 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1566 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1567 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1568 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1569 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1570 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1571 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1572 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1573 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1574 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1575 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1576 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1577 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001578]
1579
Marat Dukhan1b354632020-03-23 12:50:22 -07001580INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001581 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/xnnpack/argmaxpool.h",
1583 "src/xnnpack/avgpool.h",
1584 "src/xnnpack/clamp.h",
1585 "src/xnnpack/common.h",
1586 "src/xnnpack/conv.h",
1587 "src/xnnpack/dwconv.h",
1588 "src/xnnpack/gavgpool.h",
1589 "src/xnnpack/gemm.h",
1590 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001591 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001592 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001593 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001594 "src/xnnpack/lut.h",
1595 "src/xnnpack/math.h",
1596 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001597 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001598 "src/xnnpack/packx.h",
1599 "src/xnnpack/pad.h",
1600 "src/xnnpack/params.h",
1601 "src/xnnpack/pavgpool.h",
1602 "src/xnnpack/ppmm.h",
1603 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001604 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001605 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001606 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001607 "src/xnnpack/rmax.h",
1608 "src/xnnpack/scalar-utils.h",
1609 "src/xnnpack/spmm.h",
1610 "src/xnnpack/unpool.h",
1611 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001612 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001613 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001614 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001615 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001616 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001617 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001618 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001619]
1620
1621INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001622 "include/xnnpack.h",
1623 "src/xnnpack/allocator.h",
1624 "src/xnnpack/compute.h",
1625 "src/xnnpack/im2col.h",
1626 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001627 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001628 "src/xnnpack/operator.h",
1629 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001630 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001631 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001632 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001633 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001634]
1635
Marat Dukhan1b354632020-03-23 12:50:22 -07001636ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001637 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001638]
1639
Marat Dukhan1b354632020-03-23 12:50:22 -07001640MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001641 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001642 "include/xnnpack.h",
1643]
1644
Marat Dukhan1b354632020-03-23 12:50:22 -07001645MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001646 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001647 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001648 "src/xnnpack/requantization.h",
1649 "include/xnnpack.h",
1650]
1651
1652OPERATOR_TEST_PARAMS_HDRS = [
1653 "src/xnnpack/params.h",
1654 "src/xnnpack/common.h",
1655]
1656
1657WEIGHTS_PACK_HDRS = [
1658 "src/xnnpack/pack.h",
1659 "src/xnnpack/operator.h",
1660 "src/xnnpack/compute.h",
1661]
1662
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001663LOGGING_COPTS = select({
1664 # No logging in optimized mode
1665 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1666 # Full logging in debug mode
1667 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1668 # Error-only logging in default (fastbuild) mode
1669 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1670})
1671
1672LOGGING_HDRS = [
1673 "src/xnnpack/log.h",
1674]
1675
Marat Dukhan08c4a432019-10-03 09:29:21 -07001676xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001677 name = "tables",
1678 srcs = TABLE_SRCS,
1679 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001680 gcc_copts = xnnpack_gcc_std_copts(),
1681 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001682)
1683
1684xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001685 name = "scalar_ukernels",
1686 srcs = SCALAR_UKERNELS,
1687 hdrs = INTERNAL_HDRS,
1688 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001689 gcc_copts = xnnpack_gcc_std_copts(),
1690 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001691 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001692 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693 "@FP16",
1694 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001695 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001696 ],
1697)
1698
1699xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001700 name = "wasm_ukernels",
1701 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001702 gcc_copts = xnnpack_gcc_std_copts(),
1703 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001704 wasm_srcs = WASM_UKERNELS,
1705 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001706 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001707 "@FP16",
1708 "@FXdiv",
1709 "@pthreadpool",
1710 ],
1711)
1712
1713xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001714 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001715 hdrs = INTERNAL_HDRS,
1716 aarch32_copts = [
1717 "-marm",
1718 "-mfpu=neon",
1719 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001720 gcc_copts = xnnpack_gcc_std_copts(),
1721 gcc_x86_copts = ["-msse2"],
1722 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723 optimized_copts = [
1724 "-O3",
1725 "-ffast-math",
1726 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001727 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001728 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001729 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730 "@FP16",
1731 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001732 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001733 ],
1734)
1735
1736xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001737 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001738 hdrs = INTERNAL_HDRS,
1739 aarch32_copts = [
1740 "-marm",
1741 "-mfpu=neon",
1742 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001743 gcc_copts = xnnpack_gcc_std_copts(),
1744 gcc_x86_copts = ["-msse2"],
1745 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001746 optimized_copts = [
1747 "-O3",
1748 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001749 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001750 deps = [
1751 ":tables",
1752 "@FP16",
1753 "@psimd",
1754 "@pthreadpool",
1755 ],
1756)
1757
1758xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001759 name = "neon_ukernels",
1760 hdrs = INTERNAL_HDRS,
1761 aarch32_copts = [
1762 "-marm",
1763 "-mfpu=neon",
1764 ],
1765 aarch32_srcs = NEON_UKERNELS,
1766 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001767 gcc_copts = xnnpack_gcc_std_copts(),
1768 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001769 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001770 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001771 "@FP16",
1772 "@pthreadpool",
1773 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001774)
1775
1776xnnpack_cc_library(
1777 name = "neonfma_ukernels",
1778 hdrs = INTERNAL_HDRS,
1779 aarch32_copts = [
1780 "-marm",
1781 "-mfpu=neon-vfpv4",
1782 ],
1783 aarch32_srcs = NEONFMA_UKERNELS,
1784 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001785 gcc_copts = xnnpack_gcc_std_copts(),
1786 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001787 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001788 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001789 "@FP16",
1790 "@pthreadpool",
1791 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001792)
1793
1794xnnpack_cc_library(
1795 name = "neonfp16arith_ukernels",
1796 hdrs = INTERNAL_HDRS,
1797 aarch64_copts = ["-march=armv8.2-a+fp16"],
1798 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001799 gcc_copts = xnnpack_gcc_std_copts(),
1800 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001801 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001802 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001803 "@FP16",
1804 "@pthreadpool",
1805 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001806)
1807
1808xnnpack_cc_library(
1809 name = "sse2_ukernels",
1810 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001811 gcc_copts = xnnpack_gcc_std_copts(),
1812 gcc_x86_copts = ["-msse2"],
1813 msvc_copts = xnnpack_msvc_std_copts(),
1814 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001815 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001816 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001817 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001818 "@FP16",
1819 "@pthreadpool",
1820 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001821)
1822
1823xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001824 name = "ssse3_ukernels",
1825 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001826 gcc_copts = xnnpack_gcc_std_copts(),
1827 gcc_x86_copts = ["-mssse3"],
1828 msvc_copts = xnnpack_msvc_std_copts(),
1829 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001830 x86_srcs = SSSE3_UKERNELS,
1831 deps = [
1832 ":tables",
1833 "@FP16",
1834 "@pthreadpool",
1835 ],
1836)
1837
1838xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001839 name = "sse41_ukernels",
1840 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001841 gcc_copts = xnnpack_gcc_std_copts(),
1842 gcc_x86_copts = ["-msse4.1"],
1843 msvc_copts = xnnpack_msvc_std_copts(),
1844 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001845 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001846 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001847 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001848 "@FP16",
1849 "@pthreadpool",
1850 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001851)
1852
1853xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001854 name = "avx_ukernels",
1855 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001856 gcc_copts = xnnpack_gcc_std_copts(),
1857 gcc_x86_copts = ["-mavx"],
1858 msvc_copts = xnnpack_msvc_std_copts(),
1859 msvc_x86_32_copts = ["/arch:AVX"],
1860 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001861 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001862 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001863 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001864 "@FP16",
1865 "@pthreadpool",
1866 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867)
1868
1869xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001870 name = "fma3_ukernels",
1871 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001872 gcc_copts = xnnpack_gcc_std_copts(),
1873 gcc_x86_copts = ["-mfma"],
1874 msvc_copts = xnnpack_msvc_std_copts(),
1875 msvc_x86_32_copts = ["/arch:AVX"],
1876 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001877 x86_srcs = FMA3_UKERNELS,
1878 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001879 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001880 "@FP16",
1881 "@pthreadpool",
1882 ],
1883)
1884
1885xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001886 name = "avx2_ukernels",
1887 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001888 gcc_copts = xnnpack_gcc_std_copts(),
1889 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001890 "-mfma",
1891 "-mavx2",
1892 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001893 msvc_copts = xnnpack_msvc_std_copts(),
1894 msvc_x86_32_copts = ["/arch:AVX2"],
1895 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001896 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001897 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001898 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001899 "@FP16",
1900 "@pthreadpool",
1901 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001902)
1903
1904xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001905 name = "avx512f_ukernels",
1906 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001907 gcc_copts = xnnpack_gcc_std_copts(),
1908 gcc_x86_copts = ["-mavx512f"],
1909 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1910 msvc_copts = xnnpack_msvc_std_copts(),
1911 msvc_x86_32_copts = ["/arch:AVX512"],
1912 msvc_x86_64_copts = ["/arch:AVX512"],
1913 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001914 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001915 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001916 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001917 "@FP16",
1918 "@pthreadpool",
1919 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001920)
1921
1922xnnpack_cc_library(
1923 name = "asm_ukernels",
1924 hdrs = ["src/xnnpack/assembly.h"],
1925 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001926 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001927 aarch64_srcs = AARCH64_ASM_UKERNELS,
1928)
1929
1930xnnpack_aggregate_library(
1931 name = "ukernels",
1932 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001933 ":neon_ukernels",
1934 ":neonfma_ukernels",
1935 ":asm_ukernels",
1936 ],
1937 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001938 ":neon_ukernels",
1939 ":neonfma_ukernels",
1940 ":neonfp16arith_ukernels",
1941 ":asm_ukernels",
1942 ],
1943 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07001944 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001945 ":psimd_fastmath_ukernels",
1946 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001947 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001948 wasm_deps = [":wasm_ukernels"],
1949 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001950 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001951 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001952 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001953 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001954 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001955 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001956 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001957 ":avx512f_ukernels",
1958 ],
1959)
1960
1961xnnpack_cc_library(
1962 name = "im2col",
1963 srcs = ["src/im2col.c"],
1964 hdrs = [
1965 "src/xnnpack/common.h",
1966 "src/xnnpack/im2col.h",
1967 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001968 gcc_copts = xnnpack_gcc_std_copts(),
1969 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001970)
1971
1972xnnpack_cc_library(
1973 name = "indirection",
1974 srcs = ["src/indirection.c"],
1975 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001976 gcc_copts = xnnpack_gcc_std_copts(),
1977 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001978 deps = [
1979 "@FP16",
1980 "@FXdiv",
1981 "@pthreadpool",
1982 ],
1983)
1984
1985xnnpack_cc_library(
1986 name = "operator_run",
1987 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001988 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001989 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07001990 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1991 "//conditions:default": [],
1992 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07001993 gcc_copts = xnnpack_gcc_std_copts(),
1994 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001995 deps = [
1996 "@FP16",
1997 "@FXdiv",
1998 "@clog",
1999 "@pthreadpool",
2000 ],
2001)
2002
2003cc_library(
2004 name = "enable_assembly",
2005 defines = select({
2006 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2007 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002008 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002009 }),
2010)
2011
Marat Dukhancf056b22019-10-07 10:26:29 -07002012xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002013 name = "operators",
2014 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002015 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002016 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002017 ],
2018 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002019 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002020 "-Isrc",
2021 "-Iinclude",
2022 ] + select({
2023 ":debug_build": [],
2024 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002025 }) + select({
2026 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2027 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002028 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002029 gcc_copts = xnnpack_gcc_std_copts(),
2030 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002031 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002032 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002033 "@FP16",
2034 "@FXdiv",
2035 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002036 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002037 ],
2038)
2039
Marat Dukhan10a38082020-04-17 03:58:35 -07002040xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002041 name = "XNNPACK",
2042 srcs = [
2043 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002044 "src/runtime.c",
2045 "src/subgraph.c",
2046 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002047 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002048 hdrs = ["include/xnnpack.h"],
2049 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002050 "-Isrc",
2051 "-Iinclude",
2052 ] + select({
2053 ":debug_build": [],
2054 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002055 }) + select({
2056 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2057 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002058 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002059 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002060 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002061 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002062 visibility = xnnpack_visibility(),
2063 deps = [
2064 ":enable_assembly",
2065 ":ukernels",
2066 ":operator_run",
2067 ":operators",
2068 "@clog",
2069 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002070 ] + select({
2071 ":emscripten": [],
2072 "//conditions:default": ["@cpuinfo"],
2073 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002074)
2075
Marat Dukhan10a38082020-04-17 03:58:35 -07002076xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002077 name = "xnnpack_operators_nhwc_f32",
2078 srcs = [
2079 "src/init.c",
2080 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002081 hdrs = ["include/xnnpack.h"],
2082 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002083 "-Isrc",
2084 "-Iinclude",
2085 ] + select({
2086 ":debug_build": [],
2087 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002088 }) + select({
2089 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2090 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002091 }),
2092 defines = [
2093 "XNN_NO_Q8_OPERATORS",
2094 "XNN_NO_U8_OPERATORS",
2095 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002096 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002097 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002098 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002099 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002100 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002101 visibility = xnnpack_visibility(),
2102 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002103 ":enable_assembly",
2104 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002105 ":operator_run",
2106 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002107 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002108 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002109 ] + select({
2110 ":emscripten": [],
2111 "//conditions:default": ["@cpuinfo"],
2112 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002113)
2114
Marat Dukhancf056b22019-10-07 10:26:29 -07002115xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002116 name = "bench_utils",
2117 srcs = ["bench/utils.cc"],
2118 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002119 deps = [
2120 "@com_google_benchmark//:benchmark",
2121 "@cpuinfo",
2122 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002123)
2124
Frank Barchard7e955972019-10-11 10:34:25 -07002125######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002126
2127xnnpack_benchmark(
2128 name = "q8_gemm_bench",
2129 srcs = [
2130 "bench/gemm.h",
2131 "bench/q8-gemm.cc",
2132 "src/xnnpack/AlignedAllocator.h",
2133 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002134 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002135 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2136)
2137
2138xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002139 name = "f16_igemm_bench",
2140 srcs = [
2141 "bench/f16-igemm.cc",
2142 "bench/conv.h",
2143 "bench/google/conv.h",
2144 "src/xnnpack/AlignedAllocator.h",
2145 ] + MICROKERNEL_BENCHMARK_HDRS,
2146 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2147)
2148
2149xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002150 name = "f16_gemm_bench",
2151 srcs = [
2152 "bench/f16-gemm.cc",
2153 "bench/gemm.h",
2154 "src/xnnpack/AlignedAllocator.h",
2155 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002156 deps = MICROKERNEL_BENCHMARK_DEPS,
2157)
2158
2159xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002160 name = "f16_spmm_bench",
2161 srcs = [
2162 "bench/f16-spmm.cc",
2163 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002164 "src/xnnpack/AlignedAllocator.h",
2165 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002166 deps = MICROKERNEL_BENCHMARK_DEPS,
2167)
2168
2169xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002170 name = "f32_igemm_bench",
2171 srcs = [
2172 "bench/f32-igemm.cc",
2173 "bench/conv.h",
2174 "src/xnnpack/AlignedAllocator.h",
2175 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002176 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002177)
2178
2179xnnpack_benchmark(
2180 name = "f32_conv_hwc_bench",
2181 srcs = [
2182 "bench/f32-conv-hwc.cc",
2183 "bench/dconv.h",
2184 "src/xnnpack/AlignedAllocator.h",
2185 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002186 deps = MICROKERNEL_BENCHMARK_DEPS,
2187)
2188
2189xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002190 name = "f32_conv_hwc2spchw_bench",
2191 srcs = [
2192 "bench/f32-conv-hwc2spchw.cc",
2193 "bench/dconv.h",
2194 "src/xnnpack/AlignedAllocator.h",
2195 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002196 deps = MICROKERNEL_BENCHMARK_DEPS,
2197)
2198
2199xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002200 name = "f32_dwconv_bench",
2201 srcs = [
2202 "bench/f32-dwconv.cc",
2203 "bench/dwconv.h",
2204 "src/xnnpack/AlignedAllocator.h",
2205 ] + MICROKERNEL_BENCHMARK_HDRS,
2206 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2207)
2208
2209xnnpack_benchmark(
2210 name = "f32_dwconv_spchw_bench",
2211 srcs = [
2212 "bench/f32-dwconv-spchw.cc",
2213 "bench/dwconv.h",
2214 "src/xnnpack/AlignedAllocator.h",
2215 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2216 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2217)
2218
2219xnnpack_benchmark(
2220 name = "f32_gemm_bench",
2221 srcs = [
2222 "bench/f32-gemm.cc",
2223 "bench/gemm.h",
2224 "src/xnnpack/AlignedAllocator.h",
2225 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002226 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002227 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002228)
2229
2230xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002231 name = "f32_raddexpminusmax_bench",
2232 srcs = [
2233 "bench/f32-raddexpminusmax.cc",
2234 "src/xnnpack/AlignedAllocator.h",
2235 ] + MICROKERNEL_BENCHMARK_HDRS,
2236 deps = MICROKERNEL_BENCHMARK_DEPS,
2237)
2238
2239xnnpack_benchmark(
2240 name = "f32_raddextexp_bench",
2241 srcs = [
2242 "bench/f32-raddextexp.cc",
2243 "src/xnnpack/AlignedAllocator.h",
2244 ] + MICROKERNEL_BENCHMARK_HDRS,
2245 deps = MICROKERNEL_BENCHMARK_DEPS,
2246)
2247
2248xnnpack_benchmark(
2249 name = "f32_raddstoreexpminusmax_bench",
2250 srcs = [
2251 "bench/f32-raddstoreexpminusmax.cc",
2252 "src/xnnpack/AlignedAllocator.h",
2253 ] + MICROKERNEL_BENCHMARK_HDRS,
2254 deps = MICROKERNEL_BENCHMARK_DEPS,
2255)
2256
2257xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002258 name = "f32_rmax_bench",
2259 srcs = [
2260 "bench/f32-rmax.cc",
2261 "src/xnnpack/AlignedAllocator.h",
2262 ] + MICROKERNEL_BENCHMARK_HDRS,
2263 deps = MICROKERNEL_BENCHMARK_DEPS,
2264)
2265
2266xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002267 name = "f32_sigmoid_bench",
2268 srcs = [
2269 "bench/f32-sigmoid.cc",
2270 "src/xnnpack/AlignedAllocator.h",
2271 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002272 deps = MICROKERNEL_BENCHMARK_DEPS,
2273)
2274
2275xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002276 name = "f32_spmm_bench",
2277 srcs = [
2278 "bench/f32-spmm.cc",
2279 "bench/gemm.h",
2280 "src/xnnpack/AlignedAllocator.h",
2281 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002282 deps = MICROKERNEL_BENCHMARK_DEPS,
2283)
2284
2285xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002286 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002287 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002288 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002289 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002290 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002291 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002292)
2293
2294xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002295 name = "f32_vscaleexpminusmax_bench",
2296 srcs = [
2297 "bench/f32-vscaleexpminusmax.cc",
2298 "src/xnnpack/AlignedAllocator.h",
2299 ] + MICROKERNEL_BENCHMARK_HDRS,
2300 deps = MICROKERNEL_BENCHMARK_DEPS,
2301)
2302
2303xnnpack_benchmark(
2304 name = "f32_vscaleextexp_bench",
2305 srcs = [
2306 "bench/f32-vscaleextexp.cc",
2307 "src/xnnpack/AlignedAllocator.h",
2308 ] + MICROKERNEL_BENCHMARK_HDRS,
2309 deps = MICROKERNEL_BENCHMARK_DEPS,
2310)
2311
2312xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002313 name = "f32_im2col_gemm_bench",
2314 srcs = [
2315 "bench/f32-im2col-gemm.cc",
2316 "bench/conv.h",
2317 "src/xnnpack/AlignedAllocator.h",
2318 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2319 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2320)
2321
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002322xnnpack_benchmark(
2323 name = "requantization_bench",
2324 srcs = [
2325 "bench/requantization.cc",
2326 "src/xnnpack/requantization-stubs.h",
2327 "src/xnnpack/AlignedAllocator.h",
2328 ] + MICROKERNEL_BENCHMARK_HDRS,
2329 deps = MICROKERNEL_BENCHMARK_DEPS,
2330)
2331
Marat Dukhan08c4a432019-10-03 09:29:21 -07002332########################### Benchmarks for operators ###########################
2333
2334xnnpack_benchmark(
2335 name = "add_bench",
2336 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002337 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002338)
2339
2340xnnpack_benchmark(
2341 name = "average_pooling_bench",
2342 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002343 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002344 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002345 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002346)
2347
2348xnnpack_benchmark(
2349 name = "channel_shuffle_bench",
2350 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002351 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352)
2353
2354xnnpack_benchmark(
2355 name = "convolution_bench",
2356 srcs = ["bench/convolution.cc"],
2357 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002358 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002359 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002360)
2361
2362xnnpack_benchmark(
2363 name = "deconvolution_bench",
2364 srcs = ["bench/deconvolution.cc"],
2365 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002366 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002367 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002368)
2369
2370xnnpack_benchmark(
2371 name = "global_average_pooling_bench",
2372 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002373 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002374)
2375
2376xnnpack_benchmark(
2377 name = "max_pooling_bench",
2378 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002379 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002380)
2381
2382xnnpack_benchmark(
2383 name = "sigmoid_bench",
2384 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002385 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002386 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002387 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002388)
2389
2390xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002391 name = "prelu_bench",
2392 srcs = ["bench/prelu.cc"],
2393 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002394 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002395 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002396)
2397
2398xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002399 name = "softmax_bench",
2400 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002401 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002402 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002403 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404)
2405
Marat Dukhanc068bb62019-10-04 13:24:39 -07002406############################# End-to-end benchmarks ############################
2407
2408cc_library(
2409 name = "mobilenet_v1",
2410 srcs = ["models/mobilenet-v1.cc"],
2411 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002412 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002413 linkstatic = True,
2414 deps = [
2415 ":XNNPACK",
2416 "@pthreadpool",
2417 ],
2418)
2419
2420cc_library(
2421 name = "mobilenet_v2",
2422 srcs = ["models/mobilenet-v2.cc"],
2423 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002424 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002425 linkstatic = True,
2426 deps = [
2427 ":XNNPACK",
2428 "@pthreadpool",
2429 ],
2430)
2431
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002432cc_library(
2433 name = "mobilenet_v3_large",
2434 srcs = ["models/mobilenet-v3-large.cc"],
2435 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002436 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002437 linkstatic = True,
2438 deps = [
2439 ":XNNPACK",
2440 "@pthreadpool",
2441 ],
2442)
2443
2444cc_library(
2445 name = "mobilenet_v3_small",
2446 srcs = ["models/mobilenet-v3-small.cc"],
2447 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002448 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002449 linkstatic = True,
2450 deps = [
2451 ":XNNPACK",
2452 "@pthreadpool",
2453 ],
2454)
2455
Marat Dukhanc068bb62019-10-04 13:24:39 -07002456xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002457 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002458 srcs = [
2459 "bench/f32-dwconv-e2e.cc",
2460 "bench/end2end.h",
2461 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002462 deps = MICROKERNEL_BENCHMARK_DEPS + [
2463 ":XNNPACK",
2464 ":mobilenet_v1",
2465 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002466 ":mobilenet_v3_large",
2467 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002468 ],
2469)
2470
2471xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002472 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002473 srcs = [
2474 "bench/f32-gemm-e2e.cc",
2475 "bench/end2end.h",
2476 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002477 deps = MICROKERNEL_BENCHMARK_DEPS + [
2478 ":XNNPACK",
2479 ":mobilenet_v1",
2480 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002481 ":mobilenet_v3_large",
2482 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002483 ],
2484)
2485
2486xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002487 name = "end2end_bench",
2488 srcs = ["bench/end2end.cc"],
2489 deps = [
2490 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002491 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002492 ":mobilenet_v1",
2493 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002494 ":mobilenet_v3_large",
2495 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002496 "@pthreadpool",
2497 ],
2498)
2499
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002500#################### Accuracy evaluation for math functions ####################
2501
2502xnnpack_benchmark(
2503 name = "f32_exp_eval",
2504 srcs = [
2505 "eval/f32-exp.cc",
2506 "src/xnnpack/AlignedAllocator.h",
2507 ] + ACCURACY_EVAL_HDRS,
2508 deps = ACCURACY_EVAL_DEPS,
2509)
2510
Marat Dukhan515c9772019-10-17 18:07:57 -07002511xnnpack_benchmark(
2512 name = "f32_expminus_eval",
2513 srcs = [
2514 "eval/f32-expminus.cc",
2515 "src/xnnpack/AlignedAllocator.h",
2516 ] + ACCURACY_EVAL_HDRS,
2517 deps = ACCURACY_EVAL_DEPS,
2518)
2519
Marat Dukhan98ba4412019-10-23 02:14:28 -07002520xnnpack_benchmark(
2521 name = "f32_extexp_eval",
2522 srcs = [
2523 "eval/f32-extexp.cc",
2524 "src/xnnpack/AlignedAllocator.h",
2525 ] + ACCURACY_EVAL_HDRS,
2526 deps = ACCURACY_EVAL_DEPS,
2527)
2528
Marat Dukhan346a9e52019-11-15 09:06:30 -08002529xnnpack_benchmark(
2530 name = "f32_sigmoid_eval",
2531 srcs = [
2532 "eval/f32-sigmoid.cc",
2533 "src/xnnpack/AlignedAllocator.h",
2534 ] + ACCURACY_EVAL_HDRS,
2535 deps = ACCURACY_EVAL_DEPS,
2536)
2537
Marat Dukhan08c4a432019-10-03 09:29:21 -07002538######################### Unit tests for micro-kernels #########################
2539
2540xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002541 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002543 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002544 "test/gemm-microkernel-tester.h",
2545 "src/xnnpack/AlignedAllocator.h",
2546 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2547 deps = MICROKERNEL_TEST_DEPS,
2548)
2549
2550xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002551 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002552 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002553 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002554 "test/spmm-microkernel-tester.h",
2555 "src/xnnpack/AlignedAllocator.h",
2556 ] + MICROKERNEL_TEST_HDRS,
2557 deps = MICROKERNEL_TEST_DEPS,
2558)
2559
2560xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002561 name = "f32_argmaxpool_test",
2562 srcs = [
2563 "test/f32-argmaxpool.cc",
2564 "test/argmaxpool-microkernel-tester.h",
2565 "src/xnnpack/AlignedAllocator.h",
2566 ] + MICROKERNEL_TEST_HDRS,
2567 deps = MICROKERNEL_TEST_DEPS,
2568)
2569
2570xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002571 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002572 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002573 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002574 "test/avgpool-microkernel-tester.h",
2575 "src/xnnpack/AlignedAllocator.h",
2576 ] + MICROKERNEL_TEST_HDRS,
2577 deps = MICROKERNEL_TEST_DEPS,
2578)
2579
2580xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002581 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002582 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002583 "test/f32-ibilinear.cc",
2584 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002585 "src/xnnpack/AlignedAllocator.h",
2586 ] + MICROKERNEL_TEST_HDRS,
2587 deps = MICROKERNEL_TEST_DEPS,
2588)
2589
2590xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002591 name = "f32_clamp_test",
2592 srcs = [
2593 "test/f32-clamp.cc",
2594 "test/clamp-microkernel-tester.h",
2595 ] + MICROKERNEL_TEST_HDRS,
2596 deps = MICROKERNEL_TEST_DEPS,
2597)
2598
2599xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002600 name = "f32_igemm_test",
2601 srcs = [
2602 "test/f32-igemm.cc",
2603 "test/gemm-microkernel-tester.h",
2604 "src/xnnpack/AlignedAllocator.h",
2605 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2606 deps = MICROKERNEL_TEST_DEPS,
2607)
2608
2609xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002610 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002611 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002612 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002613 "test/gemm-microkernel-tester.h",
2614 "src/xnnpack/AlignedAllocator.h",
2615 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002616 # Work-around for "too many sections" error
2617 mingw_copts = ["-Wa,-mbig-obj"],
2618 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002619 deps = MICROKERNEL_TEST_DEPS,
2620)
2621
2622xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002623 name = "f16_igemm_minmax_test",
2624 srcs = [
2625 "test/f16-igemm-minmax.cc",
2626 "test/gemm-microkernel-tester.h",
2627 "src/xnnpack/AlignedAllocator.h",
2628 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2629 deps = MICROKERNEL_TEST_DEPS,
2630)
2631
2632xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002633 name = "f32_conv_hwc_test",
2634 srcs = [
2635 "test/f32-conv-hwc.cc",
2636 "test/conv-hwc-microkernel-tester.h",
2637 "src/xnnpack/AlignedAllocator.h",
2638 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2639 deps = MICROKERNEL_TEST_DEPS,
2640)
2641
2642xnnpack_unit_test(
2643 name = "f32_conv_hwc2spchw_test",
2644 srcs = [
2645 "test/f32-conv-hwc2spchw.cc",
2646 "test/conv-hwc2spchw-microkernel-tester.h",
2647 "src/xnnpack/AlignedAllocator.h",
2648 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2649 deps = MICROKERNEL_TEST_DEPS,
2650)
2651
2652xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002653 name = "f32_dwconv_test",
2654 srcs = [
2655 "test/f32-dwconv.cc",
2656 "test/dwconv-microkernel-tester.h",
2657 "src/xnnpack/AlignedAllocator.h",
2658 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2659 deps = MICROKERNEL_TEST_DEPS,
2660)
2661
2662xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002663 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002664 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002665 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002666 "test/dwconv-microkernel-tester.h",
2667 "src/xnnpack/AlignedAllocator.h",
2668 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2669 deps = MICROKERNEL_TEST_DEPS,
2670)
2671
2672xnnpack_unit_test(
2673 name = "f32_dwconv_spchw_test",
2674 srcs = [
2675 "test/f32-dwconv-spchw.cc",
2676 "test/dwconv-spchw-microkernel-tester.h",
2677 "src/xnnpack/AlignedAllocator.h",
2678 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2679 deps = MICROKERNEL_TEST_DEPS,
2680)
2681
2682xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002683 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002684 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002685 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686 "test/gavgpool-microkernel-tester.h",
2687 "src/xnnpack/AlignedAllocator.h",
2688 ] + MICROKERNEL_TEST_HDRS,
2689 deps = MICROKERNEL_TEST_DEPS,
2690)
2691
2692xnnpack_unit_test(
2693 name = "f32_gavgpool_spchw_test",
2694 srcs = [
2695 "test/f32-gavgpool-spchw.cc",
2696 "test/gavgpool-spchw-microkernel-tester.h",
2697 "src/xnnpack/AlignedAllocator.h",
2698 ] + MICROKERNEL_TEST_HDRS,
2699 deps = MICROKERNEL_TEST_DEPS,
2700)
2701
2702xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002703 name = "f32_gemm_test",
2704 srcs = [
2705 "test/f32-gemm.cc",
2706 "test/gemm-microkernel-tester.h",
2707 "src/xnnpack/AlignedAllocator.h",
2708 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2709 deps = MICROKERNEL_TEST_DEPS,
2710)
2711
2712xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002713 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002714 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002715 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002716 "test/gemm-microkernel-tester.h",
2717 "src/xnnpack/AlignedAllocator.h",
2718 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2719 deps = MICROKERNEL_TEST_DEPS,
2720)
2721
2722xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002723 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002724 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002725 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002726 "test/gemm-microkernel-tester.h",
2727 "src/xnnpack/AlignedAllocator.h",
2728 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2729 deps = MICROKERNEL_TEST_DEPS,
2730)
2731
2732xnnpack_unit_test(
2733 name = "f32_hswish_test",
2734 srcs = [
2735 "test/f32-hswish.cc",
2736 "test/hswish-microkernel-tester.h",
2737 ] + MICROKERNEL_TEST_HDRS,
2738 deps = MICROKERNEL_TEST_DEPS,
2739)
2740
2741xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002742 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002743 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002744 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002745 "test/maxpool-microkernel-tester.h",
2746 ] + MICROKERNEL_TEST_HDRS,
2747 deps = MICROKERNEL_TEST_DEPS,
2748)
2749
2750xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002751 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002752 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002753 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002754 "test/avgpool-microkernel-tester.h",
2755 "src/xnnpack/AlignedAllocator.h",
2756 ] + MICROKERNEL_TEST_HDRS,
2757 deps = MICROKERNEL_TEST_DEPS,
2758)
2759
2760xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002761 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002762 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002763 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002764 "test/gemm-microkernel-tester.h",
2765 "src/xnnpack/AlignedAllocator.h",
2766 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2767 deps = MICROKERNEL_TEST_DEPS,
2768)
2769
2770xnnpack_unit_test(
2771 name = "f32_prelu_test",
2772 srcs = [
2773 "test/f32-prelu.cc",
2774 "test/prelu-microkernel-tester.h",
2775 "src/xnnpack/AlignedAllocator.h",
2776 ] + MICROKERNEL_TEST_HDRS,
2777 deps = MICROKERNEL_TEST_DEPS,
2778)
2779
2780xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002781 name = "f32_raddexpminusmax_test",
2782 srcs = [
2783 "test/f32-raddexpminusmax.cc",
2784 "test/raddexpminusmax-microkernel-tester.h",
2785 ] + MICROKERNEL_TEST_HDRS,
2786 deps = MICROKERNEL_TEST_DEPS,
2787)
2788
2789xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002790 name = "f32_raddextexp_test",
2791 srcs = [
2792 "test/f32-raddextexp.cc",
2793 "test/raddextexp-microkernel-tester.h",
2794 ] + MICROKERNEL_TEST_HDRS,
2795 deps = MICROKERNEL_TEST_DEPS,
2796)
2797
2798xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002799 name = "f32_raddstoreexpminusmax_test",
2800 srcs = [
2801 "test/f32-raddstoreexpminusmax.cc",
2802 "test/raddstoreexpminusmax-microkernel-tester.h",
2803 ] + MICROKERNEL_TEST_HDRS,
2804 deps = MICROKERNEL_TEST_DEPS,
2805)
2806
2807xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002808 name = "f32_rmax_test",
2809 srcs = [
2810 "test/f32-rmax.cc",
2811 "test/rmax-microkernel-tester.h",
2812 ] + MICROKERNEL_TEST_HDRS,
2813 deps = MICROKERNEL_TEST_DEPS,
2814)
2815
2816xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002817 name = "f32_sigmoid_test",
2818 srcs = [
2819 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002820 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002821 ] + MICROKERNEL_TEST_HDRS,
2822 deps = MICROKERNEL_TEST_DEPS,
2823)
2824
2825xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002826 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002827 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002828 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002829 "test/spmm-microkernel-tester.h",
2830 "src/xnnpack/AlignedAllocator.h",
2831 ] + MICROKERNEL_TEST_HDRS,
2832 deps = MICROKERNEL_TEST_DEPS,
2833)
2834
2835xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002836 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002837 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002838 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002839 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002840 ] + MICROKERNEL_TEST_HDRS,
2841 deps = MICROKERNEL_TEST_DEPS,
2842)
2843
2844xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002845 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002846 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002847 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002848 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002849 ] + MICROKERNEL_TEST_HDRS,
2850 deps = MICROKERNEL_TEST_DEPS,
2851)
2852
2853xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002854 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002855 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002856 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002857 "test/vbinary-microkernel-tester.h",
2858 ] + MICROKERNEL_TEST_HDRS,
2859 deps = MICROKERNEL_TEST_DEPS,
2860)
2861
2862xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002863 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002864 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002865 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002866 "test/vbinaryc-microkernel-tester.h",
2867 ] + MICROKERNEL_TEST_HDRS,
2868 deps = MICROKERNEL_TEST_DEPS,
2869)
2870
2871xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002872 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002873 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002874 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002875 "test/vbinaryc-microkernel-tester.h",
2876 ] + MICROKERNEL_TEST_HDRS,
2877 deps = MICROKERNEL_TEST_DEPS,
2878)
2879
2880xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002881 name = "f32_vmax_test",
2882 srcs = [
2883 "test/f32-vmax.cc",
2884 "test/vbinary-microkernel-tester.h",
2885 ] + MICROKERNEL_TEST_HDRS,
2886 deps = MICROKERNEL_TEST_DEPS,
2887)
2888
2889xnnpack_unit_test(
2890 name = "f32_vmaxc_test",
2891 srcs = [
2892 "test/f32-vmaxc.cc",
2893 "test/vbinaryc-microkernel-tester.h",
2894 ] + MICROKERNEL_TEST_HDRS,
2895 deps = MICROKERNEL_TEST_DEPS,
2896)
2897
2898xnnpack_unit_test(
2899 name = "f32_vmin_test",
2900 srcs = [
2901 "test/f32-vmin.cc",
2902 "test/vbinary-microkernel-tester.h",
2903 ] + MICROKERNEL_TEST_HDRS,
2904 deps = MICROKERNEL_TEST_DEPS,
2905)
2906
2907xnnpack_unit_test(
2908 name = "f32_vminc_test",
2909 srcs = [
2910 "test/f32-vminc.cc",
2911 "test/vbinaryc-microkernel-tester.h",
2912 ] + MICROKERNEL_TEST_HDRS,
2913 deps = MICROKERNEL_TEST_DEPS,
2914)
2915
2916xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002917 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002918 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002919 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002920 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002921 ] + MICROKERNEL_TEST_HDRS,
2922 deps = MICROKERNEL_TEST_DEPS,
2923)
2924
2925xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002926 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002927 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002928 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002929 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 ] + MICROKERNEL_TEST_HDRS,
2931 deps = MICROKERNEL_TEST_DEPS,
2932)
2933
2934xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002935 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002937 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002938 "test/vmulcaddc-microkernel-tester.h",
2939 "src/xnnpack/AlignedAllocator.h",
2940 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2941 deps = MICROKERNEL_TEST_DEPS,
2942)
2943
2944xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002945 name = "f32_vscale_test",
2946 srcs = [
2947 "test/f32-vscale.cc",
2948 "test/vscale-microkernel-tester.h",
2949 ] + MICROKERNEL_TEST_HDRS,
2950 deps = MICROKERNEL_TEST_DEPS,
2951)
2952
2953xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002954 name = "f32_vscaleexpminusmax_test",
2955 srcs = [
2956 "test/f32-vscaleexpminusmax.cc",
2957 "test/vscaleexpminusmax-microkernel-tester.h",
2958 ] + MICROKERNEL_TEST_HDRS,
2959 deps = MICROKERNEL_TEST_DEPS,
2960)
2961
2962xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002963 name = "f32_vscaleextexp_test",
2964 srcs = [
2965 "test/f32-vscaleextexp.cc",
2966 "test/vscaleextexp-microkernel-tester.h",
2967 ] + MICROKERNEL_TEST_HDRS,
2968 deps = MICROKERNEL_TEST_DEPS,
2969)
2970
2971xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002972 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07002973 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002974 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002975 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002976 ] + MICROKERNEL_TEST_HDRS,
2977 deps = MICROKERNEL_TEST_DEPS,
2978)
2979
2980xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002981 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002982 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002983 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002984 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002985 ] + MICROKERNEL_TEST_HDRS,
2986 deps = MICROKERNEL_TEST_DEPS,
2987)
2988
2989xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002990 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002991 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002992 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002993 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002994 ] + MICROKERNEL_TEST_HDRS,
2995 deps = MICROKERNEL_TEST_DEPS,
2996)
2997
2998xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002999 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003001 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003002 "test/avgpool-microkernel-tester.h",
3003 "src/xnnpack/AlignedAllocator.h",
3004 ] + MICROKERNEL_TEST_HDRS,
3005 deps = MICROKERNEL_TEST_DEPS,
3006)
3007
3008xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003009 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003011 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003012 "test/gemm-microkernel-tester.h",
3013 "src/xnnpack/AlignedAllocator.h",
3014 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3015 deps = MICROKERNEL_TEST_DEPS,
3016)
3017
3018xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003019 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003020 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003021 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 "test/dwconv-microkernel-tester.h",
3023 "src/xnnpack/AlignedAllocator.h",
3024 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3025 deps = MICROKERNEL_TEST_DEPS,
3026)
3027
3028xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003029 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003031 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003032 "test/gavgpool-microkernel-tester.h",
3033 "src/xnnpack/AlignedAllocator.h",
3034 ] + MICROKERNEL_TEST_HDRS,
3035 deps = MICROKERNEL_TEST_DEPS,
3036)
3037
3038xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003039 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003040 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003041 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003042 "test/gemm-microkernel-tester.h",
3043 "src/xnnpack/AlignedAllocator.h",
3044 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3045 deps = MICROKERNEL_TEST_DEPS,
3046)
3047
3048xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003049 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003050 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003051 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003052 "test/vadd-microkernel-tester.h",
3053 ] + MICROKERNEL_TEST_HDRS,
3054 deps = MICROKERNEL_TEST_DEPS,
3055)
3056
3057xnnpack_unit_test(
3058 name = "u8_clamp_test",
3059 srcs = [
3060 "test/u8-clamp.cc",
3061 "test/clamp-microkernel-tester.h",
3062 ] + MICROKERNEL_TEST_HDRS,
3063 deps = MICROKERNEL_TEST_DEPS,
3064)
3065
3066xnnpack_unit_test(
3067 name = "u8_lut32norm_test",
3068 srcs = [
3069 "test/u8-lut32norm.cc",
3070 "test/lut-norm-microkernel-tester.h",
3071 ] + MICROKERNEL_TEST_HDRS,
3072 deps = MICROKERNEL_TEST_DEPS,
3073)
3074
3075xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003076 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003077 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003078 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003079 "test/maxpool-microkernel-tester.h",
3080 ] + MICROKERNEL_TEST_HDRS,
3081 deps = MICROKERNEL_TEST_DEPS,
3082)
3083
3084xnnpack_unit_test(
3085 name = "u8_rmax_test",
3086 srcs = [
3087 "test/u8-rmax.cc",
3088 "test/rmax-microkernel-tester.h",
3089 ] + MICROKERNEL_TEST_HDRS,
3090 deps = MICROKERNEL_TEST_DEPS,
3091)
3092
3093xnnpack_unit_test(
3094 name = "x32_packx_test",
3095 srcs = [
3096 "test/x32-packx.cc",
3097 "test/pack-microkernel-tester.h",
3098 "src/xnnpack/AlignedAllocator.h",
3099 ] + MICROKERNEL_TEST_HDRS,
3100 deps = MICROKERNEL_TEST_DEPS,
3101)
3102
3103xnnpack_unit_test(
3104 name = "x32_pad_test",
3105 srcs = [
3106 "test/x32-pad.cc",
3107 "test/pad-microkernel-tester.h",
3108 ] + MICROKERNEL_TEST_HDRS,
3109 deps = MICROKERNEL_TEST_DEPS,
3110)
3111
3112xnnpack_unit_test(
3113 name = "x32_unpool_test",
3114 srcs = [
3115 "test/x32-unpool.cc",
3116 "test/unpool-microkernel-tester.h",
3117 ] + MICROKERNEL_TEST_HDRS,
3118 deps = MICROKERNEL_TEST_DEPS,
3119)
3120
3121xnnpack_unit_test(
3122 name = "x32_zip_test",
3123 srcs = [
3124 "test/x32-zip.cc",
3125 "test/zip-microkernel-tester.h",
3126 ] + MICROKERNEL_TEST_HDRS,
3127 deps = MICROKERNEL_TEST_DEPS,
3128)
3129
3130xnnpack_unit_test(
3131 name = "x8_lut_test",
3132 srcs = [
3133 "test/x8-lut.cc",
3134 "test/lut-microkernel-tester.h",
3135 ] + MICROKERNEL_TEST_HDRS,
3136 deps = MICROKERNEL_TEST_DEPS,
3137)
3138
3139xnnpack_unit_test(
3140 name = "x8_zip_test",
3141 srcs = [
3142 "test/x8-zip.cc",
3143 "test/zip-microkernel-tester.h",
3144 ] + MICROKERNEL_TEST_HDRS,
3145 deps = MICROKERNEL_TEST_DEPS,
3146)
3147
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003148xnnpack_unit_test(
3149 name = "requantization_test",
3150 srcs = [
3151 "src/xnnpack/requantization-stubs.h",
3152 "test/requantization.cc",
3153 "test/requantization-tester.h",
3154 ] + MICROKERNEL_TEST_HDRS,
3155 deps = MICROKERNEL_TEST_DEPS,
3156)
3157
Marat Dukhan20c3b922020-03-10 03:45:06 -07003158########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003159
3160xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003161 name = "operator_size_test",
3162 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003163 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003164)
3165
Marat Dukhan20c3b922020-03-10 03:45:06 -07003166xnnpack_binary(
3167 name = "subgraph_size_test",
3168 srcs = ["test/subgraph-size.c"],
3169 deps = [":XNNPACK"],
3170)
3171
3172########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173
3174xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003175 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003176 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003177 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003178 "test/add-operator-tester.h",
3179 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003180 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181)
3182
3183xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003184 name = "add_nd_test",
3185 srcs = [
3186 "test/add-nd.cc",
3187 "test/binary-elementwise-operator-tester.h",
3188 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003189 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003190)
3191
3192xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003193 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003194 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003195 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003196 "test/argmax-pooling-operator-tester.h",
3197 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003198 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003199)
3200
3201xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003202 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003204 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003205 "test/average-pooling-operator-tester.h",
3206 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003207 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003208)
3209
3210xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003211 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003213 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003214 "test/channel-pad-operator-tester.h",
3215 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003216 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003217)
3218
3219xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003220 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003221 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003222 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003223 "test/channel-shuffle-operator-tester.h",
3224 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003225 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003226)
3227
3228xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003229 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003230 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003231 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003232 "test/clamp-operator-tester.h",
3233 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003234 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003235)
3236
3237xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003238 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003239 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003240 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003241 "test/convolution-operator-tester.h",
3242 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003243 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003244)
3245
3246xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003247 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003248 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003249 "test/convolution-nchw.cc",
3250 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003251 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003252 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253)
3254
3255xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003256 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003257 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003258 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003259 "test/deconvolution-operator-tester.h",
3260 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003261 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003262)
3263
3264xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003265 name = "divide_nd_test",
3266 srcs = [
3267 "test/binary-elementwise-operator-tester.h",
3268 "test/divide-nd.cc",
3269 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003270 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003271)
3272
3273xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003274 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003275 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003276 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003277 "test/fully-connected-operator-tester.h",
3278 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003279 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003280)
3281
3282xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003283 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003284 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003285 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003286 "test/global-average-pooling-operator-tester.h",
3287 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003288 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003289)
3290
3291xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003292 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003293 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003294 "test/global-average-pooling-ncw.cc",
3295 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003296 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003297 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003298)
3299
3300xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003301 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003302 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003303 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003304 "test/hardswish-operator-tester.h",
3305 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003306 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003307)
3308
3309xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003310 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003311 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003312 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003313 "test/leaky-relu-operator-tester.h",
3314 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003315 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003316)
3317
3318xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003319 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003320 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003321 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003322 "test/max-pooling-operator-tester.h",
3323 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003324 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003325)
3326
3327xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003328 name = "maximum_nd_test",
3329 srcs = [
3330 "test/binary-elementwise-operator-tester.h",
3331 "test/maximum-nd.cc",
3332 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003333 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003334)
3335
3336xnnpack_unit_test(
3337 name = "minimum_nd_test",
3338 srcs = [
3339 "test/binary-elementwise-operator-tester.h",
3340 "test/minimum-nd.cc",
3341 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003342 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003343)
3344
3345xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003346 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003347 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003348 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003349 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003350 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003351 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003352)
3353
3354xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003355 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003356 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003357 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003358 "test/prelu-operator-tester.h",
3359 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003360 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003361)
3362
3363xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003364 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003365 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003366 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003367 "test/resize-bilinear-operator-tester.h",
3368 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003369 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003370)
3371
3372xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003373 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003374 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003375 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003376 "test/sigmoid-operator-tester.h",
3377 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003378 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003379)
3380
3381xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003382 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003383 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003384 "test/softmax-nc.cc",
3385 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003386 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003387 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003388)
3389
3390xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003391 name = "subtract_nd_test",
3392 srcs = [
3393 "test/binary-elementwise-operator-tester.h",
3394 "test/subtract-nd.cc",
3395 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003396 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003397)
3398
3399xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003400 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003401 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003402 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003403 "test/unpooling-operator-tester.h",
3404 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003405 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406)
3407
3408############################# Build configurations #############################
3409
Marat Dukhanb8642352019-10-30 15:43:02 -07003410# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003411config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003412 name = "xnn_enable_assembly_explicit_true",
3413 define_values = {"xnn_enable_assembly": "true"},
3414)
3415
3416# Disables usage of assembly kernels.
3417config_setting(
3418 name = "xnn_enable_assembly_explicit_false",
3419 define_values = {"xnn_enable_assembly": "false"},
3420)
3421
Marat Dukhan05702cf2020-03-26 15:41:33 -07003422# Disables usage of HMP-aware optimizations.
3423config_setting(
3424 name = "xnn_enable_hmp_explicit_false",
3425 define_values = {"xnn_enable_hmp": "false"},
3426)
3427
Marat Dukhanb8642352019-10-30 15:43:02 -07003428# Builds with -c dbg
3429config_setting(
3430 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003431 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003432 "compilation_mode": "dbg",
3433 },
3434)
3435
3436# Builds with -c opt
3437config_setting(
3438 name = "optimized_build",
3439 values = {
3440 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003441 },
3442)
3443
3444config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003445 name = "linux_k8",
3446 values = {"cpu": "k8"},
3447)
3448
3449config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003450 name = "linux_arm",
3451 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003452)
3453
3454config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003455 name = "linux_armhf",
3456 values = {"cpu": "armhf"},
3457)
3458
3459config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003460 name = "linux_aarch64",
3461 values = {"cpu": "aarch64"},
3462)
3463
3464config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003465 name = "android",
3466 values = {"crosstool_top": "//external:android/crosstool"},
3467)
3468
3469config_setting(
3470 name = "android_armv7",
3471 values = {
3472 "crosstool_top": "//external:android/crosstool",
3473 "cpu": "armeabi-v7a",
3474 },
3475)
3476
3477config_setting(
3478 name = "android_arm64",
3479 values = {
3480 "crosstool_top": "//external:android/crosstool",
3481 "cpu": "arm64-v8a",
3482 },
3483)
3484
3485config_setting(
3486 name = "android_x86",
3487 values = {
3488 "crosstool_top": "//external:android/crosstool",
3489 "cpu": "x86",
3490 },
3491)
3492
3493config_setting(
3494 name = "android_x86_64",
3495 values = {
3496 "crosstool_top": "//external:android/crosstool",
3497 "cpu": "x86_64",
3498 },
3499)
3500
3501config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003502 name = "windows_x86_64",
3503 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003504)
3505
3506config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003507 name = "windows_x86_64_clang",
3508 values = {
3509 "compiler": "clang-cl",
3510 "cpu": "x64_windows",
3511 },
3512)
3513
3514config_setting(
3515 name = "windows_x86_64_mingw",
3516 values = {
3517 "compiler": "mingw-gcc",
3518 "cpu": "x64_windows",
3519 },
3520)
3521
3522config_setting(
3523 name = "windows_x86_64_msys",
3524 values = {
3525 "compiler": "msys-gcc",
3526 "cpu": "x64_windows",
3527 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003528)
3529
3530config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003531 name = "macos_x86_64",
3532 values = {
3533 "apple_platform_type": "macos",
3534 "cpu": "darwin",
3535 },
3536)
3537
3538config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003540 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003541)
3542
3543config_setting(
3544 name = "emscripten_wasm",
3545 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003546 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003547 "cpu": "wasm",
3548 },
3549)
3550
3551config_setting(
3552 name = "emscripten_wasmsimd",
3553 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003554 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003556 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557 },
3558)
3559
3560config_setting(
3561 name = "emscripten_asmjs",
3562 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003563 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 "cpu": "asmjs",
3565 },
3566)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003567
3568config_setting(
3569 name = "ios_armv7",
3570 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003571 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003572 "cpu": "ios_armv7",
3573 },
3574)
3575
3576config_setting(
3577 name = "ios_arm64",
3578 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003579 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003580 "cpu": "ios_arm64",
3581 },
3582)
3583
3584config_setting(
3585 name = "ios_arm64e",
3586 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003587 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003588 "cpu": "ios_arm64e",
3589 },
3590)
3591
3592config_setting(
3593 name = "ios_x86",
3594 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003595 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003596 "cpu": "ios_i386",
3597 },
3598)
3599
3600config_setting(
3601 name = "ios_x86_64",
3602 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003603 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003604 "cpu": "ios_x86_64",
3605 },
3606)
3607
3608config_setting(
3609 name = "watchos_armv7k",
3610 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003611 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003612 "cpu": "watchos_armv7k",
3613 },
3614)
3615
3616config_setting(
3617 name = "watchos_arm64_32",
3618 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003619 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003620 "cpu": "watchos_arm64_32",
3621 },
3622)
3623
3624config_setting(
3625 name = "watchos_x86",
3626 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003627 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003628 "cpu": "watchos_i386",
3629 },
3630)
3631
3632config_setting(
3633 name = "watchos_x86_64",
3634 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003635 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003636 "cpu": "watchos_x86_64",
3637 },
3638)
3639
3640config_setting(
3641 name = "tvos_arm64",
3642 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003643 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003644 "cpu": "tvos_arm64",
3645 },
3646)
3647
3648config_setting(
3649 name = "tvos_x86_64",
3650 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003651 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003652 "cpu": "tvos_x86_64",
3653 },
3654)