blob: 1edf8c4b91d44971afc87ec1c8bffe22020db72c [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan8d3c6932020-03-06 20:27:27 -08009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_copts", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat 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 Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
Marat Dukhanfd8e6892020-01-27 15:25:25 -080072 "src/softmax-nc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080073 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -080085 "src/f32-avgpool/9p8x-scalar-c1.c",
86 "src/f32-avgpool/9x-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -070087 "src/f32-clamp/gen/scalar-x1.c",
88 "src/f32-clamp/gen/scalar-x2.c",
89 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080090 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080091 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070092 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070093 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070094 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080095 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080096 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -070097 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
99 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
100 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
101 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
102 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
103 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
105 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
106 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
107 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
108 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700109 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700110 "src/f32-gavgpool/7p7x-scalar-c1.c",
111 "src/f32-gavgpool/7x-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700112 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
113 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
114 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
115 "src/f32-gemm/gen/1x4-minmax-scalar.c",
116 "src/f32-gemm/gen/2x4-minmax-scalar.c",
117 "src/f32-gemm/gen/4x2-minmax-scalar.c",
118 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800119 "src/f32-hswish/gen/scalar-x1.c",
120 "src/f32-hswish/gen/scalar-x2.c",
121 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700122 "src/f32-ibilinear/gen/scalar-c1.c",
123 "src/f32-ibilinear/gen/scalar-c2.c",
124 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700125 "src/f32-igemm/gen/1x4-minmax-scalar.c",
126 "src/f32-igemm/gen/2x4-minmax-scalar.c",
127 "src/f32-igemm/gen/4x2-minmax-scalar.c",
128 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800129 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800130 "src/f32-pavgpool/9p8x-scalar-c1.c",
131 "src/f32-pavgpool/9x-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700132 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
133 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
134 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
135 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800136 "src/f32-prelu/gen/scalar-2x1.c",
137 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800138 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
139 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
140 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
141 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
142 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
143 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
144 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
145 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
146 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
147 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
148 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
149 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700150 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800151 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
152 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
153 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
154 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
155 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
156 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
157 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
158 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
159 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800160 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
161 "src/f32-spmm/gen/1x1-scalar.c",
162 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
163 "src/f32-spmm/gen/2x1-scalar.c",
164 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
165 "src/f32-spmm/gen/4x1-scalar.c",
166 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
167 "src/f32-spmm/gen/8x1-scalar.c",
168 "src/f32-spmm/gen/8x2-scalar.c",
169 "src/f32-spmm/gen/8x4-scalar.c",
170 "src/f32-vbinary/gen/vadd-scalar-x1.c",
171 "src/f32-vbinary/gen/vadd-scalar-x2.c",
172 "src/f32-vbinary/gen/vadd-scalar-x4.c",
173 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
174 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
175 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800176 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
177 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
178 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
179 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
180 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
181 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800182 "src/f32-vbinary/gen/vmax-scalar-x1.c",
183 "src/f32-vbinary/gen/vmax-scalar-x2.c",
184 "src/f32-vbinary/gen/vmax-scalar-x4.c",
185 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
186 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
187 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
188 "src/f32-vbinary/gen/vmin-scalar-x1.c",
189 "src/f32-vbinary/gen/vmin-scalar-x2.c",
190 "src/f32-vbinary/gen/vmin-scalar-x4.c",
191 "src/f32-vbinary/gen/vminc-scalar-x1.c",
192 "src/f32-vbinary/gen/vminc-scalar-x2.c",
193 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800194 "src/f32-vbinary/gen/vmul-scalar-x1.c",
195 "src/f32-vbinary/gen/vmul-scalar-x2.c",
196 "src/f32-vbinary/gen/vmul-scalar-x4.c",
197 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
198 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
199 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800200 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
201 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
202 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800203 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
204 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
205 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
206 "src/f32-vbinary/gen/vsub-scalar-x1.c",
207 "src/f32-vbinary/gen/vsub-scalar-x2.c",
208 "src/f32-vbinary/gen/vsub-scalar-x4.c",
209 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
210 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
211 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
212 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
213 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
214 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800215 "src/math/expminus-scalar-lut2048-p1.c",
216 "src/math/expminus-scalar-lut64-p2.c",
217 "src/math/expminus-scalar-p5.c",
218 "src/math/sigmoid-scalar-lut2048-p1-div.c",
219 "src/math/sigmoid-scalar-lut64-p2-div.c",
220 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800221 "src/q8-avgpool/9p8x-scalar-c1.c",
222 "src/q8-avgpool/9x-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700223 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700224 "src/q8-gavgpool/7p7x-scalar-c1.c",
225 "src/q8-gavgpool/7x-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700226 "src/q8-gemm/2x2-minmax-scalar.c",
227 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700228 "src/q8-vadd/scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700229 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700230 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800231 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700232 "src/u8-rmax/scalar.c",
233 "src/x32-packx/x2-scalar.c",
234 "src/x32-packx/x3-scalar.c",
235 "src/x32-packx/x4-scalar.c",
236 "src/x32-pad/x2-scalar.c",
237 "src/x32-unpool/scalar.c",
238 "src/x32-zip/x2-scalar.c",
239 "src/x32-zip/x3-scalar.c",
240 "src/x32-zip/x4-scalar.c",
241 "src/x32-zip/xm-scalar.c",
242 "src/x8-lut/scalar.c",
243 "src/x8-zip/x2-scalar.c",
244 "src/x8-zip/x3-scalar.c",
245 "src/x8-zip/x4-scalar.c",
246 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700247 "src/requantization/precise-scalar.c",
248 "src/requantization/fp32-scalar.c",
249 "src/requantization/q31-scalar.c",
250 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700251]
252
Marat Dukhan436ebe62019-12-04 15:10:12 -0800253WASM_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800254 "src/f32-avgpool/9p8x-wasm-c1.c",
255 "src/f32-avgpool/9x-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700256 "src/f32-clamp/gen/wasm-x1.c",
257 "src/f32-clamp/gen/wasm-x2.c",
258 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700259 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
260 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
261 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
262 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
263 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
264 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
265 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
266 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
267 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
268 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
269 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
270 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700271 "src/f32-gavgpool/7p7x-wasm-c1.c",
272 "src/f32-gavgpool/7x-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700273 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
274 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
275 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
276 "src/f32-gemm/gen/1x4-minmax-wasm.c",
277 "src/f32-gemm/gen/2x4-minmax-wasm.c",
278 "src/f32-gemm/gen/4x2-minmax-wasm.c",
279 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800280 "src/f32-hswish/gen/wasm-x1.c",
281 "src/f32-hswish/gen/wasm-x2.c",
282 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700283 "src/f32-igemm/gen/1x4-minmax-wasm.c",
284 "src/f32-igemm/gen/2x4-minmax-wasm.c",
285 "src/f32-igemm/gen/4x2-minmax-wasm.c",
286 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800287 "src/f32-maxpool/9p8x-wasm-c1.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800288 "src/f32-pavgpool/9p8x-wasm-c1.c",
289 "src/f32-pavgpool/9x-wasm-c1.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800290 "src/f32-vbinary/gen/vadd-wasm-x1.c",
291 "src/f32-vbinary/gen/vadd-wasm-x2.c",
292 "src/f32-vbinary/gen/vadd-wasm-x4.c",
293 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
294 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
295 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800296 "src/f32-vbinary/gen/vdiv-wasm-x1.c",
297 "src/f32-vbinary/gen/vdiv-wasm-x2.c",
298 "src/f32-vbinary/gen/vdiv-wasm-x4.c",
299 "src/f32-vbinary/gen/vdivc-wasm-x1.c",
300 "src/f32-vbinary/gen/vdivc-wasm-x2.c",
301 "src/f32-vbinary/gen/vdivc-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800302 "src/f32-vbinary/gen/vmax-wasm-x1.c",
303 "src/f32-vbinary/gen/vmax-wasm-x2.c",
304 "src/f32-vbinary/gen/vmax-wasm-x4.c",
305 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
306 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
307 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
308 "src/f32-vbinary/gen/vmin-wasm-x1.c",
309 "src/f32-vbinary/gen/vmin-wasm-x2.c",
310 "src/f32-vbinary/gen/vmin-wasm-x4.c",
311 "src/f32-vbinary/gen/vminc-wasm-x1.c",
312 "src/f32-vbinary/gen/vminc-wasm-x2.c",
313 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800314 "src/f32-vbinary/gen/vmul-wasm-x1.c",
315 "src/f32-vbinary/gen/vmul-wasm-x2.c",
316 "src/f32-vbinary/gen/vmul-wasm-x4.c",
317 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
318 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
319 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800320 "src/f32-vbinary/gen/vrdivc-wasm-x1.c",
321 "src/f32-vbinary/gen/vrdivc-wasm-x2.c",
322 "src/f32-vbinary/gen/vrdivc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800323 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
324 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
325 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
326 "src/f32-vbinary/gen/vsub-wasm-x1.c",
327 "src/f32-vbinary/gen/vsub-wasm-x2.c",
328 "src/f32-vbinary/gen/vsub-wasm-x4.c",
329 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
330 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
331 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
332 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
333 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
334 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
335]
336
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800337PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800338 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800339 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800340 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800341 "src/f32-avgpool/9p8x-psimd-c4.c",
342 "src/f32-avgpool/9x-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700343 "src/f32-clamp/gen/psimd-x4.c",
344 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700345 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
346 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
347 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
348 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
349 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
350 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
351 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
352 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
353 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
354 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
355 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
356 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700357 "src/f32-gavgpool/7p7x-psimd-c4.c",
358 "src/f32-gavgpool/7x-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700359 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
360 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
361 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
362 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
363 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
364 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
365 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
366 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
367 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
368 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
369 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
370 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
371 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
372 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
373 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
374 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
375 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
376 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
377 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800378 "src/f32-hswish/gen/psimd-x4.c",
379 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700380 "src/f32-ibilinear/gen/psimd-c4.c",
381 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700382 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
383 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
384 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
385 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
386 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
387 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
388 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
389 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
390 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
391 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800392 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800393 "src/f32-pavgpool/9p8x-psimd-c4.c",
394 "src/f32-pavgpool/9x-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700395 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800396 "src/f32-prelu/gen/psimd-2x4.c",
397 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800398 "src/f32-rmax/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800399 "src/f32-vbinary/gen/vadd-psimd-x4.c",
400 "src/f32-vbinary/gen/vadd-psimd-x8.c",
401 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
402 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800403 "src/f32-vbinary/gen/vdiv-psimd-x4.c",
404 "src/f32-vbinary/gen/vdiv-psimd-x8.c",
405 "src/f32-vbinary/gen/vdivc-psimd-x4.c",
406 "src/f32-vbinary/gen/vdivc-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800407 "src/f32-vbinary/gen/vmax-psimd-x4.c",
408 "src/f32-vbinary/gen/vmax-psimd-x8.c",
409 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
410 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
411 "src/f32-vbinary/gen/vmin-psimd-x4.c",
412 "src/f32-vbinary/gen/vmin-psimd-x8.c",
413 "src/f32-vbinary/gen/vminc-psimd-x4.c",
414 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800415 "src/f32-vbinary/gen/vmul-psimd-x4.c",
416 "src/f32-vbinary/gen/vmul-psimd-x8.c",
417 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
418 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800419 "src/f32-vbinary/gen/vrdivc-psimd-x4.c",
420 "src/f32-vbinary/gen/vrdivc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800421 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
422 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
423 "src/f32-vbinary/gen/vsub-psimd-x4.c",
424 "src/f32-vbinary/gen/vsub-psimd-x8.c",
425 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
426 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
427 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
428 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700429 "src/x32-packx/x4-psimd.c",
430 "src/x32-pad/x2-psimd.c",
431 "src/x32-unpool/psimd.c",
432 "src/x32-zip/x2-psimd.c",
433 "src/x32-zip/x3-psimd.c",
434 "src/x32-zip/x4-psimd.c",
435 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700436 "src/requantization/precise-psimd.c",
437 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700438]
439
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800440PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800441 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
442 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
443 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
444 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
445 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
446 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
447 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
448 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
449 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
450 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
451 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
452 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800453 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
454 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
455 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
456 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
457 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
458 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
459 "src/math/sigmoid-psimd-p5-div.c",
460]
461
Marat Dukhan08c4a432019-10-03 09:29:21 -0700462# ISA-specific micro-kernels
463NEON_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800464 "src/f32-avgpool/9p8x-neon-c4.c",
465 "src/f32-avgpool/9x-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700466 "src/f32-clamp/gen/neon-x4.c",
467 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700468 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
469 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
470 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
471 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700472 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700473 "src/f32-gavgpool/7p7x-neon-c4.c",
474 "src/f32-gavgpool/7x-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700475 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
476 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
477 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
478 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
479 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
480 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
481 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
482 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
483 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
484 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
485 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
486 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
487 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
488 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
489 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
490 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
491 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
492 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
493 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
494 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
495 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
496 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
497 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
498 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
499 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
500 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
501 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
502 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
503 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
504 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
505 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800506 "src/f32-hswish/gen/neon-x4.c",
507 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700508 "src/f32-ibilinear/gen/neon-c4.c",
509 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700510 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
511 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
512 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
513 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
514 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
515 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
516 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
517 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
518 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
519 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
520 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
521 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
522 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
523 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
524 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
525 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Frank Barchardf092a4a2020-03-03 14:22:46 -0800526 "src/f32-maxpool/9p8x-neon-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800527 "src/f32-pavgpool/9p8x-neon-c4.c",
528 "src/f32-pavgpool/9x-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700529 "src/f32-ppmm/gen/4x8-minmax-neon.c",
530 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800531 "src/f32-prelu/gen/neon-2x4.c",
532 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800533 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
534 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
535 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
536 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
537 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
538 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
539 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
540 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
541 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
542 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
543 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
544 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
545 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
546 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
547 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
548 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
549 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
550 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
551 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
552 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
553 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
554 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
555 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700557 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800558 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800559 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
560 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
561 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
562 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
563 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
564 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
565 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
566 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
567 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
568 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
569 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
570 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
571 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
572 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
573 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
574 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
575 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
576 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800577 "src/f32-vbinary/gen/vadd-neon-x4.c",
578 "src/f32-vbinary/gen/vadd-neon-x8.c",
579 "src/f32-vbinary/gen/vaddc-neon-x4.c",
580 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800581 "src/f32-vbinary/gen/vmax-neon-x4.c",
582 "src/f32-vbinary/gen/vmax-neon-x8.c",
583 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
584 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
585 "src/f32-vbinary/gen/vmin-neon-x4.c",
586 "src/f32-vbinary/gen/vmin-neon-x8.c",
587 "src/f32-vbinary/gen/vminc-neon-x4.c",
588 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800589 "src/f32-vbinary/gen/vmul-neon-x4.c",
590 "src/f32-vbinary/gen/vmul-neon-x8.c",
591 "src/f32-vbinary/gen/vmulc-neon-x4.c",
592 "src/f32-vbinary/gen/vmulc-neon-x8.c",
593 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
594 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
595 "src/f32-vbinary/gen/vsub-neon-x4.c",
596 "src/f32-vbinary/gen/vsub-neon-x8.c",
597 "src/f32-vbinary/gen/vsubc-neon-x4.c",
598 "src/f32-vbinary/gen/vsubc-neon-x8.c",
599 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
600 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800601 "src/q8-avgpool/9p8x-neon-c8.c",
602 "src/q8-avgpool/9x-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700603 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700604 "src/q8-gavgpool/7p7x-neon-c8.c",
605 "src/q8-gavgpool/7x-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700606 "src/q8-gemm/4x8-minmax-neon.c",
607 "src/q8-gemm/8x8-minmax-neon.c",
608 "src/q8-igemm/4x8-minmax-neon.c",
609 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700610 "src/q8-vadd/neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700611 "src/u8-clamp/neon-x64.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800612 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700613 "src/u8-rmax/neon.c",
614 "src/x32-packx/x4-neon-st4.c",
615 "src/x32-pad/x2-neon.c",
616 "src/x32-zip/x2-neon.c",
617 "src/x32-zip/x3-neon.c",
618 "src/x32-zip/x4-neon.c",
619 "src/x32-zip/xm-neon.c",
620 "src/x8-zip/x2-neon.c",
621 "src/x8-zip/x3-neon.c",
622 "src/x8-zip/x4-neon.c",
623 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800624 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800625 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
626 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
627 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
628 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
629 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
630 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700631 "src/requantization/precise-neon.c",
632 "src/requantization/fp32-neon.c",
633 "src/requantization/q31-neon.c",
634 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700635]
636
637NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700638 "src/f32-ibilinear/gen/neonfma-c4.c",
639 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700640 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
641 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
642 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
643 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
644 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
645 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
646 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
647 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
648 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
649 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
650 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
651 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
652 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
653 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
654 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
655 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
656 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
657 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
658 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
659 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
660 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
661 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
662 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
663 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
664 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
665 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
666 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
667 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
668 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
669 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
670 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800671 "src/f32-hswish/gen/neonfma-x4.c",
672 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700673 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
674 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800675 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
676 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
677 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
678 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
679 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
680 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
681 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
682 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
683 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
684 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
685 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
686 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
687 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
688 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
689 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
690 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
691 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
692 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
693 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
694 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
695 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
696 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
697 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
698 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800699 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
700 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
701 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
702 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
703 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
704 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
705 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
706 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
707 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
708 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
709 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
710 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
711 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
712 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
713 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
714 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
715 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
716 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
717 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
718 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
719 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
720 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
721 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
722 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
723 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
724 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
725 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
726 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
727 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
728 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
729 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
730 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
731 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
732 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
733 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
734 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
735 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
736 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
737 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
738 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
739 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
740 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
741 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
742 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
743 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
744 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
745 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
746 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
747 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
748 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
749 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
750 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
751 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
752 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800753 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
754 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800755 "src/math/exp-neonfma-lut64-p2.c",
756 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800757 "src/math/expminus-neonfma-lut2048-p1.c",
758 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800759 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800760 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
761 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
762 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
763 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
764 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
765 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
766 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
767 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
768 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
769 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
770 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
771 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
772 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
773 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
774 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
775 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
776 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
777 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700778]
779
780AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800781 "src/f32-vbinary/gen/vdiv-neon-x4.c",
782 "src/f32-vbinary/gen/vdiv-neon-x8.c",
783 "src/f32-vbinary/gen/vdivc-neon-x4.c",
784 "src/f32-vbinary/gen/vdivc-neon-x8.c",
785 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
786 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700787 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
788 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
789 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
790 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
791 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
792 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
793 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
794 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
795 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
796 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
797 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
798 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
799 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
800 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
801 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
802 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
803 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
804 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
805 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
806 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700807 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
808 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
809 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
810 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
811 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
812 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
813 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800814 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
816 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
817 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
818 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
819 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
820 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
821 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
822 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
823 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
824 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
825 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
826 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
827 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
828 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
829 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
830 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
831 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800832 "src/f32-spmm/gen/12x1-neonfma.c",
833 "src/f32-spmm/gen/12x2-neonfma.c",
834 "src/f32-spmm/gen/12x4-neonfma.c",
835 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
836 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
837 "src/f32-spmm/gen/16x1-neonfma.c",
838 "src/f32-spmm/gen/16x2-neonfma.c",
839 "src/f32-spmm/gen/16x4-neonfma.c",
840 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
841 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
842 "src/f32-spmm/gen/4x1-neonfma.c",
843 "src/f32-spmm/gen/4x2-neonfma.c",
844 "src/f32-spmm/gen/4x4-neonfma.c",
845 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
846 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
847 "src/f32-spmm/gen/8x1-neonfma.c",
848 "src/f32-spmm/gen/8x2-neonfma.c",
849 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800850 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
851 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
852 "src/math/sigmoid-neonfma-rr1-p5-div.c",
853 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
854 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
855 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700856]
857
858AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800859 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
860 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
861 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhanbdb56f52020-02-05 21:42:49 -0800862 "src/f16-spmm/gen/8x1-neonfp16arith.c",
863 "src/f16-spmm/gen/8x1-neonfp16arith-unroll2.c",
864 "src/f16-spmm/gen/16x1-neonfp16arith.c",
865 "src/f16-spmm/gen/16x1-neonfp16arith-unroll2.c",
866 "src/f16-spmm/gen/24x1-neonfp16arith.c",
867 "src/f16-spmm/gen/24x1-neonfp16arith-unroll2.c",
868 "src/f16-spmm/gen/32x1-neonfp16arith.c",
869 "src/f16-spmm/gen/32x1-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700870]
871
872SSE_UKERNELS = [
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800873 "src/f32-avgpool/9p8x-sse-c4.c",
874 "src/f32-avgpool/9x-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700875 "src/f32-clamp/gen/sse-x4.c",
876 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800877 "src/f32-dwconv-spchw/3x3p1-sse.c",
878 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700879 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
880 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
881 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
882 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
883 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
884 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
885 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
886 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
887 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
888 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
889 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
890 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700891 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700892 "src/f32-gavgpool/7p7x-sse-c4.c",
893 "src/f32-gavgpool/7x-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700894 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
895 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
896 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
897 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
898 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
899 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
900 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
901 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
902 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
903 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
904 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
905 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
906 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800907 "src/f32-hswish/gen/sse-x4.c",
908 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700909 "src/f32-ibilinear/gen/sse-c4.c",
910 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700911 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
912 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
913 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
914 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
915 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
916 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
917 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800918 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800919 "src/f32-pavgpool/9p8x-sse-c4.c",
920 "src/f32-pavgpool/9x-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700921 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700922 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800923 "src/f32-spmm/gen/4x1-sse.c",
924 "src/f32-spmm/gen/8x1-sse.c",
925 "src/f32-vbinary/gen/vadd-sse-x4.c",
926 "src/f32-vbinary/gen/vadd-sse-x8.c",
927 "src/f32-vbinary/gen/vaddc-sse-x4.c",
928 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800929 "src/f32-vbinary/gen/vdiv-sse-x4.c",
930 "src/f32-vbinary/gen/vdiv-sse-x8.c",
931 "src/f32-vbinary/gen/vdivc-sse-x4.c",
932 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800933 "src/f32-vbinary/gen/vmax-sse-x4.c",
934 "src/f32-vbinary/gen/vmax-sse-x8.c",
935 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
936 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
937 "src/f32-vbinary/gen/vmin-sse-x4.c",
938 "src/f32-vbinary/gen/vmin-sse-x8.c",
939 "src/f32-vbinary/gen/vminc-sse-x4.c",
940 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800941 "src/f32-vbinary/gen/vmul-sse-x4.c",
942 "src/f32-vbinary/gen/vmul-sse-x8.c",
943 "src/f32-vbinary/gen/vmulc-sse-x4.c",
944 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800945 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
946 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800947 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
948 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
949 "src/f32-vbinary/gen/vsub-sse-x4.c",
950 "src/f32-vbinary/gen/vsub-sse-x8.c",
951 "src/f32-vbinary/gen/vsubc-sse-x4.c",
952 "src/f32-vbinary/gen/vsubc-sse-x8.c",
953 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
954 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700955 "src/x32-packx/x4-sse.c",
956]
957
958SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800959 "src/f32-argmaxpool/9p8x-sse2-c4.c",
960 "src/f32-argmaxpool/4x-sse2-c4.c",
961 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800962 "src/f32-prelu/gen/sse2-2x4.c",
963 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800964 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
965 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
966 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
967 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
968 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
969 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
970 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
971 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
972 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
973 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
974 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
975 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800976 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800977 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800978 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800979 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800980 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
981 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan6ee435a2020-02-26 22:33:38 -0800982 "src/q8-avgpool/9p8x-sse2-c8.c",
983 "src/q8-avgpool/9x-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700984 "src/q8-igemm/4x4c2-minmax-sse2.c",
985 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhana63a6fc2020-03-10 06:12:48 -0700986 "src/q8-gavgpool/7p7x-sse2-c8.c",
987 "src/q8-gavgpool/7x-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700988 "src/q8-gemm/2x4c8-minmax-sse2.c",
989 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700990 "src/q8-vadd/sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700991 "src/u8-clamp/sse2-x64.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800992 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700993 "src/u8-rmax/sse2.c",
994 "src/x32-pad/x2-sse2.c",
995 "src/x32-zip/x2-sse2.c",
996 "src/x32-zip/x3-sse2.c",
997 "src/x32-zip/x4-sse2.c",
998 "src/x32-zip/xm-sse2.c",
999 "src/x8-zip/x2-sse2.c",
1000 "src/x8-zip/x3-sse2.c",
1001 "src/x8-zip/x4-sse2.c",
1002 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001003 "src/math/exp-sse2-p5.c",
1004 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001005 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001006 "src/requantization/precise-sse2.c",
1007 "src/requantization/fp32-sse2.c",
1008 "src/requantization/q31-sse2.c",
1009 "src/requantization/gemmlowp-sse2.c",
1010]
1011
1012SSSE3_UKERNELS = [
1013 "src/requantization/precise-ssse3.c",
1014 "src/requantization/q31-ssse3.c",
1015 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001016]
1017
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001018SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001019 "src/f32-prelu/gen/sse41-2x4.c",
1020 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001021 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001022 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001023 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001024 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001025 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1026 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001027 "src/requantization/precise-sse4.c",
1028 "src/requantization/q31-sse4.c",
1029 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001030]
1031
Marat Dukhan08c4a432019-10-03 09:29:21 -07001032AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001033 "src/f32-clamp/gen/avx-x8.c",
1034 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001035 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1036 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1037 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1038 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1039 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1040 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1041 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1042 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1043 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1044 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1045 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1046 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1047 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1048 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1049 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1050 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1051 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1052 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1053 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1054 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1055 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1056 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1057 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1058 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1059 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1060 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1061 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1062 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1063 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1064 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001065 "src/f32-hswish/gen/avx-x8.c",
1066 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001067 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1068 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1069 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1070 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1071 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1072 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1073 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1074 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1075 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001076 "src/f32-prelu/gen/avx-2x8.c",
1077 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001078 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001079 "src/f32-vbinary/gen/vadd-avx-x8.c",
1080 "src/f32-vbinary/gen/vadd-avx-x16.c",
1081 "src/f32-vbinary/gen/vaddc-avx-x8.c",
1082 "src/f32-vbinary/gen/vaddc-avx-x16.c",
1083 "src/f32-vbinary/gen/vdiv-avx-x8.c",
1084 "src/f32-vbinary/gen/vdiv-avx-x16.c",
1085 "src/f32-vbinary/gen/vdivc-avx-x8.c",
1086 "src/f32-vbinary/gen/vdivc-avx-x16.c",
1087 "src/f32-vbinary/gen/vmax-avx-x8.c",
1088 "src/f32-vbinary/gen/vmax-avx-x16.c",
1089 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1090 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1091 "src/f32-vbinary/gen/vmin-avx-x8.c",
1092 "src/f32-vbinary/gen/vmin-avx-x16.c",
1093 "src/f32-vbinary/gen/vminc-avx-x8.c",
1094 "src/f32-vbinary/gen/vminc-avx-x16.c",
1095 "src/f32-vbinary/gen/vmul-avx-x8.c",
1096 "src/f32-vbinary/gen/vmul-avx-x16.c",
1097 "src/f32-vbinary/gen/vmulc-avx-x8.c",
1098 "src/f32-vbinary/gen/vmulc-avx-x16.c",
1099 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
1100 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
1101 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
1102 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
1103 "src/f32-vbinary/gen/vsub-avx-x8.c",
1104 "src/f32-vbinary/gen/vsub-avx-x16.c",
1105 "src/f32-vbinary/gen/vsubc-avx-x8.c",
1106 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001107 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001108]
1109
Marat Dukhanfda12b82019-11-21 12:27:59 -08001110FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001111 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1112 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1113 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1114 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1115 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1116 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1117 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1118 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1119 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1120 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1121 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1122 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1123 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1124 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1125 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1126 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1127 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1128 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1129 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1130 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1131 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1132 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1133 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1134 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1135 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1136 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1137 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1138 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1139 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1140 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1141 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1142 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1143 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1144 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1145 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1146 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1147 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1148 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1149 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1150 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001151 "src/f32-hswish/gen/fma3-x8.c",
1152 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1154 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1155 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1156 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1157 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1158 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1159 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1160 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1161 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1162 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1163 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1164 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1165 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1166 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001167]
1168
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001169AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001170 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1171 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1172 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1173 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1174 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1175 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1176 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1177 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1178 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1179 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1180 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1181 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1182 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1183 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1184 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1185 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1186 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1187 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1188 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1189 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1190 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1191 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1192 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1193 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1194 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1195 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1196 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1197 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1198 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1199 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1200 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1201 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1202 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1203 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1204 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1205 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001206 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1207 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1208 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1209 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1210 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1211 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1212 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1213 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1214 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1215 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1216 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1217 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1218 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1219 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1220 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1221 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1222 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1223 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1224 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1225 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1226 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1227 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1228 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1229 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1230 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1231 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1232 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1233 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1234 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1235 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001236 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1237 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1238 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1239 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1240 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1241 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1242 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1243 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1244 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1245 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1246 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1247 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1248 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1249 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1250 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1251 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1252 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1253 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1254 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1255 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1256 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1257 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1258 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1259 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001260 "src/math/exp-avx2-p5.c",
1261 "src/math/exp-avx2-perm-p3.c",
1262 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001263 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001264 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001265 "src/math/sigmoid-avx2-rr2-p5-div.c",
1266 "src/math/sigmoid-avx2-rr1-p5-div.c",
1267 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1268 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1269 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1270 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001271]
1272
Marat Dukhan08c4a432019-10-03 09:29:21 -07001273AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001274 "src/f32-clamp/gen/avx512f-x16.c",
1275 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001276 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1277 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1278 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1279 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1280 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1281 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1282 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1283 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1284 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1285 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1286 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1287 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1288 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1289 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1290 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1291 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1292 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1293 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1294 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1295 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1296 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1297 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1298 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1299 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001300 "src/f32-hswish/gen/avx512f-x16.c",
1301 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001302 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1303 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1304 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1305 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1306 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1307 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001308 "src/f32-prelu/gen/avx512f-2x16.c",
1309 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001310 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1311 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1312 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1313 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1314 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1315 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1316 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1317 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1318 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1319 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1320 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1321 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1322 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1323 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1324 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1325 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1326 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1327 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1328 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1329 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1330 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1331 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1332 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1333 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1334 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1335 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1336 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1337 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1338 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1339 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1340 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1341 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1342 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1343 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1344 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1345 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001346 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001347 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1348 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1349 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1350 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1351 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1352 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1353 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1354 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1355 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1356 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1357 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1358 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1359 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1360 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1361 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1362 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1363 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1364 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1365 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1366 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1367 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1368 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1369 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1370 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1371 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1372 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1373 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1374 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001375 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001376 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1377 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1378 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1379 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1380 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1381 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1382 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1383 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1384 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1385 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1386 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1387 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1388 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1389 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1390 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1391 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1392 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1393 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1394 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1395 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1396 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1397 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1398 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1399 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001400 "src/math/exp-avx512f-p5-scalef.c",
1401 "src/math/exp-avx512f-p5.c",
1402 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001403 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001404 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001405]
1406
1407AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001408 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001409 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1410 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1411 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1412 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1413 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1414 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1415 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1416 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1417 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1418 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001419]
1420
1421AARCH64_ASM_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001422 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1423 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
1424 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1425 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1426 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1427 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1428 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1429 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1430 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1431 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1432 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1433 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1434 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1435 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1436 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1437 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1438 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1439 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1440 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1441 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1442 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1443 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1444 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
1445 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1446 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1447 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1448 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1449 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1450 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1451 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1452 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1453 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1454 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1455 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1456 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1457 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1458 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1459 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1460 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1461 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1462 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1463 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1464 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1465 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1466 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1467 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1468 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1469 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1470 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1471 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1472 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1473 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1474 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1475 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1476 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1477 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1478 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1479 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1480 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1481 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1482 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001483]
1484
Marat Dukhan1b354632020-03-23 12:50:22 -07001485INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001486 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001487 "src/xnnpack/argmaxpool.h",
1488 "src/xnnpack/avgpool.h",
1489 "src/xnnpack/clamp.h",
1490 "src/xnnpack/common.h",
1491 "src/xnnpack/conv.h",
1492 "src/xnnpack/dwconv.h",
1493 "src/xnnpack/gavgpool.h",
1494 "src/xnnpack/gemm.h",
1495 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001496 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001497 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001498 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001499 "src/xnnpack/lut.h",
1500 "src/xnnpack/math.h",
1501 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001502 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001503 "src/xnnpack/packx.h",
1504 "src/xnnpack/pad.h",
1505 "src/xnnpack/params.h",
1506 "src/xnnpack/pavgpool.h",
1507 "src/xnnpack/ppmm.h",
1508 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001509 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001510 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001511 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001512 "src/xnnpack/rmax.h",
1513 "src/xnnpack/scalar-utils.h",
1514 "src/xnnpack/spmm.h",
1515 "src/xnnpack/unpool.h",
1516 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001517 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001518 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001519 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001520 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001521 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001522 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001523 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001524]
1525
1526INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001527 "include/xnnpack.h",
1528 "src/xnnpack/allocator.h",
1529 "src/xnnpack/compute.h",
1530 "src/xnnpack/im2col.h",
1531 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001532 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001533 "src/xnnpack/operator.h",
1534 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001535 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001536 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001537 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001538 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001539]
1540
Marat Dukhan1b354632020-03-23 12:50:22 -07001541ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001542 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001543]
1544
Marat Dukhan1b354632020-03-23 12:50:22 -07001545MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001546 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001547 "include/xnnpack.h",
1548]
1549
Marat Dukhan1b354632020-03-23 12:50:22 -07001550MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001551 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001552 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001553 "src/xnnpack/requantization.h",
1554 "include/xnnpack.h",
1555]
1556
1557OPERATOR_TEST_PARAMS_HDRS = [
1558 "src/xnnpack/params.h",
1559 "src/xnnpack/common.h",
1560]
1561
1562WEIGHTS_PACK_HDRS = [
1563 "src/xnnpack/pack.h",
1564 "src/xnnpack/operator.h",
1565 "src/xnnpack/compute.h",
1566]
1567
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001568LOGGING_COPTS = select({
1569 # No logging in optimized mode
1570 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1571 # Full logging in debug mode
1572 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1573 # Error-only logging in default (fastbuild) mode
1574 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1575})
1576
1577LOGGING_HDRS = [
1578 "src/xnnpack/log.h",
1579]
1580
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001582 name = "tables",
1583 srcs = TABLE_SRCS,
1584 hdrs = INTERNAL_HDRS,
1585 copts = xnnpack_std_copts(),
1586)
1587
1588xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001589 name = "scalar_ukernels",
1590 srcs = SCALAR_UKERNELS,
1591 hdrs = INTERNAL_HDRS,
1592 aarch32_copts = ["-marm"],
1593 copts = xnnpack_std_copts(),
1594 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001595 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001596 "@FP16",
1597 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001598 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001599 ],
1600)
1601
1602xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001603 name = "wasm_ukernels",
1604 hdrs = INTERNAL_HDRS,
1605 copts = xnnpack_std_copts(),
1606 wasm_srcs = WASM_UKERNELS,
1607 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001608 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001609 "@FP16",
1610 "@FXdiv",
1611 "@pthreadpool",
1612 ],
1613)
1614
1615xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001616 name = "psimd_fastmath_ukernels",
1617 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001618 hdrs = INTERNAL_HDRS,
1619 aarch32_copts = [
1620 "-marm",
1621 "-mfpu=neon",
1622 ],
1623 copts = xnnpack_std_copts(),
1624 optimized_copts = [
1625 "-O3",
1626 "-ffast-math",
1627 ],
1628 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001629 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001630 "@FP16",
1631 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001632 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001633 ],
1634)
1635
1636xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001637 name = "psimd_accmath_ukernels",
1638 srcs = PSIMD_ACCMATH_UKERNELS,
1639 hdrs = INTERNAL_HDRS,
1640 aarch32_copts = [
1641 "-marm",
1642 "-mfpu=neon",
1643 ],
1644 copts = xnnpack_std_copts(),
1645 optimized_copts = [
1646 "-O3",
1647 ],
1648 deps = [
1649 ":tables",
1650 "@FP16",
1651 "@psimd",
1652 "@pthreadpool",
1653 ],
1654)
1655
1656xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001657 name = "neon_ukernels",
1658 hdrs = INTERNAL_HDRS,
1659 aarch32_copts = [
1660 "-marm",
1661 "-mfpu=neon",
1662 ],
1663 aarch32_srcs = NEON_UKERNELS,
1664 aarch64_srcs = NEON_UKERNELS,
1665 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001666 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001667 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001668 "@FP16",
1669 "@pthreadpool",
1670 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001671)
1672
1673xnnpack_cc_library(
1674 name = "neonfma_ukernels",
1675 hdrs = INTERNAL_HDRS,
1676 aarch32_copts = [
1677 "-marm",
1678 "-mfpu=neon-vfpv4",
1679 ],
1680 aarch32_srcs = NEONFMA_UKERNELS,
1681 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1682 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001683 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001684 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001685 "@FP16",
1686 "@pthreadpool",
1687 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001688)
1689
1690xnnpack_cc_library(
1691 name = "neonfp16arith_ukernels",
1692 hdrs = INTERNAL_HDRS,
1693 aarch64_copts = ["-march=armv8.2-a+fp16"],
1694 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1695 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001696 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001697 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001698 "@FP16",
1699 "@pthreadpool",
1700 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001701)
1702
1703xnnpack_cc_library(
1704 name = "sse2_ukernels",
1705 hdrs = INTERNAL_HDRS,
1706 copts = xnnpack_std_copts(),
1707 x86_copts = ["-msse2"],
1708 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001709 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001710 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001711 "@FP16",
1712 "@pthreadpool",
1713 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001714)
1715
1716xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001717 name = "ssse3_ukernels",
1718 hdrs = INTERNAL_HDRS,
1719 copts = xnnpack_std_copts(),
1720 x86_copts = ["-mssse3"],
1721 x86_srcs = SSSE3_UKERNELS,
1722 deps = [
1723 ":tables",
1724 "@FP16",
1725 "@pthreadpool",
1726 ],
1727)
1728
1729xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001730 name = "sse41_ukernels",
1731 hdrs = INTERNAL_HDRS,
1732 copts = xnnpack_std_copts(),
1733 x86_copts = ["-msse4.1"],
1734 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001735 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001736 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001737 "@FP16",
1738 "@pthreadpool",
1739 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001740)
1741
1742xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001743 name = "avx_ukernels",
1744 hdrs = INTERNAL_HDRS,
1745 copts = xnnpack_std_copts(),
1746 x86_copts = ["-mavx"],
1747 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001748 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001749 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001750 "@FP16",
1751 "@pthreadpool",
1752 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001753)
1754
1755xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001756 name = "fma3_ukernels",
1757 hdrs = INTERNAL_HDRS,
1758 copts = xnnpack_std_copts(),
1759 x86_copts = [
1760 "-mfma",
1761 ],
1762 x86_srcs = FMA3_UKERNELS,
1763 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001764 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001765 "@FP16",
1766 "@pthreadpool",
1767 ],
1768)
1769
1770xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001771 name = "avx2_ukernels",
1772 hdrs = INTERNAL_HDRS,
1773 copts = xnnpack_std_copts(),
1774 x86_copts = [
1775 "-mfma",
1776 "-mavx2",
1777 ],
1778 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001779 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001780 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001781 "@FP16",
1782 "@pthreadpool",
1783 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001784)
1785
1786xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001787 name = "avx512f_ukernels",
1788 hdrs = INTERNAL_HDRS,
1789 copts = xnnpack_std_copts(),
1790 x86_copts = ["-mavx512f"],
1791 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001792 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001793 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001794 "@FP16",
1795 "@pthreadpool",
1796 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001797)
1798
1799xnnpack_cc_library(
1800 name = "asm_ukernels",
1801 hdrs = ["src/xnnpack/assembly.h"],
1802 aarch32_srcs = AARCH32_ASM_UKERNELS,
1803 aarch64_srcs = AARCH64_ASM_UKERNELS,
1804)
1805
1806xnnpack_aggregate_library(
1807 name = "ukernels",
1808 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001809 ":psimd_fastmath_ukernels",
1810 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001811 ":neon_ukernels",
1812 ":neonfma_ukernels",
1813 ":asm_ukernels",
1814 ],
1815 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001816 ":psimd_fastmath_ukernels",
1817 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001818 ":neon_ukernels",
1819 ":neonfma_ukernels",
1820 ":neonfp16arith_ukernels",
1821 ":asm_ukernels",
1822 ],
1823 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001824 wasm_deps = [
1825 ":wasm_ukernels",
1826 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001827 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001828 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001829 ":psimd_fastmath_ukernels",
1830 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001831 ],
1832 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001833 ":psimd_fastmath_ukernels",
1834 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001836 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001837 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001839 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001840 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 ":avx512f_ukernels",
1842 ],
1843)
1844
1845xnnpack_cc_library(
1846 name = "im2col",
1847 srcs = ["src/im2col.c"],
1848 hdrs = [
1849 "src/xnnpack/common.h",
1850 "src/xnnpack/im2col.h",
1851 ],
1852 copts = xnnpack_std_copts(),
1853)
1854
1855xnnpack_cc_library(
1856 name = "indirection",
1857 srcs = ["src/indirection.c"],
1858 hdrs = INTERNAL_HDRS,
1859 copts = xnnpack_std_copts(),
1860 deps = [
1861 "@FP16",
1862 "@FXdiv",
1863 "@pthreadpool",
1864 ],
1865)
1866
1867xnnpack_cc_library(
1868 name = "operator_run",
1869 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001870 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1871 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001872 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1873 "-Wno-vla",
Marat Dukhan05702cf2020-03-26 15:41:33 -07001874 ] + select({
1875 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1876 "//conditions:default": [],
1877 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001878 deps = [
1879 "@FP16",
1880 "@FXdiv",
1881 "@clog",
1882 "@pthreadpool",
1883 ],
1884)
1885
1886cc_library(
1887 name = "enable_assembly",
1888 defines = select({
1889 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1890 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001891 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001892 }),
1893)
1894
Marat Dukhancf056b22019-10-07 10:26:29 -07001895xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001896 name = "operators",
1897 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001898 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001899 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001900 ],
1901 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001902 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001903 "-Isrc",
1904 "-Iinclude",
1905 ] + select({
1906 ":debug_build": [],
1907 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001908 }) + select({
1909 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1910 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001911 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001912 wasm_srcs = ["src/wasm-stubs.c"],
1913 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001914 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001915 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001916 "@FP16",
1917 "@FXdiv",
1918 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001919 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001920 ],
1921)
1922
1923cc_library(
1924 name = "XNNPACK",
1925 srcs = [
1926 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08001927 "src/runtime.c",
1928 "src/subgraph.c",
1929 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001930 ],
1931 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1932 "-Isrc",
1933 "-Iinclude",
1934 ] + select({
1935 ":debug_build": [],
1936 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001937 }) + select({
1938 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1939 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001940 }),
1941 includes = ["include"],
1942 linkstatic = True,
1943 textual_hdrs = ["include/xnnpack.h"],
1944 visibility = xnnpack_visibility(),
1945 deps = [
1946 ":enable_assembly",
1947 ":ukernels",
1948 ":operator_run",
1949 ":operators",
1950 "@clog",
1951 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001952 ] + select({
1953 ":emscripten": [],
1954 "//conditions:default": ["@cpuinfo"],
1955 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001956)
1957
1958cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001959 name = "xnnpack_operators_nhwc_f32",
1960 srcs = [
1961 "src/init.c",
1962 ],
1963 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1964 "-Isrc",
1965 "-Iinclude",
1966 ] + select({
1967 ":debug_build": [],
1968 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001969 }) + select({
1970 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1971 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001972 }),
1973 defines = [
1974 "XNN_NO_Q8_OPERATORS",
1975 "XNN_NO_U8_OPERATORS",
1976 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001977 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001978 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001979 includes = ["include"],
1980 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001981 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001982 visibility = xnnpack_visibility(),
1983 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001984 ":enable_assembly",
1985 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001986 ":operator_run",
1987 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001988 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001989 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001990 ] + select({
1991 ":emscripten": [],
1992 "//conditions:default": ["@cpuinfo"],
1993 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001994)
1995
Marat Dukhancf056b22019-10-07 10:26:29 -07001996xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997 name = "bench_utils",
1998 srcs = ["bench/utils.cc"],
1999 hdrs = ["bench/utils.h"],
2000 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002001 deps = [
2002 "@com_google_benchmark//:benchmark",
2003 "@cpuinfo",
2004 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002005)
2006
Frank Barchard7e955972019-10-11 10:34:25 -07002007######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002008
2009xnnpack_benchmark(
2010 name = "q8_gemm_bench",
2011 srcs = [
2012 "bench/gemm.h",
2013 "bench/q8-gemm.cc",
2014 "src/xnnpack/AlignedAllocator.h",
2015 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2016 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
2017 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2018)
2019
2020xnnpack_benchmark(
2021 name = "f16_gemm_bench",
2022 srcs = [
2023 "bench/f16-gemm.cc",
2024 "bench/gemm.h",
2025 "src/xnnpack/AlignedAllocator.h",
2026 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2027 copts = ["-Wno-unused-function"],
2028 deps = MICROKERNEL_BENCHMARK_DEPS,
2029)
2030
2031xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002032 name = "f16_spmm_bench",
2033 srcs = [
2034 "bench/f16-spmm.cc",
2035 "bench/gemm.h",
2036 "bench/google/gemm.h",
2037 "src/xnnpack/AlignedAllocator.h",
2038 ] + MICROKERNEL_BENCHMARK_HDRS,
2039 copts = ["-Wno-unused-function"],
2040 deps = MICROKERNEL_BENCHMARK_DEPS,
2041)
2042
2043xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002044 name = "f32_igemm_bench",
2045 srcs = [
2046 "bench/f32-igemm.cc",
2047 "bench/conv.h",
2048 "src/xnnpack/AlignedAllocator.h",
2049 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002050 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051)
2052
2053xnnpack_benchmark(
2054 name = "f32_conv_hwc_bench",
2055 srcs = [
2056 "bench/f32-conv-hwc.cc",
2057 "bench/dconv.h",
2058 "src/xnnpack/AlignedAllocator.h",
2059 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2060 copts = ["-Wno-unused-function"],
2061 deps = MICROKERNEL_BENCHMARK_DEPS,
2062)
2063
2064xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002065 name = "f32_conv_hwc2spchw_bench",
2066 srcs = [
2067 "bench/f32-conv-hwc2spchw.cc",
2068 "bench/dconv.h",
2069 "src/xnnpack/AlignedAllocator.h",
2070 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2071 copts = ["-Wno-unused-function"],
2072 deps = MICROKERNEL_BENCHMARK_DEPS,
2073)
2074
2075xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002076 name = "f32_dwconv_bench",
2077 srcs = [
2078 "bench/f32-dwconv.cc",
2079 "bench/dwconv.h",
2080 "src/xnnpack/AlignedAllocator.h",
2081 ] + MICROKERNEL_BENCHMARK_HDRS,
2082 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2083)
2084
2085xnnpack_benchmark(
2086 name = "f32_dwconv_spchw_bench",
2087 srcs = [
2088 "bench/f32-dwconv-spchw.cc",
2089 "bench/dwconv.h",
2090 "src/xnnpack/AlignedAllocator.h",
2091 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2092 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2093)
2094
2095xnnpack_benchmark(
2096 name = "f32_gemm_bench",
2097 srcs = [
2098 "bench/f32-gemm.cc",
2099 "bench/gemm.h",
2100 "src/xnnpack/AlignedAllocator.h",
2101 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2102 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002103 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002104)
2105
2106xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002107 name = "f32_raddexpminusmax_bench",
2108 srcs = [
2109 "bench/f32-raddexpminusmax.cc",
2110 "src/xnnpack/AlignedAllocator.h",
2111 ] + MICROKERNEL_BENCHMARK_HDRS,
2112 deps = MICROKERNEL_BENCHMARK_DEPS,
2113)
2114
2115xnnpack_benchmark(
2116 name = "f32_raddextexp_bench",
2117 srcs = [
2118 "bench/f32-raddextexp.cc",
2119 "src/xnnpack/AlignedAllocator.h",
2120 ] + MICROKERNEL_BENCHMARK_HDRS,
2121 deps = MICROKERNEL_BENCHMARK_DEPS,
2122)
2123
2124xnnpack_benchmark(
2125 name = "f32_raddstoreexpminusmax_bench",
2126 srcs = [
2127 "bench/f32-raddstoreexpminusmax.cc",
2128 "src/xnnpack/AlignedAllocator.h",
2129 ] + MICROKERNEL_BENCHMARK_HDRS,
2130 deps = MICROKERNEL_BENCHMARK_DEPS,
2131)
2132
2133xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002134 name = "f32_rmax_bench",
2135 srcs = [
2136 "bench/f32-rmax.cc",
2137 "src/xnnpack/AlignedAllocator.h",
2138 ] + MICROKERNEL_BENCHMARK_HDRS,
2139 deps = MICROKERNEL_BENCHMARK_DEPS,
2140)
2141
2142xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002143 name = "f32_sigmoid_bench",
2144 srcs = [
2145 "bench/f32-sigmoid.cc",
2146 "src/xnnpack/AlignedAllocator.h",
2147 ] + MICROKERNEL_BENCHMARK_HDRS,
2148 copts = ["-Wno-unused-function"],
2149 deps = MICROKERNEL_BENCHMARK_DEPS,
2150)
2151
2152xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002153 name = "f32_spmm_bench",
2154 srcs = [
2155 "bench/f32-spmm.cc",
2156 "bench/gemm.h",
2157 "src/xnnpack/AlignedAllocator.h",
2158 ] + MICROKERNEL_BENCHMARK_HDRS,
2159 copts = ["-Wno-unused-function"],
2160 deps = MICROKERNEL_BENCHMARK_DEPS,
2161)
2162
2163xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002164 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002165 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002166 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002167 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002168 copts = ["-Wno-unused-function"] + xnnpack_optional_dnnl_copts(),
2169 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002170)
2171
2172xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002173 name = "f32_vscaleexpminusmax_bench",
2174 srcs = [
2175 "bench/f32-vscaleexpminusmax.cc",
2176 "src/xnnpack/AlignedAllocator.h",
2177 ] + MICROKERNEL_BENCHMARK_HDRS,
2178 deps = MICROKERNEL_BENCHMARK_DEPS,
2179)
2180
2181xnnpack_benchmark(
2182 name = "f32_vscaleextexp_bench",
2183 srcs = [
2184 "bench/f32-vscaleextexp.cc",
2185 "src/xnnpack/AlignedAllocator.h",
2186 ] + MICROKERNEL_BENCHMARK_HDRS,
2187 deps = MICROKERNEL_BENCHMARK_DEPS,
2188)
2189
2190xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002191 name = "f32_im2col_gemm_bench",
2192 srcs = [
2193 "bench/f32-im2col-gemm.cc",
2194 "bench/conv.h",
2195 "src/xnnpack/AlignedAllocator.h",
2196 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2197 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2198)
2199
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002200xnnpack_benchmark(
2201 name = "requantization_bench",
2202 srcs = [
2203 "bench/requantization.cc",
2204 "src/xnnpack/requantization-stubs.h",
2205 "src/xnnpack/AlignedAllocator.h",
2206 ] + MICROKERNEL_BENCHMARK_HDRS,
2207 deps = MICROKERNEL_BENCHMARK_DEPS,
2208)
2209
Marat Dukhan08c4a432019-10-03 09:29:21 -07002210########################### Benchmarks for operators ###########################
2211
2212xnnpack_benchmark(
2213 name = "add_bench",
2214 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002215 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002216)
2217
2218xnnpack_benchmark(
2219 name = "average_pooling_bench",
2220 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002221 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002222 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002223)
2224
2225xnnpack_benchmark(
2226 name = "channel_shuffle_bench",
2227 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002228 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002229)
2230
2231xnnpack_benchmark(
2232 name = "convolution_bench",
2233 srcs = ["bench/convolution.cc"],
2234 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002235 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002236)
2237
2238xnnpack_benchmark(
2239 name = "deconvolution_bench",
2240 srcs = ["bench/deconvolution.cc"],
2241 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002242 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243)
2244
2245xnnpack_benchmark(
2246 name = "global_average_pooling_bench",
2247 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002248 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002249)
2250
2251xnnpack_benchmark(
2252 name = "max_pooling_bench",
2253 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002254 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002255)
2256
2257xnnpack_benchmark(
2258 name = "sigmoid_bench",
2259 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002260 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002261 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002262)
2263
2264xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002265 name = "prelu_bench",
2266 srcs = ["bench/prelu.cc"],
2267 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002268 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002269)
2270
2271xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002272 name = "softmax_bench",
2273 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002274 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan1b354632020-03-23 12:50:22 -07002275 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002276)
2277
Marat Dukhanc068bb62019-10-04 13:24:39 -07002278############################# End-to-end benchmarks ############################
2279
2280cc_library(
2281 name = "mobilenet_v1",
2282 srcs = ["models/mobilenet-v1.cc"],
2283 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002284 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002285 linkstatic = True,
2286 deps = [
2287 ":XNNPACK",
2288 "@pthreadpool",
2289 ],
2290)
2291
2292cc_library(
2293 name = "mobilenet_v2",
2294 srcs = ["models/mobilenet-v2.cc"],
2295 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002296 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002297 linkstatic = True,
2298 deps = [
2299 ":XNNPACK",
2300 "@pthreadpool",
2301 ],
2302)
2303
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002304cc_library(
2305 name = "mobilenet_v3_large",
2306 srcs = ["models/mobilenet-v3-large.cc"],
2307 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002308 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002309 linkstatic = True,
2310 deps = [
2311 ":XNNPACK",
2312 "@pthreadpool",
2313 ],
2314)
2315
2316cc_library(
2317 name = "mobilenet_v3_small",
2318 srcs = ["models/mobilenet-v3-small.cc"],
2319 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002320 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002321 linkstatic = True,
2322 deps = [
2323 ":XNNPACK",
2324 "@pthreadpool",
2325 ],
2326)
2327
Marat Dukhanc068bb62019-10-04 13:24:39 -07002328xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002329 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002330 srcs = [
2331 "bench/f32-dwconv-e2e.cc",
2332 "bench/end2end.h",
2333 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002334 copts = ["-Wno-unused-function"],
2335 deps = MICROKERNEL_BENCHMARK_DEPS + [
2336 ":XNNPACK",
2337 ":mobilenet_v1",
2338 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002339 ":mobilenet_v3_large",
2340 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002341 ],
2342)
2343
2344xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002345 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002346 srcs = [
2347 "bench/f32-gemm-e2e.cc",
2348 "bench/end2end.h",
2349 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002350 copts = ["-Wno-unused-function"],
2351 deps = MICROKERNEL_BENCHMARK_DEPS + [
2352 ":XNNPACK",
2353 ":mobilenet_v1",
2354 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002355 ":mobilenet_v3_large",
2356 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002357 ],
2358)
2359
2360xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002361 name = "end2end_bench",
2362 srcs = ["bench/end2end.cc"],
2363 deps = [
2364 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002365 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002366 ":mobilenet_v1",
2367 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002368 ":mobilenet_v3_large",
2369 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002370 "@pthreadpool",
2371 ],
2372)
2373
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002374#################### Accuracy evaluation for math functions ####################
2375
2376xnnpack_benchmark(
2377 name = "f32_exp_eval",
2378 srcs = [
2379 "eval/f32-exp.cc",
2380 "src/xnnpack/AlignedAllocator.h",
2381 ] + ACCURACY_EVAL_HDRS,
2382 deps = ACCURACY_EVAL_DEPS,
2383)
2384
Marat Dukhan515c9772019-10-17 18:07:57 -07002385xnnpack_benchmark(
2386 name = "f32_expminus_eval",
2387 srcs = [
2388 "eval/f32-expminus.cc",
2389 "src/xnnpack/AlignedAllocator.h",
2390 ] + ACCURACY_EVAL_HDRS,
2391 deps = ACCURACY_EVAL_DEPS,
2392)
2393
Marat Dukhan98ba4412019-10-23 02:14:28 -07002394xnnpack_benchmark(
2395 name = "f32_extexp_eval",
2396 srcs = [
2397 "eval/f32-extexp.cc",
2398 "src/xnnpack/AlignedAllocator.h",
2399 ] + ACCURACY_EVAL_HDRS,
2400 deps = ACCURACY_EVAL_DEPS,
2401)
2402
Marat Dukhan346a9e52019-11-15 09:06:30 -08002403xnnpack_benchmark(
2404 name = "f32_sigmoid_eval",
2405 srcs = [
2406 "eval/f32-sigmoid.cc",
2407 "src/xnnpack/AlignedAllocator.h",
2408 ] + ACCURACY_EVAL_HDRS,
2409 deps = ACCURACY_EVAL_DEPS,
2410)
2411
Marat Dukhan08c4a432019-10-03 09:29:21 -07002412######################### Unit tests for micro-kernels #########################
2413
2414xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002415 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002416 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002417 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002418 "test/gemm-microkernel-tester.h",
2419 "src/xnnpack/AlignedAllocator.h",
2420 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2421 deps = MICROKERNEL_TEST_DEPS,
2422)
2423
2424xnnpack_unit_test(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002425 name = "f16_spmm_test",
2426 srcs = [
2427 "test/f16-spmm.cc",
2428 "test/spmm-microkernel-tester.h",
2429 "src/xnnpack/AlignedAllocator.h",
2430 ] + MICROKERNEL_TEST_HDRS,
2431 deps = MICROKERNEL_TEST_DEPS,
2432)
2433
2434xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002435 name = "f32_argmaxpool_test",
2436 srcs = [
2437 "test/f32-argmaxpool.cc",
2438 "test/argmaxpool-microkernel-tester.h",
2439 "src/xnnpack/AlignedAllocator.h",
2440 ] + MICROKERNEL_TEST_HDRS,
2441 deps = MICROKERNEL_TEST_DEPS,
2442)
2443
2444xnnpack_unit_test(
2445 name = "f32_avgpool_test",
2446 srcs = [
2447 "test/f32-avgpool.cc",
2448 "test/avgpool-microkernel-tester.h",
2449 "src/xnnpack/AlignedAllocator.h",
2450 ] + MICROKERNEL_TEST_HDRS,
2451 deps = MICROKERNEL_TEST_DEPS,
2452)
2453
2454xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002455 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002456 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002457 "test/f32-ibilinear.cc",
2458 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002459 "src/xnnpack/AlignedAllocator.h",
2460 ] + MICROKERNEL_TEST_HDRS,
2461 deps = MICROKERNEL_TEST_DEPS,
2462)
2463
2464xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465 name = "f32_clamp_test",
2466 srcs = [
2467 "test/f32-clamp.cc",
2468 "test/clamp-microkernel-tester.h",
2469 ] + MICROKERNEL_TEST_HDRS,
2470 deps = MICROKERNEL_TEST_DEPS,
2471)
2472
2473xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002474 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002475 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002476 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002477 "test/gemm-microkernel-tester.h",
2478 "src/xnnpack/AlignedAllocator.h",
2479 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2480 deps = MICROKERNEL_TEST_DEPS,
2481)
2482
2483xnnpack_unit_test(
2484 name = "f32_conv_hwc_test",
2485 srcs = [
2486 "test/f32-conv-hwc.cc",
2487 "test/conv-hwc-microkernel-tester.h",
2488 "src/xnnpack/AlignedAllocator.h",
2489 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2490 deps = MICROKERNEL_TEST_DEPS,
2491)
2492
2493xnnpack_unit_test(
2494 name = "f32_conv_hwc2spchw_test",
2495 srcs = [
2496 "test/f32-conv-hwc2spchw.cc",
2497 "test/conv-hwc2spchw-microkernel-tester.h",
2498 "src/xnnpack/AlignedAllocator.h",
2499 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2500 deps = MICROKERNEL_TEST_DEPS,
2501)
2502
2503xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002504 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002505 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002506 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002507 "test/dwconv-microkernel-tester.h",
2508 "src/xnnpack/AlignedAllocator.h",
2509 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2510 deps = MICROKERNEL_TEST_DEPS,
2511)
2512
2513xnnpack_unit_test(
2514 name = "f32_dwconv_spchw_test",
2515 srcs = [
2516 "test/f32-dwconv-spchw.cc",
2517 "test/dwconv-spchw-microkernel-tester.h",
2518 "src/xnnpack/AlignedAllocator.h",
2519 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2520 deps = MICROKERNEL_TEST_DEPS,
2521)
2522
2523xnnpack_unit_test(
2524 name = "f32_gavgpool_test",
2525 srcs = [
2526 "test/f32-gavgpool.cc",
2527 "test/gavgpool-microkernel-tester.h",
2528 "src/xnnpack/AlignedAllocator.h",
2529 ] + MICROKERNEL_TEST_HDRS,
2530 deps = MICROKERNEL_TEST_DEPS,
2531)
2532
2533xnnpack_unit_test(
2534 name = "f32_gavgpool_spchw_test",
2535 srcs = [
2536 "test/f32-gavgpool-spchw.cc",
2537 "test/gavgpool-spchw-microkernel-tester.h",
2538 "src/xnnpack/AlignedAllocator.h",
2539 ] + MICROKERNEL_TEST_HDRS,
2540 deps = MICROKERNEL_TEST_DEPS,
2541)
2542
2543xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002544 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002545 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002546 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002547 "test/gemm-microkernel-tester.h",
2548 "src/xnnpack/AlignedAllocator.h",
2549 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2550 deps = MICROKERNEL_TEST_DEPS,
2551)
2552
2553xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002554 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002555 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002556 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002557 "test/gemm-microkernel-tester.h",
2558 "src/xnnpack/AlignedAllocator.h",
2559 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2560 deps = MICROKERNEL_TEST_DEPS,
2561)
2562
2563xnnpack_unit_test(
2564 name = "f32_hswish_test",
2565 srcs = [
2566 "test/f32-hswish.cc",
2567 "test/hswish-microkernel-tester.h",
2568 ] + MICROKERNEL_TEST_HDRS,
2569 deps = MICROKERNEL_TEST_DEPS,
2570)
2571
2572xnnpack_unit_test(
2573 name = "f32_maxpool_test",
2574 srcs = [
2575 "test/f32-maxpool.cc",
2576 "test/maxpool-microkernel-tester.h",
2577 ] + MICROKERNEL_TEST_HDRS,
2578 deps = MICROKERNEL_TEST_DEPS,
2579)
2580
2581xnnpack_unit_test(
2582 name = "f32_pavgpool_test",
2583 srcs = [
2584 "test/f32-pavgpool.cc",
2585 "test/avgpool-microkernel-tester.h",
2586 "src/xnnpack/AlignedAllocator.h",
2587 ] + MICROKERNEL_TEST_HDRS,
2588 deps = MICROKERNEL_TEST_DEPS,
2589)
2590
2591xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002592 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002593 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002594 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002595 "test/gemm-microkernel-tester.h",
2596 "src/xnnpack/AlignedAllocator.h",
2597 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2598 deps = MICROKERNEL_TEST_DEPS,
2599)
2600
2601xnnpack_unit_test(
2602 name = "f32_prelu_test",
2603 srcs = [
2604 "test/f32-prelu.cc",
2605 "test/prelu-microkernel-tester.h",
2606 "src/xnnpack/AlignedAllocator.h",
2607 ] + MICROKERNEL_TEST_HDRS,
2608 deps = MICROKERNEL_TEST_DEPS,
2609)
2610
2611xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002612 name = "f32_raddexpminusmax_test",
2613 srcs = [
2614 "test/f32-raddexpminusmax.cc",
2615 "test/raddexpminusmax-microkernel-tester.h",
2616 ] + MICROKERNEL_TEST_HDRS,
2617 deps = MICROKERNEL_TEST_DEPS,
2618)
2619
2620xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002621 name = "f32_raddextexp_test",
2622 srcs = [
2623 "test/f32-raddextexp.cc",
2624 "test/raddextexp-microkernel-tester.h",
2625 ] + MICROKERNEL_TEST_HDRS,
2626 deps = MICROKERNEL_TEST_DEPS,
2627)
2628
2629xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002630 name = "f32_raddstoreexpminusmax_test",
2631 srcs = [
2632 "test/f32-raddstoreexpminusmax.cc",
2633 "test/raddstoreexpminusmax-microkernel-tester.h",
2634 ] + MICROKERNEL_TEST_HDRS,
2635 deps = MICROKERNEL_TEST_DEPS,
2636)
2637
2638xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002639 name = "f32_rmax_test",
2640 srcs = [
2641 "test/f32-rmax.cc",
2642 "test/rmax-microkernel-tester.h",
2643 ] + MICROKERNEL_TEST_HDRS,
2644 deps = MICROKERNEL_TEST_DEPS,
2645)
2646
2647xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002648 name = "f32_sigmoid_test",
2649 srcs = [
2650 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002651 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002652 ] + MICROKERNEL_TEST_HDRS,
2653 deps = MICROKERNEL_TEST_DEPS,
2654)
2655
2656xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002657 name = "f32_spmm_test",
2658 srcs = [
2659 "test/f32-spmm.cc",
2660 "test/spmm-microkernel-tester.h",
2661 "src/xnnpack/AlignedAllocator.h",
2662 ] + MICROKERNEL_TEST_HDRS,
2663 deps = MICROKERNEL_TEST_DEPS,
2664)
2665
2666xnnpack_unit_test(
2667 name = "f32_vadd_test",
2668 srcs = [
2669 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002670 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002671 ] + MICROKERNEL_TEST_HDRS,
2672 deps = MICROKERNEL_TEST_DEPS,
2673)
2674
2675xnnpack_unit_test(
2676 name = "f32_vaddc_test",
2677 srcs = [
2678 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002679 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002680 ] + MICROKERNEL_TEST_HDRS,
2681 deps = MICROKERNEL_TEST_DEPS,
2682)
2683
2684xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002685 name = "f32_vdiv_test",
2686 srcs = [
2687 "test/f32-vdiv.cc",
2688 "test/vbinary-microkernel-tester.h",
2689 ] + MICROKERNEL_TEST_HDRS,
2690 deps = MICROKERNEL_TEST_DEPS,
2691)
2692
2693xnnpack_unit_test(
2694 name = "f32_vdivc_test",
2695 srcs = [
2696 "test/f32-vdivc.cc",
2697 "test/vbinaryc-microkernel-tester.h",
2698 ] + MICROKERNEL_TEST_HDRS,
2699 deps = MICROKERNEL_TEST_DEPS,
2700)
2701
2702xnnpack_unit_test(
2703 name = "f32_vrdivc_test",
2704 srcs = [
2705 "test/f32-vrdivc.cc",
2706 "test/vbinaryc-microkernel-tester.h",
2707 ] + MICROKERNEL_TEST_HDRS,
2708 deps = MICROKERNEL_TEST_DEPS,
2709)
2710
2711xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002712 name = "f32_vmax_test",
2713 srcs = [
2714 "test/f32-vmax.cc",
2715 "test/vbinary-microkernel-tester.h",
2716 ] + MICROKERNEL_TEST_HDRS,
2717 deps = MICROKERNEL_TEST_DEPS,
2718)
2719
2720xnnpack_unit_test(
2721 name = "f32_vmaxc_test",
2722 srcs = [
2723 "test/f32-vmaxc.cc",
2724 "test/vbinaryc-microkernel-tester.h",
2725 ] + MICROKERNEL_TEST_HDRS,
2726 deps = MICROKERNEL_TEST_DEPS,
2727)
2728
2729xnnpack_unit_test(
2730 name = "f32_vmin_test",
2731 srcs = [
2732 "test/f32-vmin.cc",
2733 "test/vbinary-microkernel-tester.h",
2734 ] + MICROKERNEL_TEST_HDRS,
2735 deps = MICROKERNEL_TEST_DEPS,
2736)
2737
2738xnnpack_unit_test(
2739 name = "f32_vminc_test",
2740 srcs = [
2741 "test/f32-vminc.cc",
2742 "test/vbinaryc-microkernel-tester.h",
2743 ] + MICROKERNEL_TEST_HDRS,
2744 deps = MICROKERNEL_TEST_DEPS,
2745)
2746
2747xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002748 name = "f32_vmul_test",
2749 srcs = [
2750 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002751 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002752 ] + MICROKERNEL_TEST_HDRS,
2753 deps = MICROKERNEL_TEST_DEPS,
2754)
2755
2756xnnpack_unit_test(
2757 name = "f32_vmulc_test",
2758 srcs = [
2759 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002760 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002761 ] + MICROKERNEL_TEST_HDRS,
2762 deps = MICROKERNEL_TEST_DEPS,
2763)
2764
2765xnnpack_unit_test(
2766 name = "f32_vmulcaddc_test",
2767 srcs = [
2768 "test/f32-vmulcaddc.cc",
2769 "test/vmulcaddc-microkernel-tester.h",
2770 "src/xnnpack/AlignedAllocator.h",
2771 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2772 deps = MICROKERNEL_TEST_DEPS,
2773)
2774
2775xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002776 name = "f32_vscale_test",
2777 srcs = [
2778 "test/f32-vscale.cc",
2779 "test/vscale-microkernel-tester.h",
2780 ] + MICROKERNEL_TEST_HDRS,
2781 deps = MICROKERNEL_TEST_DEPS,
2782)
2783
2784xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002785 name = "f32_vscaleexpminusmax_test",
2786 srcs = [
2787 "test/f32-vscaleexpminusmax.cc",
2788 "test/vscaleexpminusmax-microkernel-tester.h",
2789 ] + MICROKERNEL_TEST_HDRS,
2790 deps = MICROKERNEL_TEST_DEPS,
2791)
2792
2793xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002794 name = "f32_vscaleextexp_test",
2795 srcs = [
2796 "test/f32-vscaleextexp.cc",
2797 "test/vscaleextexp-microkernel-tester.h",
2798 ] + MICROKERNEL_TEST_HDRS,
2799 deps = MICROKERNEL_TEST_DEPS,
2800)
2801
2802xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002803 name = "f32_vsub_test",
2804 srcs = [
2805 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002806 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002807 ] + MICROKERNEL_TEST_HDRS,
2808 deps = MICROKERNEL_TEST_DEPS,
2809)
2810
2811xnnpack_unit_test(
2812 name = "f32_vsubc_test",
2813 srcs = [
2814 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002815 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002816 ] + MICROKERNEL_TEST_HDRS,
2817 deps = MICROKERNEL_TEST_DEPS,
2818)
2819
2820xnnpack_unit_test(
2821 name = "f32_vrsubc_test",
2822 srcs = [
2823 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002824 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002825 ] + MICROKERNEL_TEST_HDRS,
2826 deps = MICROKERNEL_TEST_DEPS,
2827)
2828
2829xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002830 name = "q8_avgpool_test",
2831 srcs = [
2832 "test/q8-avgpool.cc",
2833 "test/avgpool-microkernel-tester.h",
2834 "src/xnnpack/AlignedAllocator.h",
2835 ] + MICROKERNEL_TEST_HDRS,
2836 deps = MICROKERNEL_TEST_DEPS,
2837)
2838
2839xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002840 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002841 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002842 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002843 "test/gemm-microkernel-tester.h",
2844 "src/xnnpack/AlignedAllocator.h",
2845 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2846 deps = MICROKERNEL_TEST_DEPS,
2847)
2848
2849xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002850 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002851 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002852 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002853 "test/dwconv-microkernel-tester.h",
2854 "src/xnnpack/AlignedAllocator.h",
2855 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2856 deps = MICROKERNEL_TEST_DEPS,
2857)
2858
2859xnnpack_unit_test(
2860 name = "q8_gavgpool_test",
2861 srcs = [
2862 "test/q8-gavgpool.cc",
2863 "test/gavgpool-microkernel-tester.h",
2864 "src/xnnpack/AlignedAllocator.h",
2865 ] + MICROKERNEL_TEST_HDRS,
2866 deps = MICROKERNEL_TEST_DEPS,
2867)
2868
2869xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002870 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002871 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002872 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002873 "test/gemm-microkernel-tester.h",
2874 "src/xnnpack/AlignedAllocator.h",
2875 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2876 deps = MICROKERNEL_TEST_DEPS,
2877)
2878
2879xnnpack_unit_test(
2880 name = "q8_vadd_test",
2881 srcs = [
2882 "test/q8-vadd.cc",
2883 "test/vadd-microkernel-tester.h",
2884 ] + MICROKERNEL_TEST_HDRS,
2885 deps = MICROKERNEL_TEST_DEPS,
2886)
2887
2888xnnpack_unit_test(
2889 name = "u8_clamp_test",
2890 srcs = [
2891 "test/u8-clamp.cc",
2892 "test/clamp-microkernel-tester.h",
2893 ] + MICROKERNEL_TEST_HDRS,
2894 deps = MICROKERNEL_TEST_DEPS,
2895)
2896
2897xnnpack_unit_test(
2898 name = "u8_lut32norm_test",
2899 srcs = [
2900 "test/u8-lut32norm.cc",
2901 "test/lut-norm-microkernel-tester.h",
2902 ] + MICROKERNEL_TEST_HDRS,
2903 deps = MICROKERNEL_TEST_DEPS,
2904)
2905
2906xnnpack_unit_test(
2907 name = "u8_maxpool_test",
2908 srcs = [
2909 "test/u8-maxpool.cc",
2910 "test/maxpool-microkernel-tester.h",
2911 ] + MICROKERNEL_TEST_HDRS,
2912 deps = MICROKERNEL_TEST_DEPS,
2913)
2914
2915xnnpack_unit_test(
2916 name = "u8_rmax_test",
2917 srcs = [
2918 "test/u8-rmax.cc",
2919 "test/rmax-microkernel-tester.h",
2920 ] + MICROKERNEL_TEST_HDRS,
2921 deps = MICROKERNEL_TEST_DEPS,
2922)
2923
2924xnnpack_unit_test(
2925 name = "x32_packx_test",
2926 srcs = [
2927 "test/x32-packx.cc",
2928 "test/pack-microkernel-tester.h",
2929 "src/xnnpack/AlignedAllocator.h",
2930 ] + MICROKERNEL_TEST_HDRS,
2931 deps = MICROKERNEL_TEST_DEPS,
2932)
2933
2934xnnpack_unit_test(
2935 name = "x32_pad_test",
2936 srcs = [
2937 "test/x32-pad.cc",
2938 "test/pad-microkernel-tester.h",
2939 ] + MICROKERNEL_TEST_HDRS,
2940 deps = MICROKERNEL_TEST_DEPS,
2941)
2942
2943xnnpack_unit_test(
2944 name = "x32_unpool_test",
2945 srcs = [
2946 "test/x32-unpool.cc",
2947 "test/unpool-microkernel-tester.h",
2948 ] + MICROKERNEL_TEST_HDRS,
2949 deps = MICROKERNEL_TEST_DEPS,
2950)
2951
2952xnnpack_unit_test(
2953 name = "x32_zip_test",
2954 srcs = [
2955 "test/x32-zip.cc",
2956 "test/zip-microkernel-tester.h",
2957 ] + MICROKERNEL_TEST_HDRS,
2958 deps = MICROKERNEL_TEST_DEPS,
2959)
2960
2961xnnpack_unit_test(
2962 name = "x8_lut_test",
2963 srcs = [
2964 "test/x8-lut.cc",
2965 "test/lut-microkernel-tester.h",
2966 ] + MICROKERNEL_TEST_HDRS,
2967 deps = MICROKERNEL_TEST_DEPS,
2968)
2969
2970xnnpack_unit_test(
2971 name = "x8_zip_test",
2972 srcs = [
2973 "test/x8-zip.cc",
2974 "test/zip-microkernel-tester.h",
2975 ] + MICROKERNEL_TEST_HDRS,
2976 deps = MICROKERNEL_TEST_DEPS,
2977)
2978
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002979xnnpack_unit_test(
2980 name = "requantization_test",
2981 srcs = [
2982 "src/xnnpack/requantization-stubs.h",
2983 "test/requantization.cc",
2984 "test/requantization-tester.h",
2985 ] + MICROKERNEL_TEST_HDRS,
2986 deps = MICROKERNEL_TEST_DEPS,
2987)
2988
Marat Dukhan20c3b922020-03-10 03:45:06 -07002989########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002990
2991xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07002992 name = "operator_size_test",
2993 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002994 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995)
2996
Marat Dukhan20c3b922020-03-10 03:45:06 -07002997xnnpack_binary(
2998 name = "subgraph_size_test",
2999 srcs = ["test/subgraph-size.c"],
3000 deps = [":XNNPACK"],
3001)
3002
3003########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003004
3005xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003006 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003008 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009 "test/add-operator-tester.h",
3010 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003011 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003012)
3013
3014xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003015 name = "add_nd_test",
3016 srcs = [
3017 "test/add-nd.cc",
3018 "test/binary-elementwise-operator-tester.h",
3019 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003020 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003021)
3022
3023xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003024 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003026 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003027 "test/argmax-pooling-operator-tester.h",
3028 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003029 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030)
3031
3032xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003033 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003034 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003035 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003036 "test/average-pooling-operator-tester.h",
3037 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003038 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039)
3040
3041xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003042 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003044 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003045 "test/channel-pad-operator-tester.h",
3046 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003047 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003048)
3049
3050xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003051 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003052 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003053 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003054 "test/channel-shuffle-operator-tester.h",
3055 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003056 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003057)
3058
3059xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003060 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003061 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003062 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003063 "test/clamp-operator-tester.h",
3064 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003065 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003066)
3067
3068xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003069 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003070 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003071 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003072 "test/convolution-operator-tester.h",
3073 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003074 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003075)
3076
3077xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003078 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003079 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003080 "test/convolution-nchw.cc",
3081 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003082 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003083 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003084)
3085
3086xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003087 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003088 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003089 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003090 "test/deconvolution-operator-tester.h",
3091 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003092 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003093)
3094
3095xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003096 name = "divide_nd_test",
3097 srcs = [
3098 "test/binary-elementwise-operator-tester.h",
3099 "test/divide-nd.cc",
3100 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003101 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003102)
3103
3104xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003105 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003107 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003108 "test/fully-connected-operator-tester.h",
3109 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003110 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003111)
3112
3113xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003114 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003115 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003116 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003117 "test/global-average-pooling-operator-tester.h",
3118 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003119 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120)
3121
3122xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003123 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003124 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003125 "test/global-average-pooling-ncw.cc",
3126 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003127 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003128 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003129)
3130
3131xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003132 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003133 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003134 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003135 "test/hardswish-operator-tester.h",
3136 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003137 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003138)
3139
3140xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003141 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003143 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003144 "test/leaky-relu-operator-tester.h",
3145 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003146 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003147)
3148
3149xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003150 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003151 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003152 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003153 "test/max-pooling-operator-tester.h",
3154 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003155 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156)
3157
3158xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003159 name = "maximum_nd_test",
3160 srcs = [
3161 "test/binary-elementwise-operator-tester.h",
3162 "test/maximum-nd.cc",
3163 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003164 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003165)
3166
3167xnnpack_unit_test(
3168 name = "minimum_nd_test",
3169 srcs = [
3170 "test/binary-elementwise-operator-tester.h",
3171 "test/minimum-nd.cc",
3172 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003173 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003174)
3175
3176xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003177 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003178 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003179 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003180 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003181 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003182 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003183)
3184
3185xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003186 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003188 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189 "test/prelu-operator-tester.h",
3190 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003191 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003192)
3193
3194xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003195 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003196 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003197 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003198 "test/resize-bilinear-operator-tester.h",
3199 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003200 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003201)
3202
3203xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003204 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003205 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003206 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 "test/sigmoid-operator-tester.h",
3208 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003209 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210)
3211
3212xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003213 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003214 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003215 "test/softmax-nc.cc",
3216 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003217 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003218 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219)
3220
3221xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003222 name = "subtract_nd_test",
3223 srcs = [
3224 "test/binary-elementwise-operator-tester.h",
3225 "test/subtract-nd.cc",
3226 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003227 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003228)
3229
3230xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003231 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003232 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003233 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003234 "test/unpooling-operator-tester.h",
3235 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003236 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237)
3238
3239############################# Build configurations #############################
3240
Marat Dukhanb8642352019-10-30 15:43:02 -07003241# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003242config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003243 name = "xnn_enable_assembly_explicit_true",
3244 define_values = {"xnn_enable_assembly": "true"},
3245)
3246
3247# Disables usage of assembly kernels.
3248config_setting(
3249 name = "xnn_enable_assembly_explicit_false",
3250 define_values = {"xnn_enable_assembly": "false"},
3251)
3252
Marat Dukhan05702cf2020-03-26 15:41:33 -07003253# Disables usage of HMP-aware optimizations.
3254config_setting(
3255 name = "xnn_enable_hmp_explicit_false",
3256 define_values = {"xnn_enable_hmp": "false"},
3257)
3258
Marat Dukhanb8642352019-10-30 15:43:02 -07003259# Builds with -c dbg
3260config_setting(
3261 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003262 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003263 "compilation_mode": "dbg",
3264 },
3265)
3266
3267# Builds with -c opt
3268config_setting(
3269 name = "optimized_build",
3270 values = {
3271 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272 },
3273)
3274
3275config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003276 name = "linux_k8",
3277 values = {"cpu": "k8"},
3278)
3279
3280config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003281 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003282 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003283)
3284
3285config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003286 name = "android",
3287 values = {"crosstool_top": "//external:android/crosstool"},
3288)
3289
3290config_setting(
3291 name = "android_armv7",
3292 values = {
3293 "crosstool_top": "//external:android/crosstool",
3294 "cpu": "armeabi-v7a",
3295 },
3296)
3297
3298config_setting(
3299 name = "android_arm64",
3300 values = {
3301 "crosstool_top": "//external:android/crosstool",
3302 "cpu": "arm64-v8a",
3303 },
3304)
3305
3306config_setting(
3307 name = "android_x86",
3308 values = {
3309 "crosstool_top": "//external:android/crosstool",
3310 "cpu": "x86",
3311 },
3312)
3313
3314config_setting(
3315 name = "android_x86_64",
3316 values = {
3317 "crosstool_top": "//external:android/crosstool",
3318 "cpu": "x86_64",
3319 },
3320)
3321
3322config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003323 name = "macos_x86_64",
3324 values = {
3325 "apple_platform_type": "macos",
3326 "cpu": "darwin",
3327 },
3328)
3329
3330config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003331 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003332 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003333)
3334
3335config_setting(
3336 name = "emscripten_wasm",
3337 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003338 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003339 "cpu": "wasm",
3340 },
3341)
3342
3343config_setting(
3344 name = "emscripten_wasmsimd",
3345 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003346 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003347 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003348 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003349 },
3350)
3351
3352config_setting(
3353 name = "emscripten_asmjs",
3354 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003355 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003356 "cpu": "asmjs",
3357 },
3358)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003359
3360config_setting(
3361 name = "ios_armv7",
3362 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003363 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003364 "cpu": "ios_armv7",
3365 },
3366)
3367
3368config_setting(
3369 name = "ios_arm64",
3370 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003371 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003372 "cpu": "ios_arm64",
3373 },
3374)
3375
3376config_setting(
3377 name = "ios_arm64e",
3378 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003379 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003380 "cpu": "ios_arm64e",
3381 },
3382)
3383
3384config_setting(
3385 name = "ios_x86",
3386 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003387 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003388 "cpu": "ios_i386",
3389 },
3390)
3391
3392config_setting(
3393 name = "ios_x86_64",
3394 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003395 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003396 "cpu": "ios_x86_64",
3397 },
3398)
3399
3400config_setting(
3401 name = "watchos_armv7k",
3402 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003403 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003404 "cpu": "watchos_armv7k",
3405 },
3406)
3407
3408config_setting(
3409 name = "watchos_arm64_32",
3410 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003411 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003412 "cpu": "watchos_arm64_32",
3413 },
3414)
3415
3416config_setting(
3417 name = "watchos_x86",
3418 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003419 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003420 "cpu": "watchos_i386",
3421 },
3422)
3423
3424config_setting(
3425 name = "watchos_x86_64",
3426 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003427 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003428 "cpu": "watchos_x86_64",
3429 },
3430)
3431
3432config_setting(
3433 name = "tvos_arm64",
3434 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003435 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003436 "cpu": "tvos_arm64",
3437 },
3438)
3439
3440config_setting(
3441 name = "tvos_x86_64",
3442 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003443 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003444 "cpu": "tvos_x86_64",
3445 },
3446)