blob: 423d8108d15fca50910a1a65f78636291d10f5da [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
Marat Dukhan08c4a432019-10-03 09:29:21 -070022MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
Marat Dukhan1b354632020-03-23 12:50:22 -070046OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070053 "src/operators/add-nc.c",
54 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
57 "src/operators/channel-pad-nc.c",
58 "src/operators/channel-shuffle-nc.c",
59 "src/operators/clamp-nc.c",
60 "src/operators/convolution-nchw.c",
61 "src/operators/convolution-nhwc.c",
62 "src/operators/deconvolution-nhwc.c",
63 "src/operators/fully-connected-nc.c",
64 "src/operators/global-average-pooling-ncw.c",
65 "src/operators/global-average-pooling-nwc.c",
66 "src/operators/hardswish-nc.c",
67 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
70 "src/operators/resize-bilinear-nhwc.c",
71 "src/operators/sigmoid-nc.c",
72 "src/operators/softmax-nc.c",
73 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -070085 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
86 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -070087 "src/f32-clamp/gen/scalar-x1.c",
88 "src/f32-clamp/gen/scalar-x2.c",
89 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080090 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080091 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070092 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070093 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070094 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080095 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080096 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -070097 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x4-scalar.c",
99 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
100 "src/f32-dwconv/gen/up1x9-scalar.c",
101 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
102 "src/f32-dwconv/gen/up1x25-scalar.c",
103 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x4-scalar.c",
105 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
106 "src/f32-dwconv/gen/up2x9-scalar.c",
107 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
108 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700109 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
110 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
111 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
112 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700113 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
114 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700115 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
116 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
117 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
118 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700119 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
120 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700121 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700122 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
123 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700124 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
125 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
126 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700127 "src/f32-gemm/gen/1x4-scalar.c",
128 "src/f32-gemm/gen/2x4-scalar.c",
129 "src/f32-gemm/gen/4x2-scalar.c",
130 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700131 "src/f32-gemm/gen/1x4-minmax-scalar.c",
132 "src/f32-gemm/gen/2x4-minmax-scalar.c",
133 "src/f32-gemm/gen/4x2-minmax-scalar.c",
134 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800135 "src/f32-hswish/gen/scalar-x1.c",
136 "src/f32-hswish/gen/scalar-x2.c",
137 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700138 "src/f32-ibilinear/gen/scalar-c1.c",
139 "src/f32-ibilinear/gen/scalar-c2.c",
140 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-igemm/gen/1x4-scalar.c",
142 "src/f32-igemm/gen/2x4-scalar.c",
143 "src/f32-igemm/gen/4x2-scalar.c",
144 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-igemm/gen/1x4-minmax-scalar.c",
146 "src/f32-igemm/gen/2x4-minmax-scalar.c",
147 "src/f32-igemm/gen/4x2-minmax-scalar.c",
148 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700149 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
150 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
151 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700152 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
153 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
154 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
155 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800156 "src/f32-prelu/gen/scalar-2x1.c",
157 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800158 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
159 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
160 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
161 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
162 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
163 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
164 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
165 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
166 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
167 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
168 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
169 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700170 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800171 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
172 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
173 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
174 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
175 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
176 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
177 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
178 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
179 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700180 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
181 "src/f32-spmm/gen/1x1-minmax-scalar.c",
182 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
183 "src/f32-spmm/gen/2x1-minmax-scalar.c",
184 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
185 "src/f32-spmm/gen/4x1-minmax-scalar.c",
186 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
187 "src/f32-spmm/gen/8x1-minmax-scalar.c",
188 "src/f32-spmm/gen/8x2-minmax-scalar.c",
189 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700190 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
191 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
192 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
193 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
194 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
195 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
196 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
197 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
198 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
199 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
200 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
201 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800202 "src/f32-vbinary/gen/vmax-scalar-x1.c",
203 "src/f32-vbinary/gen/vmax-scalar-x2.c",
204 "src/f32-vbinary/gen/vmax-scalar-x4.c",
205 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
206 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
207 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
208 "src/f32-vbinary/gen/vmin-scalar-x1.c",
209 "src/f32-vbinary/gen/vmin-scalar-x2.c",
210 "src/f32-vbinary/gen/vmin-scalar-x4.c",
211 "src/f32-vbinary/gen/vminc-scalar-x1.c",
212 "src/f32-vbinary/gen/vminc-scalar-x2.c",
213 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700214 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
215 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
216 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
217 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
218 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
219 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
220 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
221 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
222 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
223 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
224 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
225 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
226 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
227 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
228 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
229 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700232 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
233 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
234 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800235 "src/math/expminus-scalar-lut2048-p1.c",
236 "src/math/expminus-scalar-lut64-p2.c",
237 "src/math/expminus-scalar-p5.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700238 "src/math/roundne-scalar.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800239 "src/math/sigmoid-scalar-lut2048-p1-div.c",
240 "src/math/sigmoid-scalar-lut64-p2-div.c",
241 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700242 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
243 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700244 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700245 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
246 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700247 "src/q8-gemm/2x2-minmax-scalar.c",
248 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700249 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700250 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700251 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700252 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700253 "src/u8-rmax/scalar.c",
254 "src/x32-packx/x2-scalar.c",
255 "src/x32-packx/x3-scalar.c",
256 "src/x32-packx/x4-scalar.c",
257 "src/x32-pad/x2-scalar.c",
258 "src/x32-unpool/scalar.c",
259 "src/x32-zip/x2-scalar.c",
260 "src/x32-zip/x3-scalar.c",
261 "src/x32-zip/x4-scalar.c",
262 "src/x32-zip/xm-scalar.c",
263 "src/x8-lut/scalar.c",
264 "src/x8-zip/x2-scalar.c",
265 "src/x8-zip/x3-scalar.c",
266 "src/x8-zip/x4-scalar.c",
267 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700268 "src/requantization/precise-scalar.c",
269 "src/requantization/fp32-scalar.c",
270 "src/requantization/q31-scalar.c",
271 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700272]
273
Marat Dukhan436ebe62019-12-04 15:10:12 -0800274WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700275 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
276 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700277 "src/f32-clamp/gen/wasm-x1.c",
278 "src/f32-clamp/gen/wasm-x2.c",
279 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700280 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
281 "src/f32-dwconv/gen/up1x4-wasm.c",
282 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
283 "src/f32-dwconv/gen/up1x9-wasm.c",
284 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
285 "src/f32-dwconv/gen/up1x25-wasm.c",
286 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
287 "src/f32-dwconv/gen/up2x4-wasm.c",
288 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
289 "src/f32-dwconv/gen/up2x9-wasm.c",
290 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
291 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700292 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
293 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
294 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
295 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700296 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
297 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700298 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
299 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
300 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
301 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700302 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
303 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700304 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
305 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700306 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
307 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
308 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700309 "src/f32-gemm/gen/1x4-wasm.c",
310 "src/f32-gemm/gen/2x4-wasm.c",
311 "src/f32-gemm/gen/4x2-wasm.c",
312 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700313 "src/f32-gemm/gen/1x4-minmax-wasm.c",
314 "src/f32-gemm/gen/2x4-minmax-wasm.c",
315 "src/f32-gemm/gen/4x2-minmax-wasm.c",
316 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800317 "src/f32-hswish/gen/wasm-x1.c",
318 "src/f32-hswish/gen/wasm-x2.c",
319 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700320 "src/f32-igemm/gen/1x4-wasm.c",
321 "src/f32-igemm/gen/2x4-wasm.c",
322 "src/f32-igemm/gen/4x2-wasm.c",
323 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700324 "src/f32-igemm/gen/1x4-minmax-wasm.c",
325 "src/f32-igemm/gen/2x4-minmax-wasm.c",
326 "src/f32-igemm/gen/4x2-minmax-wasm.c",
327 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700328 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
329 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
330 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700331 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
332 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
333 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
334 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
335 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
336 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
337 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
338 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
339 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
340 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
341 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
342 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vmax-wasm-x1.c",
344 "src/f32-vbinary/gen/vmax-wasm-x2.c",
345 "src/f32-vbinary/gen/vmax-wasm-x4.c",
346 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
347 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
348 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
349 "src/f32-vbinary/gen/vmin-wasm-x1.c",
350 "src/f32-vbinary/gen/vmin-wasm-x2.c",
351 "src/f32-vbinary/gen/vmin-wasm-x4.c",
352 "src/f32-vbinary/gen/vminc-wasm-x1.c",
353 "src/f32-vbinary/gen/vminc-wasm-x2.c",
354 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700355 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
356 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
357 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
358 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
359 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
360 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
361 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
362 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
363 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
364 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
365 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
366 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
367 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
368 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
369 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
370 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
371 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
372 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700373 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
374 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
375 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800376]
377
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800378PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800379 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800380 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800381 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700382 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
383 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700384 "src/f32-clamp/gen/psimd-x4.c",
385 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700386 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
387 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
388 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
389 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
390 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
391 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
392 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
393 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
394 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
395 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
396 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
397 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700398 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
399 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700400 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
401 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
402 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
403 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
404 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
405 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
406 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
407 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
408 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
409 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
410 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
411 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
412 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
413 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
414 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
415 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
416 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
417 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
418 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800419 "src/f32-hswish/gen/psimd-x4.c",
420 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700421 "src/f32-ibilinear/gen/psimd-c4.c",
422 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700423 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
424 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
425 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
426 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
427 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
428 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
429 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
430 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
431 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
432 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700433 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
434 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
435 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700436 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800437 "src/f32-prelu/gen/psimd-2x4.c",
438 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800439 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700440 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
441 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
442 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
443 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
444 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
445 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
446 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
447 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800448 "src/f32-vbinary/gen/vmax-psimd-x4.c",
449 "src/f32-vbinary/gen/vmax-psimd-x8.c",
450 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
451 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
452 "src/f32-vbinary/gen/vmin-psimd-x4.c",
453 "src/f32-vbinary/gen/vmin-psimd-x8.c",
454 "src/f32-vbinary/gen/vminc-psimd-x4.c",
455 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700456 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
457 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
458 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
459 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
460 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
461 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
462 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
463 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
464 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
465 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
466 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
467 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700468 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
469 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700470 "src/x32-packx/x4-psimd.c",
471 "src/x32-pad/x2-psimd.c",
472 "src/x32-unpool/psimd.c",
473 "src/x32-zip/x2-psimd.c",
474 "src/x32-zip/x3-psimd.c",
475 "src/x32-zip/x4-psimd.c",
476 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700477 "src/requantization/precise-psimd.c",
478 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700479]
480
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800481PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800482 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
483 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
484 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
485 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
486 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
487 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
488 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
489 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
490 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
491 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
492 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
493 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800494 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
495 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
496 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
497 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
498 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
499 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700500 "src/math/roundne-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800501 "src/math/sigmoid-psimd-p5-div.c",
502]
503
Marat Dukhan08c4a432019-10-03 09:29:21 -0700504# ISA-specific micro-kernels
505NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700506 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
507 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700508 "src/f32-clamp/gen/neon-x4.c",
509 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700510 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
511 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
512 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
513 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700514 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
515 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
516 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
517 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700518 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
519 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
520 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
521 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700522 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700523 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
524 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700525 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
526 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
527 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
528 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
529 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
530 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
531 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
532 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
533 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
534 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
535 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
536 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
537 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
538 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
539 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
540 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
541 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
542 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
543 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
544 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
545 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
546 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
547 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
548 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
549 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
550 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
551 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
552 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
553 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
554 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
555 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800556 "src/f32-hswish/gen/neon-x4.c",
557 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700558 "src/f32-ibilinear/gen/neon-c4.c",
559 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700560 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
561 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
562 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
563 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
564 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
565 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
566 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
567 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
568 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
569 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
570 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
571 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
572 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
573 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
574 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
575 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700576 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
577 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
578 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700579 "src/f32-ppmm/gen/4x8-minmax-neon.c",
580 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800581 "src/f32-prelu/gen/neon-2x4.c",
582 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800583 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
584 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
585 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
586 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
587 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
588 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
589 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
590 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
591 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
592 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
593 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
594 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
595 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
596 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
597 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
598 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
599 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
600 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
601 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
602 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
603 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
604 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
605 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
606 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700607 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800608 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800609 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
610 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
611 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
612 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
613 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
614 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
615 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
616 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
617 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
618 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
619 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
620 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
621 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
622 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
623 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
624 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
625 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
626 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700627 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
628 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
629 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
630 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800631 "src/f32-vbinary/gen/vmax-neon-x4.c",
632 "src/f32-vbinary/gen/vmax-neon-x8.c",
633 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
634 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
635 "src/f32-vbinary/gen/vmin-neon-x4.c",
636 "src/f32-vbinary/gen/vmin-neon-x8.c",
637 "src/f32-vbinary/gen/vminc-neon-x4.c",
638 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700639 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
640 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
641 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
642 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
643 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
644 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
645 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
646 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
647 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
648 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700649 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
650 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
651 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
652 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700653 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700654 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
655 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700656 "src/q8-gemm/4x8-minmax-neon.c",
657 "src/q8-gemm/8x8-minmax-neon.c",
658 "src/q8-igemm/4x8-minmax-neon.c",
659 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700660 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700661 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700662 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700663 "src/u8-rmax/neon.c",
664 "src/x32-packx/x4-neon-st4.c",
665 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700666 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700667 "src/x32-zip/x2-neon.c",
668 "src/x32-zip/x3-neon.c",
669 "src/x32-zip/x4-neon.c",
670 "src/x32-zip/xm-neon.c",
671 "src/x8-zip/x2-neon.c",
672 "src/x8-zip/x3-neon.c",
673 "src/x8-zip/x4-neon.c",
674 "src/x8-zip/xm-neon.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700675 "src/math/roundne-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800676 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800677 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
678 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
679 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
681 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
682 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700683 "src/requantization/precise-neon.c",
684 "src/requantization/fp32-neon.c",
685 "src/requantization/q31-neon.c",
686 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700687]
688
689NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700690 "src/f32-ibilinear/gen/neonfma-c4.c",
691 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700692 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
693 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
694 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
695 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
696 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
697 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
698 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
699 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
700 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700701 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
702 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
703 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
704 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700705 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
706 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
707 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
708 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700709 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
710 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
711 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
712 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700713 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
714 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
715 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
716 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
717 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
718 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
719 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
720 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
721 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
722 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
723 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
724 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
725 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
726 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
727 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
728 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
729 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
730 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800731 "src/f32-hswish/gen/neonfma-x4.c",
732 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700733 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
734 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800735 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
736 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
737 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
738 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
739 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
740 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
741 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
742 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
743 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
744 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
745 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
746 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
747 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
748 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
749 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
750 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
751 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
752 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
753 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
754 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
755 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
756 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
757 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
758 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800759 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
768 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
769 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
770 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
771 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
772 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
773 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
774 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
775 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
776 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
777 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
778 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
779 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700813 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
814 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800815 "src/math/exp-neonfma-lut64-p2.c",
816 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800817 "src/math/expminus-neonfma-lut2048-p1.c",
818 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800819 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800820 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
821 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
822 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
823 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
824 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
825 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
826 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
827 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
828 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
829 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
830 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
831 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
832 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
833 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
834 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
835 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
836 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
837 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700838]
839
840AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700841 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
842 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
843 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
844 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
845 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
846 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700847 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
848 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
849 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
850 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
851 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
852 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
853 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
854 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
855 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
856 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
857 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
858 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
859 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
860 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
861 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
862 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
863 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
864 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
865 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
866 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700867 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
868 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
869 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
870 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
871 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
872 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
873 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800874 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
875 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
876 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
877 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
878 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
879 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
880 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
881 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
882 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
883 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
884 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
885 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
886 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
887 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
888 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
889 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
890 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
891 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700892 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
893 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
894 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
895 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
896 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
897 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
898 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
899 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
900 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
901 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
902 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
903 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
904 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
905 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
906 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
907 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
908 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
909 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800910 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
911 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
912 "src/math/sigmoid-neonfma-rr1-p5-div.c",
913 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
914 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
915 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700916]
917
Marat Dukhan8853b822020-05-07 12:19:01 -0700918NEONV8_UKERNELS = [
919 "src/math/roundne-neonv8.c",
920]
921
Marat Dukhan08c4a432019-10-03 09:29:21 -0700922AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard83ea9392020-04-20 20:15:14 -0700923 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
924 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
925 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
926 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
927 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
928 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
929 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
930 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -0700931 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
932 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
933 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
934 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -0700935 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
936 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
937 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
938 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
939 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
940 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
941 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
942 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
943 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
944 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
945 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
946 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700947 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
948 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
949 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
950 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
951 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
952 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
953 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
954 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700955]
956
957SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700958 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
959 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700960 "src/f32-clamp/gen/sse-x4.c",
961 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800962 "src/f32-dwconv-spchw/3x3p1-sse.c",
963 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700964 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
965 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
966 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
967 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
968 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
969 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
970 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
971 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
972 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
973 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
974 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
975 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700976 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700977 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
978 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700979 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
980 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
981 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
982 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
983 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
984 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
985 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
986 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
987 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
988 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
989 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
990 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
991 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800992 "src/f32-hswish/gen/sse-x4.c",
993 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700994 "src/f32-ibilinear/gen/sse-c4.c",
995 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700996 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
997 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
998 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
999 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1000 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1001 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1002 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001003 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1004 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1005 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001006 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001007 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001008 "src/f32-spmm/gen/4x1-minmax-sse.c",
1009 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001010 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1011 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1012 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1013 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1014 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1015 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1016 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1017 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001018 "src/f32-vbinary/gen/vmax-sse-x4.c",
1019 "src/f32-vbinary/gen/vmax-sse-x8.c",
1020 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1021 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1022 "src/f32-vbinary/gen/vmin-sse-x4.c",
1023 "src/f32-vbinary/gen/vmin-sse-x8.c",
1024 "src/f32-vbinary/gen/vminc-sse-x4.c",
1025 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001026 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1027 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1028 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1029 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1030 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1031 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1032 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1033 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1034 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1035 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1036 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1037 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001038 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1039 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001040 "src/x32-packx/x4-sse.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001041 "src/math/roundne-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001042]
1043
1044SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001045 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1046 "src/f32-argmaxpool/4x-sse2-c4.c",
1047 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001048 "src/f32-prelu/gen/sse2-2x4.c",
1049 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001050 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1051 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1052 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1053 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1054 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1055 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1056 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1057 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1058 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1059 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1060 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1061 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001062 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001063 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001064 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001065 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001066 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1067 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001068 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1069 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001070 "src/q8-igemm/4x4c2-minmax-sse2.c",
1071 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001072 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1073 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001074 "src/q8-gemm/2x4c8-minmax-sse2.c",
1075 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001076 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001077 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001078 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001079 "src/u8-rmax/sse2.c",
1080 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001081 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001082 "src/x32-zip/x2-sse2.c",
1083 "src/x32-zip/x3-sse2.c",
1084 "src/x32-zip/x4-sse2.c",
1085 "src/x32-zip/xm-sse2.c",
1086 "src/x8-zip/x2-sse2.c",
1087 "src/x8-zip/x3-sse2.c",
1088 "src/x8-zip/x4-sse2.c",
1089 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001090 "src/math/exp-sse2-p5.c",
1091 "src/math/expminus-sse2-p5.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001092 "src/math/roundne-sse2.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001093 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001094 "src/requantization/precise-sse2.c",
1095 "src/requantization/fp32-sse2.c",
1096 "src/requantization/q31-sse2.c",
1097 "src/requantization/gemmlowp-sse2.c",
1098]
1099
1100SSSE3_UKERNELS = [
1101 "src/requantization/precise-ssse3.c",
1102 "src/requantization/q31-ssse3.c",
1103 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001104]
1105
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001106SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001107 "src/f32-prelu/gen/sse41-2x4.c",
1108 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001109 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001110 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001111 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001112 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001113 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1114 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001115 "src/math/roundne-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001116 "src/requantization/precise-sse4.c",
1117 "src/requantization/q31-sse4.c",
1118 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001119]
1120
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001122 "src/f32-clamp/gen/avx-x8.c",
1123 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001124 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1125 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1126 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1127 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1128 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1129 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1130 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1131 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1132 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1133 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1134 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1135 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1136 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1137 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1138 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1139 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1140 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1141 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1142 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1143 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1144 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1145 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1146 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1147 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1148 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1149 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1150 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1151 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1152 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1153 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001154 "src/f32-hswish/gen/avx-x8.c",
1155 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001156 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1157 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1158 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1159 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1160 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1161 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1162 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1163 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1164 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001165 "src/f32-prelu/gen/avx-2x8.c",
1166 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001167 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001168 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1169 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1170 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1171 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1172 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1173 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1174 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1175 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001176 "src/f32-vbinary/gen/vmax-avx-x8.c",
1177 "src/f32-vbinary/gen/vmax-avx-x16.c",
1178 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1179 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1180 "src/f32-vbinary/gen/vmin-avx-x8.c",
1181 "src/f32-vbinary/gen/vmin-avx-x16.c",
1182 "src/f32-vbinary/gen/vminc-avx-x8.c",
1183 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001184 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1185 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1186 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1187 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1188 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1189 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1190 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1191 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1192 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1193 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1194 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1195 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001196 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001197]
1198
Marat Dukhanfda12b82019-11-21 12:27:59 -08001199FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001200 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1201 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1202 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1203 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1204 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1205 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1206 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1207 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1208 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1209 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1210 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1211 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1212 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1213 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1214 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1215 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1216 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1217 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1218 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1219 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1220 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1221 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1222 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1223 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1224 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1225 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1226 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1227 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1228 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1229 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1230 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1231 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1232 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1233 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1234 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1235 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1236 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1237 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1238 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1239 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001240 "src/f32-hswish/gen/fma3-x8.c",
1241 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001242 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1243 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1244 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1245 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1246 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1247 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1248 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1249 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1250 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1251 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1252 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1253 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1254 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1255 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001256]
1257
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001258AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001259 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1260 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1261 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1262 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1263 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1264 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1265 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1266 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1267 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1268 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1269 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1270 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1271 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1272 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1273 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1274 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1275 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1276 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1277 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1278 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1279 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1280 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1281 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1282 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1283 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1284 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1285 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1286 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1287 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1288 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1289 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1290 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1291 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1292 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1293 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1294 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001295 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1296 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1297 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1298 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1299 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1300 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1301 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1302 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1303 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1304 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1305 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1306 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1307 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1308 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1309 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1310 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1311 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1312 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1313 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1314 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1315 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1316 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1317 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1318 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1319 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1320 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1321 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1322 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1323 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1324 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001325 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1326 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1327 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1328 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1329 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1330 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1331 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1332 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1333 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1334 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1335 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1336 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1337 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1338 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1339 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1340 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1341 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1342 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1343 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1344 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1345 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1346 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1347 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1348 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001349 "src/math/exp-avx2-p5.c",
1350 "src/math/exp-avx2-perm-p3.c",
1351 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001352 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001353 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001354 "src/math/sigmoid-avx2-rr2-p5-div.c",
1355 "src/math/sigmoid-avx2-rr1-p5-div.c",
1356 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1357 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1358 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1359 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001360]
1361
Marat Dukhan08c4a432019-10-03 09:29:21 -07001362AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001363 "src/f32-clamp/gen/avx512f-x16.c",
1364 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001365 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1366 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1367 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1368 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1369 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1370 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1371 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1372 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1373 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1374 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1375 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1376 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1377 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1378 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1379 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1380 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1381 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1382 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1383 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1384 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1385 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1386 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1387 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1388 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001389 "src/f32-hswish/gen/avx512f-x16.c",
1390 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001391 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1392 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1393 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1394 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1395 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1396 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001397 "src/f32-prelu/gen/avx512f-2x16.c",
1398 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001399 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1400 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1401 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1402 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1403 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1404 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1405 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1406 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1407 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1408 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1409 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1410 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1411 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1412 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1413 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1414 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1415 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1416 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1417 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1418 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1419 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1420 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1421 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1422 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1423 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1424 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1425 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1426 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1427 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1428 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1429 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1430 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1431 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1432 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1433 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1434 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001435 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001436 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1437 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1438 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1439 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1440 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1441 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1442 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1443 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001444 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1445 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1446 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1447 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1448 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1449 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1450 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1451 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001452 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1453 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1454 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1455 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1456 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1457 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1458 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1459 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1460 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1461 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1462 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1463 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001464 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001465 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1466 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1467 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1468 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1469 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1470 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1471 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1472 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1473 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1474 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1475 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1476 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1477 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1478 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1479 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1480 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1481 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1482 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1483 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1484 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1485 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1486 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1487 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1488 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001489 "src/math/exp-avx512f-p5-scalef.c",
1490 "src/math/exp-avx512f-p5.c",
1491 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001492 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001493 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001494]
1495
1496AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001497 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001498 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1499 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1500 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1501 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1502 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1503 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1504 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1505 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1506 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1507 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001508]
1509
1510AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001511 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001512 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1513 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001514 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001515 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1516 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001517 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1518 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1519 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001520 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001521 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1522 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1523 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001524 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001525 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1526 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001527 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001528 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1529 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1530 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1531 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1532 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1533 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1534 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1535 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1536 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1537 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1538 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1539 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1540 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1541 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1542 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1543 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1544 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1545 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1546 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1547 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1548 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001549 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001550 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1551 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1552 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1553 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1554 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1555 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1556 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1557 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1558 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1559 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1560 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1561 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1562 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1563 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1564 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1565 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1566 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1567 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1568 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1569 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1570 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1571 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1572 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1573 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1574 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1575 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1576 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1577 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1578 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1579 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1580 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1581 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1582 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1583 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1584 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1585 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1586 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1587 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001588]
1589
Marat Dukhan1b354632020-03-23 12:50:22 -07001590INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001591 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001592 "src/xnnpack/argmaxpool.h",
1593 "src/xnnpack/avgpool.h",
1594 "src/xnnpack/clamp.h",
1595 "src/xnnpack/common.h",
1596 "src/xnnpack/conv.h",
1597 "src/xnnpack/dwconv.h",
1598 "src/xnnpack/gavgpool.h",
1599 "src/xnnpack/gemm.h",
1600 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001601 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001602 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001603 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604 "src/xnnpack/lut.h",
1605 "src/xnnpack/math.h",
1606 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001607 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001608 "src/xnnpack/packx.h",
1609 "src/xnnpack/pad.h",
1610 "src/xnnpack/params.h",
1611 "src/xnnpack/pavgpool.h",
1612 "src/xnnpack/ppmm.h",
1613 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001614 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001615 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001616 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001617 "src/xnnpack/rmax.h",
1618 "src/xnnpack/scalar-utils.h",
1619 "src/xnnpack/spmm.h",
1620 "src/xnnpack/unpool.h",
1621 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001622 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001623 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001624 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001625 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001626 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001627 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001628 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001629]
1630
1631INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001632 "include/xnnpack.h",
1633 "src/xnnpack/allocator.h",
1634 "src/xnnpack/compute.h",
1635 "src/xnnpack/im2col.h",
1636 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001637 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001638 "src/xnnpack/operator.h",
1639 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001640 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001641 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001642 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001643 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001644]
1645
Marat Dukhan1b354632020-03-23 12:50:22 -07001646ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001647 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001648]
1649
Marat Dukhan1b354632020-03-23 12:50:22 -07001650MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001651 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001652 "include/xnnpack.h",
1653]
1654
Marat Dukhan1b354632020-03-23 12:50:22 -07001655MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001656 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001657 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001658 "src/xnnpack/requantization.h",
1659 "include/xnnpack.h",
1660]
1661
1662OPERATOR_TEST_PARAMS_HDRS = [
1663 "src/xnnpack/params.h",
1664 "src/xnnpack/common.h",
1665]
1666
1667WEIGHTS_PACK_HDRS = [
1668 "src/xnnpack/pack.h",
1669 "src/xnnpack/operator.h",
1670 "src/xnnpack/compute.h",
1671]
1672
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001673LOGGING_COPTS = select({
1674 # No logging in optimized mode
1675 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1676 # Full logging in debug mode
1677 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1678 # Error-only logging in default (fastbuild) mode
1679 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1680})
1681
1682LOGGING_HDRS = [
1683 "src/xnnpack/log.h",
1684]
1685
Marat Dukhan08c4a432019-10-03 09:29:21 -07001686xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001687 name = "tables",
1688 srcs = TABLE_SRCS,
1689 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001690 gcc_copts = xnnpack_gcc_std_copts(),
1691 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001692)
1693
1694xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001695 name = "scalar_ukernels",
1696 srcs = SCALAR_UKERNELS,
1697 hdrs = INTERNAL_HDRS,
1698 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001699 gcc_copts = xnnpack_gcc_std_copts(),
1700 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001701 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001702 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001703 "@FP16",
1704 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001705 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001706 ],
1707)
1708
1709xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001710 name = "wasm_ukernels",
1711 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001712 gcc_copts = xnnpack_gcc_std_copts(),
1713 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001714 wasm_srcs = WASM_UKERNELS,
1715 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001716 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001717 "@FP16",
1718 "@FXdiv",
1719 "@pthreadpool",
1720 ],
1721)
1722
1723xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001724 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001725 hdrs = INTERNAL_HDRS,
1726 aarch32_copts = [
1727 "-marm",
1728 "-mfpu=neon",
1729 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001730 gcc_copts = xnnpack_gcc_std_copts(),
1731 gcc_x86_copts = ["-msse2"],
1732 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001733 optimized_copts = [
1734 "-O3",
1735 "-ffast-math",
1736 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001737 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001738 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001739 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001740 "@FP16",
1741 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001742 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001743 ],
1744)
1745
1746xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001747 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001748 hdrs = INTERNAL_HDRS,
1749 aarch32_copts = [
1750 "-marm",
1751 "-mfpu=neon",
1752 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001753 gcc_copts = xnnpack_gcc_std_copts(),
1754 gcc_x86_copts = ["-msse2"],
1755 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001756 optimized_copts = [
1757 "-O3",
1758 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001759 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001760 deps = [
1761 ":tables",
1762 "@FP16",
1763 "@psimd",
1764 "@pthreadpool",
1765 ],
1766)
1767
1768xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769 name = "neon_ukernels",
1770 hdrs = INTERNAL_HDRS,
1771 aarch32_copts = [
1772 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001773 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001774 "-mfpu=neon",
1775 ],
1776 aarch32_srcs = NEON_UKERNELS,
1777 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001778 gcc_copts = xnnpack_gcc_std_copts(),
1779 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001780 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001781 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001782 "@FP16",
1783 "@pthreadpool",
1784 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001785)
1786
1787xnnpack_cc_library(
1788 name = "neonfma_ukernels",
1789 hdrs = INTERNAL_HDRS,
1790 aarch32_copts = [
1791 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001792 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001793 "-mfpu=neon-vfpv4",
1794 ],
1795 aarch32_srcs = NEONFMA_UKERNELS,
1796 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001797 gcc_copts = xnnpack_gcc_std_copts(),
1798 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001799 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001800 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001801 "@FP16",
1802 "@pthreadpool",
1803 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001804)
1805
1806xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001807 name = "neonv8_ukernels",
1808 hdrs = INTERNAL_HDRS,
1809 aarch32_copts = [
1810 "-marm",
1811 "-march=armv8-a",
1812 "-mfpu=neon-fp-armv8",
1813 ],
1814 aarch32_srcs = NEONV8_UKERNELS,
1815 aarch64_srcs = NEONV8_UKERNELS,
1816 gcc_copts = xnnpack_gcc_std_copts(),
1817 msvc_copts = xnnpack_msvc_std_copts(),
1818 deps = [
1819 ":tables",
1820 "@FP16",
1821 "@pthreadpool",
1822 ],
1823)
1824
1825xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001826 name = "neonfp16arith_ukernels",
1827 hdrs = INTERNAL_HDRS,
1828 aarch64_copts = ["-march=armv8.2-a+fp16"],
1829 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001830 gcc_copts = xnnpack_gcc_std_copts(),
1831 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001832 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001833 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001834 "@FP16",
1835 "@pthreadpool",
1836 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001837)
1838
1839xnnpack_cc_library(
1840 name = "sse2_ukernels",
1841 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001842 gcc_copts = xnnpack_gcc_std_copts(),
1843 gcc_x86_copts = ["-msse2"],
1844 msvc_copts = xnnpack_msvc_std_copts(),
1845 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001846 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001847 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001848 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001849 "@FP16",
1850 "@pthreadpool",
1851 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001852)
1853
1854xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001855 name = "ssse3_ukernels",
1856 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001857 gcc_copts = xnnpack_gcc_std_copts(),
1858 gcc_x86_copts = ["-mssse3"],
1859 msvc_copts = xnnpack_msvc_std_copts(),
1860 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001861 x86_srcs = SSSE3_UKERNELS,
1862 deps = [
1863 ":tables",
1864 "@FP16",
1865 "@pthreadpool",
1866 ],
1867)
1868
1869xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001870 name = "sse41_ukernels",
1871 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001872 gcc_copts = xnnpack_gcc_std_copts(),
1873 gcc_x86_copts = ["-msse4.1"],
1874 msvc_copts = xnnpack_msvc_std_copts(),
1875 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001876 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001877 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001878 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001879 "@FP16",
1880 "@pthreadpool",
1881 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001882)
1883
1884xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001885 name = "avx_ukernels",
1886 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001887 gcc_copts = xnnpack_gcc_std_copts(),
1888 gcc_x86_copts = ["-mavx"],
1889 msvc_copts = xnnpack_msvc_std_copts(),
1890 msvc_x86_32_copts = ["/arch:AVX"],
1891 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001892 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001893 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001894 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001895 "@FP16",
1896 "@pthreadpool",
1897 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001898)
1899
1900xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001901 name = "fma3_ukernels",
1902 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001903 gcc_copts = xnnpack_gcc_std_copts(),
1904 gcc_x86_copts = ["-mfma"],
1905 msvc_copts = xnnpack_msvc_std_copts(),
1906 msvc_x86_32_copts = ["/arch:AVX"],
1907 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001908 x86_srcs = FMA3_UKERNELS,
1909 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001910 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001911 "@FP16",
1912 "@pthreadpool",
1913 ],
1914)
1915
1916xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001917 name = "avx2_ukernels",
1918 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001919 gcc_copts = xnnpack_gcc_std_copts(),
1920 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001921 "-mfma",
1922 "-mavx2",
1923 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001924 msvc_copts = xnnpack_msvc_std_copts(),
1925 msvc_x86_32_copts = ["/arch:AVX2"],
1926 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001927 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001928 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001929 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001930 "@FP16",
1931 "@pthreadpool",
1932 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001933)
1934
1935xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001936 name = "avx512f_ukernels",
1937 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001938 gcc_copts = xnnpack_gcc_std_copts(),
1939 gcc_x86_copts = ["-mavx512f"],
1940 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1941 msvc_copts = xnnpack_msvc_std_copts(),
1942 msvc_x86_32_copts = ["/arch:AVX512"],
1943 msvc_x86_64_copts = ["/arch:AVX512"],
1944 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001945 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001946 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001947 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001948 "@FP16",
1949 "@pthreadpool",
1950 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001951)
1952
1953xnnpack_cc_library(
1954 name = "asm_ukernels",
1955 hdrs = ["src/xnnpack/assembly.h"],
1956 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001957 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001958 aarch64_srcs = AARCH64_ASM_UKERNELS,
1959)
1960
1961xnnpack_aggregate_library(
1962 name = "ukernels",
1963 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001964 ":neon_ukernels",
1965 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07001966 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001967 ":asm_ukernels",
1968 ],
1969 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001970 ":neon_ukernels",
1971 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07001972 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001973 ":neonfp16arith_ukernels",
1974 ":asm_ukernels",
1975 ],
1976 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07001977 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001978 ":psimd_fastmath_ukernels",
1979 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001980 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001981 wasm_deps = [":wasm_ukernels"],
1982 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001983 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001984 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001985 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001986 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001987 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001988 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001989 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001990 ":avx512f_ukernels",
1991 ],
1992)
1993
1994xnnpack_cc_library(
1995 name = "im2col",
1996 srcs = ["src/im2col.c"],
1997 hdrs = [
1998 "src/xnnpack/common.h",
1999 "src/xnnpack/im2col.h",
2000 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002001 gcc_copts = xnnpack_gcc_std_copts(),
2002 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002003)
2004
2005xnnpack_cc_library(
2006 name = "indirection",
2007 srcs = ["src/indirection.c"],
2008 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002009 gcc_copts = xnnpack_gcc_std_copts(),
2010 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002011 deps = [
2012 "@FP16",
2013 "@FXdiv",
2014 "@pthreadpool",
2015 ],
2016)
2017
2018xnnpack_cc_library(
2019 name = "operator_run",
2020 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002021 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002022 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002023 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2024 "//conditions:default": [],
2025 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002026 gcc_copts = xnnpack_gcc_std_copts(),
2027 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002028 deps = [
2029 "@FP16",
2030 "@FXdiv",
2031 "@clog",
2032 "@pthreadpool",
2033 ],
2034)
2035
2036cc_library(
2037 name = "enable_assembly",
2038 defines = select({
2039 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2040 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002041 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002042 }),
2043)
2044
Marat Dukhancf056b22019-10-07 10:26:29 -07002045xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002046 name = "operators",
2047 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002048 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002049 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002050 ],
2051 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002052 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002053 "-Isrc",
2054 "-Iinclude",
2055 ] + select({
2056 ":debug_build": [],
2057 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002058 }) + select({
2059 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2060 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002061 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002062 gcc_copts = xnnpack_gcc_std_copts(),
2063 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002064 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002065 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002066 "@FP16",
2067 "@FXdiv",
2068 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002069 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002070 ],
2071)
2072
Marat Dukhan10a38082020-04-17 03:58:35 -07002073xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002074 name = "XNNPACK",
2075 srcs = [
2076 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002077 "src/runtime.c",
2078 "src/subgraph.c",
2079 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002080 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002081 hdrs = ["include/xnnpack.h"],
2082 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002083 "-Isrc",
2084 "-Iinclude",
2085 ] + select({
2086 ":debug_build": [],
2087 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002088 }) + select({
2089 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2090 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002091 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002092 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002093 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002094 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002095 visibility = xnnpack_visibility(),
2096 deps = [
2097 ":enable_assembly",
2098 ":ukernels",
2099 ":operator_run",
2100 ":operators",
2101 "@clog",
2102 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002103 ] + select({
2104 ":emscripten": [],
2105 "//conditions:default": ["@cpuinfo"],
2106 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002107)
2108
Marat Dukhan10a38082020-04-17 03:58:35 -07002109xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002110 name = "xnnpack_operators_nhwc_f32",
2111 srcs = [
2112 "src/init.c",
2113 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002114 hdrs = ["include/xnnpack.h"],
2115 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002116 "-Isrc",
2117 "-Iinclude",
2118 ] + select({
2119 ":debug_build": [],
2120 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002121 }) + select({
2122 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2123 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002124 }),
2125 defines = [
2126 "XNN_NO_Q8_OPERATORS",
2127 "XNN_NO_U8_OPERATORS",
2128 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002129 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002130 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002131 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002132 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002133 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002134 visibility = xnnpack_visibility(),
2135 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002136 ":enable_assembly",
2137 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002138 ":operator_run",
2139 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002140 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002141 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002142 ] + select({
2143 ":emscripten": [],
2144 "//conditions:default": ["@cpuinfo"],
2145 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002146)
2147
Marat Dukhancf056b22019-10-07 10:26:29 -07002148xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002149 name = "bench_utils",
2150 srcs = ["bench/utils.cc"],
2151 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002152 deps = [
2153 "@com_google_benchmark//:benchmark",
2154 "@cpuinfo",
2155 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002156)
2157
Frank Barchard7e955972019-10-11 10:34:25 -07002158######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002159
2160xnnpack_benchmark(
2161 name = "q8_gemm_bench",
2162 srcs = [
2163 "bench/gemm.h",
2164 "bench/q8-gemm.cc",
2165 "src/xnnpack/AlignedAllocator.h",
2166 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002167 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002168 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2169)
2170
2171xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002172 name = "f16_igemm_bench",
2173 srcs = [
2174 "bench/f16-igemm.cc",
2175 "bench/conv.h",
2176 "bench/google/conv.h",
2177 "src/xnnpack/AlignedAllocator.h",
2178 ] + MICROKERNEL_BENCHMARK_HDRS,
2179 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2180)
2181
2182xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183 name = "f16_gemm_bench",
2184 srcs = [
2185 "bench/f16-gemm.cc",
2186 "bench/gemm.h",
2187 "src/xnnpack/AlignedAllocator.h",
2188 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002189 deps = MICROKERNEL_BENCHMARK_DEPS,
2190)
2191
2192xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002193 name = "f16_spmm_bench",
2194 srcs = [
2195 "bench/f16-spmm.cc",
2196 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002197 "src/xnnpack/AlignedAllocator.h",
2198 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002199 deps = MICROKERNEL_BENCHMARK_DEPS,
2200)
2201
2202xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002203 name = "f32_igemm_bench",
2204 srcs = [
2205 "bench/f32-igemm.cc",
2206 "bench/conv.h",
2207 "src/xnnpack/AlignedAllocator.h",
2208 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002209 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002210)
2211
2212xnnpack_benchmark(
2213 name = "f32_conv_hwc_bench",
2214 srcs = [
2215 "bench/f32-conv-hwc.cc",
2216 "bench/dconv.h",
2217 "src/xnnpack/AlignedAllocator.h",
2218 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002219 deps = MICROKERNEL_BENCHMARK_DEPS,
2220)
2221
2222xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002223 name = "f32_conv_hwc2spchw_bench",
2224 srcs = [
2225 "bench/f32-conv-hwc2spchw.cc",
2226 "bench/dconv.h",
2227 "src/xnnpack/AlignedAllocator.h",
2228 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002229 deps = MICROKERNEL_BENCHMARK_DEPS,
2230)
2231
2232xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002233 name = "f32_dwconv_bench",
2234 srcs = [
2235 "bench/f32-dwconv.cc",
2236 "bench/dwconv.h",
2237 "src/xnnpack/AlignedAllocator.h",
2238 ] + MICROKERNEL_BENCHMARK_HDRS,
2239 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2240)
2241
2242xnnpack_benchmark(
2243 name = "f32_dwconv_spchw_bench",
2244 srcs = [
2245 "bench/f32-dwconv-spchw.cc",
2246 "bench/dwconv.h",
2247 "src/xnnpack/AlignedAllocator.h",
2248 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2249 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2250)
2251
2252xnnpack_benchmark(
2253 name = "f32_gemm_bench",
2254 srcs = [
2255 "bench/f32-gemm.cc",
2256 "bench/gemm.h",
2257 "src/xnnpack/AlignedAllocator.h",
2258 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002259 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002260 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002261)
2262
2263xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002264 name = "f32_raddexpminusmax_bench",
2265 srcs = [
2266 "bench/f32-raddexpminusmax.cc",
2267 "src/xnnpack/AlignedAllocator.h",
2268 ] + MICROKERNEL_BENCHMARK_HDRS,
2269 deps = MICROKERNEL_BENCHMARK_DEPS,
2270)
2271
2272xnnpack_benchmark(
2273 name = "f32_raddextexp_bench",
2274 srcs = [
2275 "bench/f32-raddextexp.cc",
2276 "src/xnnpack/AlignedAllocator.h",
2277 ] + MICROKERNEL_BENCHMARK_HDRS,
2278 deps = MICROKERNEL_BENCHMARK_DEPS,
2279)
2280
2281xnnpack_benchmark(
2282 name = "f32_raddstoreexpminusmax_bench",
2283 srcs = [
2284 "bench/f32-raddstoreexpminusmax.cc",
2285 "src/xnnpack/AlignedAllocator.h",
2286 ] + MICROKERNEL_BENCHMARK_HDRS,
2287 deps = MICROKERNEL_BENCHMARK_DEPS,
2288)
2289
2290xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002291 name = "f32_rmax_bench",
2292 srcs = [
2293 "bench/f32-rmax.cc",
2294 "src/xnnpack/AlignedAllocator.h",
2295 ] + MICROKERNEL_BENCHMARK_HDRS,
2296 deps = MICROKERNEL_BENCHMARK_DEPS,
2297)
2298
2299xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002300 name = "f32_sigmoid_bench",
2301 srcs = [
2302 "bench/f32-sigmoid.cc",
2303 "src/xnnpack/AlignedAllocator.h",
2304 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002305 deps = MICROKERNEL_BENCHMARK_DEPS,
2306)
2307
2308xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002309 name = "f32_spmm_bench",
2310 srcs = [
2311 "bench/f32-spmm.cc",
2312 "bench/gemm.h",
2313 "src/xnnpack/AlignedAllocator.h",
2314 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002315 deps = MICROKERNEL_BENCHMARK_DEPS,
2316)
2317
2318xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002319 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002320 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002321 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002322 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002323 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002324 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002325)
2326
2327xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002328 name = "f32_vscaleexpminusmax_bench",
2329 srcs = [
2330 "bench/f32-vscaleexpminusmax.cc",
2331 "src/xnnpack/AlignedAllocator.h",
2332 ] + MICROKERNEL_BENCHMARK_HDRS,
2333 deps = MICROKERNEL_BENCHMARK_DEPS,
2334)
2335
2336xnnpack_benchmark(
2337 name = "f32_vscaleextexp_bench",
2338 srcs = [
2339 "bench/f32-vscaleextexp.cc",
2340 "src/xnnpack/AlignedAllocator.h",
2341 ] + MICROKERNEL_BENCHMARK_HDRS,
2342 deps = MICROKERNEL_BENCHMARK_DEPS,
2343)
2344
2345xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002346 name = "f32_im2col_gemm_bench",
2347 srcs = [
2348 "bench/f32-im2col-gemm.cc",
2349 "bench/conv.h",
2350 "src/xnnpack/AlignedAllocator.h",
2351 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2352 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2353)
2354
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002355xnnpack_benchmark(
2356 name = "requantization_bench",
2357 srcs = [
2358 "bench/requantization.cc",
2359 "src/xnnpack/requantization-stubs.h",
2360 "src/xnnpack/AlignedAllocator.h",
2361 ] + MICROKERNEL_BENCHMARK_HDRS,
2362 deps = MICROKERNEL_BENCHMARK_DEPS,
2363)
2364
Marat Dukhan08c4a432019-10-03 09:29:21 -07002365########################### Benchmarks for operators ###########################
2366
2367xnnpack_benchmark(
2368 name = "add_bench",
2369 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002370 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002371)
2372
2373xnnpack_benchmark(
2374 name = "average_pooling_bench",
2375 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002376 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002377 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002378 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002379)
2380
2381xnnpack_benchmark(
2382 name = "channel_shuffle_bench",
2383 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002384 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002385)
2386
2387xnnpack_benchmark(
2388 name = "convolution_bench",
2389 srcs = ["bench/convolution.cc"],
2390 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002391 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002392 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002393)
2394
2395xnnpack_benchmark(
2396 name = "deconvolution_bench",
2397 srcs = ["bench/deconvolution.cc"],
2398 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002399 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002400 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002401)
2402
2403xnnpack_benchmark(
2404 name = "global_average_pooling_bench",
2405 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002406 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002407)
2408
2409xnnpack_benchmark(
2410 name = "max_pooling_bench",
2411 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002412 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002413)
2414
2415xnnpack_benchmark(
2416 name = "sigmoid_bench",
2417 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002418 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002419 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002420 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002421)
2422
2423xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002424 name = "prelu_bench",
2425 srcs = ["bench/prelu.cc"],
2426 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002427 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002428 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002429)
2430
2431xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002432 name = "softmax_bench",
2433 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002434 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002435 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002436 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002437)
2438
Marat Dukhanc068bb62019-10-04 13:24:39 -07002439############################# End-to-end benchmarks ############################
2440
2441cc_library(
2442 name = "mobilenet_v1",
2443 srcs = ["models/mobilenet-v1.cc"],
2444 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002445 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002446 linkstatic = True,
2447 deps = [
2448 ":XNNPACK",
2449 "@pthreadpool",
2450 ],
2451)
2452
2453cc_library(
2454 name = "mobilenet_v2",
2455 srcs = ["models/mobilenet-v2.cc"],
2456 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002457 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002458 linkstatic = True,
2459 deps = [
2460 ":XNNPACK",
2461 "@pthreadpool",
2462 ],
2463)
2464
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002465cc_library(
2466 name = "mobilenet_v3_large",
2467 srcs = ["models/mobilenet-v3-large.cc"],
2468 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002469 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002470 linkstatic = True,
2471 deps = [
2472 ":XNNPACK",
2473 "@pthreadpool",
2474 ],
2475)
2476
2477cc_library(
2478 name = "mobilenet_v3_small",
2479 srcs = ["models/mobilenet-v3-small.cc"],
2480 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002481 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002482 linkstatic = True,
2483 deps = [
2484 ":XNNPACK",
2485 "@pthreadpool",
2486 ],
2487)
2488
Marat Dukhanc068bb62019-10-04 13:24:39 -07002489xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002490 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002491 srcs = [
2492 "bench/f32-dwconv-e2e.cc",
2493 "bench/end2end.h",
2494 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002495 deps = MICROKERNEL_BENCHMARK_DEPS + [
2496 ":XNNPACK",
2497 ":mobilenet_v1",
2498 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002499 ":mobilenet_v3_large",
2500 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002501 ],
2502)
2503
2504xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002505 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002506 srcs = [
2507 "bench/f32-gemm-e2e.cc",
2508 "bench/end2end.h",
2509 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002510 deps = MICROKERNEL_BENCHMARK_DEPS + [
2511 ":XNNPACK",
2512 ":mobilenet_v1",
2513 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002514 ":mobilenet_v3_large",
2515 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002516 ],
2517)
2518
2519xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002520 name = "end2end_bench",
2521 srcs = ["bench/end2end.cc"],
2522 deps = [
2523 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002524 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002525 ":mobilenet_v1",
2526 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002527 ":mobilenet_v3_large",
2528 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002529 "@pthreadpool",
2530 ],
2531)
2532
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002533#################### Accuracy evaluation for math functions ####################
2534
2535xnnpack_benchmark(
2536 name = "f32_exp_eval",
2537 srcs = [
2538 "eval/f32-exp.cc",
2539 "src/xnnpack/AlignedAllocator.h",
2540 ] + ACCURACY_EVAL_HDRS,
2541 deps = ACCURACY_EVAL_DEPS,
2542)
2543
Marat Dukhan515c9772019-10-17 18:07:57 -07002544xnnpack_benchmark(
2545 name = "f32_expminus_eval",
2546 srcs = [
2547 "eval/f32-expminus.cc",
2548 "src/xnnpack/AlignedAllocator.h",
2549 ] + ACCURACY_EVAL_HDRS,
2550 deps = ACCURACY_EVAL_DEPS,
2551)
2552
Marat Dukhan98ba4412019-10-23 02:14:28 -07002553xnnpack_benchmark(
2554 name = "f32_extexp_eval",
2555 srcs = [
2556 "eval/f32-extexp.cc",
2557 "src/xnnpack/AlignedAllocator.h",
2558 ] + ACCURACY_EVAL_HDRS,
2559 deps = ACCURACY_EVAL_DEPS,
2560)
2561
Marat Dukhan8853b822020-05-07 12:19:01 -07002562xnnpack_unit_test(
2563 name = "f32_roundne_eval",
2564 srcs = [
2565 "eval/f32-roundne.cc",
2566 "src/xnnpack/AlignedAllocator.h",
2567 "src/xnnpack/math-stubs.h",
2568 ] + MICROKERNEL_TEST_HDRS,
2569 tags = [
2570 "notap",
2571 ],
2572 deps = MICROKERNEL_TEST_DEPS,
2573)
2574
Marat Dukhan346a9e52019-11-15 09:06:30 -08002575xnnpack_benchmark(
2576 name = "f32_sigmoid_eval",
2577 srcs = [
2578 "eval/f32-sigmoid.cc",
2579 "src/xnnpack/AlignedAllocator.h",
2580 ] + ACCURACY_EVAL_HDRS,
2581 deps = ACCURACY_EVAL_DEPS,
2582)
2583
Marat Dukhan08c4a432019-10-03 09:29:21 -07002584######################### Unit tests for micro-kernels #########################
2585
2586xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002587 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002588 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002589 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002590 "test/gemm-microkernel-tester.h",
2591 "src/xnnpack/AlignedAllocator.h",
2592 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2593 deps = MICROKERNEL_TEST_DEPS,
2594)
2595
2596xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002597 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002598 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002599 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002600 "test/spmm-microkernel-tester.h",
2601 "src/xnnpack/AlignedAllocator.h",
2602 ] + MICROKERNEL_TEST_HDRS,
2603 deps = MICROKERNEL_TEST_DEPS,
2604)
2605
2606xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002607 name = "f32_argmaxpool_test",
2608 srcs = [
2609 "test/f32-argmaxpool.cc",
2610 "test/argmaxpool-microkernel-tester.h",
2611 "src/xnnpack/AlignedAllocator.h",
2612 ] + MICROKERNEL_TEST_HDRS,
2613 deps = MICROKERNEL_TEST_DEPS,
2614)
2615
2616xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002617 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002618 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002619 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002620 "test/avgpool-microkernel-tester.h",
2621 "src/xnnpack/AlignedAllocator.h",
2622 ] + MICROKERNEL_TEST_HDRS,
2623 deps = MICROKERNEL_TEST_DEPS,
2624)
2625
2626xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002627 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002628 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002629 "test/f32-ibilinear.cc",
2630 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002631 "src/xnnpack/AlignedAllocator.h",
2632 ] + MICROKERNEL_TEST_HDRS,
2633 deps = MICROKERNEL_TEST_DEPS,
2634)
2635
2636xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002637 name = "f32_clamp_test",
2638 srcs = [
2639 "test/f32-clamp.cc",
2640 "test/clamp-microkernel-tester.h",
2641 ] + MICROKERNEL_TEST_HDRS,
2642 deps = MICROKERNEL_TEST_DEPS,
2643)
2644
2645xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002646 name = "f32_igemm_test",
2647 srcs = [
2648 "test/f32-igemm.cc",
2649 "test/gemm-microkernel-tester.h",
2650 "src/xnnpack/AlignedAllocator.h",
2651 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2652 deps = MICROKERNEL_TEST_DEPS,
2653)
2654
2655xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002656 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002657 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002658 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002659 "test/gemm-microkernel-tester.h",
2660 "src/xnnpack/AlignedAllocator.h",
2661 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002662 # Work-around for "too many sections" error
2663 mingw_copts = ["-Wa,-mbig-obj"],
2664 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002665 deps = MICROKERNEL_TEST_DEPS,
2666)
2667
2668xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002669 name = "f16_igemm_minmax_test",
2670 srcs = [
2671 "test/f16-igemm-minmax.cc",
2672 "test/gemm-microkernel-tester.h",
2673 "src/xnnpack/AlignedAllocator.h",
2674 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2675 deps = MICROKERNEL_TEST_DEPS,
2676)
2677
2678xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002679 name = "f32_conv_hwc_test",
2680 srcs = [
2681 "test/f32-conv-hwc.cc",
2682 "test/conv-hwc-microkernel-tester.h",
2683 "src/xnnpack/AlignedAllocator.h",
2684 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2685 deps = MICROKERNEL_TEST_DEPS,
2686)
2687
2688xnnpack_unit_test(
2689 name = "f32_conv_hwc2spchw_test",
2690 srcs = [
2691 "test/f32-conv-hwc2spchw.cc",
2692 "test/conv-hwc2spchw-microkernel-tester.h",
2693 "src/xnnpack/AlignedAllocator.h",
2694 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2695 deps = MICROKERNEL_TEST_DEPS,
2696)
2697
2698xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002699 name = "f32_dwconv_test",
2700 srcs = [
2701 "test/f32-dwconv.cc",
2702 "test/dwconv-microkernel-tester.h",
2703 "src/xnnpack/AlignedAllocator.h",
2704 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2705 deps = MICROKERNEL_TEST_DEPS,
2706)
2707
2708xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002709 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002710 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002711 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002712 "test/dwconv-microkernel-tester.h",
2713 "src/xnnpack/AlignedAllocator.h",
2714 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2715 deps = MICROKERNEL_TEST_DEPS,
2716)
2717
2718xnnpack_unit_test(
2719 name = "f32_dwconv_spchw_test",
2720 srcs = [
2721 "test/f32-dwconv-spchw.cc",
2722 "test/dwconv-spchw-microkernel-tester.h",
2723 "src/xnnpack/AlignedAllocator.h",
2724 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2725 deps = MICROKERNEL_TEST_DEPS,
2726)
2727
2728xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002729 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002730 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002731 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732 "test/gavgpool-microkernel-tester.h",
2733 "src/xnnpack/AlignedAllocator.h",
2734 ] + MICROKERNEL_TEST_HDRS,
2735 deps = MICROKERNEL_TEST_DEPS,
2736)
2737
2738xnnpack_unit_test(
2739 name = "f32_gavgpool_spchw_test",
2740 srcs = [
2741 "test/f32-gavgpool-spchw.cc",
2742 "test/gavgpool-spchw-microkernel-tester.h",
2743 "src/xnnpack/AlignedAllocator.h",
2744 ] + MICROKERNEL_TEST_HDRS,
2745 deps = MICROKERNEL_TEST_DEPS,
2746)
2747
2748xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002749 name = "f32_gemm_test",
2750 srcs = [
2751 "test/f32-gemm.cc",
2752 "test/gemm-microkernel-tester.h",
2753 "src/xnnpack/AlignedAllocator.h",
2754 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2755 deps = MICROKERNEL_TEST_DEPS,
2756)
2757
2758xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002759 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002760 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002761 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002762 "test/gemm-microkernel-tester.h",
2763 "src/xnnpack/AlignedAllocator.h",
2764 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2765 deps = MICROKERNEL_TEST_DEPS,
2766)
2767
2768xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002769 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002770 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002771 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002772 "test/gemm-microkernel-tester.h",
2773 "src/xnnpack/AlignedAllocator.h",
2774 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2775 deps = MICROKERNEL_TEST_DEPS,
2776)
2777
2778xnnpack_unit_test(
2779 name = "f32_hswish_test",
2780 srcs = [
2781 "test/f32-hswish.cc",
2782 "test/hswish-microkernel-tester.h",
2783 ] + MICROKERNEL_TEST_HDRS,
2784 deps = MICROKERNEL_TEST_DEPS,
2785)
2786
2787xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002788 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002789 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002790 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002791 "test/maxpool-microkernel-tester.h",
2792 ] + MICROKERNEL_TEST_HDRS,
2793 deps = MICROKERNEL_TEST_DEPS,
2794)
2795
2796xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002797 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002798 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002799 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002800 "test/avgpool-microkernel-tester.h",
2801 "src/xnnpack/AlignedAllocator.h",
2802 ] + MICROKERNEL_TEST_HDRS,
2803 deps = MICROKERNEL_TEST_DEPS,
2804)
2805
2806xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002807 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002808 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002809 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002810 "test/gemm-microkernel-tester.h",
2811 "src/xnnpack/AlignedAllocator.h",
2812 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2813 deps = MICROKERNEL_TEST_DEPS,
2814)
2815
2816xnnpack_unit_test(
2817 name = "f32_prelu_test",
2818 srcs = [
2819 "test/f32-prelu.cc",
2820 "test/prelu-microkernel-tester.h",
2821 "src/xnnpack/AlignedAllocator.h",
2822 ] + MICROKERNEL_TEST_HDRS,
2823 deps = MICROKERNEL_TEST_DEPS,
2824)
2825
2826xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002827 name = "f32_raddexpminusmax_test",
2828 srcs = [
2829 "test/f32-raddexpminusmax.cc",
2830 "test/raddexpminusmax-microkernel-tester.h",
2831 ] + MICROKERNEL_TEST_HDRS,
2832 deps = MICROKERNEL_TEST_DEPS,
2833)
2834
2835xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002836 name = "f32_raddextexp_test",
2837 srcs = [
2838 "test/f32-raddextexp.cc",
2839 "test/raddextexp-microkernel-tester.h",
2840 ] + MICROKERNEL_TEST_HDRS,
2841 deps = MICROKERNEL_TEST_DEPS,
2842)
2843
2844xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002845 name = "f32_raddstoreexpminusmax_test",
2846 srcs = [
2847 "test/f32-raddstoreexpminusmax.cc",
2848 "test/raddstoreexpminusmax-microkernel-tester.h",
2849 ] + MICROKERNEL_TEST_HDRS,
2850 deps = MICROKERNEL_TEST_DEPS,
2851)
2852
2853xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002854 name = "f32_rmax_test",
2855 srcs = [
2856 "test/f32-rmax.cc",
2857 "test/rmax-microkernel-tester.h",
2858 ] + MICROKERNEL_TEST_HDRS,
2859 deps = MICROKERNEL_TEST_DEPS,
2860)
2861
2862xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002863 name = "f32_sigmoid_test",
2864 srcs = [
2865 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002866 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002867 ] + MICROKERNEL_TEST_HDRS,
2868 deps = MICROKERNEL_TEST_DEPS,
2869)
2870
2871xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002872 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002873 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002874 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002875 "test/spmm-microkernel-tester.h",
2876 "src/xnnpack/AlignedAllocator.h",
2877 ] + MICROKERNEL_TEST_HDRS,
2878 deps = MICROKERNEL_TEST_DEPS,
2879)
2880
2881xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002882 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002883 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002884 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002885 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002886 ] + MICROKERNEL_TEST_HDRS,
2887 deps = MICROKERNEL_TEST_DEPS,
2888)
2889
2890xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002891 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002892 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002893 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002894 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002895 ] + MICROKERNEL_TEST_HDRS,
2896 deps = MICROKERNEL_TEST_DEPS,
2897)
2898
2899xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002900 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002901 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002902 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002903 "test/vbinary-microkernel-tester.h",
2904 ] + MICROKERNEL_TEST_HDRS,
2905 deps = MICROKERNEL_TEST_DEPS,
2906)
2907
2908xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002909 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002910 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002911 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002912 "test/vbinaryc-microkernel-tester.h",
2913 ] + MICROKERNEL_TEST_HDRS,
2914 deps = MICROKERNEL_TEST_DEPS,
2915)
2916
2917xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002918 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002919 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002920 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002921 "test/vbinaryc-microkernel-tester.h",
2922 ] + MICROKERNEL_TEST_HDRS,
2923 deps = MICROKERNEL_TEST_DEPS,
2924)
2925
2926xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002927 name = "f32_vmax_test",
2928 srcs = [
2929 "test/f32-vmax.cc",
2930 "test/vbinary-microkernel-tester.h",
2931 ] + MICROKERNEL_TEST_HDRS,
2932 deps = MICROKERNEL_TEST_DEPS,
2933)
2934
2935xnnpack_unit_test(
2936 name = "f32_vmaxc_test",
2937 srcs = [
2938 "test/f32-vmaxc.cc",
2939 "test/vbinaryc-microkernel-tester.h",
2940 ] + MICROKERNEL_TEST_HDRS,
2941 deps = MICROKERNEL_TEST_DEPS,
2942)
2943
2944xnnpack_unit_test(
2945 name = "f32_vmin_test",
2946 srcs = [
2947 "test/f32-vmin.cc",
2948 "test/vbinary-microkernel-tester.h",
2949 ] + MICROKERNEL_TEST_HDRS,
2950 deps = MICROKERNEL_TEST_DEPS,
2951)
2952
2953xnnpack_unit_test(
2954 name = "f32_vminc_test",
2955 srcs = [
2956 "test/f32-vminc.cc",
2957 "test/vbinaryc-microkernel-tester.h",
2958 ] + MICROKERNEL_TEST_HDRS,
2959 deps = MICROKERNEL_TEST_DEPS,
2960)
2961
2962xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002963 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002965 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002966 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002967 ] + MICROKERNEL_TEST_HDRS,
2968 deps = MICROKERNEL_TEST_DEPS,
2969)
2970
2971xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002972 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002973 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002974 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002975 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976 ] + MICROKERNEL_TEST_HDRS,
2977 deps = MICROKERNEL_TEST_DEPS,
2978)
2979
2980xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002981 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002983 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002984 "test/vmulcaddc-microkernel-tester.h",
2985 "src/xnnpack/AlignedAllocator.h",
2986 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2987 deps = MICROKERNEL_TEST_DEPS,
2988)
2989
2990xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002991 name = "f32_vscale_test",
2992 srcs = [
2993 "test/f32-vscale.cc",
2994 "test/vscale-microkernel-tester.h",
2995 ] + MICROKERNEL_TEST_HDRS,
2996 deps = MICROKERNEL_TEST_DEPS,
2997)
2998
2999xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003000 name = "f32_vscaleexpminusmax_test",
3001 srcs = [
3002 "test/f32-vscaleexpminusmax.cc",
3003 "test/vscaleexpminusmax-microkernel-tester.h",
3004 ] + MICROKERNEL_TEST_HDRS,
3005 deps = MICROKERNEL_TEST_DEPS,
3006)
3007
3008xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003009 name = "f32_vscaleextexp_test",
3010 srcs = [
3011 "test/f32-vscaleextexp.cc",
3012 "test/vscaleextexp-microkernel-tester.h",
3013 ] + MICROKERNEL_TEST_HDRS,
3014 deps = MICROKERNEL_TEST_DEPS,
3015)
3016
3017xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003018 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003019 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003020 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003021 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003022 ] + MICROKERNEL_TEST_HDRS,
3023 deps = MICROKERNEL_TEST_DEPS,
3024)
3025
3026xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003027 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003028 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003029 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003030 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003031 ] + MICROKERNEL_TEST_HDRS,
3032 deps = MICROKERNEL_TEST_DEPS,
3033)
3034
3035xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003036 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003037 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003038 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003039 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003040 ] + MICROKERNEL_TEST_HDRS,
3041 deps = MICROKERNEL_TEST_DEPS,
3042)
3043
3044xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003045 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003046 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003047 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003048 "test/avgpool-microkernel-tester.h",
3049 "src/xnnpack/AlignedAllocator.h",
3050 ] + MICROKERNEL_TEST_HDRS,
3051 deps = MICROKERNEL_TEST_DEPS,
3052)
3053
3054xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003055 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003056 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003057 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003058 "test/gemm-microkernel-tester.h",
3059 "src/xnnpack/AlignedAllocator.h",
3060 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3061 deps = MICROKERNEL_TEST_DEPS,
3062)
3063
3064xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003065 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003066 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003067 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003068 "test/dwconv-microkernel-tester.h",
3069 "src/xnnpack/AlignedAllocator.h",
3070 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3071 deps = MICROKERNEL_TEST_DEPS,
3072)
3073
3074xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003075 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003076 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003077 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003078 "test/gavgpool-microkernel-tester.h",
3079 "src/xnnpack/AlignedAllocator.h",
3080 ] + MICROKERNEL_TEST_HDRS,
3081 deps = MICROKERNEL_TEST_DEPS,
3082)
3083
3084xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003085 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003087 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003088 "test/gemm-microkernel-tester.h",
3089 "src/xnnpack/AlignedAllocator.h",
3090 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3091 deps = MICROKERNEL_TEST_DEPS,
3092)
3093
3094xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003095 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003096 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003097 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003098 "test/vadd-microkernel-tester.h",
3099 ] + MICROKERNEL_TEST_HDRS,
3100 deps = MICROKERNEL_TEST_DEPS,
3101)
3102
3103xnnpack_unit_test(
3104 name = "u8_clamp_test",
3105 srcs = [
3106 "test/u8-clamp.cc",
3107 "test/clamp-microkernel-tester.h",
3108 ] + MICROKERNEL_TEST_HDRS,
3109 deps = MICROKERNEL_TEST_DEPS,
3110)
3111
3112xnnpack_unit_test(
3113 name = "u8_lut32norm_test",
3114 srcs = [
3115 "test/u8-lut32norm.cc",
3116 "test/lut-norm-microkernel-tester.h",
3117 ] + MICROKERNEL_TEST_HDRS,
3118 deps = MICROKERNEL_TEST_DEPS,
3119)
3120
3121xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003122 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003123 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003124 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003125 "test/maxpool-microkernel-tester.h",
3126 ] + MICROKERNEL_TEST_HDRS,
3127 deps = MICROKERNEL_TEST_DEPS,
3128)
3129
3130xnnpack_unit_test(
3131 name = "u8_rmax_test",
3132 srcs = [
3133 "test/u8-rmax.cc",
3134 "test/rmax-microkernel-tester.h",
3135 ] + MICROKERNEL_TEST_HDRS,
3136 deps = MICROKERNEL_TEST_DEPS,
3137)
3138
3139xnnpack_unit_test(
3140 name = "x32_packx_test",
3141 srcs = [
3142 "test/x32-packx.cc",
3143 "test/pack-microkernel-tester.h",
3144 "src/xnnpack/AlignedAllocator.h",
3145 ] + MICROKERNEL_TEST_HDRS,
3146 deps = MICROKERNEL_TEST_DEPS,
3147)
3148
3149xnnpack_unit_test(
3150 name = "x32_pad_test",
3151 srcs = [
3152 "test/x32-pad.cc",
3153 "test/pad-microkernel-tester.h",
3154 ] + MICROKERNEL_TEST_HDRS,
3155 deps = MICROKERNEL_TEST_DEPS,
3156)
3157
3158xnnpack_unit_test(
3159 name = "x32_unpool_test",
3160 srcs = [
3161 "test/x32-unpool.cc",
3162 "test/unpool-microkernel-tester.h",
3163 ] + MICROKERNEL_TEST_HDRS,
3164 deps = MICROKERNEL_TEST_DEPS,
3165)
3166
3167xnnpack_unit_test(
3168 name = "x32_zip_test",
3169 srcs = [
3170 "test/x32-zip.cc",
3171 "test/zip-microkernel-tester.h",
3172 ] + MICROKERNEL_TEST_HDRS,
3173 deps = MICROKERNEL_TEST_DEPS,
3174)
3175
3176xnnpack_unit_test(
3177 name = "x8_lut_test",
3178 srcs = [
3179 "test/x8-lut.cc",
3180 "test/lut-microkernel-tester.h",
3181 ] + MICROKERNEL_TEST_HDRS,
3182 deps = MICROKERNEL_TEST_DEPS,
3183)
3184
3185xnnpack_unit_test(
3186 name = "x8_zip_test",
3187 srcs = [
3188 "test/x8-zip.cc",
3189 "test/zip-microkernel-tester.h",
3190 ] + MICROKERNEL_TEST_HDRS,
3191 deps = MICROKERNEL_TEST_DEPS,
3192)
3193
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003194xnnpack_unit_test(
3195 name = "requantization_test",
3196 srcs = [
3197 "src/xnnpack/requantization-stubs.h",
3198 "test/requantization.cc",
3199 "test/requantization-tester.h",
3200 ] + MICROKERNEL_TEST_HDRS,
3201 deps = MICROKERNEL_TEST_DEPS,
3202)
3203
Marat Dukhan20c3b922020-03-10 03:45:06 -07003204########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003205
3206xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003207 name = "operator_size_test",
3208 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003209 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210)
3211
Marat Dukhan20c3b922020-03-10 03:45:06 -07003212xnnpack_binary(
3213 name = "subgraph_size_test",
3214 srcs = ["test/subgraph-size.c"],
3215 deps = [":XNNPACK"],
3216)
3217
3218########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219
3220xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003221 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003222 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003223 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003224 "test/add-operator-tester.h",
3225 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003226 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227)
3228
3229xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003230 name = "add_nd_test",
3231 srcs = [
3232 "test/add-nd.cc",
3233 "test/binary-elementwise-operator-tester.h",
3234 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003235 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003236)
3237
3238xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003239 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003241 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003242 "test/argmax-pooling-operator-tester.h",
3243 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003244 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245)
3246
3247xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003248 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003249 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003250 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003251 "test/average-pooling-operator-tester.h",
3252 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003253 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003254)
3255
3256xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003257 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003258 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003259 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003260 "test/channel-pad-operator-tester.h",
3261 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003262 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003263)
3264
3265xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003266 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003267 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003268 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003269 "test/channel-shuffle-operator-tester.h",
3270 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003271 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272)
3273
3274xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003275 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003276 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003277 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278 "test/clamp-operator-tester.h",
3279 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003280 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281)
3282
3283xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003284 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003285 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003286 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003287 "test/convolution-operator-tester.h",
3288 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003289 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003290)
3291
3292xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003293 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003294 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003295 "test/convolution-nchw.cc",
3296 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003297 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003299)
3300
3301xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003302 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003303 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003304 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003305 "test/deconvolution-operator-tester.h",
3306 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003307 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003308)
3309
3310xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003311 name = "divide_nd_test",
3312 srcs = [
3313 "test/binary-elementwise-operator-tester.h",
3314 "test/divide-nd.cc",
3315 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003316 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003317)
3318
3319xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003320 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003321 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003322 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003323 "test/fully-connected-operator-tester.h",
3324 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003325 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003326)
3327
3328xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003329 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003330 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003331 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003332 "test/global-average-pooling-operator-tester.h",
3333 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003334 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003335)
3336
3337xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003338 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003339 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003340 "test/global-average-pooling-ncw.cc",
3341 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003342 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003343 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003344)
3345
3346xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003347 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003348 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003349 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350 "test/hardswish-operator-tester.h",
3351 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003352 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003353)
3354
3355xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003356 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003357 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003358 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003359 "test/leaky-relu-operator-tester.h",
3360 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003361 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003362)
3363
3364xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003365 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003366 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003367 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003368 "test/max-pooling-operator-tester.h",
3369 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003371)
3372
3373xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003374 name = "maximum_nd_test",
3375 srcs = [
3376 "test/binary-elementwise-operator-tester.h",
3377 "test/maximum-nd.cc",
3378 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003379 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003380)
3381
3382xnnpack_unit_test(
3383 name = "minimum_nd_test",
3384 srcs = [
3385 "test/binary-elementwise-operator-tester.h",
3386 "test/minimum-nd.cc",
3387 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003388 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003389)
3390
3391xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003392 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003393 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003394 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003395 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003396 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003397 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003398)
3399
3400xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003401 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003403 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003404 "test/prelu-operator-tester.h",
3405 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003406 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003407)
3408
3409xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003410 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003411 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003412 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003413 "test/resize-bilinear-operator-tester.h",
3414 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003415 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003416)
3417
3418xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003419 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003420 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003421 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003422 "test/sigmoid-operator-tester.h",
3423 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003424 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003425)
3426
3427xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003428 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003429 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003430 "test/softmax-nc.cc",
3431 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003432 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003433 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003434)
3435
3436xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003437 name = "subtract_nd_test",
3438 srcs = [
3439 "test/binary-elementwise-operator-tester.h",
3440 "test/subtract-nd.cc",
3441 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003442 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003443)
3444
3445xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003446 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003447 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003448 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003449 "test/unpooling-operator-tester.h",
3450 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003451 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003452)
3453
3454############################# Build configurations #############################
3455
Marat Dukhanb8642352019-10-30 15:43:02 -07003456# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003457config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003458 name = "xnn_enable_assembly_explicit_true",
3459 define_values = {"xnn_enable_assembly": "true"},
3460)
3461
3462# Disables usage of assembly kernels.
3463config_setting(
3464 name = "xnn_enable_assembly_explicit_false",
3465 define_values = {"xnn_enable_assembly": "false"},
3466)
3467
Marat Dukhan05702cf2020-03-26 15:41:33 -07003468# Disables usage of HMP-aware optimizations.
3469config_setting(
3470 name = "xnn_enable_hmp_explicit_false",
3471 define_values = {"xnn_enable_hmp": "false"},
3472)
3473
Marat Dukhanb8642352019-10-30 15:43:02 -07003474# Builds with -c dbg
3475config_setting(
3476 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003477 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003478 "compilation_mode": "dbg",
3479 },
3480)
3481
3482# Builds with -c opt
3483config_setting(
3484 name = "optimized_build",
3485 values = {
3486 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003487 },
3488)
3489
3490config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003491 name = "linux_k8",
3492 values = {"cpu": "k8"},
3493)
3494
3495config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003496 name = "linux_arm",
3497 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003498)
3499
3500config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003501 name = "linux_armhf",
3502 values = {"cpu": "armhf"},
3503)
3504
3505config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003506 name = "linux_aarch64",
3507 values = {"cpu": "aarch64"},
3508)
3509
3510config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511 name = "android",
3512 values = {"crosstool_top": "//external:android/crosstool"},
3513)
3514
3515config_setting(
3516 name = "android_armv7",
3517 values = {
3518 "crosstool_top": "//external:android/crosstool",
3519 "cpu": "armeabi-v7a",
3520 },
3521)
3522
3523config_setting(
3524 name = "android_arm64",
3525 values = {
3526 "crosstool_top": "//external:android/crosstool",
3527 "cpu": "arm64-v8a",
3528 },
3529)
3530
3531config_setting(
3532 name = "android_x86",
3533 values = {
3534 "crosstool_top": "//external:android/crosstool",
3535 "cpu": "x86",
3536 },
3537)
3538
3539config_setting(
3540 name = "android_x86_64",
3541 values = {
3542 "crosstool_top": "//external:android/crosstool",
3543 "cpu": "x86_64",
3544 },
3545)
3546
3547config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003548 name = "windows_x86_64",
3549 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003550)
3551
3552config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003553 name = "windows_x86_64_clang",
3554 values = {
3555 "compiler": "clang-cl",
3556 "cpu": "x64_windows",
3557 },
3558)
3559
3560config_setting(
3561 name = "windows_x86_64_mingw",
3562 values = {
3563 "compiler": "mingw-gcc",
3564 "cpu": "x64_windows",
3565 },
3566)
3567
3568config_setting(
3569 name = "windows_x86_64_msys",
3570 values = {
3571 "compiler": "msys-gcc",
3572 "cpu": "x64_windows",
3573 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003574)
3575
3576config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003577 name = "macos_x86_64",
3578 values = {
3579 "apple_platform_type": "macos",
3580 "cpu": "darwin",
3581 },
3582)
3583
3584config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003586 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003587)
3588
3589config_setting(
3590 name = "emscripten_wasm",
3591 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003592 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593 "cpu": "wasm",
3594 },
3595)
3596
3597config_setting(
3598 name = "emscripten_wasmsimd",
3599 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003600 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003602 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603 },
3604)
3605
3606config_setting(
3607 name = "emscripten_asmjs",
3608 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003609 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610 "cpu": "asmjs",
3611 },
3612)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003613
3614config_setting(
3615 name = "ios_armv7",
3616 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003617 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003618 "cpu": "ios_armv7",
3619 },
3620)
3621
3622config_setting(
3623 name = "ios_arm64",
3624 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003625 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003626 "cpu": "ios_arm64",
3627 },
3628)
3629
3630config_setting(
3631 name = "ios_arm64e",
3632 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003633 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003634 "cpu": "ios_arm64e",
3635 },
3636)
3637
3638config_setting(
3639 name = "ios_x86",
3640 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003641 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003642 "cpu": "ios_i386",
3643 },
3644)
3645
3646config_setting(
3647 name = "ios_x86_64",
3648 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003649 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003650 "cpu": "ios_x86_64",
3651 },
3652)
3653
3654config_setting(
3655 name = "watchos_armv7k",
3656 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003657 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003658 "cpu": "watchos_armv7k",
3659 },
3660)
3661
3662config_setting(
3663 name = "watchos_arm64_32",
3664 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003665 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003666 "cpu": "watchos_arm64_32",
3667 },
3668)
3669
3670config_setting(
3671 name = "watchos_x86",
3672 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003673 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003674 "cpu": "watchos_i386",
3675 },
3676)
3677
3678config_setting(
3679 name = "watchos_x86_64",
3680 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003681 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003682 "cpu": "watchos_x86_64",
3683 },
3684)
3685
3686config_setting(
3687 name = "tvos_arm64",
3688 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003689 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003690 "cpu": "tvos_arm64",
3691 },
3692)
3693
3694config_setting(
3695 name = "tvos_x86_64",
3696 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003697 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003698 "cpu": "tvos_x86_64",
3699 },
3700)