blob: b9bf34ae3aaa016578e0d6b408813f54167eaed1 [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 = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070039 ":ukernels_test_mode",
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 = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070047 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070048 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070053 "src/operators/argmax-pooling-nhwc.c",
54 "src/operators/average-pooling-nhwc.c",
55 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070056 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070057 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070058 "src/operators/convolution-nchw.c",
59 "src/operators/convolution-nhwc.c",
60 "src/operators/deconvolution-nhwc.c",
61 "src/operators/fully-connected-nc.c",
62 "src/operators/global-average-pooling-ncw.c",
63 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/leaky-relu-nc.c",
65 "src/operators/max-pooling-nhwc.c",
66 "src/operators/prelu-nc.c",
67 "src/operators/resize-bilinear-nhwc.c",
68 "src/operators/sigmoid-nc.c",
69 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070070 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070072]
73
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070074SUBGRAPH_SRCS = [
75 "src/subgraph/add2.c",
76 "src/subgraph/argmax-pooling-2d.c",
77 "src/subgraph/average-pooling-2d.c",
78 "src/subgraph/clamp.c",
79 "src/subgraph/convolution-2d.c",
80 "src/subgraph/deconvolution-2d.c",
81 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070082 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070083 "src/subgraph/fully-connected.c",
84 "src/subgraph/hardswish.c",
85 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/maximum2.c",
87 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070088 "src/subgraph/multiply2.c",
89 "src/subgraph/prelu.c",
90 "src/subgraph/sigmoid.c",
91 "src/subgraph/softmax.c",
92 "src/subgraph/static-constant-pad.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/squared-difference.c",
94 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/unpooling-2d.c",
96]
97
Marat Dukhan3a77ea72019-12-23 12:10:24 -080098TABLE_SRCS = [
99 "src/tables/exp2-k-over-64.c",
100 "src/tables/exp2-k-over-2048.c",
101]
102
Marat Dukhan08c4a432019-10-03 09:29:21 -0700103SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800104 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800105 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800106 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700107 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
108 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700109 "src/f32-clamp/gen/scalar-x1.c",
110 "src/f32-clamp/gen/scalar-x2.c",
111 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800112 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800113 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700114 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
115 "src/f32-dwconv-chw/3x3p1-scalar.c",
116 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
117 "src/f32-dwconv-chw/5x5p2-scalar.c",
118 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700119 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
120 "src/f32-dwconv/gen/up1x4-scalar.c",
121 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
122 "src/f32-dwconv/gen/up1x9-scalar.c",
123 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
124 "src/f32-dwconv/gen/up1x25-scalar.c",
125 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
126 "src/f32-dwconv/gen/up2x4-scalar.c",
127 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
128 "src/f32-dwconv/gen/up2x9-scalar.c",
129 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
130 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700131 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
133 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700137 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700143 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700144 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
145 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700146 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
147 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
148 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700149 "src/f32-gemm/gen/1x4-scalar.c",
150 "src/f32-gemm/gen/2x4-scalar.c",
151 "src/f32-gemm/gen/4x2-scalar.c",
152 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700153 "src/f32-gemm/gen/1x4-relu-scalar.c",
154 "src/f32-gemm/gen/2x4-relu-scalar.c",
155 "src/f32-gemm/gen/4x2-relu-scalar.c",
156 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700157 "src/f32-gemm/gen/1x4-minmax-scalar.c",
158 "src/f32-gemm/gen/2x4-minmax-scalar.c",
159 "src/f32-gemm/gen/4x2-minmax-scalar.c",
160 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800161 "src/f32-hswish/gen/scalar-x1.c",
162 "src/f32-hswish/gen/scalar-x2.c",
163 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700164 "src/f32-ibilinear/gen/scalar-c1.c",
165 "src/f32-ibilinear/gen/scalar-c2.c",
166 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700167 "src/f32-igemm/gen/1x4-scalar.c",
168 "src/f32-igemm/gen/2x4-scalar.c",
169 "src/f32-igemm/gen/4x2-scalar.c",
170 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700171 "src/f32-igemm/gen/1x4-relu-scalar.c",
172 "src/f32-igemm/gen/2x4-relu-scalar.c",
173 "src/f32-igemm/gen/4x2-relu-scalar.c",
174 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700175 "src/f32-igemm/gen/1x4-minmax-scalar.c",
176 "src/f32-igemm/gen/2x4-minmax-scalar.c",
177 "src/f32-igemm/gen/4x2-minmax-scalar.c",
178 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700179 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
180 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
181 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700182 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
183 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
184 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
185 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800186 "src/f32-prelu/gen/scalar-2x1.c",
187 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800188 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
189 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
190 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
191 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
192 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
193 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
194 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
195 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
196 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
197 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
198 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
199 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700200 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800201 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
202 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
203 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
204 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
205 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
206 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
207 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
208 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
209 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700210 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
211 "src/f32-spmm/gen/1x1-minmax-scalar.c",
212 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
213 "src/f32-spmm/gen/2x1-minmax-scalar.c",
214 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
215 "src/f32-spmm/gen/4x1-minmax-scalar.c",
216 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
217 "src/f32-spmm/gen/8x1-minmax-scalar.c",
218 "src/f32-spmm/gen/8x2-minmax-scalar.c",
219 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700220 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
221 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
222 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
223 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
224 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
225 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
226 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
227 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
228 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
229 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800232 "src/f32-vbinary/gen/vmax-scalar-x1.c",
233 "src/f32-vbinary/gen/vmax-scalar-x2.c",
234 "src/f32-vbinary/gen/vmax-scalar-x4.c",
235 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
236 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
237 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
238 "src/f32-vbinary/gen/vmin-scalar-x1.c",
239 "src/f32-vbinary/gen/vmin-scalar-x2.c",
240 "src/f32-vbinary/gen/vmin-scalar-x4.c",
241 "src/f32-vbinary/gen/vminc-scalar-x1.c",
242 "src/f32-vbinary/gen/vminc-scalar-x2.c",
243 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700244 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
245 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
246 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
247 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
250 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
251 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
252 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700253 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x1.c",
254 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x2.c",
255 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700256 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700259 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
260 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
261 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
262 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
263 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
264 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
268 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
269 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
270 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700271 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
272 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
273 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700274 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
275 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
276 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
277 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
278 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
279 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
280 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
281 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
282 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
283 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
284 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
285 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700286 "src/f32-vunary/gen/vabs-scalar-x1.c",
287 "src/f32-vunary/gen/vabs-scalar-x2.c",
288 "src/f32-vunary/gen/vabs-scalar-x4.c",
289 "src/f32-vunary/gen/vneg-scalar-x1.c",
290 "src/f32-vunary/gen/vneg-scalar-x2.c",
291 "src/f32-vunary/gen/vneg-scalar-x4.c",
292 "src/f32-vunary/gen/vsqr-scalar-x1.c",
293 "src/f32-vunary/gen/vsqr-scalar-x2.c",
294 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800295 "src/math/expminus-scalar-lut2048-p1.c",
296 "src/math/expminus-scalar-lut64-p2.c",
297 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700298 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700299 "src/math/roundne-scalar-nearbyint.c",
300 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700301 "src/math/roundd-scalar-addsub.c",
302 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700303 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700304 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700305 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700306 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700307 "src/math/roundz-scalar-addsub.c",
308 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700309 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800310 "src/math/sigmoid-scalar-lut2048-p1-div.c",
311 "src/math/sigmoid-scalar-lut64-p2-div.c",
312 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700313 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
314 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700315 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
317 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700318 "src/q8-gemm/2x2-minmax-scalar.c",
319 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700320 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700321 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700322 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700323 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700324 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700325 "src/x32-fill/scalar-float.c",
326 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700327 "src/x32-packx/x2-scalar.c",
328 "src/x32-packx/x3-scalar.c",
329 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700330 "src/x32-pad/scalar-int.c",
331 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700332 "src/x32-unpool/scalar.c",
333 "src/x32-zip/x2-scalar.c",
334 "src/x32-zip/x3-scalar.c",
335 "src/x32-zip/x4-scalar.c",
336 "src/x32-zip/xm-scalar.c",
337 "src/x8-lut/scalar.c",
338 "src/x8-zip/x2-scalar.c",
339 "src/x8-zip/x3-scalar.c",
340 "src/x8-zip/x4-scalar.c",
341 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700342 "src/requantization/precise-scalar.c",
343 "src/requantization/fp32-scalar.c",
344 "src/requantization/q31-scalar.c",
345 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700346]
347
Marat Dukhan436ebe62019-12-04 15:10:12 -0800348WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700349 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
350 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700351 "src/f32-clamp/gen/wasm-x1.c",
352 "src/f32-clamp/gen/wasm-x2.c",
353 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700354 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
355 "src/f32-dwconv/gen/up1x4-wasm.c",
356 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
357 "src/f32-dwconv/gen/up1x9-wasm.c",
358 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
359 "src/f32-dwconv/gen/up1x25-wasm.c",
360 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
361 "src/f32-dwconv/gen/up2x4-wasm.c",
362 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
363 "src/f32-dwconv/gen/up2x9-wasm.c",
364 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
365 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700366 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
367 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
368 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
369 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700370 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
371 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700372 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
373 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
374 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
375 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700376 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
377 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700378 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
379 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700380 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
381 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
382 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700383 "src/f32-gemm/gen/1x4-wasm.c",
384 "src/f32-gemm/gen/2x4-wasm.c",
385 "src/f32-gemm/gen/4x2-wasm.c",
386 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700387 "src/f32-gemm/gen/1x4-relu-wasm.c",
388 "src/f32-gemm/gen/2x4-relu-wasm.c",
389 "src/f32-gemm/gen/4x2-relu-wasm.c",
390 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700391 "src/f32-gemm/gen/1x4-minmax-wasm.c",
392 "src/f32-gemm/gen/2x4-minmax-wasm.c",
393 "src/f32-gemm/gen/4x2-minmax-wasm.c",
394 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800395 "src/f32-hswish/gen/wasm-x1.c",
396 "src/f32-hswish/gen/wasm-x2.c",
397 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700398 "src/f32-igemm/gen/1x4-wasm.c",
399 "src/f32-igemm/gen/2x4-wasm.c",
400 "src/f32-igemm/gen/4x2-wasm.c",
401 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700402 "src/f32-igemm/gen/1x4-relu-wasm.c",
403 "src/f32-igemm/gen/2x4-relu-wasm.c",
404 "src/f32-igemm/gen/4x2-relu-wasm.c",
405 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700406 "src/f32-igemm/gen/1x4-minmax-wasm.c",
407 "src/f32-igemm/gen/2x4-minmax-wasm.c",
408 "src/f32-igemm/gen/4x2-minmax-wasm.c",
409 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700410 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
411 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
412 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700413 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
414 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
415 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
416 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
417 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
418 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
419 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
420 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
421 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
422 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
423 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
424 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800425 "src/f32-vbinary/gen/vmax-wasm-x1.c",
426 "src/f32-vbinary/gen/vmax-wasm-x2.c",
427 "src/f32-vbinary/gen/vmax-wasm-x4.c",
428 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
429 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
430 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
431 "src/f32-vbinary/gen/vmin-wasm-x1.c",
432 "src/f32-vbinary/gen/vmin-wasm-x2.c",
433 "src/f32-vbinary/gen/vmin-wasm-x4.c",
434 "src/f32-vbinary/gen/vminc-wasm-x1.c",
435 "src/f32-vbinary/gen/vminc-wasm-x2.c",
436 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700437 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
438 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
439 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
440 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
441 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
442 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
443 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
444 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
445 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
446 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
447 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
448 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
449 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
450 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
451 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
452 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
453 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
454 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700455 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
456 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
457 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800458]
459
Marat Dukhan290055c2020-06-09 12:24:29 -0700460WASMSIMD_UKERNELS = [
461 "src/requantization/fp32-wasmsimd.c",
462]
463
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800464PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800465 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800466 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800467 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700468 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
469 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700470 "src/f32-clamp/gen/psimd-x4.c",
471 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700472 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
473 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
474 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
475 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
476 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
477 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
478 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
479 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
480 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
481 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
482 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
483 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700484 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
485 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700486 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
487 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
488 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
489 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
490 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
491 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
492 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
493 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
494 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
495 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
496 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
497 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
498 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
499 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
500 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
501 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
502 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
503 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
504 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800505 "src/f32-hswish/gen/psimd-x4.c",
506 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700507 "src/f32-ibilinear/gen/psimd-c4.c",
508 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700509 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
510 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
511 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
512 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
513 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
514 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
515 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
516 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
517 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
518 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700520 "src/f32-spmm/gen/4x1-minmax-psimd.c",
521 "src/f32-spmm/gen/8x1-minmax-psimd.c",
522 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700523 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
524 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700525 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800526 "src/f32-prelu/gen/psimd-2x4.c",
527 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800528 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700529 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
530 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
531 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
532 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
533 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
534 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
535 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
536 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800537 "src/f32-vbinary/gen/vmax-psimd-x4.c",
538 "src/f32-vbinary/gen/vmax-psimd-x8.c",
539 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
540 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
541 "src/f32-vbinary/gen/vmin-psimd-x4.c",
542 "src/f32-vbinary/gen/vmin-psimd-x8.c",
543 "src/f32-vbinary/gen/vminc-psimd-x4.c",
544 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700545 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
546 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
547 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
548 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
549 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
550 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700551 "src/f32-vbinary/gen/vrsqrdiffc-psimd-x4.c",
552 "src/f32-vbinary/gen/vrsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700553 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
554 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700555 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
556 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
557 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
558 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700559 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
560 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
561 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
562 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700563 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
564 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700565 "src/f32-vunary/gen/vabs-psimd-x4.c",
566 "src/f32-vunary/gen/vabs-psimd-x8.c",
567 "src/f32-vunary/gen/vneg-psimd-x4.c",
568 "src/f32-vunary/gen/vneg-psimd-x8.c",
569 "src/f32-vunary/gen/vsqr-psimd-x4.c",
570 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700571 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700572 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700573 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700574 "src/x32-unpool/psimd.c",
575 "src/x32-zip/x2-psimd.c",
576 "src/x32-zip/x3-psimd.c",
577 "src/x32-zip/x4-psimd.c",
578 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700579 "src/requantization/precise-psimd.c",
580 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700581]
582
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800583PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800584 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
585 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
586 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
587 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
588 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
589 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
590 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
591 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
592 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
593 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
594 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
595 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800596 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
597 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
598 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
599 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
600 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
601 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700602 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
603 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
604 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
605 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
606 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
607 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
608 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
609 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700610 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700611 "src/math/roundd-psimd-addsub.c",
612 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700613 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800614 "src/math/sigmoid-psimd-p5-div.c",
615]
616
Marat Dukhan08c4a432019-10-03 09:29:21 -0700617# ISA-specific micro-kernels
618NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700619 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
620 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700621 "src/f32-clamp/gen/neon-x4.c",
622 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700623 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
624 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700625 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
626 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700627 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
628 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700629 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
630 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700631 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
632 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
633 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
634 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700635 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
636 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
637 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
638 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700639 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
640 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
641 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
642 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700643 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700644 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
645 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700646 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
647 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
648 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
649 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
650 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
651 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
652 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
653 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
654 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
655 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
656 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
657 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
658 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
659 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
660 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
661 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
662 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
663 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
664 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
665 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
666 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
667 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
668 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
669 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
670 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
671 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
672 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
673 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
674 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
675 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
676 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800677 "src/f32-hswish/gen/neon-x4.c",
678 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700679 "src/f32-ibilinear/gen/neon-c4.c",
680 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700681 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
682 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
683 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
684 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
685 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
686 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
687 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
688 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
689 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
690 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
691 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
692 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
693 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
694 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
695 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
696 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700697 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
698 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
699 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700700 "src/f32-ppmm/gen/4x8-minmax-neon.c",
701 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800702 "src/f32-prelu/gen/neon-2x4.c",
703 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800704 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
705 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
706 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
707 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
708 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
709 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
710 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
711 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
712 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
713 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
714 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
715 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
716 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
717 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
718 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
719 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
720 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
722 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
723 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
725 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
726 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
727 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700728 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800729 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800730 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
731 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
732 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
733 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
734 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
735 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
736 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
737 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
738 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
739 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
740 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
741 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
742 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
743 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
744 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
745 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
746 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
747 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700748 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
749 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
750 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
751 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800752 "src/f32-vbinary/gen/vmax-neon-x4.c",
753 "src/f32-vbinary/gen/vmax-neon-x8.c",
754 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
755 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
756 "src/f32-vbinary/gen/vmin-neon-x4.c",
757 "src/f32-vbinary/gen/vmin-neon-x8.c",
758 "src/f32-vbinary/gen/vminc-neon-x4.c",
759 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700760 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
761 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
762 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
763 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700764 "src/f32-vbinary/gen/vrsqrdiffc-neon-x4.c",
765 "src/f32-vbinary/gen/vrsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700766 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
767 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700768 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
769 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
770 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
771 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700772 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
773 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
774 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
775 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700776 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
777 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700778 "src/f32-vrnd/gen/vrndne-neon-x4.c",
779 "src/f32-vrnd/gen/vrndne-neon-x8.c",
780 "src/f32-vrnd/gen/vrndz-neon-x4.c",
781 "src/f32-vrnd/gen/vrndz-neon-x8.c",
782 "src/f32-vrnd/gen/vrndu-neon-x4.c",
783 "src/f32-vrnd/gen/vrndu-neon-x8.c",
784 "src/f32-vrnd/gen/vrndd-neon-x4.c",
785 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700786 "src/f32-vunary/gen/vabs-neon-x4.c",
787 "src/f32-vunary/gen/vabs-neon-x8.c",
788 "src/f32-vunary/gen/vneg-neon-x4.c",
789 "src/f32-vunary/gen/vneg-neon-x8.c",
790 "src/f32-vunary/gen/vsqr-neon-x4.c",
791 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700792 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
793 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700794 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700795 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
796 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700797 "src/q8-gemm/4x8-minmax-neon.c",
798 "src/q8-gemm/8x8-minmax-neon.c",
799 "src/q8-igemm/4x8-minmax-neon.c",
800 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700801 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700802 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700803 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700804 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700805 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700806 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700807 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700808 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700809 "src/x32-zip/x2-neon.c",
810 "src/x32-zip/x3-neon.c",
811 "src/x32-zip/x4-neon.c",
812 "src/x32-zip/xm-neon.c",
813 "src/x8-zip/x2-neon.c",
814 "src/x8-zip/x3-neon.c",
815 "src/x8-zip/x4-neon.c",
816 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700817 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700818 "src/math/roundd-neon-addsub.c",
819 "src/math/roundd-neon-cvt.c",
820 "src/math/roundu-neon-addsub.c",
821 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700822 "src/math/roundz-neon-addsub.c",
823 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800824 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800825 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
826 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
827 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
828 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
829 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
830 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700831 "src/requantization/precise-neon.c",
832 "src/requantization/fp32-neon.c",
833 "src/requantization/q31-neon.c",
834 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700835]
836
837NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700838 "src/f32-ibilinear/gen/neonfma-c4.c",
839 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700840 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
841 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
842 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
843 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
844 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
845 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
846 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
847 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
848 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700849 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
850 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
851 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
852 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700853 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
854 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
855 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
856 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700857 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
858 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
859 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
860 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700861 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
862 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
863 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
864 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
865 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
866 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
867 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
868 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
869 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
870 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
871 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
872 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
873 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
874 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
875 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
876 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
877 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
878 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800879 "src/f32-hswish/gen/neonfma-x4.c",
880 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700881 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
882 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800883 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
884 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
885 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
886 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
887 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
888 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
889 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
890 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
891 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
892 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
893 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
894 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
895 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
896 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
897 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
898 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
899 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
900 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
901 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
902 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
903 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
904 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
905 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
906 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800907 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
908 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
909 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
910 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
911 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
912 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
913 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
914 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
915 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
916 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
917 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
918 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
919 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
920 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
921 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
922 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
923 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
924 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
925 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
926 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
927 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
928 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
929 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
930 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
931 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
932 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
933 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
934 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
935 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
936 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
937 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
938 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
939 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
940 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
941 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
942 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
943 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
948 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
949 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
950 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
956 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
957 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
958 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
959 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
960 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700961 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
962 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800963 "src/math/exp-neonfma-lut64-p2.c",
964 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800965 "src/math/expminus-neonfma-lut2048-p1.c",
966 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800967 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800968 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
969 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
970 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
971 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
972 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
973 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
974 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
975 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
976 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
977 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
978 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
979 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
980 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
981 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
982 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
983 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
984 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
985 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700986]
987
988AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700989 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
990 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
991 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
992 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
993 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
994 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700995 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
996 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
997 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
998 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
999 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1000 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1001 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1002 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1003 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1004 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1005 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1006 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1007 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1008 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1009 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1010 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1011 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1012 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1013 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1014 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001015 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1016 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001017 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1018 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001019 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1020 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001021 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1022 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001023 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1024 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1025 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1026 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1027 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001028 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1029 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1030 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1031 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1032 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1033 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1034 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1035 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1036 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1037 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1038 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1039 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1040 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1041 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1042 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1043 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1044 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1045 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001046 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1047 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1048 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1049 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1050 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1051 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1052 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1053 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1054 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1055 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1056 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1057 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1058 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1059 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1060 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1061 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1062 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1063 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001064 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1065 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1066 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1067 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1068 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1069 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001070]
1071
Marat Dukhan8853b822020-05-07 12:19:01 -07001072NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001073 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1074 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1075 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1076 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1077 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1078 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1079 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1080 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001081 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001082 "src/math/roundd-neonv8.c",
1083 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001084 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001085]
1086
Marat Dukhan08c4a432019-10-03 09:29:21 -07001087AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001088 "src/f16-clamp/gen/neonfp16arith-x16.c",
1089 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001090 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1091 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1092 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1093 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1094 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1095 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1096 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1097 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1098 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1099 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1100 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1101 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001102 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1103 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001104 "src/f16-hswish/gen/neonfp16arith-x16.c",
1105 "src/f16-hswish/gen/neonfp16arith-x8.c",
1106 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1107 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001108 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1109 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1110 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1111 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1112 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1113 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1114 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1115 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1116 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1117 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1118 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1119 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1120 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1121 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1122 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1123 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1124 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1125 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1126 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1127 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1128 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1129 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1130 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1131 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1132 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1133 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1134 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1135 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001136 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1137 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001138 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1139 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1140 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1141 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1142 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1143 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1144 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1145 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001146 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1147 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1148 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1149 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001150 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1151 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1152 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1153 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1154 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1155 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1156 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1157 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1158 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1159 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1160 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1161 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001162 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1163 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1164 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1165 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1166 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1167 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1168 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1169 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001170]
1171
1172SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001173 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1174 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001175 "src/f32-clamp/gen/sse-x4.c",
1176 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001177 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1178 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001179 "src/f32-dwconv-chw/3x3p1-sse.c",
1180 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001181 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1182 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1183 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1184 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1185 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1186 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1187 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1188 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1189 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1190 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1191 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1192 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001193 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001194 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1195 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001196 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1197 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1198 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1199 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1200 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1201 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1202 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1203 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1204 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1205 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1206 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1207 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1208 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001209 "src/f32-hswish/gen/sse-x4.c",
1210 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001211 "src/f32-ibilinear/gen/sse-c4.c",
1212 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001213 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1214 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1215 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1216 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1217 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1218 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1219 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001220 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1221 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1222 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001223 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001224 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001225 "src/f32-spmm/gen/4x1-minmax-sse.c",
1226 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001227 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001228 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1229 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1230 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1231 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1232 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1233 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1234 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1235 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001236 "src/f32-vbinary/gen/vmax-sse-x4.c",
1237 "src/f32-vbinary/gen/vmax-sse-x8.c",
1238 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1239 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1240 "src/f32-vbinary/gen/vmin-sse-x4.c",
1241 "src/f32-vbinary/gen/vmin-sse-x8.c",
1242 "src/f32-vbinary/gen/vminc-sse-x4.c",
1243 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001244 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1245 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1246 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1247 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1248 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1249 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001250 "src/f32-vbinary/gen/vrsqrdiffc-sse-x4.c",
1251 "src/f32-vbinary/gen/vrsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001252 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1253 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001254 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1255 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1256 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1257 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001258 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1259 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1260 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1261 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001262 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1263 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001264 "src/f32-vunary/gen/vabs-sse-x4.c",
1265 "src/f32-vunary/gen/vabs-sse-x8.c",
1266 "src/f32-vunary/gen/vneg-sse-x4.c",
1267 "src/f32-vunary/gen/vneg-sse-x8.c",
1268 "src/f32-vunary/gen/vsqr-sse-x4.c",
1269 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001270 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001271 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001272 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001273 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001274 "src/math/roundd-sse-addsub.c",
1275 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001276 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001277]
1278
1279SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001280 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1281 "src/f32-argmaxpool/4x-sse2-c4.c",
1282 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001283 "src/f32-prelu/gen/sse2-2x4.c",
1284 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001285 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1286 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1287 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1288 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1289 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1290 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1291 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1292 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1293 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1294 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1295 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1296 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001297 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001298 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001299 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001300 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001301 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1302 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001303 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1304 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1305 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1306 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1307 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1308 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1309 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1310 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001311 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1312 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001313 "src/q8-igemm/4x4c2-minmax-sse2.c",
1314 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001315 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1316 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001317 "src/q8-gemm/2x4c8-minmax-sse2.c",
1318 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001319 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001320 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001321 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001322 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001323 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001324 "src/x32-zip/x2-sse2.c",
1325 "src/x32-zip/x3-sse2.c",
1326 "src/x32-zip/x4-sse2.c",
1327 "src/x32-zip/xm-sse2.c",
1328 "src/x8-zip/x2-sse2.c",
1329 "src/x8-zip/x3-sse2.c",
1330 "src/x8-zip/x4-sse2.c",
1331 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001332 "src/math/exp-sse2-p5.c",
1333 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001334 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001335 "src/math/roundd-sse2-cvt.c",
1336 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001337 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001338 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001339 "src/requantization/precise-sse2.c",
1340 "src/requantization/fp32-sse2.c",
1341 "src/requantization/q31-sse2.c",
1342 "src/requantization/gemmlowp-sse2.c",
1343]
1344
1345SSSE3_UKERNELS = [
1346 "src/requantization/precise-ssse3.c",
1347 "src/requantization/q31-ssse3.c",
1348 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001349]
1350
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001351SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001352 "src/f32-prelu/gen/sse41-2x4.c",
1353 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001354 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001355 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001356 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001357 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001358 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1359 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001360 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1361 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1362 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1363 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1364 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1365 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1366 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1367 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001368 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001369 "src/math/roundd-sse41.c",
1370 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001371 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001372 "src/requantization/precise-sse4.c",
1373 "src/requantization/q31-sse4.c",
1374 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001375]
1376
Marat Dukhan08c4a432019-10-03 09:29:21 -07001377AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001378 "src/f32-clamp/gen/avx-x8.c",
1379 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001380 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1381 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1382 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1383 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1384 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1385 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1386 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1387 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1388 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1389 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1390 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1391 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1392 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1393 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1394 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1395 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1396 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1397 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1398 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1399 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1400 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1401 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1402 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1403 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1404 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1405 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1406 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1407 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1408 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1409 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001410 "src/f32-hswish/gen/avx-x8.c",
1411 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001412 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1413 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1414 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1415 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1416 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1417 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1418 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1419 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1420 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001421 "src/f32-prelu/gen/avx-2x8.c",
1422 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001423 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001424 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1425 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1426 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1427 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1428 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1429 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1430 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1431 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001432 "src/f32-vbinary/gen/vmax-avx-x8.c",
1433 "src/f32-vbinary/gen/vmax-avx-x16.c",
1434 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1435 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1436 "src/f32-vbinary/gen/vmin-avx-x8.c",
1437 "src/f32-vbinary/gen/vmin-avx-x16.c",
1438 "src/f32-vbinary/gen/vminc-avx-x8.c",
1439 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001440 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1441 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1442 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1443 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1444 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1445 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001446 "src/f32-vbinary/gen/vrsqrdiffc-avx-x8.c",
1447 "src/f32-vbinary/gen/vrsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001448 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1449 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001450 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1451 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1452 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1453 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001454 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1455 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1456 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1457 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001458 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001459 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1460 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1461 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1462 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1463 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1464 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1465 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1466 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001467 "src/f32-vunary/gen/vabs-avx-x8.c",
1468 "src/f32-vunary/gen/vabs-avx-x16.c",
1469 "src/f32-vunary/gen/vneg-avx-x8.c",
1470 "src/f32-vunary/gen/vneg-avx-x16.c",
1471 "src/f32-vunary/gen/vsqr-avx-x8.c",
1472 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001473]
1474
Marat Dukhanfda12b82019-11-21 12:27:59 -08001475FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001476 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1477 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1478 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1479 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1480 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1481 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1482 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1483 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1484 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1485 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1486 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1487 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1488 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1489 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1490 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1491 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1492 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1493 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1494 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1495 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1496 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1497 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1498 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1499 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1500 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1501 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1502 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1503 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1504 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1505 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1506 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1507 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1508 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1509 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1510 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1511 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1512 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1513 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1514 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1515 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001516 "src/f32-hswish/gen/fma3-x8.c",
1517 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001518 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1519 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1520 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1521 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1522 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1523 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1524 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1525 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1526 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1527 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1528 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1529 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1530 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1531 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001532]
1533
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001534AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001535 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1536 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1537 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1538 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1539 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1540 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1541 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1542 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1543 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1544 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1545 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1546 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1547 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1548 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1549 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1550 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1551 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1552 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1553 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1554 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1555 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1556 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1557 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1558 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1559 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1560 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1561 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1562 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1563 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1564 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1565 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1566 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1567 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1568 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1569 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1570 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001571 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1572 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1573 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1574 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1575 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1576 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1577 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1578 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1579 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1580 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1581 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1582 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1583 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1584 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1585 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1586 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1587 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1588 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1589 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1590 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1591 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1592 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1593 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1594 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1595 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1596 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1597 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1598 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1599 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1600 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001601 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1602 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1603 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1604 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1605 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1606 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1607 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1608 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1609 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1610 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1611 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1612 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1613 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1614 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1615 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1616 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1617 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1618 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1619 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1620 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1621 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1622 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1623 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1624 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001625 "src/math/exp-avx2-p5.c",
1626 "src/math/exp-avx2-perm-p3.c",
1627 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001628 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001629 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001630 "src/math/sigmoid-avx2-rr2-p5-div.c",
1631 "src/math/sigmoid-avx2-rr1-p5-div.c",
1632 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1633 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1634 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1635 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001636]
1637
Marat Dukhan08c4a432019-10-03 09:29:21 -07001638AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001639 "src/f32-clamp/gen/avx512f-x16.c",
1640 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001641 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1642 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1643 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1644 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1645 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1646 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1647 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1648 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1649 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1650 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1651 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1652 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1653 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1654 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1655 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1656 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1657 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1658 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1659 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1660 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1661 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1662 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1663 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1664 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001665 "src/f32-hswish/gen/avx512f-x16.c",
1666 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001667 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1668 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1669 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1670 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1671 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1672 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001673 "src/f32-prelu/gen/avx512f-2x16.c",
1674 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001675 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1676 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1677 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1678 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1679 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1680 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1681 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1682 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1683 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1684 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1685 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1686 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1687 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1688 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1689 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1690 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1691 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1692 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1693 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1694 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1695 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1696 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1697 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1698 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1699 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1700 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1701 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1702 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1703 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1704 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1705 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1706 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1707 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1708 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1709 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1710 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001711 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001712 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1713 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1714 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1715 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1716 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1717 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1718 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1719 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001720 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1721 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1722 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1723 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1724 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1725 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1726 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1727 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001728 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1729 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1730 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1731 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1732 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1733 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001734 "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x16.c",
1735 "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001736 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1737 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001738 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
1739 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
1740 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
1741 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001742 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1743 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1744 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1745 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001746 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001747 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1748 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1749 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1750 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1751 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1752 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1753 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1754 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1755 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1756 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1757 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1758 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1759 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1760 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1761 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1762 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1763 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1764 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1765 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1766 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1767 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1768 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1769 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1770 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001771 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
1772 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
1773 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
1774 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
1775 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
1776 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
1777 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
1778 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001779 "src/f32-vunary/gen/vabs-avx512f-x16.c",
1780 "src/f32-vunary/gen/vabs-avx512f-x32.c",
1781 "src/f32-vunary/gen/vneg-avx512f-x16.c",
1782 "src/f32-vunary/gen/vneg-avx512f-x32.c",
1783 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
1784 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001785 "src/math/exp-avx512f-p5-scalef.c",
1786 "src/math/exp-avx512f-p5.c",
1787 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001788 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001789 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001790]
1791
1792AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07001793 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07001794 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001795 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1796 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1797 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1798 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1799 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1800 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1801 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1802 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1803 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1804 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001805]
1806
1807AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001808 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001809 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1810 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001811 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001812 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1813 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001814 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1815 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1816 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001817 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001818 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1819 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1820 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001821 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001822 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1823 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001824 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001825 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1826 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1827 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1828 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1829 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1830 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1831 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1832 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1833 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1834 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1835 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1836 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1837 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1838 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1839 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1840 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1841 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1842 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1843 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1844 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1845 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001846 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001847 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1848 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1849 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1850 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1851 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1852 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1853 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1854 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1855 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1856 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1857 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1858 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1859 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1860 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1861 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1862 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1863 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1864 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1865 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1866 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1867 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1868 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1869 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1870 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1871 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1872 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1873 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1874 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1875 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1876 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1877 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1878 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1879 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1880 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1881 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1882 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1883 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1884 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001885]
1886
Marat Dukhan1b354632020-03-23 12:50:22 -07001887INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001888 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001889 "src/xnnpack/argmaxpool.h",
1890 "src/xnnpack/avgpool.h",
1891 "src/xnnpack/clamp.h",
1892 "src/xnnpack/common.h",
1893 "src/xnnpack/conv.h",
1894 "src/xnnpack/dwconv.h",
1895 "src/xnnpack/gavgpool.h",
1896 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001897 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001898 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001899 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001900 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001901 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001902 "src/xnnpack/lut.h",
1903 "src/xnnpack/math.h",
1904 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001905 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001906 "src/xnnpack/packx.h",
1907 "src/xnnpack/pad.h",
1908 "src/xnnpack/params.h",
1909 "src/xnnpack/pavgpool.h",
1910 "src/xnnpack/ppmm.h",
1911 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001912 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001913 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001914 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001915 "src/xnnpack/rmax.h",
1916 "src/xnnpack/scalar-utils.h",
1917 "src/xnnpack/spmm.h",
1918 "src/xnnpack/unpool.h",
1919 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001920 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001921 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001922 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001923 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001924 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001925 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001926 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001927]
1928
1929INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001930 "include/xnnpack.h",
1931 "src/xnnpack/allocator.h",
1932 "src/xnnpack/compute.h",
1933 "src/xnnpack/im2col.h",
1934 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001935 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07001936 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001937 "src/xnnpack/operator.h",
1938 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001939 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001940 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001941 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001942 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001943]
1944
Marat Dukhan1b354632020-03-23 12:50:22 -07001945ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001946 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001947]
1948
Marat Dukhan1b354632020-03-23 12:50:22 -07001949MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001950 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001951 "include/xnnpack.h",
1952]
1953
Marat Dukhan1b354632020-03-23 12:50:22 -07001954MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001955 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001956 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001957 "src/xnnpack/requantization.h",
1958 "include/xnnpack.h",
1959]
1960
1961OPERATOR_TEST_PARAMS_HDRS = [
1962 "src/xnnpack/params.h",
1963 "src/xnnpack/common.h",
1964]
1965
1966WEIGHTS_PACK_HDRS = [
1967 "src/xnnpack/pack.h",
1968 "src/xnnpack/operator.h",
1969 "src/xnnpack/compute.h",
1970]
1971
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001972LOGGING_COPTS = select({
1973 # No logging in optimized mode
1974 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1975 # Full logging in debug mode
1976 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1977 # Error-only logging in default (fastbuild) mode
1978 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1979})
1980
Marat Dukhan3b59de22020-06-03 20:15:19 -07001981LOGGING_SRCS = select({
1982 # No logging in optimized mode
1983 ":optimized_build": [],
1984 "//conditions:default": [
1985 "src/operator-strings.c",
1986 "src/subgraph-strings.c",
1987 ],
1988})
1989
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001990LOGGING_HDRS = [
1991 "src/xnnpack/log.h",
1992]
1993
Marat Dukhan08c4a432019-10-03 09:29:21 -07001994xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001995 name = "tables",
1996 srcs = TABLE_SRCS,
1997 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001998 gcc_copts = xnnpack_gcc_std_copts(),
1999 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002000)
2001
2002xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002003 name = "scalar_ukernels",
2004 srcs = SCALAR_UKERNELS,
2005 hdrs = INTERNAL_HDRS,
2006 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002007 gcc_copts = xnnpack_gcc_std_copts(),
2008 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002009 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002010 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002011 "@FP16",
2012 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002013 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002014 ],
2015)
2016
2017xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002018 name = "scalar_ukernels_test_mode",
2019 srcs = SCALAR_UKERNELS,
2020 hdrs = INTERNAL_HDRS,
2021 aarch32_copts = ["-marm"],
2022 copts = [
2023 "-UNDEBUG",
2024 "-DXNN_TEST_MODE=1",
2025 ],
2026 gcc_copts = xnnpack_gcc_std_copts(),
2027 msvc_copts = xnnpack_msvc_std_copts(),
2028 deps = [
2029 ":tables",
2030 "@FP16",
2031 "@FXdiv",
2032 "@pthreadpool",
2033 ],
2034)
2035
2036xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002037 name = "wasm_ukernels",
2038 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002039 gcc_copts = xnnpack_gcc_std_copts(),
2040 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002041 wasm_srcs = WASM_UKERNELS,
Marat Dukhan290055c2020-06-09 12:24:29 -07002042 wasmsimd_srcs = WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002043 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002044 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002045 "@FP16",
2046 "@FXdiv",
2047 "@pthreadpool",
2048 ],
2049)
2050
2051xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002052 name = "wasm_ukernels_test_mode",
2053 hdrs = INTERNAL_HDRS,
2054 copts = [
2055 "-UNDEBUG",
2056 "-DXNN_TEST_MODE=1",
2057 ],
2058 gcc_copts = xnnpack_gcc_std_copts(),
2059 msvc_copts = xnnpack_msvc_std_copts(),
2060 wasm_srcs = WASM_UKERNELS,
Marat Dukhan290055c2020-06-09 12:24:29 -07002061 wasmsimd_srcs = WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002062 deps = [
2063 ":tables",
2064 "@FP16",
2065 "@FXdiv",
2066 "@pthreadpool",
2067 ],
2068)
2069
2070xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002071 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072 hdrs = INTERNAL_HDRS,
2073 aarch32_copts = [
2074 "-marm",
2075 "-mfpu=neon",
2076 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002077 gcc_copts = xnnpack_gcc_std_copts(),
2078 gcc_x86_copts = ["-msse2"],
2079 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002080 optimized_copts = [
2081 "-O3",
2082 "-ffast-math",
2083 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002084 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002085 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002086 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002087 "@FP16",
2088 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002089 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002090 ],
2091)
2092
2093xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002094 name = "psimd_fastmath_ukernels_test_mode",
2095 hdrs = INTERNAL_HDRS,
2096 aarch32_copts = [
2097 "-marm",
2098 "-mfpu=neon",
2099 ],
2100 copts = [
2101 "-UNDEBUG",
2102 "-DXNN_TEST_MODE=1",
2103 ],
2104 gcc_copts = xnnpack_gcc_std_copts(),
2105 gcc_x86_copts = ["-msse2"],
2106 msvc_copts = xnnpack_msvc_std_copts(),
2107 optimized_copts = [
2108 "-O3",
2109 "-ffast-math",
2110 ],
2111 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2112 deps = [
2113 ":tables",
2114 "@FP16",
2115 "@psimd",
2116 "@pthreadpool",
2117 ],
2118)
2119
2120xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002121 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002122 hdrs = INTERNAL_HDRS,
2123 aarch32_copts = [
2124 "-marm",
2125 "-mfpu=neon",
2126 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002127 gcc_copts = xnnpack_gcc_std_copts(),
2128 gcc_x86_copts = ["-msse2"],
2129 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002130 optimized_copts = [
2131 "-O3",
2132 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002133 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002134 deps = [
2135 ":tables",
2136 "@FP16",
2137 "@psimd",
2138 "@pthreadpool",
2139 ],
2140)
2141
2142xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002143 name = "psimd_accmath_ukernels_test_mode",
2144 hdrs = INTERNAL_HDRS,
2145 aarch32_copts = [
2146 "-marm",
2147 "-mfpu=neon",
2148 ],
2149 copts = [
2150 "-UNDEBUG",
2151 "-DXNN_TEST_MODE=1",
2152 ],
2153 gcc_copts = xnnpack_gcc_std_copts(),
2154 gcc_x86_copts = ["-msse2"],
2155 msvc_copts = xnnpack_msvc_std_copts(),
2156 optimized_copts = [
2157 "-O3",
2158 ],
2159 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2160 deps = [
2161 ":tables",
2162 "@FP16",
2163 "@psimd",
2164 "@pthreadpool",
2165 ],
2166)
2167
2168xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002169 name = "neon_ukernels",
2170 hdrs = INTERNAL_HDRS,
2171 aarch32_copts = [
2172 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002173 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002174 "-mfpu=neon",
2175 ],
2176 aarch32_srcs = NEON_UKERNELS,
2177 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002178 gcc_copts = xnnpack_gcc_std_copts(),
2179 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002180 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002181 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002182 "@FP16",
2183 "@pthreadpool",
2184 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002185)
2186
2187xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002188 name = "neon_ukernels_test_mode",
2189 hdrs = INTERNAL_HDRS,
2190 aarch32_copts = [
2191 "-marm",
2192 "-march=armv7-a",
2193 "-mfpu=neon",
2194 ],
2195 aarch32_srcs = NEON_UKERNELS,
2196 aarch64_srcs = NEON_UKERNELS,
2197 copts = [
2198 "-UNDEBUG",
2199 "-DXNN_TEST_MODE=1",
2200 ],
2201 gcc_copts = xnnpack_gcc_std_copts(),
2202 msvc_copts = xnnpack_msvc_std_copts(),
2203 deps = [
2204 ":tables",
2205 "@FP16",
2206 "@pthreadpool",
2207 ],
2208)
2209
2210xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002211 name = "neonfma_ukernels",
2212 hdrs = INTERNAL_HDRS,
2213 aarch32_copts = [
2214 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002215 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002216 "-mfpu=neon-vfpv4",
2217 ],
2218 aarch32_srcs = NEONFMA_UKERNELS,
2219 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002220 gcc_copts = xnnpack_gcc_std_copts(),
2221 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002222 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002223 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002224 "@FP16",
2225 "@pthreadpool",
2226 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002227)
2228
2229xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002230 name = "neonfma_ukernels_test_mode",
2231 hdrs = INTERNAL_HDRS,
2232 aarch32_copts = [
2233 "-marm",
2234 "-march=armv7-a",
2235 "-mfpu=neon-vfpv4",
2236 ],
2237 aarch32_srcs = NEONFMA_UKERNELS,
2238 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
2239 copts = [
2240 "-UNDEBUG",
2241 "-DXNN_TEST_MODE=1",
2242 ],
2243 gcc_copts = xnnpack_gcc_std_copts(),
2244 msvc_copts = xnnpack_msvc_std_copts(),
2245 deps = [
2246 ":tables",
2247 "@FP16",
2248 "@pthreadpool",
2249 ],
2250)
2251
2252xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002253 name = "neonv8_ukernels",
2254 hdrs = INTERNAL_HDRS,
2255 aarch32_copts = [
2256 "-marm",
2257 "-march=armv8-a",
2258 "-mfpu=neon-fp-armv8",
2259 ],
2260 aarch32_srcs = NEONV8_UKERNELS,
2261 aarch64_srcs = NEONV8_UKERNELS,
2262 gcc_copts = xnnpack_gcc_std_copts(),
2263 msvc_copts = xnnpack_msvc_std_copts(),
2264 deps = [
2265 ":tables",
2266 "@FP16",
2267 "@pthreadpool",
2268 ],
2269)
2270
2271xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002272 name = "neonv8_ukernels_test_mode",
2273 hdrs = INTERNAL_HDRS,
2274 aarch32_copts = [
2275 "-marm",
2276 "-march=armv8-a",
2277 "-mfpu=neon-fp-armv8",
2278 ],
2279 aarch32_srcs = NEONV8_UKERNELS,
2280 aarch64_srcs = NEONV8_UKERNELS,
2281 copts = [
2282 "-UNDEBUG",
2283 "-DXNN_TEST_MODE=1",
2284 ],
2285 gcc_copts = xnnpack_gcc_std_copts(),
2286 msvc_copts = xnnpack_msvc_std_copts(),
2287 deps = [
2288 ":tables",
2289 "@FP16",
2290 "@pthreadpool",
2291 ],
2292)
2293
2294xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002295 name = "neonfp16arith_ukernels",
2296 hdrs = INTERNAL_HDRS,
2297 aarch64_copts = ["-march=armv8.2-a+fp16"],
2298 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002299 gcc_copts = xnnpack_gcc_std_copts(),
2300 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002301 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002302 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002303 "@FP16",
2304 "@pthreadpool",
2305 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002306)
2307
2308xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002309 name = "neonfp16arith_ukernels_test_mode",
2310 hdrs = INTERNAL_HDRS,
2311 aarch64_copts = ["-march=armv8.2-a+fp16"],
2312 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2313 copts = [
2314 "-UNDEBUG",
2315 "-DXNN_TEST_MODE=1",
2316 ],
2317 gcc_copts = xnnpack_gcc_std_copts(),
2318 msvc_copts = xnnpack_msvc_std_copts(),
2319 deps = [
2320 ":tables",
2321 "@FP16",
2322 "@pthreadpool",
2323 ],
2324)
2325
2326xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002327 name = "sse2_ukernels",
2328 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002329 gcc_copts = xnnpack_gcc_std_copts(),
2330 gcc_x86_copts = ["-msse2"],
2331 msvc_copts = xnnpack_msvc_std_copts(),
2332 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002333 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002334 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002335 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002336 "@FP16",
2337 "@pthreadpool",
2338 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002339)
2340
2341xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002342 name = "sse2_ukernels_test_mode",
2343 hdrs = INTERNAL_HDRS,
2344 copts = [
2345 "-UNDEBUG",
2346 "-DXNN_TEST_MODE=1",
2347 ],
2348 gcc_copts = xnnpack_gcc_std_copts(),
2349 gcc_x86_copts = ["-msse2"],
2350 msvc_copts = xnnpack_msvc_std_copts(),
2351 msvc_x86_32_copts = ["/arch:SSE2"],
2352 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2353 deps = [
2354 ":tables",
2355 "@FP16",
2356 "@pthreadpool",
2357 ],
2358)
2359
2360xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002361 name = "ssse3_ukernels",
2362 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002363 gcc_copts = xnnpack_gcc_std_copts(),
2364 gcc_x86_copts = ["-mssse3"],
2365 msvc_copts = xnnpack_msvc_std_copts(),
2366 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002367 x86_srcs = SSSE3_UKERNELS,
2368 deps = [
2369 ":tables",
2370 "@FP16",
2371 "@pthreadpool",
2372 ],
2373)
2374
2375xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002376 name = "ssse3_ukernels_test_mode",
2377 hdrs = INTERNAL_HDRS,
2378 copts = [
2379 "-UNDEBUG",
2380 "-DXNN_TEST_MODE=1",
2381 ],
2382 gcc_copts = xnnpack_gcc_std_copts(),
2383 gcc_x86_copts = ["-mssse3"],
2384 msvc_copts = xnnpack_msvc_std_copts(),
2385 msvc_x86_32_copts = ["/arch:SSE2"],
2386 x86_srcs = SSSE3_UKERNELS,
2387 deps = [
2388 ":tables",
2389 "@FP16",
2390 "@pthreadpool",
2391 ],
2392)
2393
2394xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002395 name = "sse41_ukernels",
2396 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002397 gcc_copts = xnnpack_gcc_std_copts(),
2398 gcc_x86_copts = ["-msse4.1"],
2399 msvc_copts = xnnpack_msvc_std_copts(),
2400 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002401 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002402 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002403 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002404 "@FP16",
2405 "@pthreadpool",
2406 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002407)
2408
2409xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002410 name = "sse41_ukernels_test_mode",
2411 hdrs = INTERNAL_HDRS,
2412 copts = [
2413 "-UNDEBUG",
2414 "-DXNN_TEST_MODE=1",
2415 ],
2416 gcc_copts = xnnpack_gcc_std_copts(),
2417 gcc_x86_copts = ["-msse4.1"],
2418 msvc_copts = xnnpack_msvc_std_copts(),
2419 msvc_x86_32_copts = ["/arch:SSE2"],
2420 x86_srcs = SSE41_UKERNELS,
2421 deps = [
2422 ":tables",
2423 "@FP16",
2424 "@pthreadpool",
2425 ],
2426)
2427
2428xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002429 name = "avx_ukernels",
2430 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002431 gcc_copts = xnnpack_gcc_std_copts(),
2432 gcc_x86_copts = ["-mavx"],
2433 msvc_copts = xnnpack_msvc_std_copts(),
2434 msvc_x86_32_copts = ["/arch:AVX"],
2435 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002436 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002437 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002438 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002439 "@FP16",
2440 "@pthreadpool",
2441 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002442)
2443
2444xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002445 name = "avx_ukernels_test_mode",
2446 hdrs = INTERNAL_HDRS,
2447 copts = [
2448 "-UNDEBUG",
2449 "-DXNN_TEST_MODE=1",
2450 ],
2451 gcc_copts = xnnpack_gcc_std_copts(),
2452 gcc_x86_copts = ["-mavx"],
2453 msvc_copts = xnnpack_msvc_std_copts(),
2454 msvc_x86_32_copts = ["/arch:AVX"],
2455 msvc_x86_64_copts = ["/arch:AVX"],
2456 x86_srcs = AVX_UKERNELS,
2457 deps = [
2458 ":tables",
2459 "@FP16",
2460 "@pthreadpool",
2461 ],
2462)
2463
2464xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002465 name = "fma3_ukernels",
2466 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002467 gcc_copts = xnnpack_gcc_std_copts(),
2468 gcc_x86_copts = ["-mfma"],
2469 msvc_copts = xnnpack_msvc_std_copts(),
2470 msvc_x86_32_copts = ["/arch:AVX"],
2471 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002472 x86_srcs = FMA3_UKERNELS,
2473 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002474 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002475 "@FP16",
2476 "@pthreadpool",
2477 ],
2478)
2479
2480xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002481 name = "fma3_ukernels_test_mode",
2482 hdrs = INTERNAL_HDRS,
2483 copts = [
2484 "-UNDEBUG",
2485 "-DXNN_TEST_MODE=1",
2486 ],
2487 gcc_copts = xnnpack_gcc_std_copts(),
2488 gcc_x86_copts = ["-mfma"],
2489 msvc_copts = xnnpack_msvc_std_copts(),
2490 msvc_x86_32_copts = ["/arch:AVX"],
2491 msvc_x86_64_copts = ["/arch:AVX"],
2492 x86_srcs = FMA3_UKERNELS,
2493 deps = [
2494 ":tables",
2495 "@FP16",
2496 "@pthreadpool",
2497 ],
2498)
2499
2500xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002501 name = "avx2_ukernels",
2502 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002503 gcc_copts = xnnpack_gcc_std_copts(),
2504 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002505 "-mfma",
2506 "-mavx2",
2507 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002508 msvc_copts = xnnpack_msvc_std_copts(),
2509 msvc_x86_32_copts = ["/arch:AVX2"],
2510 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002511 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002512 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002513 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002514 "@FP16",
2515 "@pthreadpool",
2516 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002517)
2518
2519xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002520 name = "avx2_ukernels_test_mode",
2521 hdrs = INTERNAL_HDRS,
2522 copts = [
2523 "-UNDEBUG",
2524 "-DXNN_TEST_MODE=1",
2525 ],
2526 gcc_copts = xnnpack_gcc_std_copts(),
2527 gcc_x86_copts = [
2528 "-mfma",
2529 "-mavx2",
2530 ],
2531 msvc_copts = xnnpack_msvc_std_copts(),
2532 msvc_x86_32_copts = ["/arch:AVX2"],
2533 msvc_x86_64_copts = ["/arch:AVX2"],
2534 x86_srcs = AVX2_UKERNELS,
2535 deps = [
2536 ":tables",
2537 "@FP16",
2538 "@pthreadpool",
2539 ],
2540)
2541
2542xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002543 name = "avx512f_ukernels",
2544 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002545 gcc_copts = xnnpack_gcc_std_copts(),
2546 gcc_x86_copts = ["-mavx512f"],
2547 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2548 msvc_copts = xnnpack_msvc_std_copts(),
2549 msvc_x86_32_copts = ["/arch:AVX512"],
2550 msvc_x86_64_copts = ["/arch:AVX512"],
2551 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002552 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002553 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002554 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002555 "@FP16",
2556 "@pthreadpool",
2557 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002558)
2559
2560xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002561 name = "avx512f_ukernels_test_mode",
2562 hdrs = INTERNAL_HDRS,
2563 copts = [
2564 "-UNDEBUG",
2565 "-DXNN_TEST_MODE=1",
2566 ],
2567 gcc_copts = xnnpack_gcc_std_copts(),
2568 gcc_x86_copts = ["-mavx512f"],
2569 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2570 msvc_copts = xnnpack_msvc_std_copts(),
2571 msvc_x86_32_copts = ["/arch:AVX512"],
2572 msvc_x86_64_copts = ["/arch:AVX512"],
2573 msys_copts = ["-fno-asynchronous-unwind-tables"],
2574 x86_srcs = AVX512F_UKERNELS,
2575 deps = [
2576 ":tables",
2577 "@FP16",
2578 "@pthreadpool",
2579 ],
2580)
2581
2582xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002583 name = "asm_ukernels",
2584 hdrs = ["src/xnnpack/assembly.h"],
2585 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002586 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002587 aarch64_srcs = AARCH64_ASM_UKERNELS,
2588)
2589
Marat Dukhan3b59de22020-06-03 20:15:19 -07002590xnnpack_cc_library(
2591 name = "logging_utils",
2592 srcs = LOGGING_SRCS,
2593 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2594 copts = LOGGING_COPTS + [
2595 "-Isrc",
2596 "-Iinclude",
2597 ] + select({
2598 ":debug_build": [],
2599 "//conditions:default": xnnpack_min_size_copts(),
2600 }),
2601 gcc_copts = xnnpack_gcc_std_copts(),
2602 msvc_copts = xnnpack_msvc_std_copts(),
2603 visibility = xnnpack_visibility(),
2604 deps = [
2605 "@FP16",
2606 "@clog",
2607 "@pthreadpool",
2608 ],
2609)
2610
Marat Dukhan08c4a432019-10-03 09:29:21 -07002611xnnpack_aggregate_library(
2612 name = "ukernels",
2613 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002614 ":neon_ukernels",
2615 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002616 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002617 ":asm_ukernels",
2618 ],
2619 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002620 ":neon_ukernels",
2621 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002622 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002623 ":neonfp16arith_ukernels",
2624 ":asm_ukernels",
2625 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002626 generic_deps = [
2627 ":scalar_ukernels",
2628 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002629 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002630 ":psimd_fastmath_ukernels",
2631 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002632 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002633 wasm_deps = [
2634 ":wasm_ukernels",
2635 ],
2636 wasmsimd_deps = [
2637 ":wasm_ukernels",
2638 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002639 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002640 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002641 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002642 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002643 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002644 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002645 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002646 ":avx512f_ukernels",
2647 ],
2648)
2649
Marat Dukhan33fcf782020-05-24 14:27:15 -07002650xnnpack_aggregate_library(
2651 name = "ukernels_test_mode",
2652 aarch32_deps = [
2653 ":neon_ukernels_test_mode",
2654 ":neonfma_ukernels_test_mode",
2655 ":neonv8_ukernels_test_mode",
2656 ":asm_ukernels",
2657 ],
2658 aarch64_deps = [
2659 ":neon_ukernels_test_mode",
2660 ":neonfma_ukernels_test_mode",
2661 ":neonv8_ukernels_test_mode",
2662 ":neonfp16arith_ukernels_test_mode",
2663 ":asm_ukernels",
2664 ],
2665 generic_deps = [
2666 ":scalar_ukernels_test_mode",
2667 ],
2668 psimd_deps = [
2669 ":psimd_fastmath_ukernels_test_mode",
2670 ":psimd_accmath_ukernels_test_mode",
2671 ],
2672 wasm_deps = [
2673 ":wasm_ukernels_test_mode",
2674 ],
2675 wasmsimd_deps = [
2676 ":wasm_ukernels_test_mode",
2677 ],
2678 x86_deps = [
2679 ":sse2_ukernels_test_mode",
2680 ":ssse3_ukernels_test_mode",
2681 ":sse41_ukernels_test_mode",
2682 ":avx_ukernels_test_mode",
2683 ":fma3_ukernels_test_mode",
2684 ":avx2_ukernels_test_mode",
2685 ":avx512f_ukernels_test_mode",
2686 ],
2687)
2688
Marat Dukhan08c4a432019-10-03 09:29:21 -07002689xnnpack_cc_library(
2690 name = "im2col",
2691 srcs = ["src/im2col.c"],
2692 hdrs = [
2693 "src/xnnpack/common.h",
2694 "src/xnnpack/im2col.h",
2695 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002696 gcc_copts = xnnpack_gcc_std_copts(),
2697 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002698)
2699
2700xnnpack_cc_library(
2701 name = "indirection",
2702 srcs = ["src/indirection.c"],
2703 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002704 gcc_copts = xnnpack_gcc_std_copts(),
2705 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002706 deps = [
2707 "@FP16",
2708 "@FXdiv",
2709 "@pthreadpool",
2710 ],
2711)
2712
2713xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002714 name = "indirection_test_mode",
2715 srcs = ["src/indirection.c"],
2716 hdrs = INTERNAL_HDRS,
2717 copts = [
2718 "-UNDEBUG",
2719 "-DXNN_TEST_MODE=1",
2720 ],
2721 gcc_copts = xnnpack_gcc_std_copts(),
2722 msvc_copts = xnnpack_msvc_std_copts(),
2723 deps = [
2724 "@FP16",
2725 "@FXdiv",
2726 "@pthreadpool",
2727 ],
2728)
2729
2730xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002731 name = "operator_run",
2732 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002733 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002734 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002735 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2736 "//conditions:default": [],
2737 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002738 gcc_copts = xnnpack_gcc_std_copts(),
2739 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002741 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002742 "@FP16",
2743 "@FXdiv",
2744 "@clog",
2745 "@pthreadpool",
2746 ],
2747)
2748
Chao Mei6ddfc602020-05-13 22:29:36 -07002749xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002750 name = "operator_run_test_mode",
2751 srcs = ["src/operator-run.c"],
2752 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2753 copts = LOGGING_COPTS + [
2754 "-UNDEBUG",
2755 "-DXNN_TEST_MODE=1",
2756 ] + select({
2757 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2758 "//conditions:default": [],
2759 }),
2760 gcc_copts = xnnpack_gcc_std_copts(),
2761 msvc_copts = xnnpack_msvc_std_copts(),
2762 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002763 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002764 "@FP16",
2765 "@FXdiv",
2766 "@clog",
2767 "@pthreadpool",
2768 ],
2769)
2770
2771xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07002772 name = "memory_planner",
2773 srcs = ["src/memory-planner.c"],
2774 hdrs = INTERNAL_HDRS,
2775 defines = select({
2776 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2777 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2778 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2779 }),
2780 gcc_copts = xnnpack_gcc_std_copts(),
2781 msvc_copts = xnnpack_msvc_std_copts(),
2782 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002783 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07002784 "@pthreadpool",
2785 ],
2786)
2787
Marat Dukhan33fcf782020-05-24 14:27:15 -07002788xnnpack_cc_library(
2789 name = "memory_planner_test_mode",
2790 srcs = ["src/memory-planner.c"],
2791 hdrs = INTERNAL_HDRS,
2792 copts = [
2793 "-UNDEBUG",
2794 "-DXNN_TEST_MODE=1",
2795 ],
2796 defines = select({
2797 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2798 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2799 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2800 }),
2801 gcc_copts = xnnpack_gcc_std_copts(),
2802 msvc_copts = xnnpack_msvc_std_copts(),
2803 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002804 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002805 "@pthreadpool",
2806 ],
2807)
2808
Marat Dukhan08c4a432019-10-03 09:29:21 -07002809cc_library(
2810 name = "enable_assembly",
2811 defines = select({
2812 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2813 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002814 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002815 }),
2816)
2817
Marat Dukhancf056b22019-10-07 10:26:29 -07002818xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002819 name = "operators",
2820 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002821 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002822 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002823 ],
2824 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002825 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002826 "-Isrc",
2827 "-Iinclude",
2828 ] + select({
2829 ":debug_build": [],
2830 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002831 }) + select({
2832 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2833 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002834 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002835 gcc_copts = xnnpack_gcc_std_copts(),
2836 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002837 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002838 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002839 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840 "@FP16",
2841 "@FXdiv",
2842 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002843 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002844 ],
2845)
2846
Marat Dukhan10a38082020-04-17 03:58:35 -07002847xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002848 name = "operators_test_mode",
2849 srcs = OPERATOR_SRCS + [
2850 "src/memory.c",
2851 "src/operator-delete.c",
2852 ],
2853 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2854 copts = LOGGING_COPTS + [
2855 "-Isrc",
2856 "-Iinclude",
2857 "-UNDEBUG",
2858 "-DXNN_TEST_MODE=1",
2859 ] + select({
2860 ":debug_build": [],
2861 "//conditions:default": xnnpack_min_size_copts(),
2862 }) + select({
2863 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2864 "//conditions:default": [],
2865 }),
2866 gcc_copts = xnnpack_gcc_std_copts(),
2867 msvc_copts = xnnpack_msvc_std_copts(),
2868 deps = [
2869 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002870 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002871 "@FP16",
2872 "@FXdiv",
2873 "@clog",
2874 "@pthreadpool",
2875 ],
2876)
2877
2878xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002879 name = "XNNPACK",
2880 srcs = [
2881 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002882 "src/runtime.c",
2883 "src/subgraph.c",
2884 "src/tensor.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002885 ] + SUBGRAPH_SRCS + LOGGING_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002886 hdrs = ["include/xnnpack.h"],
2887 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002888 "-Isrc",
2889 "-Iinclude",
2890 ] + select({
2891 ":debug_build": [],
2892 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002893 }) + select({
2894 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2895 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002896 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002897 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002898 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002899 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002900 visibility = xnnpack_visibility(),
2901 deps = [
2902 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002903 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07002904 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002905 ":operator_run",
2906 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002907 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002908 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07002909 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002910 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002911 ] + select({
2912 ":emscripten": [],
2913 "//conditions:default": ["@cpuinfo"],
2914 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002915)
2916
Marat Dukhan10a38082020-04-17 03:58:35 -07002917xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002918 name = "XNNPACK_test_mode",
2919 srcs = [
2920 "src/init.c",
2921 "src/runtime.c",
2922 "src/subgraph.c",
2923 "src/tensor.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002924 ] + SUBGRAPH_SRCS + LOGGING_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002925 hdrs = ["include/xnnpack.h"],
2926 copts = LOGGING_COPTS + [
2927 "-Isrc",
2928 "-Iinclude",
2929 "-UNDEBUG",
2930 "-DXNN_TEST_MODE=1",
2931 ] + select({
2932 ":debug_build": [],
2933 "//conditions:default": xnnpack_min_size_copts(),
2934 }) + select({
2935 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2936 "//conditions:default": [],
2937 }),
2938 gcc_copts = xnnpack_gcc_std_copts(),
2939 includes = ["include"],
2940 msvc_copts = xnnpack_msvc_std_copts(),
2941 visibility = xnnpack_visibility(),
2942 deps = [
2943 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002944 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002945 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002946 ":operator_run_test_mode",
2947 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002948 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002949 "@clog",
2950 "@FP16",
2951 "@pthreadpool",
2952 ] + select({
2953 ":emscripten": [],
2954 "//conditions:default": ["@cpuinfo"],
2955 }),
2956)
2957
2958xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002959 name = "xnnpack_operators_nhwc_f32",
2960 srcs = [
2961 "src/init.c",
2962 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002963 hdrs = ["include/xnnpack.h"],
2964 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002965 "-Isrc",
2966 "-Iinclude",
2967 ] + select({
2968 ":debug_build": [],
2969 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002970 }) + select({
2971 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2972 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002973 }),
2974 defines = [
2975 "XNN_NO_Q8_OPERATORS",
2976 "XNN_NO_U8_OPERATORS",
2977 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002978 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002979 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002980 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002981 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002982 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002983 visibility = xnnpack_visibility(),
2984 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002985 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002986 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 ":operator_run",
2988 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002989 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002990 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002992 ] + select({
2993 ":emscripten": [],
2994 "//conditions:default": ["@cpuinfo"],
2995 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002996)
2997
Marat Dukhancf056b22019-10-07 10:26:29 -07002998xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 name = "bench_utils",
3000 srcs = ["bench/utils.cc"],
3001 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003002 deps = [
3003 "@com_google_benchmark//:benchmark",
3004 "@cpuinfo",
3005 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003006)
3007
Frank Barchard7e955972019-10-11 10:34:25 -07003008######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009
3010xnnpack_benchmark(
3011 name = "q8_gemm_bench",
3012 srcs = [
3013 "bench/gemm.h",
3014 "bench/q8-gemm.cc",
3015 "src/xnnpack/AlignedAllocator.h",
3016 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003017 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003018 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
3019)
3020
3021xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003022 name = "f16_igemm_bench",
3023 srcs = [
3024 "bench/f16-igemm.cc",
3025 "bench/conv.h",
3026 "bench/google/conv.h",
3027 "src/xnnpack/AlignedAllocator.h",
3028 ] + MICROKERNEL_BENCHMARK_HDRS,
3029 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3030)
3031
3032xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003033 name = "f16_gemm_bench",
3034 srcs = [
3035 "bench/f16-gemm.cc",
3036 "bench/gemm.h",
3037 "src/xnnpack/AlignedAllocator.h",
3038 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039 deps = MICROKERNEL_BENCHMARK_DEPS,
3040)
3041
3042xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003043 name = "f16_spmm_bench",
3044 srcs = [
3045 "bench/f16-spmm.cc",
3046 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003047 "src/xnnpack/AlignedAllocator.h",
3048 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003049 deps = MICROKERNEL_BENCHMARK_DEPS,
3050)
3051
3052xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003053 name = "f32_igemm_bench",
3054 srcs = [
3055 "bench/f32-igemm.cc",
3056 "bench/conv.h",
3057 "src/xnnpack/AlignedAllocator.h",
3058 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07003059 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003060)
3061
3062xnnpack_benchmark(
3063 name = "f32_conv_hwc_bench",
3064 srcs = [
3065 "bench/f32-conv-hwc.cc",
3066 "bench/dconv.h",
3067 "src/xnnpack/AlignedAllocator.h",
3068 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069 deps = MICROKERNEL_BENCHMARK_DEPS,
3070)
3071
3072xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003073 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003074 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003075 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003076 "bench/dconv.h",
3077 "src/xnnpack/AlignedAllocator.h",
3078 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07003079 deps = MICROKERNEL_BENCHMARK_DEPS,
3080)
3081
3082xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003083 name = "f16_dwconv_bench",
3084 srcs = [
3085 "bench/f16-dwconv.cc",
3086 "bench/dwconv.h",
3087 "bench/google/dwconv.h",
3088 "src/xnnpack/AlignedAllocator.h",
3089 ] + MICROKERNEL_BENCHMARK_HDRS,
3090 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3091)
3092
3093xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003094 name = "f32_dwconv_bench",
3095 srcs = [
3096 "bench/f32-dwconv.cc",
3097 "bench/dwconv.h",
3098 "src/xnnpack/AlignedAllocator.h",
3099 ] + MICROKERNEL_BENCHMARK_HDRS,
3100 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3101)
3102
3103xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003104 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003105 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003106 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003107 "bench/dwconv.h",
3108 "src/xnnpack/AlignedAllocator.h",
3109 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3110 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3111)
3112
3113xnnpack_benchmark(
3114 name = "f32_gemm_bench",
3115 srcs = [
3116 "bench/f32-gemm.cc",
3117 "bench/gemm.h",
3118 "src/xnnpack/AlignedAllocator.h",
3119 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003120 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07003121 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122)
3123
3124xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003125 name = "f32_raddexpminusmax_bench",
3126 srcs = [
3127 "bench/f32-raddexpminusmax.cc",
3128 "src/xnnpack/AlignedAllocator.h",
3129 ] + MICROKERNEL_BENCHMARK_HDRS,
3130 deps = MICROKERNEL_BENCHMARK_DEPS,
3131)
3132
3133xnnpack_benchmark(
3134 name = "f32_raddextexp_bench",
3135 srcs = [
3136 "bench/f32-raddextexp.cc",
3137 "src/xnnpack/AlignedAllocator.h",
3138 ] + MICROKERNEL_BENCHMARK_HDRS,
3139 deps = MICROKERNEL_BENCHMARK_DEPS,
3140)
3141
3142xnnpack_benchmark(
3143 name = "f32_raddstoreexpminusmax_bench",
3144 srcs = [
3145 "bench/f32-raddstoreexpminusmax.cc",
3146 "src/xnnpack/AlignedAllocator.h",
3147 ] + MICROKERNEL_BENCHMARK_HDRS,
3148 deps = MICROKERNEL_BENCHMARK_DEPS,
3149)
3150
3151xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003152 name = "f32_rmax_bench",
3153 srcs = [
3154 "bench/f32-rmax.cc",
3155 "src/xnnpack/AlignedAllocator.h",
3156 ] + MICROKERNEL_BENCHMARK_HDRS,
3157 deps = MICROKERNEL_BENCHMARK_DEPS,
3158)
3159
3160xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003161 name = "f32_sigmoid_bench",
3162 srcs = [
3163 "bench/f32-sigmoid.cc",
3164 "src/xnnpack/AlignedAllocator.h",
3165 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003166 deps = MICROKERNEL_BENCHMARK_DEPS,
3167)
3168
3169xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170 name = "f32_spmm_bench",
3171 srcs = [
3172 "bench/f32-spmm.cc",
3173 "bench/gemm.h",
3174 "src/xnnpack/AlignedAllocator.h",
3175 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003176 deps = MICROKERNEL_BENCHMARK_DEPS,
3177)
3178
3179xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003180 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003181 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003182 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003183 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003184 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003185 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003186)
3187
3188xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003189 name = "f32_vscaleexpminusmax_bench",
3190 srcs = [
3191 "bench/f32-vscaleexpminusmax.cc",
3192 "src/xnnpack/AlignedAllocator.h",
3193 ] + MICROKERNEL_BENCHMARK_HDRS,
3194 deps = MICROKERNEL_BENCHMARK_DEPS,
3195)
3196
3197xnnpack_benchmark(
3198 name = "f32_vscaleextexp_bench",
3199 srcs = [
3200 "bench/f32-vscaleextexp.cc",
3201 "src/xnnpack/AlignedAllocator.h",
3202 ] + MICROKERNEL_BENCHMARK_HDRS,
3203 deps = MICROKERNEL_BENCHMARK_DEPS,
3204)
3205
3206xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 name = "f32_im2col_gemm_bench",
3208 srcs = [
3209 "bench/f32-im2col-gemm.cc",
3210 "bench/conv.h",
3211 "src/xnnpack/AlignedAllocator.h",
3212 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3213 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
3214)
3215
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003216xnnpack_benchmark(
3217 name = "requantization_bench",
3218 srcs = [
3219 "bench/requantization.cc",
3220 "src/xnnpack/requantization-stubs.h",
3221 "src/xnnpack/AlignedAllocator.h",
3222 ] + MICROKERNEL_BENCHMARK_HDRS,
3223 deps = MICROKERNEL_BENCHMARK_DEPS,
3224)
3225
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003226xnnpack_benchmark(
3227 name = "rounding_bench",
3228 srcs = [
3229 "bench/rounding.cc",
3230 "src/xnnpack/math-stubs.h",
3231 "src/xnnpack/AlignedAllocator.h",
3232 ] + MICROKERNEL_BENCHMARK_HDRS,
3233 deps = MICROKERNEL_BENCHMARK_DEPS,
3234)
3235
Marat Dukhan08c4a432019-10-03 09:29:21 -07003236########################### Benchmarks for operators ###########################
3237
3238xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003239 name = "average_pooling_bench",
3240 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003241 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003242 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003243 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003244)
3245
3246xnnpack_benchmark(
3247 name = "channel_shuffle_bench",
3248 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003249 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250)
3251
3252xnnpack_benchmark(
3253 name = "convolution_bench",
3254 srcs = ["bench/convolution.cc"],
3255 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003256 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003257 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003258)
3259
3260xnnpack_benchmark(
3261 name = "deconvolution_bench",
3262 srcs = ["bench/deconvolution.cc"],
3263 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003264 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003265 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003266)
3267
3268xnnpack_benchmark(
3269 name = "global_average_pooling_bench",
3270 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003271 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272)
3273
3274xnnpack_benchmark(
3275 name = "max_pooling_bench",
3276 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003277 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278)
3279
3280xnnpack_benchmark(
3281 name = "sigmoid_bench",
3282 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003283 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003284 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003285 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003286)
3287
3288xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003289 name = "prelu_bench",
3290 srcs = ["bench/prelu.cc"],
3291 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003292 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003293 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003294)
3295
3296xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003297 name = "softmax_bench",
3298 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003299 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003300 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003301 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003302)
3303
Marat Dukhanc068bb62019-10-04 13:24:39 -07003304############################# End-to-end benchmarks ############################
3305
3306cc_library(
3307 name = "mobilenet_v1",
3308 srcs = ["models/mobilenet-v1.cc"],
3309 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003310 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003311 linkstatic = True,
3312 deps = [
3313 ":XNNPACK",
3314 "@pthreadpool",
3315 ],
3316)
3317
3318cc_library(
3319 name = "mobilenet_v2",
3320 srcs = ["models/mobilenet-v2.cc"],
3321 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003322 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003323 linkstatic = True,
3324 deps = [
3325 ":XNNPACK",
3326 "@pthreadpool",
3327 ],
3328)
3329
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003330cc_library(
3331 name = "mobilenet_v3_large",
3332 srcs = ["models/mobilenet-v3-large.cc"],
3333 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003334 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003335 linkstatic = True,
3336 deps = [
3337 ":XNNPACK",
3338 "@pthreadpool",
3339 ],
3340)
3341
3342cc_library(
3343 name = "mobilenet_v3_small",
3344 srcs = ["models/mobilenet-v3-small.cc"],
3345 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003346 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003347 linkstatic = True,
3348 deps = [
3349 ":XNNPACK",
3350 "@pthreadpool",
3351 ],
3352)
3353
Marat Dukhanc068bb62019-10-04 13:24:39 -07003354xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003355 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003356 srcs = [
3357 "bench/f32-dwconv-e2e.cc",
3358 "bench/end2end.h",
3359 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003360 deps = MICROKERNEL_BENCHMARK_DEPS + [
3361 ":XNNPACK",
3362 ":mobilenet_v1",
3363 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003364 ":mobilenet_v3_large",
3365 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003366 ],
3367)
3368
3369xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003370 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003371 srcs = [
3372 "bench/f32-gemm-e2e.cc",
3373 "bench/end2end.h",
3374 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003375 deps = MICROKERNEL_BENCHMARK_DEPS + [
3376 ":XNNPACK",
3377 ":mobilenet_v1",
3378 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003379 ":mobilenet_v3_large",
3380 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003381 ],
3382)
3383
3384xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003385 name = "end2end_bench",
3386 srcs = ["bench/end2end.cc"],
3387 deps = [
3388 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003389 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003390 ":mobilenet_v1",
3391 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003392 ":mobilenet_v3_large",
3393 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003394 "@pthreadpool",
3395 ],
3396)
3397
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003398#################### Accuracy evaluation for math functions ####################
3399
3400xnnpack_benchmark(
3401 name = "f32_exp_eval",
3402 srcs = [
3403 "eval/f32-exp.cc",
3404 "src/xnnpack/AlignedAllocator.h",
3405 ] + ACCURACY_EVAL_HDRS,
3406 deps = ACCURACY_EVAL_DEPS,
3407)
3408
Marat Dukhan515c9772019-10-17 18:07:57 -07003409xnnpack_benchmark(
3410 name = "f32_expminus_eval",
3411 srcs = [
3412 "eval/f32-expminus.cc",
3413 "src/xnnpack/AlignedAllocator.h",
3414 ] + ACCURACY_EVAL_HDRS,
3415 deps = ACCURACY_EVAL_DEPS,
3416)
3417
Marat Dukhan98ba4412019-10-23 02:14:28 -07003418xnnpack_benchmark(
3419 name = "f32_extexp_eval",
3420 srcs = [
3421 "eval/f32-extexp.cc",
3422 "src/xnnpack/AlignedAllocator.h",
3423 ] + ACCURACY_EVAL_HDRS,
3424 deps = ACCURACY_EVAL_DEPS,
3425)
3426
Marat Dukhan8853b822020-05-07 12:19:01 -07003427xnnpack_unit_test(
3428 name = "f32_roundne_eval",
3429 srcs = [
3430 "eval/f32-roundne.cc",
3431 "src/xnnpack/AlignedAllocator.h",
3432 "src/xnnpack/math-stubs.h",
3433 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07003434 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07003435 deps = MICROKERNEL_TEST_DEPS,
3436)
3437
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003438xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003439 name = "f32_roundd_eval",
3440 srcs = [
3441 "eval/f32-roundd.cc",
3442 "src/xnnpack/AlignedAllocator.h",
3443 "src/xnnpack/math-stubs.h",
3444 ] + MICROKERNEL_TEST_HDRS,
3445 automatic = False,
3446 deps = MICROKERNEL_TEST_DEPS,
3447)
3448
3449xnnpack_unit_test(
3450 name = "f32_roundu_eval",
3451 srcs = [
3452 "eval/f32-roundu.cc",
3453 "src/xnnpack/AlignedAllocator.h",
3454 "src/xnnpack/math-stubs.h",
3455 ] + MICROKERNEL_TEST_HDRS,
3456 automatic = False,
3457 deps = MICROKERNEL_TEST_DEPS,
3458)
3459
3460xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003461 name = "f32_roundz_eval",
3462 srcs = [
3463 "eval/f32-roundz.cc",
3464 "src/xnnpack/AlignedAllocator.h",
3465 "src/xnnpack/math-stubs.h",
3466 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003467 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003468 deps = MICROKERNEL_TEST_DEPS,
3469)
3470
Marat Dukhan346a9e52019-11-15 09:06:30 -08003471xnnpack_benchmark(
3472 name = "f32_sigmoid_eval",
3473 srcs = [
3474 "eval/f32-sigmoid.cc",
3475 "src/xnnpack/AlignedAllocator.h",
3476 ] + ACCURACY_EVAL_HDRS,
3477 deps = ACCURACY_EVAL_DEPS,
3478)
3479
Marat Dukhan08c4a432019-10-03 09:29:21 -07003480######################### Unit tests for micro-kernels #########################
3481
3482xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003483 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003485 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003486 "test/gemm-microkernel-tester.h",
3487 "src/xnnpack/AlignedAllocator.h",
3488 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3489 deps = MICROKERNEL_TEST_DEPS,
3490)
3491
3492xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003493 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003494 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003495 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003496 "test/spmm-microkernel-tester.h",
3497 "src/xnnpack/AlignedAllocator.h",
3498 ] + MICROKERNEL_TEST_HDRS,
3499 deps = MICROKERNEL_TEST_DEPS,
3500)
3501
3502xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003503 name = "f32_argmaxpool_test",
3504 srcs = [
3505 "test/f32-argmaxpool.cc",
3506 "test/argmaxpool-microkernel-tester.h",
3507 "src/xnnpack/AlignedAllocator.h",
3508 ] + MICROKERNEL_TEST_HDRS,
3509 deps = MICROKERNEL_TEST_DEPS,
3510)
3511
3512xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003513 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003514 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003515 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003516 "test/avgpool-microkernel-tester.h",
3517 "src/xnnpack/AlignedAllocator.h",
3518 ] + MICROKERNEL_TEST_HDRS,
3519 deps = MICROKERNEL_TEST_DEPS,
3520)
3521
3522xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07003523 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003524 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07003525 "test/f32-ibilinear.cc",
3526 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003527 "src/xnnpack/AlignedAllocator.h",
3528 ] + MICROKERNEL_TEST_HDRS,
3529 deps = MICROKERNEL_TEST_DEPS,
3530)
3531
3532xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003533 name = "f16_clamp_test",
3534 srcs = [
3535 "test/f16-clamp.cc",
3536 "test/clamp-microkernel-tester.h",
3537 ] + MICROKERNEL_TEST_HDRS,
3538 deps = MICROKERNEL_TEST_DEPS,
3539)
3540
3541xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003542 name = "f32_clamp_test",
3543 srcs = [
3544 "test/f32-clamp.cc",
3545 "test/clamp-microkernel-tester.h",
3546 ] + MICROKERNEL_TEST_HDRS,
3547 deps = MICROKERNEL_TEST_DEPS,
3548)
3549
3550xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003551 name = "f32_igemm_test",
3552 srcs = [
3553 "test/f32-igemm.cc",
3554 "test/gemm-microkernel-tester.h",
3555 "src/xnnpack/AlignedAllocator.h",
3556 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3557 deps = MICROKERNEL_TEST_DEPS,
3558)
3559
3560xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07003561 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003562 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07003563 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 "test/gemm-microkernel-tester.h",
3565 "src/xnnpack/AlignedAllocator.h",
3566 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3567 deps = MICROKERNEL_TEST_DEPS,
3568)
3569
3570xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07003571 name = "f32_igemm_minmax_test",
3572 srcs = [
3573 "test/f32-igemm-minmax.cc",
3574 "test/gemm-microkernel-tester.h",
3575 "src/xnnpack/AlignedAllocator.h",
3576 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3577 deps = MICROKERNEL_TEST_DEPS,
3578)
3579
3580xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07003581 name = "f16_igemm_minmax_test",
3582 srcs = [
3583 "test/f16-igemm-minmax.cc",
3584 "test/gemm-microkernel-tester.h",
3585 "src/xnnpack/AlignedAllocator.h",
3586 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3587 deps = MICROKERNEL_TEST_DEPS,
3588)
3589
3590xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003591 name = "f32_conv_hwc_test",
3592 srcs = [
3593 "test/f32-conv-hwc.cc",
3594 "test/conv-hwc-microkernel-tester.h",
3595 "src/xnnpack/AlignedAllocator.h",
3596 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3597 deps = MICROKERNEL_TEST_DEPS,
3598)
3599
3600xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003601 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003603 "test/f32-conv-hwc2chw.cc",
3604 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003605 "src/xnnpack/AlignedAllocator.h",
3606 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3607 deps = MICROKERNEL_TEST_DEPS,
3608)
3609
3610xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003611 name = "f32_dwconv_test",
3612 srcs = [
3613 "test/f32-dwconv.cc",
3614 "test/dwconv-microkernel-tester.h",
3615 "src/xnnpack/AlignedAllocator.h",
3616 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3617 deps = MICROKERNEL_TEST_DEPS,
3618)
3619
3620xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07003621 name = "f16_dwconv_minmax_test",
3622 srcs = [
3623 "test/f16-dwconv-minmax.cc",
3624 "test/dwconv-microkernel-tester.h",
3625 "src/xnnpack/AlignedAllocator.h",
3626 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3627 deps = MICROKERNEL_TEST_DEPS,
3628)
3629
3630xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003631 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003633 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 "test/dwconv-microkernel-tester.h",
3635 "src/xnnpack/AlignedAllocator.h",
3636 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3637 deps = MICROKERNEL_TEST_DEPS,
3638)
3639
3640xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003641 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003643 "test/f32-dwconv-chw.cc",
3644 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 "src/xnnpack/AlignedAllocator.h",
3646 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3647 deps = MICROKERNEL_TEST_DEPS,
3648)
3649
3650xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07003651 name = "f16_gavgpool_minmax_test",
3652 srcs = [
3653 "test/f16-gavgpool-minmax.cc",
3654 "test/gavgpool-microkernel-tester.h",
3655 "src/xnnpack/AlignedAllocator.h",
3656 ] + MICROKERNEL_TEST_HDRS,
3657 deps = MICROKERNEL_TEST_DEPS,
3658)
3659
3660xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003661 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003663 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003664 "test/gavgpool-microkernel-tester.h",
3665 "src/xnnpack/AlignedAllocator.h",
3666 ] + MICROKERNEL_TEST_HDRS,
3667 deps = MICROKERNEL_TEST_DEPS,
3668)
3669
3670xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003671 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003672 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003673 "test/f32-gavgpool-cw.cc",
3674 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003675 "src/xnnpack/AlignedAllocator.h",
3676 ] + MICROKERNEL_TEST_HDRS,
3677 deps = MICROKERNEL_TEST_DEPS,
3678)
3679
3680xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003681 name = "f32_gemm_test",
3682 srcs = [
3683 "test/f32-gemm.cc",
3684 "test/gemm-microkernel-tester.h",
3685 "src/xnnpack/AlignedAllocator.h",
3686 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3687 deps = MICROKERNEL_TEST_DEPS,
3688)
3689
3690xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07003691 name = "f32_gemm_relu_test",
3692 srcs = [
3693 "test/f32-gemm-relu.cc",
3694 "test/gemm-microkernel-tester.h",
3695 "src/xnnpack/AlignedAllocator.h",
3696 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3697 deps = MICROKERNEL_TEST_DEPS,
3698)
3699
3700xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003701 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003702 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003703 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003704 "test/gemm-microkernel-tester.h",
3705 "src/xnnpack/AlignedAllocator.h",
3706 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3707 deps = MICROKERNEL_TEST_DEPS,
3708)
3709
3710xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003711 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003712 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003713 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003714 "test/gemm-microkernel-tester.h",
3715 "src/xnnpack/AlignedAllocator.h",
3716 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3717 deps = MICROKERNEL_TEST_DEPS,
3718)
3719
3720xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003721 name = "f16_hswish_test",
3722 srcs = [
3723 "test/f16-hswish.cc",
3724 "test/hswish-microkernel-tester.h",
3725 ] + MICROKERNEL_TEST_HDRS,
3726 deps = MICROKERNEL_TEST_DEPS,
3727)
3728
3729xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003730 name = "f32_hswish_test",
3731 srcs = [
3732 "test/f32-hswish.cc",
3733 "test/hswish-microkernel-tester.h",
3734 ] + MICROKERNEL_TEST_HDRS,
3735 deps = MICROKERNEL_TEST_DEPS,
3736)
3737
3738xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003739 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003740 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003741 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003742 "test/maxpool-microkernel-tester.h",
3743 ] + MICROKERNEL_TEST_HDRS,
3744 deps = MICROKERNEL_TEST_DEPS,
3745)
3746
3747xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003748 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003749 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003750 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003751 "test/avgpool-microkernel-tester.h",
3752 "src/xnnpack/AlignedAllocator.h",
3753 ] + MICROKERNEL_TEST_HDRS,
3754 deps = MICROKERNEL_TEST_DEPS,
3755)
3756
3757xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003758 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003760 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003761 "test/gemm-microkernel-tester.h",
3762 "src/xnnpack/AlignedAllocator.h",
3763 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3764 deps = MICROKERNEL_TEST_DEPS,
3765)
3766
3767xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003768 name = "f16_prelu_test",
3769 srcs = [
3770 "test/f16-prelu.cc",
3771 "test/prelu-microkernel-tester.h",
3772 "src/xnnpack/AlignedAllocator.h",
3773 ] + MICROKERNEL_TEST_HDRS,
3774 deps = MICROKERNEL_TEST_DEPS,
3775)
3776
3777xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003778 name = "f32_prelu_test",
3779 srcs = [
3780 "test/f32-prelu.cc",
3781 "test/prelu-microkernel-tester.h",
3782 "src/xnnpack/AlignedAllocator.h",
3783 ] + MICROKERNEL_TEST_HDRS,
3784 deps = MICROKERNEL_TEST_DEPS,
3785)
3786
3787xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003788 name = "f32_raddexpminusmax_test",
3789 srcs = [
3790 "test/f32-raddexpminusmax.cc",
3791 "test/raddexpminusmax-microkernel-tester.h",
3792 ] + MICROKERNEL_TEST_HDRS,
3793 deps = MICROKERNEL_TEST_DEPS,
3794)
3795
3796xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003797 name = "f32_raddextexp_test",
3798 srcs = [
3799 "test/f32-raddextexp.cc",
3800 "test/raddextexp-microkernel-tester.h",
3801 ] + MICROKERNEL_TEST_HDRS,
3802 deps = MICROKERNEL_TEST_DEPS,
3803)
3804
3805xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003806 name = "f32_raddstoreexpminusmax_test",
3807 srcs = [
3808 "test/f32-raddstoreexpminusmax.cc",
3809 "test/raddstoreexpminusmax-microkernel-tester.h",
3810 ] + MICROKERNEL_TEST_HDRS,
3811 deps = MICROKERNEL_TEST_DEPS,
3812)
3813
3814xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003815 name = "f32_rmax_test",
3816 srcs = [
3817 "test/f32-rmax.cc",
3818 "test/rmax-microkernel-tester.h",
3819 ] + MICROKERNEL_TEST_HDRS,
3820 deps = MICROKERNEL_TEST_DEPS,
3821)
3822
3823xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08003824 name = "f32_sigmoid_test",
3825 srcs = [
3826 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003827 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08003828 ] + MICROKERNEL_TEST_HDRS,
3829 deps = MICROKERNEL_TEST_DEPS,
3830)
3831
3832xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003833 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003834 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003835 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003836 "test/spmm-microkernel-tester.h",
3837 "src/xnnpack/AlignedAllocator.h",
3838 ] + MICROKERNEL_TEST_HDRS,
3839 deps = MICROKERNEL_TEST_DEPS,
3840)
3841
3842xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07003843 name = "f16_vadd_minmax_test",
3844 srcs = [
3845 "test/f16-vadd-minmax.cc",
3846 "test/vbinary-microkernel-tester.h",
3847 ] + MICROKERNEL_TEST_HDRS,
3848 deps = MICROKERNEL_TEST_DEPS,
3849)
3850
3851xnnpack_unit_test(
3852 name = "f16_vaddc_minmax_test",
3853 srcs = [
3854 "test/f16-vaddc-minmax.cc",
3855 "test/vbinaryc-microkernel-tester.h",
3856 ] + MICROKERNEL_TEST_HDRS,
3857 deps = MICROKERNEL_TEST_DEPS,
3858)
3859
3860xnnpack_unit_test(
3861 name = "f16_vdiv_minmax_test",
3862 srcs = [
3863 "test/f16-vdiv-minmax.cc",
3864 "test/vbinary-microkernel-tester.h",
3865 ] + MICROKERNEL_TEST_HDRS,
3866 deps = MICROKERNEL_TEST_DEPS,
3867)
3868
3869xnnpack_unit_test(
3870 name = "f16_vdivc_minmax_test",
3871 srcs = [
3872 "test/f16-vdivc-minmax.cc",
3873 "test/vbinaryc-microkernel-tester.h",
3874 ] + MICROKERNEL_TEST_HDRS,
3875 deps = MICROKERNEL_TEST_DEPS,
3876)
3877
3878xnnpack_unit_test(
3879 name = "f16_vrdivc_minmax_test",
3880 srcs = [
3881 "test/f16-vrdivc-minmax.cc",
3882 "test/vbinaryc-microkernel-tester.h",
3883 ] + MICROKERNEL_TEST_HDRS,
3884 deps = MICROKERNEL_TEST_DEPS,
3885)
3886
3887xnnpack_unit_test(
3888 name = "f16_vmax_test",
3889 srcs = [
3890 "test/f16-vmax.cc",
3891 "test/vbinary-microkernel-tester.h",
3892 ] + MICROKERNEL_TEST_HDRS,
3893 deps = MICROKERNEL_TEST_DEPS,
3894)
3895
3896xnnpack_unit_test(
3897 name = "f16_vmaxc_test",
3898 srcs = [
3899 "test/f16-vmaxc.cc",
3900 "test/vbinaryc-microkernel-tester.h",
3901 ] + MICROKERNEL_TEST_HDRS,
3902 deps = MICROKERNEL_TEST_DEPS,
3903)
3904
3905xnnpack_unit_test(
3906 name = "f16_vmin_test",
3907 srcs = [
3908 "test/f16-vmin.cc",
3909 "test/vbinary-microkernel-tester.h",
3910 ] + MICROKERNEL_TEST_HDRS,
3911 deps = MICROKERNEL_TEST_DEPS,
3912)
3913
3914xnnpack_unit_test(
3915 name = "f16_vminc_test",
3916 srcs = [
3917 "test/f16-vminc.cc",
3918 "test/vbinaryc-microkernel-tester.h",
3919 ] + MICROKERNEL_TEST_HDRS,
3920 deps = MICROKERNEL_TEST_DEPS,
3921)
3922
3923xnnpack_unit_test(
3924 name = "f16_vmul_minmax_test",
3925 srcs = [
3926 "test/f16-vmul-minmax.cc",
3927 "test/vbinary-microkernel-tester.h",
3928 ] + MICROKERNEL_TEST_HDRS,
3929 deps = MICROKERNEL_TEST_DEPS,
3930)
3931
3932xnnpack_unit_test(
3933 name = "f16_vmulc_minmax_test",
3934 srcs = [
3935 "test/f16-vmulc-minmax.cc",
3936 "test/vbinaryc-microkernel-tester.h",
3937 ] + MICROKERNEL_TEST_HDRS,
3938 deps = MICROKERNEL_TEST_DEPS,
3939)
3940
3941xnnpack_unit_test(
3942 name = "f16_vsub_minmax_test",
3943 srcs = [
3944 "test/f16-vsub-minmax.cc",
3945 "test/vbinary-microkernel-tester.h",
3946 ] + MICROKERNEL_TEST_HDRS,
3947 deps = MICROKERNEL_TEST_DEPS,
3948)
3949
3950xnnpack_unit_test(
3951 name = "f16_vsubc_minmax_test",
3952 srcs = [
3953 "test/f16-vsubc-minmax.cc",
3954 "test/vbinaryc-microkernel-tester.h",
3955 ] + MICROKERNEL_TEST_HDRS,
3956 deps = MICROKERNEL_TEST_DEPS,
3957)
3958
3959xnnpack_unit_test(
3960 name = "f16_vrsubc_minmax_test",
3961 srcs = [
3962 "test/f16-vrsubc-minmax.cc",
3963 "test/vbinaryc-microkernel-tester.h",
3964 ] + MICROKERNEL_TEST_HDRS,
3965 deps = MICROKERNEL_TEST_DEPS,
3966)
3967
3968xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07003969 name = "f32_vabs_test",
3970 srcs = [
3971 "test/f32-vabs.cc",
3972 "test/vunary-microkernel-tester.h",
3973 ] + MICROKERNEL_TEST_HDRS,
3974 deps = MICROKERNEL_TEST_DEPS,
3975)
3976
3977xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003978 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003979 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003980 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003981 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003982 ] + MICROKERNEL_TEST_HDRS,
3983 deps = MICROKERNEL_TEST_DEPS,
3984)
3985
3986xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003987 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003988 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003989 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003990 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003991 ] + MICROKERNEL_TEST_HDRS,
3992 deps = MICROKERNEL_TEST_DEPS,
3993)
3994
3995xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003996 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003997 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003998 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003999 "test/vbinary-microkernel-tester.h",
4000 ] + MICROKERNEL_TEST_HDRS,
4001 deps = MICROKERNEL_TEST_DEPS,
4002)
4003
4004xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004005 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004006 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004007 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004008 "test/vbinaryc-microkernel-tester.h",
4009 ] + MICROKERNEL_TEST_HDRS,
4010 deps = MICROKERNEL_TEST_DEPS,
4011)
4012
4013xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004014 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004015 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004016 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004017 "test/vbinaryc-microkernel-tester.h",
4018 ] + MICROKERNEL_TEST_HDRS,
4019 deps = MICROKERNEL_TEST_DEPS,
4020)
4021
4022xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004023 name = "f32_vmax_test",
4024 srcs = [
4025 "test/f32-vmax.cc",
4026 "test/vbinary-microkernel-tester.h",
4027 ] + MICROKERNEL_TEST_HDRS,
4028 deps = MICROKERNEL_TEST_DEPS,
4029)
4030
4031xnnpack_unit_test(
4032 name = "f32_vmaxc_test",
4033 srcs = [
4034 "test/f32-vmaxc.cc",
4035 "test/vbinaryc-microkernel-tester.h",
4036 ] + MICROKERNEL_TEST_HDRS,
4037 deps = MICROKERNEL_TEST_DEPS,
4038)
4039
4040xnnpack_unit_test(
4041 name = "f32_vmin_test",
4042 srcs = [
4043 "test/f32-vmin.cc",
4044 "test/vbinary-microkernel-tester.h",
4045 ] + MICROKERNEL_TEST_HDRS,
4046 deps = MICROKERNEL_TEST_DEPS,
4047)
4048
4049xnnpack_unit_test(
4050 name = "f32_vminc_test",
4051 srcs = [
4052 "test/f32-vminc.cc",
4053 "test/vbinaryc-microkernel-tester.h",
4054 ] + MICROKERNEL_TEST_HDRS,
4055 deps = MICROKERNEL_TEST_DEPS,
4056)
4057
4058xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004059 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004060 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004061 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004062 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004063 ] + MICROKERNEL_TEST_HDRS,
4064 deps = MICROKERNEL_TEST_DEPS,
4065)
4066
4067xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004068 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004069 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004070 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004071 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004072 ] + MICROKERNEL_TEST_HDRS,
4073 deps = MICROKERNEL_TEST_DEPS,
4074)
4075
4076xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004077 name = "f16_vmulcaddc_minmax_test",
4078 srcs = [
4079 "test/f16-vmulcaddc-minmax.cc",
4080 "test/vmulcaddc-microkernel-tester.h",
4081 "src/xnnpack/AlignedAllocator.h",
4082 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4083 deps = MICROKERNEL_TEST_DEPS,
4084)
4085
4086xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004087 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004088 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004089 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090 "test/vmulcaddc-microkernel-tester.h",
4091 "src/xnnpack/AlignedAllocator.h",
4092 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4093 deps = MICROKERNEL_TEST_DEPS,
4094)
4095
4096xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004097 name = "f32_vneg_test",
4098 srcs = [
4099 "test/f32-vneg.cc",
4100 "test/vunary-microkernel-tester.h",
4101 ] + MICROKERNEL_TEST_HDRS,
4102 deps = MICROKERNEL_TEST_DEPS,
4103)
4104
4105xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004106 name = "f32_vrndne_test",
4107 srcs = [
4108 "test/f32-vrndne.cc",
4109 "test/vunary-microkernel-tester.h",
4110 ] + MICROKERNEL_TEST_HDRS,
4111 deps = MICROKERNEL_TEST_DEPS,
4112)
4113
4114xnnpack_unit_test(
4115 name = "f32_vrndz_test",
4116 srcs = [
4117 "test/f32-vrndz.cc",
4118 "test/vunary-microkernel-tester.h",
4119 ] + MICROKERNEL_TEST_HDRS,
4120 deps = MICROKERNEL_TEST_DEPS,
4121)
4122
4123xnnpack_unit_test(
4124 name = "f32_vrndu_test",
4125 srcs = [
4126 "test/f32-vrndu.cc",
4127 "test/vunary-microkernel-tester.h",
4128 ] + MICROKERNEL_TEST_HDRS,
4129 deps = MICROKERNEL_TEST_DEPS,
4130)
4131
4132xnnpack_unit_test(
4133 name = "f32_vrndd_test",
4134 srcs = [
4135 "test/f32-vrndd.cc",
4136 "test/vunary-microkernel-tester.h",
4137 ] + MICROKERNEL_TEST_HDRS,
4138 deps = MICROKERNEL_TEST_DEPS,
4139)
4140
4141xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004142 name = "f32_vscale_test",
4143 srcs = [
4144 "test/f32-vscale.cc",
4145 "test/vscale-microkernel-tester.h",
4146 ] + MICROKERNEL_TEST_HDRS,
4147 deps = MICROKERNEL_TEST_DEPS,
4148)
4149
4150xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004151 name = "f32_vscaleexpminusmax_test",
4152 srcs = [
4153 "test/f32-vscaleexpminusmax.cc",
4154 "test/vscaleexpminusmax-microkernel-tester.h",
4155 ] + MICROKERNEL_TEST_HDRS,
4156 deps = MICROKERNEL_TEST_DEPS,
4157)
4158
4159xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004160 name = "f32_vscaleextexp_test",
4161 srcs = [
4162 "test/f32-vscaleextexp.cc",
4163 "test/vscaleextexp-microkernel-tester.h",
4164 ] + MICROKERNEL_TEST_HDRS,
4165 deps = MICROKERNEL_TEST_DEPS,
4166)
4167
4168xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004169 name = "f32_vsqr_test",
4170 srcs = [
4171 "test/f32-vsqr.cc",
4172 "test/vunary-microkernel-tester.h",
4173 ] + MICROKERNEL_TEST_HDRS,
4174 deps = MICROKERNEL_TEST_DEPS,
4175)
4176
4177xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004178 name = "f32_vsqrdiff_test",
4179 srcs = [
4180 "test/f32-vsqrdiff.cc",
4181 "test/vbinary-microkernel-tester.h",
4182 ] + MICROKERNEL_TEST_HDRS,
4183 deps = MICROKERNEL_TEST_DEPS,
4184)
4185
4186xnnpack_unit_test(
4187 name = "f32_vsqrdiffc_test",
4188 srcs = [
4189 "test/f32-vsqrdiffc.cc",
4190 "test/vbinaryc-microkernel-tester.h",
4191 ] + MICROKERNEL_TEST_HDRS,
4192 deps = MICROKERNEL_TEST_DEPS,
4193)
4194
4195xnnpack_unit_test(
4196 name = "f32_vrsqrdiffc_test",
4197 srcs = [
4198 "test/f32-vrsqrdiffc.cc",
4199 "test/vbinaryc-microkernel-tester.h",
4200 ] + MICROKERNEL_TEST_HDRS,
4201 deps = MICROKERNEL_TEST_DEPS,
4202)
4203
4204xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004205 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004206 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004207 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004208 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004209 ] + MICROKERNEL_TEST_HDRS,
4210 deps = MICROKERNEL_TEST_DEPS,
4211)
4212
4213xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004214 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004215 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004216 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004217 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004218 ] + MICROKERNEL_TEST_HDRS,
4219 deps = MICROKERNEL_TEST_DEPS,
4220)
4221
4222xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004223 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004224 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004225 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004226 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004227 ] + MICROKERNEL_TEST_HDRS,
4228 deps = MICROKERNEL_TEST_DEPS,
4229)
4230
4231xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004232 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004233 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004234 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004235 "test/avgpool-microkernel-tester.h",
4236 "src/xnnpack/AlignedAllocator.h",
4237 ] + MICROKERNEL_TEST_HDRS,
4238 deps = MICROKERNEL_TEST_DEPS,
4239)
4240
4241xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004242 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004243 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004244 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245 "test/gemm-microkernel-tester.h",
4246 "src/xnnpack/AlignedAllocator.h",
4247 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4248 deps = MICROKERNEL_TEST_DEPS,
4249)
4250
4251xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004252 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004254 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004255 "test/dwconv-microkernel-tester.h",
4256 "src/xnnpack/AlignedAllocator.h",
4257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4258 deps = MICROKERNEL_TEST_DEPS,
4259)
4260
4261xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004262 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004264 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004265 "test/gavgpool-microkernel-tester.h",
4266 "src/xnnpack/AlignedAllocator.h",
4267 ] + MICROKERNEL_TEST_HDRS,
4268 deps = MICROKERNEL_TEST_DEPS,
4269)
4270
4271xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004272 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004273 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004274 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275 "test/gemm-microkernel-tester.h",
4276 "src/xnnpack/AlignedAllocator.h",
4277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4278 deps = MICROKERNEL_TEST_DEPS,
4279)
4280
4281xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004282 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004284 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285 "test/vadd-microkernel-tester.h",
4286 ] + MICROKERNEL_TEST_HDRS,
4287 deps = MICROKERNEL_TEST_DEPS,
4288)
4289
4290xnnpack_unit_test(
4291 name = "u8_clamp_test",
4292 srcs = [
4293 "test/u8-clamp.cc",
4294 "test/clamp-microkernel-tester.h",
4295 ] + MICROKERNEL_TEST_HDRS,
4296 deps = MICROKERNEL_TEST_DEPS,
4297)
4298
4299xnnpack_unit_test(
4300 name = "u8_lut32norm_test",
4301 srcs = [
4302 "test/u8-lut32norm.cc",
4303 "test/lut-norm-microkernel-tester.h",
4304 ] + MICROKERNEL_TEST_HDRS,
4305 deps = MICROKERNEL_TEST_DEPS,
4306)
4307
4308xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004309 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004311 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004312 "test/maxpool-microkernel-tester.h",
4313 ] + MICROKERNEL_TEST_HDRS,
4314 deps = MICROKERNEL_TEST_DEPS,
4315)
4316
4317xnnpack_unit_test(
4318 name = "u8_rmax_test",
4319 srcs = [
4320 "test/u8-rmax.cc",
4321 "test/rmax-microkernel-tester.h",
4322 ] + MICROKERNEL_TEST_HDRS,
4323 deps = MICROKERNEL_TEST_DEPS,
4324)
4325
4326xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07004327 name = "x32_fill_test",
4328 srcs = [
4329 "test/x32-fill.cc",
4330 "test/fill-microkernel-tester.h",
4331 ] + MICROKERNEL_TEST_HDRS,
4332 deps = MICROKERNEL_TEST_DEPS,
4333)
4334
4335xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004336 name = "x32_packx_test",
4337 srcs = [
4338 "test/x32-packx.cc",
4339 "test/pack-microkernel-tester.h",
4340 "src/xnnpack/AlignedAllocator.h",
4341 ] + MICROKERNEL_TEST_HDRS,
4342 deps = MICROKERNEL_TEST_DEPS,
4343)
4344
4345xnnpack_unit_test(
4346 name = "x32_pad_test",
4347 srcs = [
4348 "test/x32-pad.cc",
4349 "test/pad-microkernel-tester.h",
4350 ] + MICROKERNEL_TEST_HDRS,
4351 deps = MICROKERNEL_TEST_DEPS,
4352)
4353
4354xnnpack_unit_test(
4355 name = "x32_unpool_test",
4356 srcs = [
4357 "test/x32-unpool.cc",
4358 "test/unpool-microkernel-tester.h",
4359 ] + MICROKERNEL_TEST_HDRS,
4360 deps = MICROKERNEL_TEST_DEPS,
4361)
4362
4363xnnpack_unit_test(
4364 name = "x32_zip_test",
4365 srcs = [
4366 "test/x32-zip.cc",
4367 "test/zip-microkernel-tester.h",
4368 ] + MICROKERNEL_TEST_HDRS,
4369 deps = MICROKERNEL_TEST_DEPS,
4370)
4371
4372xnnpack_unit_test(
4373 name = "x8_lut_test",
4374 srcs = [
4375 "test/x8-lut.cc",
4376 "test/lut-microkernel-tester.h",
4377 ] + MICROKERNEL_TEST_HDRS,
4378 deps = MICROKERNEL_TEST_DEPS,
4379)
4380
4381xnnpack_unit_test(
4382 name = "x8_zip_test",
4383 srcs = [
4384 "test/x8-zip.cc",
4385 "test/zip-microkernel-tester.h",
4386 ] + MICROKERNEL_TEST_HDRS,
4387 deps = MICROKERNEL_TEST_DEPS,
4388)
4389
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004390xnnpack_unit_test(
4391 name = "requantization_test",
4392 srcs = [
4393 "src/xnnpack/requantization-stubs.h",
4394 "test/requantization.cc",
4395 "test/requantization-tester.h",
4396 ] + MICROKERNEL_TEST_HDRS,
4397 deps = MICROKERNEL_TEST_DEPS,
4398)
4399
Marat Dukhan20c3b922020-03-10 03:45:06 -07004400########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004401
4402xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07004403 name = "operator_size_test",
4404 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004405 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004406)
4407
Marat Dukhan20c3b922020-03-10 03:45:06 -07004408xnnpack_binary(
4409 name = "subgraph_size_test",
4410 srcs = ["test/subgraph-size.c"],
4411 deps = [":XNNPACK"],
4412)
4413
4414########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004415
4416xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004417 name = "abs_nc_test",
4418 srcs = [
4419 "test/abs-nc.cc",
4420 "test/abs-operator-tester.h",
4421 ],
4422 deps = OPERATOR_TEST_DEPS,
4423)
4424
4425xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004426 name = "add_nd_test",
4427 srcs = [
4428 "test/add-nd.cc",
4429 "test/binary-elementwise-operator-tester.h",
4430 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004431 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004432)
4433
4434xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004435 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004436 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004437 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004438 "test/argmax-pooling-operator-tester.h",
4439 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004440 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004441)
4442
4443xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004444 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004445 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004446 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004447 "test/average-pooling-operator-tester.h",
4448 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004449 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004450)
4451
4452xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004453 name = "bankers_rounding_nc_test",
4454 srcs = [
4455 "test/bankers-rounding-nc.cc",
4456 "test/bankers-rounding-operator-tester.h",
4457 ],
4458 deps = OPERATOR_TEST_DEPS,
4459)
4460
4461xnnpack_unit_test(
4462 name = "ceiling_nc_test",
4463 srcs = [
4464 "test/ceiling-nc.cc",
4465 "test/ceiling-operator-tester.h",
4466 ],
4467 deps = OPERATOR_TEST_DEPS,
4468)
4469
4470xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004471 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004472 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004473 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004474 "test/channel-shuffle-operator-tester.h",
4475 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004476 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004477)
4478
4479xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004480 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004481 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004482 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004483 "test/clamp-operator-tester.h",
4484 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004485 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486)
4487
4488xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07004489 name = "constant_pad_nd_test",
4490 srcs = [
4491 "test/constant-pad-nd.cc",
4492 "test/constant-pad-operator-tester.h",
4493 ],
4494 deps = OPERATOR_TEST_DEPS,
4495)
4496
4497xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004498 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004500 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004501 "test/convolution-operator-tester.h",
4502 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004503 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004504)
4505
4506xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004507 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004508 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004509 "test/convolution-nchw.cc",
4510 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004511 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004512 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004513)
4514
4515xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07004516 name = "copy_nc_test",
4517 srcs = [
4518 "test/copy-nc.cc",
4519 "test/copy-operator-tester.h",
4520 ],
4521 deps = OPERATOR_TEST_DEPS,
4522)
4523
4524xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004525 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004526 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004527 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004528 "test/deconvolution-operator-tester.h",
4529 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004530 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004531)
4532
4533xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08004534 name = "divide_nd_test",
4535 srcs = [
4536 "test/binary-elementwise-operator-tester.h",
4537 "test/divide-nd.cc",
4538 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004539 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08004540)
4541
4542xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004543 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004544 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004545 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004546 "test/fully-connected-operator-tester.h",
4547 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004548 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004549)
4550
4551xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004552 name = "floor_nc_test",
4553 srcs = [
4554 "test/floor-nc.cc",
4555 "test/floor-operator-tester.h",
4556 ],
4557 deps = OPERATOR_TEST_DEPS,
4558)
4559
4560xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004561 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004562 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004563 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004564 "test/global-average-pooling-operator-tester.h",
4565 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004566 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004567)
4568
4569xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004570 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004571 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004572 "test/global-average-pooling-ncw.cc",
4573 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004574 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004575 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004576)
4577
4578xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004579 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004580 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004581 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004582 "test/hardswish-operator-tester.h",
4583 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004584 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004585)
4586
4587xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004588 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004589 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004590 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004591 "test/leaky-relu-operator-tester.h",
4592 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004593 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004594)
4595
4596xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004597 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004598 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004599 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004600 "test/max-pooling-operator-tester.h",
4601 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004602 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004603)
4604
4605xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08004606 name = "maximum_nd_test",
4607 srcs = [
4608 "test/binary-elementwise-operator-tester.h",
4609 "test/maximum-nd.cc",
4610 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004611 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08004612)
4613
4614xnnpack_unit_test(
4615 name = "minimum_nd_test",
4616 srcs = [
4617 "test/binary-elementwise-operator-tester.h",
4618 "test/minimum-nd.cc",
4619 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004620 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08004621)
4622
4623xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004624 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08004625 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004626 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004627 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08004628 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004629 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08004630)
4631
4632xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004633 name = "negate_nc_test",
4634 srcs = [
4635 "test/negate-nc.cc",
4636 "test/negate-operator-tester.h",
4637 ],
4638 deps = OPERATOR_TEST_DEPS,
4639)
4640
4641xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004642 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004643 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004644 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004645 "test/prelu-operator-tester.h",
4646 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004647 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004648)
4649
4650xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004651 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08004652 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004653 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08004654 "test/resize-bilinear-operator-tester.h",
4655 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004656 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08004657)
4658
4659xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004660 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004661 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004662 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004663 "test/sigmoid-operator-tester.h",
4664 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004665 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004666)
4667
4668xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004669 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004670 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004671 "test/softmax-nc.cc",
4672 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004673 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004674 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675)
4676
4677xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004678 name = "square_nc_test",
4679 srcs = [
4680 "test/square-nc.cc",
4681 "test/square-operator-tester.h",
4682 ],
4683 deps = OPERATOR_TEST_DEPS,
4684)
4685
4686xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07004687 name = "squared_difference_nd_test",
4688 srcs = [
4689 "test/binary-elementwise-operator-tester.h",
4690 "test/squared-difference-nd.cc",
4691 ],
4692 deps = OPERATOR_TEST_DEPS,
4693)
4694
4695xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08004696 name = "subtract_nd_test",
4697 srcs = [
4698 "test/binary-elementwise-operator-tester.h",
4699 "test/subtract-nd.cc",
4700 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004701 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08004702)
4703
4704xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004705 name = "truncation_nc_test",
4706 srcs = [
4707 "test/truncation-nc.cc",
4708 "test/truncation-operator-tester.h",
4709 ],
4710 deps = OPERATOR_TEST_DEPS,
4711)
4712
4713xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004714 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004715 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004716 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004717 "test/unpooling-operator-tester.h",
4718 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004719 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004720)
4721
Chao Mei6ddfc602020-05-13 22:29:36 -07004722############################### Misc unit tests ###############################
4723
4724xnnpack_unit_test(
4725 name = "memory_planner_test",
4726 srcs = [
4727 "test/memory-planner-test.cc",
4728 ],
4729 deps = [
4730 ":XNNPACK",
4731 ":memory_planner",
4732 ],
4733)
4734
Marat Dukhan08c4a432019-10-03 09:29:21 -07004735############################# Build configurations #############################
4736
Marat Dukhanb8642352019-10-30 15:43:02 -07004737# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07004738config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07004739 name = "xnn_enable_assembly_explicit_true",
4740 define_values = {"xnn_enable_assembly": "true"},
4741)
4742
4743# Disables usage of assembly kernels.
4744config_setting(
4745 name = "xnn_enable_assembly_explicit_false",
4746 define_values = {"xnn_enable_assembly": "false"},
4747)
4748
Marat Dukhan05702cf2020-03-26 15:41:33 -07004749# Disables usage of HMP-aware optimizations.
4750config_setting(
4751 name = "xnn_enable_hmp_explicit_false",
4752 define_values = {"xnn_enable_hmp": "false"},
4753)
4754
Chao Mei6ddfc602020-05-13 22:29:36 -07004755# Enable usage of optimized memory allocation
4756config_setting(
4757 name = "xnn_enable_memopt_explicit_true",
4758 define_values = {"xnnpack_enable_memopt": "true"},
4759)
4760
4761# Disable usage of optimized memory allocation
4762config_setting(
4763 name = "xnn_enable_memopt_explicit_false",
4764 define_values = {"xnnpack_enable_memopt": "false"},
4765)
4766
Marat Dukhanb8642352019-10-30 15:43:02 -07004767# Builds with -c dbg
4768config_setting(
4769 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004770 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07004771 "compilation_mode": "dbg",
4772 },
4773)
4774
4775# Builds with -c opt
4776config_setting(
4777 name = "optimized_build",
4778 values = {
4779 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004780 },
4781)
4782
4783config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07004784 name = "linux_k8",
4785 values = {"cpu": "k8"},
4786)
4787
4788config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07004789 name = "linux_arm",
4790 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07004791)
4792
4793config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07004794 name = "linux_armhf",
4795 values = {"cpu": "armhf"},
4796)
4797
4798config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07004799 name = "linux_aarch64",
4800 values = {"cpu": "aarch64"},
4801)
4802
4803config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004804 name = "android",
4805 values = {"crosstool_top": "//external:android/crosstool"},
4806)
4807
4808config_setting(
4809 name = "android_armv7",
4810 values = {
4811 "crosstool_top": "//external:android/crosstool",
4812 "cpu": "armeabi-v7a",
4813 },
4814)
4815
4816config_setting(
4817 name = "android_arm64",
4818 values = {
4819 "crosstool_top": "//external:android/crosstool",
4820 "cpu": "arm64-v8a",
4821 },
4822)
4823
4824config_setting(
4825 name = "android_x86",
4826 values = {
4827 "crosstool_top": "//external:android/crosstool",
4828 "cpu": "x86",
4829 },
4830)
4831
4832config_setting(
4833 name = "android_x86_64",
4834 values = {
4835 "crosstool_top": "//external:android/crosstool",
4836 "cpu": "x86_64",
4837 },
4838)
4839
4840config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07004841 name = "windows_x86_64",
4842 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07004843)
4844
4845config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07004846 name = "windows_x86_64_clang",
4847 values = {
4848 "compiler": "clang-cl",
4849 "cpu": "x64_windows",
4850 },
4851)
4852
4853config_setting(
4854 name = "windows_x86_64_mingw",
4855 values = {
4856 "compiler": "mingw-gcc",
4857 "cpu": "x64_windows",
4858 },
4859)
4860
4861config_setting(
4862 name = "windows_x86_64_msys",
4863 values = {
4864 "compiler": "msys-gcc",
4865 "cpu": "x64_windows",
4866 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07004867)
4868
4869config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07004870 name = "macos_x86_64",
4871 values = {
4872 "apple_platform_type": "macos",
4873 "cpu": "darwin",
4874 },
4875)
4876
4877config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004878 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004879 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07004880)
4881
4882config_setting(
4883 name = "emscripten_wasm",
4884 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004885 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004886 "cpu": "wasm",
4887 },
4888)
4889
4890config_setting(
4891 name = "emscripten_wasmsimd",
4892 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004893 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004894 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07004895 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004896 },
4897)
4898
4899config_setting(
4900 name = "emscripten_asmjs",
4901 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004902 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004903 "cpu": "asmjs",
4904 },
4905)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004906
4907config_setting(
4908 name = "ios_armv7",
4909 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004910 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004911 "cpu": "ios_armv7",
4912 },
4913)
4914
4915config_setting(
4916 name = "ios_arm64",
4917 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004918 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004919 "cpu": "ios_arm64",
4920 },
4921)
4922
4923config_setting(
4924 name = "ios_arm64e",
4925 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004926 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004927 "cpu": "ios_arm64e",
4928 },
4929)
4930
4931config_setting(
4932 name = "ios_x86",
4933 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004934 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004935 "cpu": "ios_i386",
4936 },
4937)
4938
4939config_setting(
4940 name = "ios_x86_64",
4941 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004942 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004943 "cpu": "ios_x86_64",
4944 },
4945)
4946
4947config_setting(
4948 name = "watchos_armv7k",
4949 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004950 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004951 "cpu": "watchos_armv7k",
4952 },
4953)
4954
4955config_setting(
4956 name = "watchos_arm64_32",
4957 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004958 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004959 "cpu": "watchos_arm64_32",
4960 },
4961)
4962
4963config_setting(
4964 name = "watchos_x86",
4965 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004966 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004967 "cpu": "watchos_i386",
4968 },
4969)
4970
4971config_setting(
4972 name = "watchos_x86_64",
4973 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004974 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004975 "cpu": "watchos_x86_64",
4976 },
4977)
4978
4979config_setting(
4980 name = "tvos_arm64",
4981 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004982 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004983 "cpu": "tvos_arm64",
4984 },
4985)
4986
4987config_setting(
4988 name = "tvos_x86_64",
4989 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004990 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004991 "cpu": "tvos_x86_64",
4992 },
4993)