blob: 9e22c0aa27556b48ef573eb147af54d6f9bff465 [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",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/square.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700102 "src/subgraph/squared-difference.c",
103 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700104 "src/subgraph/unpooling-2d.c",
105]
106
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800107TABLE_SRCS = [
108 "src/tables/exp2-k-over-64.c",
109 "src/tables/exp2-k-over-2048.c",
110]
111
Marat Dukhan08c4a432019-10-03 09:29:21 -0700112SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800113 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800114 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800115 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700116 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
117 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700118 "src/f32-clamp/gen/scalar-x1.c",
119 "src/f32-clamp/gen/scalar-x2.c",
120 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800121 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800122 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700123 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
124 "src/f32-dwconv-chw/3x3p1-scalar.c",
125 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
126 "src/f32-dwconv-chw/5x5p2-scalar.c",
127 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700128 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
129 "src/f32-dwconv/gen/up1x4-scalar.c",
130 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
131 "src/f32-dwconv/gen/up1x9-scalar.c",
132 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
133 "src/f32-dwconv/gen/up1x25-scalar.c",
134 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
135 "src/f32-dwconv/gen/up2x4-scalar.c",
136 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
137 "src/f32-dwconv/gen/up2x9-scalar.c",
138 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
139 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700140 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
141 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
142 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700144 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700146 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
147 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
148 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700150 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700152 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700153 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
154 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700155 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
156 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
157 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800170 "src/f32-hswish/gen/scalar-x1.c",
171 "src/f32-hswish/gen/scalar-x2.c",
172 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700173 "src/f32-ibilinear/gen/scalar-c1.c",
174 "src/f32-ibilinear/gen/scalar-c2.c",
175 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700176 "src/f32-igemm/gen/1x4-scalar.c",
177 "src/f32-igemm/gen/2x4-scalar.c",
178 "src/f32-igemm/gen/4x2-scalar.c",
179 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700180 "src/f32-igemm/gen/1x4-relu-scalar.c",
181 "src/f32-igemm/gen/2x4-relu-scalar.c",
182 "src/f32-igemm/gen/4x2-relu-scalar.c",
183 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700184 "src/f32-igemm/gen/1x4-minmax-scalar.c",
185 "src/f32-igemm/gen/2x4-minmax-scalar.c",
186 "src/f32-igemm/gen/4x2-minmax-scalar.c",
187 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700188 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
189 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
190 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700191 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
192 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
193 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
194 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800195 "src/f32-prelu/gen/scalar-2x1.c",
196 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800197 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
198 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
199 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700209 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800210 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
211 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
212 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
213 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
214 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
215 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
216 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700219 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
220 "src/f32-spmm/gen/1x1-minmax-scalar.c",
221 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
222 "src/f32-spmm/gen/2x1-minmax-scalar.c",
223 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
224 "src/f32-spmm/gen/4x1-minmax-scalar.c",
225 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/8x1-minmax-scalar.c",
227 "src/f32-spmm/gen/8x2-minmax-scalar.c",
228 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700229 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700232 "src/f32-vbinary/gen/vadd-scalar-x1.c",
233 "src/f32-vbinary/gen/vadd-scalar-x2.c",
234 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
239 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
240 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
245 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
246 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
251 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
252 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800253 "src/f32-vbinary/gen/vmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vmax-scalar-x4.c",
256 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
257 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
258 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
259 "src/f32-vbinary/gen/vmin-scalar-x1.c",
260 "src/f32-vbinary/gen/vmin-scalar-x2.c",
261 "src/f32-vbinary/gen/vmin-scalar-x4.c",
262 "src/f32-vbinary/gen/vminc-scalar-x1.c",
263 "src/f32-vbinary/gen/vminc-scalar-x2.c",
264 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700268 "src/f32-vbinary/gen/vmul-scalar-x1.c",
269 "src/f32-vbinary/gen/vmul-scalar-x2.c",
270 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
275 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
276 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
281 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
282 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700289 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
290 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
291 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
292 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
293 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
294 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700298 "src/f32-vbinary/gen/vsub-scalar-x1.c",
299 "src/f32-vbinary/gen/vsub-scalar-x2.c",
300 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
305 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
306 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700307 "src/f32-vlrelu/gen/scalar-x1.c",
308 "src/f32-vlrelu/gen/scalar-x2.c",
309 "src/f32-vlrelu/gen/scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700310 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
311 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
312 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700313 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
314 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
315 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
316 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
317 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
318 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
319 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700325 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
326 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
327 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700328 "src/f32-vunary/gen/vabs-scalar-x1.c",
329 "src/f32-vunary/gen/vabs-scalar-x2.c",
330 "src/f32-vunary/gen/vabs-scalar-x4.c",
331 "src/f32-vunary/gen/vneg-scalar-x1.c",
332 "src/f32-vunary/gen/vneg-scalar-x2.c",
333 "src/f32-vunary/gen/vneg-scalar-x4.c",
334 "src/f32-vunary/gen/vsqr-scalar-x1.c",
335 "src/f32-vunary/gen/vsqr-scalar-x2.c",
336 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800337 "src/math/expminus-scalar-lut2048-p1.c",
338 "src/math/expminus-scalar-lut64-p2.c",
339 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700340 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700341 "src/math/roundne-scalar-nearbyint.c",
342 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700343 "src/math/roundd-scalar-addsub.c",
344 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700345 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700346 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700348 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700349 "src/math/roundz-scalar-addsub.c",
350 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800352 "src/math/sigmoid-scalar-lut2048-p1-div.c",
353 "src/math/sigmoid-scalar-lut64-p2-div.c",
354 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700355 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
356 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700357 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700358 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
359 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700360 "src/q8-gemm/2x2-minmax-scalar.c",
361 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700362 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700363 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700364 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700365 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700366 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700367 "src/x32-fill/scalar-float.c",
368 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700369 "src/x32-packx/x2-scalar.c",
370 "src/x32-packx/x3-scalar.c",
371 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700372 "src/x32-pad/scalar-int.c",
373 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700374 "src/x32-unpool/scalar.c",
375 "src/x32-zip/x2-scalar.c",
376 "src/x32-zip/x3-scalar.c",
377 "src/x32-zip/x4-scalar.c",
378 "src/x32-zip/xm-scalar.c",
379 "src/x8-lut/scalar.c",
380 "src/x8-zip/x2-scalar.c",
381 "src/x8-zip/x3-scalar.c",
382 "src/x8-zip/x4-scalar.c",
383 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700384 "src/requantization/precise-scalar.c",
385 "src/requantization/fp32-scalar.c",
386 "src/requantization/q31-scalar.c",
387 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700388]
389
Marat Dukhan436ebe62019-12-04 15:10:12 -0800390WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700391 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
392 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700393 "src/f32-clamp/gen/wasm-x1.c",
394 "src/f32-clamp/gen/wasm-x2.c",
395 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700396 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
397 "src/f32-dwconv/gen/up1x4-wasm.c",
398 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
399 "src/f32-dwconv/gen/up1x9-wasm.c",
400 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
401 "src/f32-dwconv/gen/up1x25-wasm.c",
402 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
403 "src/f32-dwconv/gen/up2x4-wasm.c",
404 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
405 "src/f32-dwconv/gen/up2x9-wasm.c",
406 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
407 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700408 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
409 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
410 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700412 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700414 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
415 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
416 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700418 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700420 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
421 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
423 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
424 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700425 "src/f32-gemm/gen/1x4-wasm.c",
426 "src/f32-gemm/gen/2x4-wasm.c",
427 "src/f32-gemm/gen/4x2-wasm.c",
428 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700429 "src/f32-gemm/gen/1x4-relu-wasm.c",
430 "src/f32-gemm/gen/2x4-relu-wasm.c",
431 "src/f32-gemm/gen/4x2-relu-wasm.c",
432 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700433 "src/f32-gemm/gen/1x4-minmax-wasm.c",
434 "src/f32-gemm/gen/2x4-minmax-wasm.c",
435 "src/f32-gemm/gen/4x2-minmax-wasm.c",
436 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800437 "src/f32-hswish/gen/wasm-x1.c",
438 "src/f32-hswish/gen/wasm-x2.c",
439 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700440 "src/f32-igemm/gen/1x4-wasm.c",
441 "src/f32-igemm/gen/2x4-wasm.c",
442 "src/f32-igemm/gen/4x2-wasm.c",
443 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700444 "src/f32-igemm/gen/1x4-relu-wasm.c",
445 "src/f32-igemm/gen/2x4-relu-wasm.c",
446 "src/f32-igemm/gen/4x2-relu-wasm.c",
447 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700448 "src/f32-igemm/gen/1x4-minmax-wasm.c",
449 "src/f32-igemm/gen/2x4-minmax-wasm.c",
450 "src/f32-igemm/gen/4x2-minmax-wasm.c",
451 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700452 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
453 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
454 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700455 "src/f32-prelu/gen/wasm-2x1.c",
456 "src/f32-prelu/gen/wasm-2x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700457 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
458 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
459 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
460 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
461 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
462 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
463 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
464 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
465 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
466 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
467 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
468 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800469 "src/f32-vbinary/gen/vmax-wasm-x1.c",
470 "src/f32-vbinary/gen/vmax-wasm-x2.c",
471 "src/f32-vbinary/gen/vmax-wasm-x4.c",
472 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
473 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
474 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
475 "src/f32-vbinary/gen/vmin-wasm-x1.c",
476 "src/f32-vbinary/gen/vmin-wasm-x2.c",
477 "src/f32-vbinary/gen/vmin-wasm-x4.c",
478 "src/f32-vbinary/gen/vminc-wasm-x1.c",
479 "src/f32-vbinary/gen/vminc-wasm-x2.c",
480 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700481 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
482 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
483 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
484 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
485 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
486 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
487 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
488 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
489 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
490 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
491 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
492 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
493 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
494 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
495 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
496 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
497 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
498 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700499 "src/f32-vlrelu/gen/wasm-x1.c",
500 "src/f32-vlrelu/gen/wasm-x2.c",
501 "src/f32-vlrelu/gen/wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700502 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
503 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
504 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800505]
506
Marat Dukhan290055c2020-06-09 12:24:29 -0700507WASMSIMD_UKERNELS = [
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700508 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
509 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
510 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
511 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700512 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
513 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
514 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
515 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
516 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
517 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700518 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
519 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
520 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
521 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
522 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
523 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
524 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
525 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
526 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
527 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
528 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
529 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
530 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
531 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
532 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
533 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
534 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
535 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
536 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
537 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
538 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
539 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
540 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
541 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700542 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
543 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
544 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
545 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-arm.c",
546 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat-arm.c",
547 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-x86.c",
548 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700549 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
550 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
551 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
552 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
553 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
554 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
555 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
556 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
557 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
558 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
559 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
560 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
561 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
562 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
563 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
564 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
565 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
566 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
567 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
568 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
569 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
570 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
571 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
572 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
573 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
574 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
575 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
576 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
577 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
578 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
579 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
580 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
581 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
582 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
583 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
584 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
585 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
586 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
587 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
588 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
589 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
590 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
591 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
592 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
593 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
594 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
595 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
596 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
597 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
598 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
599 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
600 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
601 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
602 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
603 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
604 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
605 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
606 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
607 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
608 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700609 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
610 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
611 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
612 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-arm.c",
613 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat-arm.c",
614 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-x86.c",
615 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700616 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
617 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
618 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
619 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
620 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
621 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
622 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
623 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
624 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
625 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
626 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
627 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
628 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
629 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
630 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
631 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
632 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
633 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
634 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
635 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
636 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
637 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
638 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
639 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
640 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
641 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
642 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
643 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
644 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
645 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700646 "src/f32-hswish/gen/wasmsimd-arm-x4.c",
647 "src/f32-hswish/gen/wasmsimd-arm-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700648 "src/f32-hswish/gen/wasmsimd-arm-x16.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700649 "src/f32-hswish/gen/wasmsimd-x86-x4.c",
650 "src/f32-hswish/gen/wasmsimd-x86-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700651 "src/f32-hswish/gen/wasmsimd-x86-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700652 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
653 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700654 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
655 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
656 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
657 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700658 "src/f32-rmax/wasmsimd-arm.c",
659 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700660 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
661 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
662 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
663 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700664 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
665 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700666 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
667 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
668 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
669 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700670 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
671 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700672 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
673 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
674 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
675 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700676 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
677 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700678 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
679 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
680 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
681 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700682 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
683 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700684 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
685 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
686 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
687 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700688 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
689 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
690 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
691 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700692 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
693 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
694 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
695 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700696 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
697 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
698 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
699 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700700 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
701 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
702 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
703 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700704 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
705 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700706 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
707 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
708 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
709 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700710 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
711 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700712 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
713 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
714 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
715 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700716 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
717 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700718 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
719 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
720 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
721 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700722 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
723 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700724 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
725 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700726 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
727 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700728 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
729 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
730 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
731 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700732 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
733 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700734 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
735 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
736 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
737 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700738 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
739 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan72b399a2020-06-28 22:41:38 -0700740 "src/f32-vrnd/gen/vrndne-wasmsimd-x4.c",
741 "src/f32-vrnd/gen/vrndne-wasmsimd-x8.c",
742 "src/f32-vrnd/gen/vrndz-wasmsimd-x4.c",
743 "src/f32-vrnd/gen/vrndz-wasmsimd-x8.c",
744 "src/f32-vrnd/gen/vrndu-wasmsimd-x4.c",
745 "src/f32-vrnd/gen/vrndu-wasmsimd-x8.c",
746 "src/f32-vrnd/gen/vrndd-wasmsimd-x4.c",
747 "src/f32-vrnd/gen/vrndd-wasmsimd-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700748 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
749 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700750 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
751 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
752 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
753 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
754 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
755 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700756 "src/math/roundne-wasmsimd-addsub.c",
757 "src/math/roundd-wasmsimd-addsub.c",
758 "src/math/roundu-wasmsimd-addsub.c",
759 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700760 "src/requantization/fp32-wasmsimd.c",
761]
762
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800763PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800764 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800765 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800766 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700767 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
768 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700769 "src/f32-clamp/gen/psimd-x4.c",
770 "src/f32-clamp/gen/psimd-x8.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700771 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700772 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700773 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700774 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700775 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700776 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
777 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
778 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
779 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
780 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
781 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
782 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
783 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
784 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
785 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
786 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
787 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700788 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700789 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
790 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700791 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
792 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
793 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
794 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
795 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
796 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
797 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
798 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
799 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
800 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
801 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
802 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
803 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
804 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
805 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
806 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
807 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
808 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
809 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800810 "src/f32-hswish/gen/psimd-x4.c",
811 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700812 "src/f32-ibilinear/gen/psimd-c4.c",
813 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700814 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
815 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
816 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
817 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
818 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
819 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
820 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
821 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
822 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
823 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700824 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700825 "src/f32-spmm/gen/4x1-minmax-psimd.c",
826 "src/f32-spmm/gen/8x1-minmax-psimd.c",
827 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700828 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
829 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700830 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800831 "src/f32-prelu/gen/psimd-2x4.c",
832 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800833 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700834 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
835 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
836 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
837 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
838 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
839 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
840 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
841 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800842 "src/f32-vbinary/gen/vmax-psimd-x4.c",
843 "src/f32-vbinary/gen/vmax-psimd-x8.c",
844 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
845 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
846 "src/f32-vbinary/gen/vmin-psimd-x4.c",
847 "src/f32-vbinary/gen/vmin-psimd-x8.c",
848 "src/f32-vbinary/gen/vminc-psimd-x4.c",
849 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700850 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
851 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
852 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
853 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
854 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
855 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
856 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
857 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700858 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
859 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
860 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
861 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700862 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
863 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
864 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
865 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700866 "src/f32-vlrelu/gen/psimd-x4.c",
867 "src/f32-vlrelu/gen/psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700868 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
869 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700870 "src/f32-vunary/gen/vabs-psimd-x4.c",
871 "src/f32-vunary/gen/vabs-psimd-x8.c",
872 "src/f32-vunary/gen/vneg-psimd-x4.c",
873 "src/f32-vunary/gen/vneg-psimd-x8.c",
874 "src/f32-vunary/gen/vsqr-psimd-x4.c",
875 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700876 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700877 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700878 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700879 "src/x32-unpool/psimd.c",
880 "src/x32-zip/x2-psimd.c",
881 "src/x32-zip/x3-psimd.c",
882 "src/x32-zip/x4-psimd.c",
883 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700884 "src/requantization/precise-psimd.c",
885 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700886]
887
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800888PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800889 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
890 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
891 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
892 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
893 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
894 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
895 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
896 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
897 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
898 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
899 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
900 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800901 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
902 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
903 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
904 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
905 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
906 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700907 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
908 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
909 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
910 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
911 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
912 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
913 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
914 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700915 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700916 "src/math/roundd-psimd-addsub.c",
917 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700918 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800919 "src/math/sigmoid-psimd-p5-div.c",
920]
921
Marat Dukhan08c4a432019-10-03 09:29:21 -0700922# ISA-specific micro-kernels
923NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700924 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
925 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700926 "src/f32-clamp/gen/neon-x4.c",
927 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700928 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
929 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700930 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
931 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700932 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
933 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700934 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
935 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700936 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
937 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
938 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
939 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700940 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
941 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
942 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
943 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700944 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
945 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
946 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
947 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700948 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700949 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
950 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700951 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
952 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
953 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
954 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
955 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
956 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
957 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
958 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
959 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
960 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
961 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
962 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
963 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
964 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
965 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
966 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
967 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
968 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
969 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
970 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
971 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
972 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
973 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
974 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
975 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
976 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
977 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
978 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
979 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
980 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
981 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800982 "src/f32-hswish/gen/neon-x4.c",
983 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700984 "src/f32-ibilinear/gen/neon-c4.c",
985 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700986 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
987 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
988 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
989 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
990 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
991 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
992 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
993 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
994 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
995 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
996 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
997 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
998 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
999 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1000 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1001 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001002 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1003 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1004 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001005 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1006 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001007 "src/f32-prelu/gen/neon-2x4.c",
1008 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001009 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1013 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1014 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1015 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1016 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1017 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1018 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1019 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1020 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1021 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1022 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1023 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1024 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1025 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1026 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1027 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1028 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1029 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1030 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1031 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1032 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001033 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001034 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001035 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1036 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1037 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1038 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1039 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1040 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1041 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1042 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1043 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1044 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1045 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1046 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1047 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1048 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1049 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1050 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1051 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1052 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001053 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1054 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1055 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1056 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001057 "src/f32-vbinary/gen/vmax-neon-x4.c",
1058 "src/f32-vbinary/gen/vmax-neon-x8.c",
1059 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1060 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1061 "src/f32-vbinary/gen/vmin-neon-x4.c",
1062 "src/f32-vbinary/gen/vmin-neon-x8.c",
1063 "src/f32-vbinary/gen/vminc-neon-x4.c",
1064 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001065 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1066 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1067 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1068 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1069 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1070 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001071 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1072 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1073 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1074 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001075 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1076 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1077 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1078 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001079 "src/f32-vlrelu/gen/neon-x4.c",
1080 "src/f32-vlrelu/gen/neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001081 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1082 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001083 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1084 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1085 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1086 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1087 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1088 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1089 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1090 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001091 "src/f32-vunary/gen/vabs-neon-x4.c",
1092 "src/f32-vunary/gen/vabs-neon-x8.c",
1093 "src/f32-vunary/gen/vneg-neon-x4.c",
1094 "src/f32-vunary/gen/vneg-neon-x8.c",
1095 "src/f32-vunary/gen/vsqr-neon-x4.c",
1096 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001097 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
1098 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001099 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001100 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
1101 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001102 "src/q8-gemm/4x8-minmax-neon.c",
1103 "src/q8-gemm/8x8-minmax-neon.c",
1104 "src/q8-igemm/4x8-minmax-neon.c",
1105 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001106 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001107 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001108 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001109 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001110 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001111 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001112 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001113 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001114 "src/x32-zip/x2-neon.c",
1115 "src/x32-zip/x3-neon.c",
1116 "src/x32-zip/x4-neon.c",
1117 "src/x32-zip/xm-neon.c",
1118 "src/x8-zip/x2-neon.c",
1119 "src/x8-zip/x3-neon.c",
1120 "src/x8-zip/x4-neon.c",
1121 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001122 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001123 "src/math/roundd-neon-addsub.c",
1124 "src/math/roundd-neon-cvt.c",
1125 "src/math/roundu-neon-addsub.c",
1126 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001127 "src/math/roundz-neon-addsub.c",
1128 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001129 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001130 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1131 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1132 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1133 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1134 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1135 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001136 "src/math/sqrt-neon-nr1rsqrts.c",
1137 "src/math/sqrt-neon-nr2rsqrts.c",
1138 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001139 "src/requantization/precise-neon.c",
1140 "src/requantization/fp32-neon.c",
1141 "src/requantization/q31-neon.c",
1142 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143]
1144
1145NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001146 "src/f32-ibilinear/gen/neonfma-c4.c",
1147 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001148 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1149 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1150 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1151 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1152 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1153 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1154 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1155 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1156 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001157 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1158 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1159 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1160 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001161 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1162 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1163 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1164 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001165 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1166 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1167 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1168 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001169 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1170 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1171 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1172 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1173 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1174 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1175 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1176 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1177 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1178 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1179 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1180 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1181 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1182 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1183 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1184 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1185 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1186 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001187 "src/f32-hswish/gen/neonfma-x4.c",
1188 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001189 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1190 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001191 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1196 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1197 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1198 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1199 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1200 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1201 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1202 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1203 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1204 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1205 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1206 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1207 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1208 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1209 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1210 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1211 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1212 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1213 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1214 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001215 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1260 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1261 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1262 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1263 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1264 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1265 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001269 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1270 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001271 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1272 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1273 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1274 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1275 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1276 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1277 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1278 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1279 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1280 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1281 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1282 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1283 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1284 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1285 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1286 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1287 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1288 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1289 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1290 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001291 "src/math/exp-neonfma-lut64-p2.c",
1292 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001293 "src/math/expminus-neonfma-lut2048-p1.c",
1294 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001295 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001296 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1297 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1298 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1299 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1300 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1301 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1302 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1303 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1304 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1305 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1306 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1307 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1308 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1309 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1310 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1311 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1312 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1313 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001314 "src/math/sqrt-neonfma-nr1fma.c",
1315 "src/math/sqrt-neonfma-nr2fma.c",
1316 "src/math/sqrt-neonfma-nr3fma.c",
1317 "src/math/sqrt-neonfma-nr2fma1adj.c",
1318 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001319]
1320
1321AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001322 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1323 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1324 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1325 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1326 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1327 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001328 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1329 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1330 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1331 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1332 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1333 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1334 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1335 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1336 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1337 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1338 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1339 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1340 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1341 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1342 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1343 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1344 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1345 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1346 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1347 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001348 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1349 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001350 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1351 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001352 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1353 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001354 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1355 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001356 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1357 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1358 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1359 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1360 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001361 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1362 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1363 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1364 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1365 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1366 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1367 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1368 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1369 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1370 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001379 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1380 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1381 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1382 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1383 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1384 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1385 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1386 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1387 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1388 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1389 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1390 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1391 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1392 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1393 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1394 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1395 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1396 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001397 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1398 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001399 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1400 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1401 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1402 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1403 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1404 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001405]
1406
Marat Dukhan8853b822020-05-07 12:19:01 -07001407NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001408 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1409 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1410 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1411 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1412 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1413 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1414 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1415 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001416 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001417 "src/math/roundd-neonv8.c",
1418 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001419 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001420]
1421
Marat Dukhan08c4a432019-10-03 09:29:21 -07001422AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001423 "src/f16-clamp/gen/neonfp16arith-x16.c",
1424 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001425 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1426 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1427 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1428 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1429 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1430 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1431 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1432 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1433 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1434 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1435 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1436 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001437 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1438 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001439 "src/f16-hswish/gen/neonfp16arith-x16.c",
1440 "src/f16-hswish/gen/neonfp16arith-x8.c",
1441 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1442 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001443 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1444 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1445 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1446 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1447 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1448 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1449 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1450 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1451 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1452 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1453 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1454 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1455 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1456 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1457 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1458 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1459 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1460 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1461 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1462 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1463 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1464 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1465 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1466 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1467 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1468 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1469 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1470 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001471 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1472 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001473 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1474 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1475 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1476 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1477 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1478 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1479 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1480 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001481 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1482 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1483 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1484 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001485 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1486 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1487 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1488 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1489 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1490 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1491 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1492 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1493 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1494 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1495 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1496 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001497 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1498 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1499 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1500 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1501 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1502 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1503 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1504 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001505]
1506
1507SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001508 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1509 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001510 "src/f32-clamp/gen/sse-x4.c",
1511 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001512 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1513 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001514 "src/f32-dwconv-chw/3x3p1-sse.c",
1515 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001516 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1517 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1518 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1519 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1520 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1521 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1522 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1523 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1524 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1525 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1526 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1527 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001528 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001529 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1530 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1532 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1533 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1534 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1535 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1536 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1537 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1538 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1539 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1540 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1541 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1542 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1543 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001544 "src/f32-hswish/gen/sse-x4.c",
1545 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001546 "src/f32-ibilinear/gen/sse-c4.c",
1547 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001548 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1549 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1550 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1551 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1552 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1553 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1554 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001555 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1556 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1557 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001558 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001559 "src/f32-prelu/gen/sse-2x4.c",
1560 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001561 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001562 "src/f32-spmm/gen/4x1-minmax-sse.c",
1563 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001564 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001565 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1566 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1567 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1568 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1569 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1570 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1571 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1572 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001573 "src/f32-vbinary/gen/vmax-sse-x4.c",
1574 "src/f32-vbinary/gen/vmax-sse-x8.c",
1575 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1576 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1577 "src/f32-vbinary/gen/vmin-sse-x4.c",
1578 "src/f32-vbinary/gen/vmin-sse-x8.c",
1579 "src/f32-vbinary/gen/vminc-sse-x4.c",
1580 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001581 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1582 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1583 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1584 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1585 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1586 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1587 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1588 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001589 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1590 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1591 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1592 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001593 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1594 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1595 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1596 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001597 "src/f32-vlrelu/gen/sse-x4.c",
1598 "src/f32-vlrelu/gen/sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001599 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1600 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001601 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1602 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001603 "src/f32-vunary/gen/vabs-sse-x4.c",
1604 "src/f32-vunary/gen/vabs-sse-x8.c",
1605 "src/f32-vunary/gen/vneg-sse-x4.c",
1606 "src/f32-vunary/gen/vneg-sse-x8.c",
1607 "src/f32-vunary/gen/vsqr-sse-x4.c",
1608 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001609 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001610 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001611 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001612 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001613 "src/math/roundd-sse-addsub.c",
1614 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001615 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001616 "src/math/sqrt-sse-hh1mac.c",
1617 "src/math/sqrt-sse-nr1mac.c",
1618 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001619]
1620
1621SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001622 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1623 "src/f32-argmaxpool/4x-sse2-c4.c",
1624 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001625 "src/f32-prelu/gen/sse2-2x4.c",
1626 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001627 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1628 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1629 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1630 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1631 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1632 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1633 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1634 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1635 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1636 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1637 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1638 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001639 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001640 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001641 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001642 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001643 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1644 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001645 "src/f32-vlrelu/gen/sse2-x4.c",
1646 "src/f32-vlrelu/gen/sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001647 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1648 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1649 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1650 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1651 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1652 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1653 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1654 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001655 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1656 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001657 "src/q8-igemm/4x4c2-minmax-sse2.c",
1658 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001659 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1660 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001661 "src/q8-gemm/2x4c8-minmax-sse2.c",
1662 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001663 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001664 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001665 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001666 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001667 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001668 "src/x32-zip/x2-sse2.c",
1669 "src/x32-zip/x3-sse2.c",
1670 "src/x32-zip/x4-sse2.c",
1671 "src/x32-zip/xm-sse2.c",
1672 "src/x8-zip/x2-sse2.c",
1673 "src/x8-zip/x3-sse2.c",
1674 "src/x8-zip/x4-sse2.c",
1675 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001676 "src/math/exp-sse2-p5.c",
1677 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001678 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001679 "src/math/roundd-sse2-cvt.c",
1680 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001681 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001682 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001683 "src/requantization/precise-sse2.c",
1684 "src/requantization/fp32-sse2.c",
1685 "src/requantization/q31-sse2.c",
1686 "src/requantization/gemmlowp-sse2.c",
1687]
1688
1689SSSE3_UKERNELS = [
1690 "src/requantization/precise-ssse3.c",
1691 "src/requantization/q31-ssse3.c",
1692 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693]
1694
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001695SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001696 "src/f32-prelu/gen/sse41-2x4.c",
1697 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001698 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001699 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001700 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001701 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001702 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1703 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001704 "src/f32-vlrelu/gen/sse41-x4.c",
1705 "src/f32-vlrelu/gen/sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001706 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1707 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1708 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1709 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1710 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1711 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1712 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1713 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001714 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001715 "src/math/roundd-sse41.c",
1716 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001717 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001718 "src/requantization/precise-sse4.c",
1719 "src/requantization/q31-sse4.c",
1720 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001721]
1722
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001724 "src/f32-clamp/gen/avx-x8.c",
1725 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001726 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1727 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1728 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1729 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1730 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1731 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1732 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1733 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1734 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1735 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1736 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1737 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1738 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1739 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1740 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1741 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1742 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1743 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1744 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1745 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1746 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1747 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1748 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1749 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1750 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1751 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1752 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1753 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1754 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1755 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001756 "src/f32-hswish/gen/avx-x8.c",
1757 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001758 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1759 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1760 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1761 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1762 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1763 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1764 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1765 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1766 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001767 "src/f32-prelu/gen/avx-2x8.c",
1768 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001770 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1771 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1772 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1773 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1774 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1775 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1776 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1777 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001778 "src/f32-vbinary/gen/vmax-avx-x8.c",
1779 "src/f32-vbinary/gen/vmax-avx-x16.c",
1780 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1781 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1782 "src/f32-vbinary/gen/vmin-avx-x8.c",
1783 "src/f32-vbinary/gen/vmin-avx-x16.c",
1784 "src/f32-vbinary/gen/vminc-avx-x8.c",
1785 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001786 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1787 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1788 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1789 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1790 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1791 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1792 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1793 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001794 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1795 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1796 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1797 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001798 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1799 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1800 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1801 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001802 "src/f32-vlrelu/gen/avx-x8.c",
1803 "src/f32-vlrelu/gen/avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001804 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001805 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1806 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1807 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1808 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1809 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1810 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1811 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1812 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001813 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1814 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001815 "src/f32-vunary/gen/vabs-avx-x8.c",
1816 "src/f32-vunary/gen/vabs-avx-x16.c",
1817 "src/f32-vunary/gen/vneg-avx-x8.c",
1818 "src/f32-vunary/gen/vneg-avx-x16.c",
1819 "src/f32-vunary/gen/vsqr-avx-x8.c",
1820 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001821]
1822
Marat Dukhanfda12b82019-11-21 12:27:59 -08001823FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001824 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1825 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1826 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1827 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1828 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1829 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1830 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1831 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1832 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1833 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1834 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1835 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1836 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1837 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1838 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1839 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1840 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1841 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1842 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1843 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1844 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1845 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1846 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1847 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1848 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1849 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1850 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1851 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1852 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1853 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1854 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1855 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1856 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1857 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1858 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1859 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1860 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1861 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1862 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1863 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001864 "src/f32-hswish/gen/fma3-x8.c",
1865 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001866 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1867 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1868 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1869 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1870 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1871 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1872 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1873 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1874 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1875 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1876 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1877 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1878 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1879 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001880 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1881 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1882 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1883 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1884 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1885 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1886 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1887 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001888 "src/math/sqrt-fma3-nr1fma.c",
1889 "src/math/sqrt-fma3-nr2fma.c",
1890 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001891]
1892
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001893AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001894 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1895 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1896 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1897 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1898 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1899 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1900 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1901 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1902 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1903 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1904 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1905 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1906 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1907 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1908 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1909 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1910 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1911 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1912 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1913 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1914 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1915 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1916 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1917 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1918 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1919 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1920 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1921 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1922 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1923 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1924 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1925 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1926 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1927 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1928 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1929 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001930 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1931 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1932 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1933 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1934 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1935 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1936 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1937 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1938 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1939 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1940 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1941 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1942 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1943 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1944 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1945 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1946 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1947 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1948 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1949 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1950 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1951 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1952 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1953 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1954 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1955 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1956 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1957 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1958 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1959 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001960 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1961 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1962 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1963 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1964 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1965 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1966 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1967 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1968 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1969 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1970 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1971 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1972 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1973 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1974 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1975 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1976 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1977 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1978 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1979 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1980 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1981 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1982 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1983 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001984 "src/math/exp-avx2-p5.c",
1985 "src/math/exp-avx2-perm-p3.c",
1986 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001987 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001988 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001989 "src/math/sigmoid-avx2-rr2-p5-div.c",
1990 "src/math/sigmoid-avx2-rr1-p5-div.c",
1991 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1992 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1993 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1994 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001995]
1996
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001998 "src/f32-clamp/gen/avx512f-x16.c",
1999 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002000 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2001 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2002 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2003 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2004 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2005 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2006 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2007 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2008 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2009 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2010 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2011 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2012 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2013 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2014 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2015 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2016 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2017 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2018 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2019 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2020 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2021 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2022 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2023 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08002024 "src/f32-hswish/gen/avx512f-x16.c",
2025 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002026 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2027 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2028 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2029 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2030 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2031 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002032 "src/f32-prelu/gen/avx512f-2x16.c",
2033 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002034 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2035 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2036 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2037 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2038 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2039 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2040 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2041 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2042 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2043 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2044 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2045 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2046 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2047 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2048 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2049 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2050 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2051 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2052 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2053 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2054 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2055 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2056 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2057 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2058 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2059 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2060 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2061 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2062 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2063 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2064 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2065 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2066 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2067 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2068 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2069 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002071 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2072 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2073 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2074 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2075 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2076 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2077 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2078 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002079 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2080 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2081 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2082 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2083 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2084 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2085 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2086 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002087 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2088 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2089 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2090 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2091 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2092 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2093 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2094 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002095 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2096 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2097 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2098 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002099 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2100 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2101 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2102 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07002103 "src/f32-vlrelu/gen/avx512f-x16.c",
2104 "src/f32-vlrelu/gen/avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002105 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002106 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2107 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2108 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2109 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2110 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2111 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2112 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2113 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2114 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2115 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2116 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2117 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2118 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2119 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2120 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2121 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2122 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2123 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2124 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2125 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2126 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2127 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2128 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2129 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002130 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2131 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2132 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2133 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2134 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2135 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2136 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2137 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002138 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2139 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2140 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2141 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2142 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2143 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2144 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2145 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002146 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2147 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2148 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2149 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2150 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2151 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002152 "src/math/exp-avx512f-p5-scalef.c",
2153 "src/math/exp-avx512f-p5.c",
2154 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002155 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002156 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002157 "src/math/sqrt-avx512f-nr1fma.c",
2158 "src/math/sqrt-avx512f-nr2fma.c",
2159 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002160]
2161
2162AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002163 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002164 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002165 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2166 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2167 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2168 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002169 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2170 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2171 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2172 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002173 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2174 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2175 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2176 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002177]
2178
2179AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002180 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002181 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2182 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002183 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002184 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2185 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002186 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2187 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2188 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002189 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002190 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2191 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2192 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002193 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002194 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2195 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002196 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002197 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2198 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2199 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2200 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2201 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2202 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2203 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2204 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2205 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2206 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2207 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2208 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2209 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2210 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2211 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2212 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2213 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2214 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002215 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2216 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002217 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002218 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2219 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2220 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2221 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2222 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2223 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2224 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2225 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2226 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2227 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2228 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2229 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2230 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2231 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2232 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2233 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2234 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2235 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002236 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2237 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2238 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2239 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2240 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2241 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2242 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2243 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2244 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2245 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2246 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2247 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2248 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2249 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2250 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2251 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2252 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2253 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002254]
2255
Marat Dukhan1b354632020-03-23 12:50:22 -07002256INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002257 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002258 "src/xnnpack/argmaxpool.h",
2259 "src/xnnpack/avgpool.h",
2260 "src/xnnpack/clamp.h",
2261 "src/xnnpack/common.h",
2262 "src/xnnpack/conv.h",
2263 "src/xnnpack/dwconv.h",
2264 "src/xnnpack/gavgpool.h",
2265 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002266 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002267 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002268 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002269 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002270 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002271 "src/xnnpack/lut.h",
2272 "src/xnnpack/math.h",
2273 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002274 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002275 "src/xnnpack/packx.h",
2276 "src/xnnpack/pad.h",
2277 "src/xnnpack/params.h",
2278 "src/xnnpack/pavgpool.h",
2279 "src/xnnpack/ppmm.h",
2280 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002281 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002282 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002283 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002284 "src/xnnpack/rmax.h",
2285 "src/xnnpack/scalar-utils.h",
2286 "src/xnnpack/spmm.h",
2287 "src/xnnpack/unpool.h",
2288 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002289 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002290 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002291 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002292 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002293 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002294 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002295 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002296]
2297
2298INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002299 "include/xnnpack.h",
2300 "src/xnnpack/allocator.h",
2301 "src/xnnpack/compute.h",
2302 "src/xnnpack/im2col.h",
2303 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002304 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002305 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002306 "src/xnnpack/operator.h",
2307 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002308 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002309 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002310 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002311 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002312]
2313
Marat Dukhan1b354632020-03-23 12:50:22 -07002314ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002315 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002316]
2317
Marat Dukhan1b354632020-03-23 12:50:22 -07002318MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002319 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002320 "include/xnnpack.h",
2321]
2322
Marat Dukhan1b354632020-03-23 12:50:22 -07002323MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002324 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002325 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002326 "src/xnnpack/requantization.h",
2327 "include/xnnpack.h",
2328]
2329
2330OPERATOR_TEST_PARAMS_HDRS = [
2331 "src/xnnpack/params.h",
2332 "src/xnnpack/common.h",
2333]
2334
2335WEIGHTS_PACK_HDRS = [
2336 "src/xnnpack/pack.h",
2337 "src/xnnpack/operator.h",
2338 "src/xnnpack/compute.h",
2339]
2340
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002341LOGGING_COPTS = select({
2342 # No logging in optimized mode
2343 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2344 # Full logging in debug mode
2345 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2346 # Error-only logging in default (fastbuild) mode
2347 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2348})
2349
Marat Dukhan3b59de22020-06-03 20:15:19 -07002350LOGGING_SRCS = select({
2351 # No logging in optimized mode
2352 ":optimized_build": [],
2353 "//conditions:default": [
2354 "src/operator-strings.c",
2355 "src/subgraph-strings.c",
2356 ],
2357})
2358
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002359LOGGING_HDRS = [
2360 "src/xnnpack/log.h",
2361]
2362
Marat Dukhan08c4a432019-10-03 09:29:21 -07002363xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002364 name = "tables",
2365 srcs = TABLE_SRCS,
2366 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002367 gcc_copts = xnnpack_gcc_std_copts(),
2368 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002369)
2370
2371xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002372 name = "scalar_ukernels",
2373 srcs = SCALAR_UKERNELS,
2374 hdrs = INTERNAL_HDRS,
2375 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002376 gcc_copts = xnnpack_gcc_std_copts(),
2377 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002378 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002379 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002380 "@FP16",
2381 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002382 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002383 ],
2384)
2385
2386xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002387 name = "scalar_ukernels_test_mode",
2388 srcs = SCALAR_UKERNELS,
2389 hdrs = INTERNAL_HDRS,
2390 aarch32_copts = ["-marm"],
2391 copts = [
2392 "-UNDEBUG",
2393 "-DXNN_TEST_MODE=1",
2394 ],
2395 gcc_copts = xnnpack_gcc_std_copts(),
2396 msvc_copts = xnnpack_msvc_std_copts(),
2397 deps = [
2398 ":tables",
2399 "@FP16",
2400 "@FXdiv",
2401 "@pthreadpool",
2402 ],
2403)
2404
2405xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002406 name = "wasm_ukernels",
2407 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002408 gcc_copts = xnnpack_gcc_std_copts(),
2409 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002410 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002411 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002412 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002413 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002414 "@FP16",
2415 "@FXdiv",
2416 "@pthreadpool",
2417 ],
2418)
2419
2420xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002421 name = "wasm_ukernels_test_mode",
2422 hdrs = INTERNAL_HDRS,
2423 copts = [
2424 "-UNDEBUG",
2425 "-DXNN_TEST_MODE=1",
2426 ],
2427 gcc_copts = xnnpack_gcc_std_copts(),
2428 msvc_copts = xnnpack_msvc_std_copts(),
2429 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002430 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002431 deps = [
2432 ":tables",
2433 "@FP16",
2434 "@FXdiv",
2435 "@pthreadpool",
2436 ],
2437)
2438
2439xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002440 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002441 hdrs = INTERNAL_HDRS,
2442 aarch32_copts = [
2443 "-marm",
2444 "-mfpu=neon",
2445 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002446 gcc_copts = xnnpack_gcc_std_copts(),
2447 gcc_x86_copts = ["-msse2"],
2448 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002449 optimized_copts = [
2450 "-O3",
2451 "-ffast-math",
2452 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002453 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002454 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002455 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002456 "@FP16",
2457 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002458 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002459 ],
2460)
2461
2462xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002463 name = "psimd_fastmath_ukernels_test_mode",
2464 hdrs = INTERNAL_HDRS,
2465 aarch32_copts = [
2466 "-marm",
2467 "-mfpu=neon",
2468 ],
2469 copts = [
2470 "-UNDEBUG",
2471 "-DXNN_TEST_MODE=1",
2472 ],
2473 gcc_copts = xnnpack_gcc_std_copts(),
2474 gcc_x86_copts = ["-msse2"],
2475 msvc_copts = xnnpack_msvc_std_copts(),
2476 optimized_copts = [
2477 "-O3",
2478 "-ffast-math",
2479 ],
2480 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2481 deps = [
2482 ":tables",
2483 "@FP16",
2484 "@psimd",
2485 "@pthreadpool",
2486 ],
2487)
2488
2489xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002490 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002491 hdrs = INTERNAL_HDRS,
2492 aarch32_copts = [
2493 "-marm",
2494 "-mfpu=neon",
2495 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002496 gcc_copts = xnnpack_gcc_std_copts(),
2497 gcc_x86_copts = ["-msse2"],
2498 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002499 optimized_copts = [
2500 "-O3",
2501 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002502 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002503 deps = [
2504 ":tables",
2505 "@FP16",
2506 "@psimd",
2507 "@pthreadpool",
2508 ],
2509)
2510
2511xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002512 name = "psimd_accmath_ukernels_test_mode",
2513 hdrs = INTERNAL_HDRS,
2514 aarch32_copts = [
2515 "-marm",
2516 "-mfpu=neon",
2517 ],
2518 copts = [
2519 "-UNDEBUG",
2520 "-DXNN_TEST_MODE=1",
2521 ],
2522 gcc_copts = xnnpack_gcc_std_copts(),
2523 gcc_x86_copts = ["-msse2"],
2524 msvc_copts = xnnpack_msvc_std_copts(),
2525 optimized_copts = [
2526 "-O3",
2527 ],
2528 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2529 deps = [
2530 ":tables",
2531 "@FP16",
2532 "@psimd",
2533 "@pthreadpool",
2534 ],
2535)
2536
2537xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002538 name = "neon_ukernels",
2539 hdrs = INTERNAL_HDRS,
2540 aarch32_copts = [
2541 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002542 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002543 "-mfpu=neon",
2544 ],
2545 aarch32_srcs = NEON_UKERNELS,
2546 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002547 gcc_copts = xnnpack_gcc_std_copts(),
2548 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002549 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002550 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002551 "@FP16",
2552 "@pthreadpool",
2553 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002554)
2555
2556xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002557 name = "neon_ukernels_test_mode",
2558 hdrs = INTERNAL_HDRS,
2559 aarch32_copts = [
2560 "-marm",
2561 "-march=armv7-a",
2562 "-mfpu=neon",
2563 ],
2564 aarch32_srcs = NEON_UKERNELS,
2565 aarch64_srcs = NEON_UKERNELS,
2566 copts = [
2567 "-UNDEBUG",
2568 "-DXNN_TEST_MODE=1",
2569 ],
2570 gcc_copts = xnnpack_gcc_std_copts(),
2571 msvc_copts = xnnpack_msvc_std_copts(),
2572 deps = [
2573 ":tables",
2574 "@FP16",
2575 "@pthreadpool",
2576 ],
2577)
2578
2579xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002580 name = "neonfma_ukernels",
2581 hdrs = INTERNAL_HDRS,
2582 aarch32_copts = [
2583 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002584 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002585 "-mfpu=neon-vfpv4",
2586 ],
2587 aarch32_srcs = NEONFMA_UKERNELS,
2588 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002589 apple_aarch32_copts = [
2590 "-mcpu=swift",
2591 "-mtune=generic",
2592 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002593 gcc_copts = xnnpack_gcc_std_copts(),
2594 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002595 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002596 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002597 "@FP16",
2598 "@pthreadpool",
2599 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002600)
2601
2602xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002603 name = "neonfma_ukernels_test_mode",
2604 hdrs = INTERNAL_HDRS,
2605 aarch32_copts = [
2606 "-marm",
2607 "-march=armv7-a",
2608 "-mfpu=neon-vfpv4",
2609 ],
2610 aarch32_srcs = NEONFMA_UKERNELS,
2611 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002612 apple_aarch32_copts = [
2613 "-mcpu=swift",
2614 "-mtune=generic",
2615 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002616 copts = [
2617 "-UNDEBUG",
2618 "-DXNN_TEST_MODE=1",
2619 ],
2620 gcc_copts = xnnpack_gcc_std_copts(),
2621 msvc_copts = xnnpack_msvc_std_copts(),
2622 deps = [
2623 ":tables",
2624 "@FP16",
2625 "@pthreadpool",
2626 ],
2627)
2628
2629xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002630 name = "neonv8_ukernels",
2631 hdrs = INTERNAL_HDRS,
2632 aarch32_copts = [
2633 "-marm",
2634 "-march=armv8-a",
2635 "-mfpu=neon-fp-armv8",
2636 ],
2637 aarch32_srcs = NEONV8_UKERNELS,
2638 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002639 apple_aarch32_copts = [
2640 "-mcpu=cyclone",
2641 "-mtune=generic",
2642 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002643 gcc_copts = xnnpack_gcc_std_copts(),
2644 msvc_copts = xnnpack_msvc_std_copts(),
2645 deps = [
2646 ":tables",
2647 "@FP16",
2648 "@pthreadpool",
2649 ],
2650)
2651
2652xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002653 name = "neonv8_ukernels_test_mode",
2654 hdrs = INTERNAL_HDRS,
2655 aarch32_copts = [
2656 "-marm",
2657 "-march=armv8-a",
2658 "-mfpu=neon-fp-armv8",
2659 ],
2660 aarch32_srcs = NEONV8_UKERNELS,
2661 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002662 apple_aarch32_copts = [
2663 "-mcpu=cyclone",
2664 "-mtune=generic",
2665 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002666 copts = [
2667 "-UNDEBUG",
2668 "-DXNN_TEST_MODE=1",
2669 ],
2670 gcc_copts = xnnpack_gcc_std_copts(),
2671 msvc_copts = xnnpack_msvc_std_copts(),
2672 deps = [
2673 ":tables",
2674 "@FP16",
2675 "@pthreadpool",
2676 ],
2677)
2678
2679xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002680 name = "neonfp16arith_ukernels",
2681 hdrs = INTERNAL_HDRS,
2682 aarch64_copts = ["-march=armv8.2-a+fp16"],
2683 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002684 gcc_copts = xnnpack_gcc_std_copts(),
2685 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002686 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002687 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002688 "@FP16",
2689 "@pthreadpool",
2690 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002691)
2692
2693xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002694 name = "neonfp16arith_ukernels_test_mode",
2695 hdrs = INTERNAL_HDRS,
2696 aarch64_copts = ["-march=armv8.2-a+fp16"],
2697 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2698 copts = [
2699 "-UNDEBUG",
2700 "-DXNN_TEST_MODE=1",
2701 ],
2702 gcc_copts = xnnpack_gcc_std_copts(),
2703 msvc_copts = xnnpack_msvc_std_copts(),
2704 deps = [
2705 ":tables",
2706 "@FP16",
2707 "@pthreadpool",
2708 ],
2709)
2710
2711xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002712 name = "sse2_ukernels",
2713 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002714 gcc_copts = xnnpack_gcc_std_copts(),
2715 gcc_x86_copts = ["-msse2"],
2716 msvc_copts = xnnpack_msvc_std_copts(),
2717 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002718 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002719 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002720 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002721 "@FP16",
2722 "@pthreadpool",
2723 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002724)
2725
2726xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002727 name = "sse2_ukernels_test_mode",
2728 hdrs = INTERNAL_HDRS,
2729 copts = [
2730 "-UNDEBUG",
2731 "-DXNN_TEST_MODE=1",
2732 ],
2733 gcc_copts = xnnpack_gcc_std_copts(),
2734 gcc_x86_copts = ["-msse2"],
2735 msvc_copts = xnnpack_msvc_std_copts(),
2736 msvc_x86_32_copts = ["/arch:SSE2"],
2737 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2738 deps = [
2739 ":tables",
2740 "@FP16",
2741 "@pthreadpool",
2742 ],
2743)
2744
2745xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002746 name = "ssse3_ukernels",
2747 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002748 gcc_copts = xnnpack_gcc_std_copts(),
2749 gcc_x86_copts = ["-mssse3"],
2750 msvc_copts = xnnpack_msvc_std_copts(),
2751 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002752 x86_srcs = SSSE3_UKERNELS,
2753 deps = [
2754 ":tables",
2755 "@FP16",
2756 "@pthreadpool",
2757 ],
2758)
2759
2760xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002761 name = "ssse3_ukernels_test_mode",
2762 hdrs = INTERNAL_HDRS,
2763 copts = [
2764 "-UNDEBUG",
2765 "-DXNN_TEST_MODE=1",
2766 ],
2767 gcc_copts = xnnpack_gcc_std_copts(),
2768 gcc_x86_copts = ["-mssse3"],
2769 msvc_copts = xnnpack_msvc_std_copts(),
2770 msvc_x86_32_copts = ["/arch:SSE2"],
2771 x86_srcs = SSSE3_UKERNELS,
2772 deps = [
2773 ":tables",
2774 "@FP16",
2775 "@pthreadpool",
2776 ],
2777)
2778
2779xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002780 name = "sse41_ukernels",
2781 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002782 gcc_copts = xnnpack_gcc_std_copts(),
2783 gcc_x86_copts = ["-msse4.1"],
2784 msvc_copts = xnnpack_msvc_std_copts(),
2785 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002786 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002787 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002788 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002789 "@FP16",
2790 "@pthreadpool",
2791 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002792)
2793
2794xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002795 name = "sse41_ukernels_test_mode",
2796 hdrs = INTERNAL_HDRS,
2797 copts = [
2798 "-UNDEBUG",
2799 "-DXNN_TEST_MODE=1",
2800 ],
2801 gcc_copts = xnnpack_gcc_std_copts(),
2802 gcc_x86_copts = ["-msse4.1"],
2803 msvc_copts = xnnpack_msvc_std_copts(),
2804 msvc_x86_32_copts = ["/arch:SSE2"],
2805 x86_srcs = SSE41_UKERNELS,
2806 deps = [
2807 ":tables",
2808 "@FP16",
2809 "@pthreadpool",
2810 ],
2811)
2812
2813xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002814 name = "avx_ukernels",
2815 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002816 gcc_copts = xnnpack_gcc_std_copts(),
2817 gcc_x86_copts = ["-mavx"],
2818 msvc_copts = xnnpack_msvc_std_copts(),
2819 msvc_x86_32_copts = ["/arch:AVX"],
2820 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002821 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002822 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002823 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002824 "@FP16",
2825 "@pthreadpool",
2826 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002827)
2828
2829xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002830 name = "avx_ukernels_test_mode",
2831 hdrs = INTERNAL_HDRS,
2832 copts = [
2833 "-UNDEBUG",
2834 "-DXNN_TEST_MODE=1",
2835 ],
2836 gcc_copts = xnnpack_gcc_std_copts(),
2837 gcc_x86_copts = ["-mavx"],
2838 msvc_copts = xnnpack_msvc_std_copts(),
2839 msvc_x86_32_copts = ["/arch:AVX"],
2840 msvc_x86_64_copts = ["/arch:AVX"],
2841 x86_srcs = AVX_UKERNELS,
2842 deps = [
2843 ":tables",
2844 "@FP16",
2845 "@pthreadpool",
2846 ],
2847)
2848
2849xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002850 name = "fma3_ukernels",
2851 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002852 gcc_copts = xnnpack_gcc_std_copts(),
2853 gcc_x86_copts = ["-mfma"],
2854 msvc_copts = xnnpack_msvc_std_copts(),
2855 msvc_x86_32_copts = ["/arch:AVX"],
2856 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002857 x86_srcs = FMA3_UKERNELS,
2858 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002859 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002860 "@FP16",
2861 "@pthreadpool",
2862 ],
2863)
2864
2865xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002866 name = "fma3_ukernels_test_mode",
2867 hdrs = INTERNAL_HDRS,
2868 copts = [
2869 "-UNDEBUG",
2870 "-DXNN_TEST_MODE=1",
2871 ],
2872 gcc_copts = xnnpack_gcc_std_copts(),
2873 gcc_x86_copts = ["-mfma"],
2874 msvc_copts = xnnpack_msvc_std_copts(),
2875 msvc_x86_32_copts = ["/arch:AVX"],
2876 msvc_x86_64_copts = ["/arch:AVX"],
2877 x86_srcs = FMA3_UKERNELS,
2878 deps = [
2879 ":tables",
2880 "@FP16",
2881 "@pthreadpool",
2882 ],
2883)
2884
2885xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002886 name = "avx2_ukernels",
2887 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002888 gcc_copts = xnnpack_gcc_std_copts(),
2889 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002890 "-mfma",
2891 "-mavx2",
2892 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002893 msvc_copts = xnnpack_msvc_std_copts(),
2894 msvc_x86_32_copts = ["/arch:AVX2"],
2895 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002896 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002897 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002898 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002899 "@FP16",
2900 "@pthreadpool",
2901 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002902)
2903
2904xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002905 name = "avx2_ukernels_test_mode",
2906 hdrs = INTERNAL_HDRS,
2907 copts = [
2908 "-UNDEBUG",
2909 "-DXNN_TEST_MODE=1",
2910 ],
2911 gcc_copts = xnnpack_gcc_std_copts(),
2912 gcc_x86_copts = [
2913 "-mfma",
2914 "-mavx2",
2915 ],
2916 msvc_copts = xnnpack_msvc_std_copts(),
2917 msvc_x86_32_copts = ["/arch:AVX2"],
2918 msvc_x86_64_copts = ["/arch:AVX2"],
2919 x86_srcs = AVX2_UKERNELS,
2920 deps = [
2921 ":tables",
2922 "@FP16",
2923 "@pthreadpool",
2924 ],
2925)
2926
2927xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002928 name = "avx512f_ukernels",
2929 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002930 gcc_copts = xnnpack_gcc_std_copts(),
2931 gcc_x86_copts = ["-mavx512f"],
2932 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2933 msvc_copts = xnnpack_msvc_std_copts(),
2934 msvc_x86_32_copts = ["/arch:AVX512"],
2935 msvc_x86_64_copts = ["/arch:AVX512"],
2936 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002937 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002938 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002939 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002940 "@FP16",
2941 "@pthreadpool",
2942 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002943)
2944
2945xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002946 name = "avx512f_ukernels_test_mode",
2947 hdrs = INTERNAL_HDRS,
2948 copts = [
2949 "-UNDEBUG",
2950 "-DXNN_TEST_MODE=1",
2951 ],
2952 gcc_copts = xnnpack_gcc_std_copts(),
2953 gcc_x86_copts = ["-mavx512f"],
2954 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2955 msvc_copts = xnnpack_msvc_std_copts(),
2956 msvc_x86_32_copts = ["/arch:AVX512"],
2957 msvc_x86_64_copts = ["/arch:AVX512"],
2958 msys_copts = ["-fno-asynchronous-unwind-tables"],
2959 x86_srcs = AVX512F_UKERNELS,
2960 deps = [
2961 ":tables",
2962 "@FP16",
2963 "@pthreadpool",
2964 ],
2965)
2966
2967xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002968 name = "asm_ukernels",
2969 hdrs = ["src/xnnpack/assembly.h"],
2970 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002971 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002972 aarch64_srcs = AARCH64_ASM_UKERNELS,
2973)
2974
Marat Dukhan3b59de22020-06-03 20:15:19 -07002975xnnpack_cc_library(
2976 name = "logging_utils",
2977 srcs = LOGGING_SRCS,
2978 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2979 copts = LOGGING_COPTS + [
2980 "-Isrc",
2981 "-Iinclude",
2982 ] + select({
2983 ":debug_build": [],
2984 "//conditions:default": xnnpack_min_size_copts(),
2985 }),
2986 gcc_copts = xnnpack_gcc_std_copts(),
2987 msvc_copts = xnnpack_msvc_std_copts(),
2988 visibility = xnnpack_visibility(),
2989 deps = [
2990 "@FP16",
2991 "@clog",
2992 "@pthreadpool",
2993 ],
2994)
2995
Marat Dukhan08c4a432019-10-03 09:29:21 -07002996xnnpack_aggregate_library(
2997 name = "ukernels",
2998 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 ":neon_ukernels",
3000 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003001 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003002 ":asm_ukernels",
3003 ],
3004 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 ":neon_ukernels",
3006 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003007 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008 ":neonfp16arith_ukernels",
3009 ":asm_ukernels",
3010 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003011 generic_deps = [
3012 ":scalar_ukernels",
3013 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003014 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003015 ":psimd_fastmath_ukernels",
3016 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003017 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003018 wasm_deps = [
3019 ":wasm_ukernels",
3020 ],
3021 wasmsimd_deps = [
3022 ":wasm_ukernels",
3023 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003026 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003027 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003028 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003029 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003030 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003031 ":avx512f_ukernels",
3032 ],
3033)
3034
Marat Dukhan33fcf782020-05-24 14:27:15 -07003035xnnpack_aggregate_library(
3036 name = "ukernels_test_mode",
3037 aarch32_deps = [
3038 ":neon_ukernels_test_mode",
3039 ":neonfma_ukernels_test_mode",
3040 ":neonv8_ukernels_test_mode",
3041 ":asm_ukernels",
3042 ],
3043 aarch64_deps = [
3044 ":neon_ukernels_test_mode",
3045 ":neonfma_ukernels_test_mode",
3046 ":neonv8_ukernels_test_mode",
3047 ":neonfp16arith_ukernels_test_mode",
3048 ":asm_ukernels",
3049 ],
3050 generic_deps = [
3051 ":scalar_ukernels_test_mode",
3052 ],
3053 psimd_deps = [
3054 ":psimd_fastmath_ukernels_test_mode",
3055 ":psimd_accmath_ukernels_test_mode",
3056 ],
3057 wasm_deps = [
3058 ":wasm_ukernels_test_mode",
3059 ],
3060 wasmsimd_deps = [
3061 ":wasm_ukernels_test_mode",
3062 ],
3063 x86_deps = [
3064 ":sse2_ukernels_test_mode",
3065 ":ssse3_ukernels_test_mode",
3066 ":sse41_ukernels_test_mode",
3067 ":avx_ukernels_test_mode",
3068 ":fma3_ukernels_test_mode",
3069 ":avx2_ukernels_test_mode",
3070 ":avx512f_ukernels_test_mode",
3071 ],
3072)
3073
Marat Dukhan86df3842020-06-17 14:03:31 -07003074# This target must be compatible with Android NDK r17 and compile with Clang 6.
3075# It is similar to :ukernels target, but excludes microkernels which require
3076# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003077xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003078 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003079 aarch32_deps = [
3080 ":neon_ukernels",
3081 ":neonfma_ukernels",
3082 ":neonv8_ukernels",
3083 ":asm_ukernels",
3084 ],
3085 aarch64_deps = [
3086 ":neon_ukernels",
3087 ":neonfma_ukernels",
3088 ":neonv8_ukernels",
3089 ":asm_ukernels",
3090 ],
3091 generic_deps = [
3092 ":scalar_ukernels",
3093 ],
3094 psimd_deps = [
3095 ":psimd_fastmath_ukernels",
3096 ":psimd_accmath_ukernels",
3097 ],
3098 wasm_deps = [
3099 ":wasm_ukernels",
3100 ],
3101 wasmsimd_deps = [
3102 ":wasm_ukernels",
3103 ],
3104 x86_deps = [
3105 ":sse2_ukernels",
3106 ":ssse3_ukernels",
3107 ":sse41_ukernels",
3108 ":avx_ukernels",
3109 ":fma3_ukernels",
3110 ":avx2_ukernels",
3111 ":avx512f_ukernels",
3112 ],
3113)
3114
Marat Dukhan08c4a432019-10-03 09:29:21 -07003115xnnpack_cc_library(
3116 name = "im2col",
3117 srcs = ["src/im2col.c"],
3118 hdrs = [
3119 "src/xnnpack/common.h",
3120 "src/xnnpack/im2col.h",
3121 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003122 gcc_copts = xnnpack_gcc_std_copts(),
3123 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003124)
3125
3126xnnpack_cc_library(
3127 name = "indirection",
3128 srcs = ["src/indirection.c"],
3129 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003130 gcc_copts = xnnpack_gcc_std_copts(),
3131 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003132 deps = [
3133 "@FP16",
3134 "@FXdiv",
3135 "@pthreadpool",
3136 ],
3137)
3138
3139xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003140 name = "indirection_test_mode",
3141 srcs = ["src/indirection.c"],
3142 hdrs = INTERNAL_HDRS,
3143 copts = [
3144 "-UNDEBUG",
3145 "-DXNN_TEST_MODE=1",
3146 ],
3147 gcc_copts = xnnpack_gcc_std_copts(),
3148 msvc_copts = xnnpack_msvc_std_copts(),
3149 deps = [
3150 "@FP16",
3151 "@FXdiv",
3152 "@pthreadpool",
3153 ],
3154)
3155
3156xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003157 name = "packing",
3158 srcs = ["src/packing.c"],
3159 hdrs = INTERNAL_HDRS,
3160 gcc_copts = xnnpack_gcc_std_copts(),
3161 msvc_copts = xnnpack_msvc_std_copts(),
3162 deps = [
3163 "@FP16",
3164 "@FXdiv",
3165 "@pthreadpool",
3166 ],
3167)
3168
3169xnnpack_cc_library(
3170 name = "packing_test_mode",
3171 srcs = ["src/packing.c"],
3172 hdrs = INTERNAL_HDRS,
3173 copts = [
3174 "-UNDEBUG",
3175 "-DXNN_TEST_MODE=1",
3176 ],
3177 gcc_copts = xnnpack_gcc_std_copts(),
3178 msvc_copts = xnnpack_msvc_std_copts(),
3179 deps = [
3180 "@FP16",
3181 "@FXdiv",
3182 "@pthreadpool",
3183 ],
3184)
3185
3186xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 name = "operator_run",
3188 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003189 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003190 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003191 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3192 "//conditions:default": [],
3193 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003194 gcc_copts = xnnpack_gcc_std_copts(),
3195 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003196 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003197 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003198 "@FP16",
3199 "@FXdiv",
3200 "@clog",
3201 "@pthreadpool",
3202 ],
3203)
3204
Chao Mei6ddfc602020-05-13 22:29:36 -07003205xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003206 name = "operator_run_test_mode",
3207 srcs = ["src/operator-run.c"],
3208 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3209 copts = LOGGING_COPTS + [
3210 "-UNDEBUG",
3211 "-DXNN_TEST_MODE=1",
3212 ] + select({
3213 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3214 "//conditions:default": [],
3215 }),
3216 gcc_copts = xnnpack_gcc_std_copts(),
3217 msvc_copts = xnnpack_msvc_std_copts(),
3218 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003219 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003220 "@FP16",
3221 "@FXdiv",
3222 "@clog",
3223 "@pthreadpool",
3224 ],
3225)
3226
3227xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003228 name = "memory_planner",
3229 srcs = ["src/memory-planner.c"],
3230 hdrs = INTERNAL_HDRS,
3231 defines = select({
3232 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3233 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3234 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3235 }),
3236 gcc_copts = xnnpack_gcc_std_copts(),
3237 msvc_copts = xnnpack_msvc_std_copts(),
3238 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003239 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003240 "@pthreadpool",
3241 ],
3242)
3243
Marat Dukhan33fcf782020-05-24 14:27:15 -07003244xnnpack_cc_library(
3245 name = "memory_planner_test_mode",
3246 srcs = ["src/memory-planner.c"],
3247 hdrs = INTERNAL_HDRS,
3248 copts = [
3249 "-UNDEBUG",
3250 "-DXNN_TEST_MODE=1",
3251 ],
3252 defines = select({
3253 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3254 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3255 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3256 }),
3257 gcc_copts = xnnpack_gcc_std_copts(),
3258 msvc_copts = xnnpack_msvc_std_copts(),
3259 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003260 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003261 "@pthreadpool",
3262 ],
3263)
3264
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265cc_library(
3266 name = "enable_assembly",
3267 defines = select({
3268 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3269 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003270 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271 }),
3272)
3273
Marat Dukhan9de90e02020-06-18 16:04:12 -07003274cc_library(
3275 name = "enable_sparse",
3276 defines = select({
3277 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3278 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003279 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003280 }),
3281)
3282
Marat Dukhancf056b22019-10-07 10:26:29 -07003283xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003284 name = "operators",
3285 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003286 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003287 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003288 ],
3289 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003290 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003291 "-Isrc",
3292 "-Iinclude",
3293 ] + select({
3294 ":debug_build": [],
3295 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003296 }) + select({
3297 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3298 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003299 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003300 gcc_copts = xnnpack_gcc_std_copts(),
3301 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003302 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003303 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003304 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003305 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003306 "@FP16",
3307 "@FXdiv",
3308 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003309 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003310 ],
3311)
3312
Marat Dukhan10a38082020-04-17 03:58:35 -07003313xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003314 name = "operators_test_mode",
3315 srcs = OPERATOR_SRCS + [
3316 "src/memory.c",
3317 "src/operator-delete.c",
3318 ],
3319 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3320 copts = LOGGING_COPTS + [
3321 "-Isrc",
3322 "-Iinclude",
3323 "-UNDEBUG",
3324 "-DXNN_TEST_MODE=1",
3325 ] + select({
3326 ":debug_build": [],
3327 "//conditions:default": xnnpack_min_size_copts(),
3328 }) + select({
3329 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3330 "//conditions:default": [],
3331 }),
3332 gcc_copts = xnnpack_gcc_std_copts(),
3333 msvc_copts = xnnpack_msvc_std_copts(),
3334 deps = [
3335 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003336 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003337 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003338 "@FP16",
3339 "@FXdiv",
3340 "@clog",
3341 "@pthreadpool",
3342 ],
3343)
3344
3345xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003346 name = "XNNPACK",
3347 srcs = [
3348 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003349 "src/runtime.c",
3350 "src/subgraph.c",
3351 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003352 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003353 hdrs = ["include/xnnpack.h"],
3354 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003355 "-Isrc",
3356 "-Iinclude",
3357 ] + select({
3358 ":debug_build": [],
3359 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003360 }) + select({
3361 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3362 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003363 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003364 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003365 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003366 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003367 visibility = xnnpack_visibility(),
3368 deps = [
3369 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003370 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003371 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003372 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003373 ":operator_run",
3374 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003375 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003376 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003377 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003378 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003379 ] + select({
3380 ":emscripten": [],
3381 "//conditions:default": ["@cpuinfo"],
3382 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003383)
3384
Marat Dukhan10a38082020-04-17 03:58:35 -07003385xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003386 name = "XNNPACK_test_mode",
3387 srcs = [
3388 "src/init.c",
3389 "src/runtime.c",
3390 "src/subgraph.c",
3391 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003392 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003393 hdrs = ["include/xnnpack.h"],
3394 copts = LOGGING_COPTS + [
3395 "-Isrc",
3396 "-Iinclude",
3397 "-UNDEBUG",
3398 "-DXNN_TEST_MODE=1",
3399 ] + select({
3400 ":debug_build": [],
3401 "//conditions:default": xnnpack_min_size_copts(),
3402 }) + select({
3403 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3404 "//conditions:default": [],
3405 }),
3406 gcc_copts = xnnpack_gcc_std_copts(),
3407 includes = ["include"],
3408 msvc_copts = xnnpack_msvc_std_copts(),
3409 visibility = xnnpack_visibility(),
3410 deps = [
3411 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003412 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003413 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003414 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003415 ":operator_run_test_mode",
3416 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003417 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003418 "@clog",
3419 "@FP16",
3420 "@pthreadpool",
3421 ] + select({
3422 ":emscripten": [],
3423 "//conditions:default": ["@cpuinfo"],
3424 }),
3425)
3426
3427xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003428 name = "xnnpack_f32",
3429 srcs = [
3430 "src/init.c",
3431 "src/runtime.c",
3432 "src/subgraph.c",
3433 "src/tensor.c",
3434 ] + SUBGRAPH_SRCS,
3435 hdrs = ["include/xnnpack.h"],
3436 copts = LOGGING_COPTS + [
3437 "-Isrc",
3438 "-Iinclude",
3439 ] + select({
3440 ":debug_build": [],
3441 "//conditions:default": xnnpack_min_size_copts(),
3442 }) + select({
3443 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3444 "//conditions:default": [],
3445 }),
3446 defines = [
3447 "XNN_NO_Q8_OPERATORS",
3448 "XNN_NO_U8_OPERATORS",
3449 "XNN_NO_X8_OPERATORS",
3450 "XNN_NO_F16_OPERATORS",
3451 "XNN_NO_X16_OPERATORS",
3452 ],
3453 gcc_copts = xnnpack_gcc_std_copts(),
3454 includes = ["include"],
3455 msvc_copts = xnnpack_msvc_std_copts(),
3456 visibility = xnnpack_visibility(),
3457 deps = [
3458 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003459 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003460 ":logging_utils",
3461 ":memory_planner",
3462 ":operator_run",
3463 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003464 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003465 "@clog",
3466 "@FP16",
3467 "@pthreadpool",
3468 ] + select({
3469 ":emscripten": [],
3470 "//conditions:default": ["@cpuinfo"],
3471 }),
3472)
3473
3474xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003475 name = "xnnpack_operators_nhwc_f32",
3476 srcs = [
3477 "src/init.c",
3478 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003479 hdrs = ["include/xnnpack.h"],
3480 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003481 "-Isrc",
3482 "-Iinclude",
3483 ] + select({
3484 ":debug_build": [],
3485 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003486 }) + select({
3487 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3488 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003489 }),
3490 defines = [
3491 "XNN_NO_Q8_OPERATORS",
3492 "XNN_NO_U8_OPERATORS",
3493 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003494 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003495 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003496 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003497 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003498 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003499 visibility = xnnpack_visibility(),
3500 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003501 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003502 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003503 ":operator_run",
3504 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003505 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003506 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003507 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003508 ] + select({
3509 ":emscripten": [],
3510 "//conditions:default": ["@cpuinfo"],
3511 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003512)
3513
Marat Dukhancf056b22019-10-07 10:26:29 -07003514xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515 name = "bench_utils",
3516 srcs = ["bench/utils.cc"],
3517 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003518 deps = [
3519 "@com_google_benchmark//:benchmark",
3520 "@cpuinfo",
3521 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003522)
3523
Frank Barchard7e955972019-10-11 10:34:25 -07003524######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003525
3526xnnpack_benchmark(
3527 name = "q8_gemm_bench",
3528 srcs = [
3529 "bench/gemm.h",
3530 "bench/q8-gemm.cc",
3531 "src/xnnpack/AlignedAllocator.h",
3532 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003533 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003534 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003535)
3536
3537xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003538 name = "f16_igemm_bench",
3539 srcs = [
3540 "bench/f16-igemm.cc",
3541 "bench/conv.h",
3542 "bench/google/conv.h",
3543 "src/xnnpack/AlignedAllocator.h",
3544 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003545 deps = MICROKERNEL_BENCHMARK_DEPS + [
3546 ":indirection",
3547 ":packing",
3548 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003549)
3550
3551xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003552 name = "f16_gemm_bench",
3553 srcs = [
3554 "bench/f16-gemm.cc",
3555 "bench/gemm.h",
3556 "src/xnnpack/AlignedAllocator.h",
3557 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003558 deps = MICROKERNEL_BENCHMARK_DEPS + [
3559 ":packing",
3560 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561)
3562
3563xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003564 name = "f16_spmm_bench",
3565 srcs = [
3566 "bench/f16-spmm.cc",
3567 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003568 "src/xnnpack/AlignedAllocator.h",
3569 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003570 deps = MICROKERNEL_BENCHMARK_DEPS,
3571)
3572
3573xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 name = "f32_igemm_bench",
3575 srcs = [
3576 "bench/f32-igemm.cc",
3577 "bench/conv.h",
3578 "src/xnnpack/AlignedAllocator.h",
3579 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003580 deps = MICROKERNEL_BENCHMARK_DEPS + [
3581 ":indirection",
3582 ":packing",
3583 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003584)
3585
3586xnnpack_benchmark(
3587 name = "f32_conv_hwc_bench",
3588 srcs = [
3589 "bench/f32-conv-hwc.cc",
3590 "bench/dconv.h",
3591 "src/xnnpack/AlignedAllocator.h",
3592 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003593 deps = MICROKERNEL_BENCHMARK_DEPS + [
3594 ":packing",
3595 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003596)
3597
3598xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003599 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003600 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003601 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003602 "bench/dconv.h",
3603 "src/xnnpack/AlignedAllocator.h",
3604 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003605 deps = MICROKERNEL_BENCHMARK_DEPS + [
3606 ":packing",
3607 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003608)
3609
3610xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003611 name = "f16_dwconv_bench",
3612 srcs = [
3613 "bench/f16-dwconv.cc",
3614 "bench/dwconv.h",
3615 "bench/google/dwconv.h",
3616 "src/xnnpack/AlignedAllocator.h",
3617 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003618 deps = MICROKERNEL_BENCHMARK_DEPS + [
3619 ":indirection",
3620 ":packing",
3621 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003622)
3623
3624xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003625 name = "f32_dwconv_bench",
3626 srcs = [
3627 "bench/f32-dwconv.cc",
3628 "bench/dwconv.h",
3629 "src/xnnpack/AlignedAllocator.h",
3630 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003631 deps = MICROKERNEL_BENCHMARK_DEPS + [
3632 ":indirection",
3633 ":packing",
3634 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003635)
3636
3637xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003638 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003639 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003640 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003641 "bench/dwconv.h",
3642 "src/xnnpack/AlignedAllocator.h",
3643 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003644 deps = MICROKERNEL_BENCHMARK_DEPS + [
3645 ":indirection",
3646 ":packing",
3647 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648)
3649
3650xnnpack_benchmark(
3651 name = "f32_gemm_bench",
3652 srcs = [
3653 "bench/f32-gemm.cc",
3654 "bench/gemm.h",
3655 "src/xnnpack/AlignedAllocator.h",
3656 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003657 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003658 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003659)
3660
3661xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003662 name = "f32_hswish_bench",
3663 srcs = [
3664 "bench/f32-hswish.cc",
3665 "src/xnnpack/AlignedAllocator.h",
3666 ] + MICROKERNEL_BENCHMARK_HDRS,
3667 deps = MICROKERNEL_BENCHMARK_DEPS,
3668)
3669
3670xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003671 name = "f32_raddexpminusmax_bench",
3672 srcs = [
3673 "bench/f32-raddexpminusmax.cc",
3674 "src/xnnpack/AlignedAllocator.h",
3675 ] + MICROKERNEL_BENCHMARK_HDRS,
3676 deps = MICROKERNEL_BENCHMARK_DEPS,
3677)
3678
3679xnnpack_benchmark(
3680 name = "f32_raddextexp_bench",
3681 srcs = [
3682 "bench/f32-raddextexp.cc",
3683 "src/xnnpack/AlignedAllocator.h",
3684 ] + MICROKERNEL_BENCHMARK_HDRS,
3685 deps = MICROKERNEL_BENCHMARK_DEPS,
3686)
3687
3688xnnpack_benchmark(
3689 name = "f32_raddstoreexpminusmax_bench",
3690 srcs = [
3691 "bench/f32-raddstoreexpminusmax.cc",
3692 "src/xnnpack/AlignedAllocator.h",
3693 ] + MICROKERNEL_BENCHMARK_HDRS,
3694 deps = MICROKERNEL_BENCHMARK_DEPS,
3695)
3696
3697xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003698 name = "f32_rmax_bench",
3699 srcs = [
3700 "bench/f32-rmax.cc",
3701 "src/xnnpack/AlignedAllocator.h",
3702 ] + MICROKERNEL_BENCHMARK_HDRS,
3703 deps = MICROKERNEL_BENCHMARK_DEPS,
3704)
3705
3706xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003707 name = "f32_sigmoid_bench",
3708 srcs = [
3709 "bench/f32-sigmoid.cc",
3710 "src/xnnpack/AlignedAllocator.h",
3711 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003712 deps = MICROKERNEL_BENCHMARK_DEPS,
3713)
3714
3715xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003716 name = "f32_spmm_bench",
3717 srcs = [
3718 "bench/f32-spmm.cc",
3719 "bench/gemm.h",
3720 "src/xnnpack/AlignedAllocator.h",
3721 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003722 deps = MICROKERNEL_BENCHMARK_DEPS,
3723)
3724
3725xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003726 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003727 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003728 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003729 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003730 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003731 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003732)
3733
3734xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003735 name = "f32_vscaleexpminusmax_bench",
3736 srcs = [
3737 "bench/f32-vscaleexpminusmax.cc",
3738 "src/xnnpack/AlignedAllocator.h",
3739 ] + MICROKERNEL_BENCHMARK_HDRS,
3740 deps = MICROKERNEL_BENCHMARK_DEPS,
3741)
3742
3743xnnpack_benchmark(
3744 name = "f32_vscaleextexp_bench",
3745 srcs = [
3746 "bench/f32-vscaleextexp.cc",
3747 "src/xnnpack/AlignedAllocator.h",
3748 ] + MICROKERNEL_BENCHMARK_HDRS,
3749 deps = MICROKERNEL_BENCHMARK_DEPS,
3750)
3751
3752xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003753 name = "f32_vsqrt_bench",
3754 srcs = [
3755 "bench/f32-vsqrt.cc",
3756 "src/xnnpack/AlignedAllocator.h",
3757 ] + MICROKERNEL_BENCHMARK_HDRS,
3758 deps = MICROKERNEL_BENCHMARK_DEPS,
3759)
3760
3761xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003762 name = "f32_im2col_gemm_bench",
3763 srcs = [
3764 "bench/f32-im2col-gemm.cc",
3765 "bench/conv.h",
3766 "src/xnnpack/AlignedAllocator.h",
3767 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003768 deps = MICROKERNEL_BENCHMARK_DEPS + [
3769 ":im2col",
3770 ":packing",
3771 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772)
3773
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003774xnnpack_benchmark(
3775 name = "requantization_bench",
3776 srcs = [
3777 "bench/requantization.cc",
3778 "src/xnnpack/requantization-stubs.h",
3779 "src/xnnpack/AlignedAllocator.h",
3780 ] + MICROKERNEL_BENCHMARK_HDRS,
3781 deps = MICROKERNEL_BENCHMARK_DEPS,
3782)
3783
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003784xnnpack_benchmark(
3785 name = "rounding_bench",
3786 srcs = [
3787 "bench/rounding.cc",
3788 "src/xnnpack/math-stubs.h",
3789 "src/xnnpack/AlignedAllocator.h",
3790 ] + MICROKERNEL_BENCHMARK_HDRS,
3791 deps = MICROKERNEL_BENCHMARK_DEPS,
3792)
3793
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794########################### Benchmarks for operators ###########################
3795
3796xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003797 name = "average_pooling_bench",
3798 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003799 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003800 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003801 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802)
3803
3804xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003805 name = "bankers_rounding_bench",
3806 srcs = ["bench/bankers-rounding.cc"],
3807 copts = xnnpack_optional_tflite_copts(),
3808 tags = ["nowin32"],
3809 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3810)
3811
3812xnnpack_benchmark(
3813 name = "ceiling_bench",
3814 srcs = ["bench/ceiling.cc"],
3815 copts = xnnpack_optional_tflite_copts(),
3816 tags = ["nowin32"],
3817 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3818)
3819
3820xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003821 name = "channel_shuffle_bench",
3822 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003823 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003824)
3825
3826xnnpack_benchmark(
3827 name = "convolution_bench",
3828 srcs = ["bench/convolution.cc"],
3829 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003830 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003831 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003832)
3833
3834xnnpack_benchmark(
3835 name = "deconvolution_bench",
3836 srcs = ["bench/deconvolution.cc"],
3837 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003838 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003839 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003840)
3841
3842xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003843 name = "floor_bench",
3844 srcs = ["bench/floor.cc"],
3845 copts = xnnpack_optional_tflite_copts(),
3846 tags = ["nowin32"],
3847 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3848)
3849
3850xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003851 name = "global_average_pooling_bench",
3852 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003853 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003854)
3855
3856xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003857 name = "hardswish_bench",
3858 srcs = ["bench/hardswish.cc"],
3859 copts = xnnpack_optional_tflite_copts(),
3860 tags = ["nowin32"],
3861 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3862)
3863
3864xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003865 name = "max_pooling_bench",
3866 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003867 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003868)
3869
3870xnnpack_benchmark(
3871 name = "sigmoid_bench",
3872 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003873 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003874 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003875 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003876)
3877
3878xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003879 name = "prelu_bench",
3880 srcs = ["bench/prelu.cc"],
3881 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003882 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003883 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003884)
3885
3886xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003887 name = "softmax_bench",
3888 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003889 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003890 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003891 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003892)
3893
Marat Dukhan87727142020-06-24 15:24:10 -07003894xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003895 name = "square_root_bench",
3896 srcs = ["bench/square-root.cc"],
3897 copts = xnnpack_optional_tflite_copts(),
3898 tags = ["nowin32"],
3899 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3900)
3901
3902xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003903 name = "truncation_bench",
3904 srcs = ["bench/truncation.cc"],
3905 deps = OPERATOR_BENCHMARK_DEPS,
3906)
3907
Marat Dukhanc068bb62019-10-04 13:24:39 -07003908############################# End-to-end benchmarks ############################
3909
3910cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003911 name = "fp32_mobilenet_v1",
3912 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003913 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003914 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003915 linkstatic = True,
3916 deps = [
3917 ":XNNPACK",
3918 "@pthreadpool",
3919 ],
3920)
3921
3922cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003923 name = "fp16_mobilenet_v1",
3924 srcs = ["models/fp16-mobilenet-v1.cc"],
3925 hdrs = ["models/models.h"],
3926 copts = xnnpack_std_cxxopts(),
3927 linkstatic = True,
3928 deps = [
3929 ":XNNPACK",
3930 "@FP16",
3931 "@pthreadpool",
3932 ],
3933)
3934
3935cc_library(
3936 name = "fp32_mobilenet_v2",
3937 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003938 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003939 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003940 linkstatic = True,
3941 deps = [
3942 ":XNNPACK",
3943 "@pthreadpool",
3944 ],
3945)
3946
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003947cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003948 name = "fp16_mobilenet_v2",
3949 srcs = ["models/fp16-mobilenet-v2.cc"],
3950 hdrs = ["models/models.h"],
3951 copts = xnnpack_std_cxxopts(),
3952 linkstatic = True,
3953 deps = [
3954 ":XNNPACK",
3955 "@FP16",
3956 "@pthreadpool",
3957 ],
3958)
3959
3960cc_library(
3961 name = "fp32_mobilenet_v3_large",
3962 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003963 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003964 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003965 linkstatic = True,
3966 deps = [
3967 ":XNNPACK",
3968 "@pthreadpool",
3969 ],
3970)
3971
3972cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003973 name = "fp32_mobilenet_v3_small",
3974 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003975 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003976 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003977 linkstatic = True,
3978 deps = [
3979 ":XNNPACK",
3980 "@pthreadpool",
3981 ],
3982)
3983
Marat Dukhanc068bb62019-10-04 13:24:39 -07003984xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003985 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003986 srcs = [
3987 "bench/f32-dwconv-e2e.cc",
3988 "bench/end2end.h",
3989 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003990 deps = MICROKERNEL_BENCHMARK_DEPS + [
3991 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003992 ":fp32_mobilenet_v1",
3993 ":fp32_mobilenet_v2",
3994 ":fp32_mobilenet_v3_large",
3995 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003996 ],
3997)
3998
3999xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004000 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004001 srcs = [
4002 "bench/f32-gemm-e2e.cc",
4003 "bench/end2end.h",
4004 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004005 deps = MICROKERNEL_BENCHMARK_DEPS + [
4006 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004007 ":fp32_mobilenet_v1",
4008 ":fp32_mobilenet_v2",
4009 ":fp32_mobilenet_v3_large",
4010 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004011 ],
4012)
4013
4014xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004015 name = "end2end_bench",
4016 srcs = ["bench/end2end.cc"],
4017 deps = [
4018 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004019 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004020 ":fp16_mobilenet_v1",
4021 ":fp16_mobilenet_v2",
4022 ":fp32_mobilenet_v1",
4023 ":fp32_mobilenet_v2",
4024 ":fp32_mobilenet_v3_large",
4025 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004026 "@pthreadpool",
4027 ],
4028)
4029
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004030#################### Accuracy evaluation for math functions ####################
4031
4032xnnpack_benchmark(
4033 name = "f32_exp_eval",
4034 srcs = [
4035 "eval/f32-exp.cc",
4036 "src/xnnpack/AlignedAllocator.h",
4037 ] + ACCURACY_EVAL_HDRS,
4038 deps = ACCURACY_EVAL_DEPS,
4039)
4040
Marat Dukhan515c9772019-10-17 18:07:57 -07004041xnnpack_benchmark(
4042 name = "f32_expminus_eval",
4043 srcs = [
4044 "eval/f32-expminus.cc",
4045 "src/xnnpack/AlignedAllocator.h",
4046 ] + ACCURACY_EVAL_HDRS,
4047 deps = ACCURACY_EVAL_DEPS,
4048)
4049
Marat Dukhan98ba4412019-10-23 02:14:28 -07004050xnnpack_benchmark(
4051 name = "f32_extexp_eval",
4052 srcs = [
4053 "eval/f32-extexp.cc",
4054 "src/xnnpack/AlignedAllocator.h",
4055 ] + ACCURACY_EVAL_HDRS,
4056 deps = ACCURACY_EVAL_DEPS,
4057)
4058
Marat Dukhan8853b822020-05-07 12:19:01 -07004059xnnpack_unit_test(
4060 name = "f32_roundne_eval",
4061 srcs = [
4062 "eval/f32-roundne.cc",
4063 "src/xnnpack/AlignedAllocator.h",
4064 "src/xnnpack/math-stubs.h",
4065 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004066 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004067 deps = MICROKERNEL_TEST_DEPS,
4068)
4069
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004070xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004071 name = "f32_roundd_eval",
4072 srcs = [
4073 "eval/f32-roundd.cc",
4074 "src/xnnpack/AlignedAllocator.h",
4075 "src/xnnpack/math-stubs.h",
4076 ] + MICROKERNEL_TEST_HDRS,
4077 automatic = False,
4078 deps = MICROKERNEL_TEST_DEPS,
4079)
4080
4081xnnpack_unit_test(
4082 name = "f32_roundu_eval",
4083 srcs = [
4084 "eval/f32-roundu.cc",
4085 "src/xnnpack/AlignedAllocator.h",
4086 "src/xnnpack/math-stubs.h",
4087 ] + MICROKERNEL_TEST_HDRS,
4088 automatic = False,
4089 deps = MICROKERNEL_TEST_DEPS,
4090)
4091
4092xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004093 name = "f32_roundz_eval",
4094 srcs = [
4095 "eval/f32-roundz.cc",
4096 "src/xnnpack/AlignedAllocator.h",
4097 "src/xnnpack/math-stubs.h",
4098 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004099 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004100 deps = MICROKERNEL_TEST_DEPS,
4101)
4102
Marat Dukhan346a9e52019-11-15 09:06:30 -08004103xnnpack_benchmark(
4104 name = "f32_sigmoid_eval",
4105 srcs = [
4106 "eval/f32-sigmoid.cc",
4107 "src/xnnpack/AlignedAllocator.h",
4108 ] + ACCURACY_EVAL_HDRS,
4109 deps = ACCURACY_EVAL_DEPS,
4110)
4111
Marat Dukhan84000762020-06-29 18:38:43 -07004112xnnpack_benchmark(
4113 name = "f32_sqrt_eval",
4114 srcs = [
4115 "eval/f32-sqrt.cc",
4116 "src/xnnpack/AlignedAllocator.h",
4117 ] + ACCURACY_EVAL_HDRS,
4118 deps = ACCURACY_EVAL_DEPS + [
4119 ":bench_utils",
4120 ],
4121)
4122
Marat Dukhan08c4a432019-10-03 09:29:21 -07004123######################### Unit tests for micro-kernels #########################
4124
4125xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004126 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004128 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 "test/gemm-microkernel-tester.h",
4130 "src/xnnpack/AlignedAllocator.h",
4131 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004132 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004133)
4134
4135xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004136 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004137 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004138 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004139 "test/spmm-microkernel-tester.h",
4140 "src/xnnpack/AlignedAllocator.h",
4141 ] + MICROKERNEL_TEST_HDRS,
4142 deps = MICROKERNEL_TEST_DEPS,
4143)
4144
4145xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004146 name = "f32_argmaxpool_test",
4147 srcs = [
4148 "test/f32-argmaxpool.cc",
4149 "test/argmaxpool-microkernel-tester.h",
4150 "src/xnnpack/AlignedAllocator.h",
4151 ] + MICROKERNEL_TEST_HDRS,
4152 deps = MICROKERNEL_TEST_DEPS,
4153)
4154
4155xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004156 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004157 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004158 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004159 "test/avgpool-microkernel-tester.h",
4160 "src/xnnpack/AlignedAllocator.h",
4161 ] + MICROKERNEL_TEST_HDRS,
4162 deps = MICROKERNEL_TEST_DEPS,
4163)
4164
4165xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004166 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004167 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004168 "test/f32-ibilinear.cc",
4169 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004170 "src/xnnpack/AlignedAllocator.h",
4171 ] + MICROKERNEL_TEST_HDRS,
4172 deps = MICROKERNEL_TEST_DEPS,
4173)
4174
4175xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004176 name = "f16_clamp_test",
4177 srcs = [
4178 "test/f16-clamp.cc",
4179 "test/clamp-microkernel-tester.h",
4180 ] + MICROKERNEL_TEST_HDRS,
4181 deps = MICROKERNEL_TEST_DEPS,
4182)
4183
4184xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004185 name = "f32_clamp_test",
4186 srcs = [
4187 "test/f32-clamp.cc",
4188 "test/clamp-microkernel-tester.h",
4189 ] + MICROKERNEL_TEST_HDRS,
4190 deps = MICROKERNEL_TEST_DEPS,
4191)
4192
4193xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004194 name = "f32_igemm_test",
4195 srcs = [
4196 "test/f32-igemm.cc",
4197 "test/gemm-microkernel-tester.h",
4198 "src/xnnpack/AlignedAllocator.h",
4199 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004200 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004201)
4202
4203xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004204 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004206 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004207 "test/gemm-microkernel-tester.h",
4208 "src/xnnpack/AlignedAllocator.h",
4209 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004210 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004211)
4212
4213xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004214 name = "f32_igemm_minmax_test",
4215 srcs = [
4216 "test/f32-igemm-minmax.cc",
4217 "test/gemm-microkernel-tester.h",
4218 "src/xnnpack/AlignedAllocator.h",
4219 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004220 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004221)
4222
4223xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004224 name = "f16_igemm_minmax_test",
4225 srcs = [
4226 "test/f16-igemm-minmax.cc",
4227 "test/gemm-microkernel-tester.h",
4228 "src/xnnpack/AlignedAllocator.h",
4229 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004230 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004231)
4232
4233xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004234 name = "f32_conv_hwc_test",
4235 srcs = [
4236 "test/f32-conv-hwc.cc",
4237 "test/conv-hwc-microkernel-tester.h",
4238 "src/xnnpack/AlignedAllocator.h",
4239 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004240 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004241)
4242
4243xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004244 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004246 "test/f32-conv-hwc2chw.cc",
4247 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004248 "src/xnnpack/AlignedAllocator.h",
4249 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004250 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251)
4252
4253xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004254 name = "f32_dwconv_test",
4255 srcs = [
4256 "test/f32-dwconv.cc",
4257 "test/dwconv-microkernel-tester.h",
4258 "src/xnnpack/AlignedAllocator.h",
4259 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004260 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004261)
4262
4263xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004264 name = "f16_dwconv_minmax_test",
4265 srcs = [
4266 "test/f16-dwconv-minmax.cc",
4267 "test/dwconv-microkernel-tester.h",
4268 "src/xnnpack/AlignedAllocator.h",
4269 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004270 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004271)
4272
4273xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004274 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004276 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004277 "test/dwconv-microkernel-tester.h",
4278 "src/xnnpack/AlignedAllocator.h",
4279 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004280 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004281)
4282
4283xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004284 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004286 "test/f32-dwconv-chw.cc",
4287 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004288 "src/xnnpack/AlignedAllocator.h",
4289 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004290 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004291)
4292
4293xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004294 name = "f16_gavgpool_minmax_test",
4295 srcs = [
4296 "test/f16-gavgpool-minmax.cc",
4297 "test/gavgpool-microkernel-tester.h",
4298 "src/xnnpack/AlignedAllocator.h",
4299 ] + MICROKERNEL_TEST_HDRS,
4300 deps = MICROKERNEL_TEST_DEPS,
4301)
4302
4303xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004304 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004306 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307 "test/gavgpool-microkernel-tester.h",
4308 "src/xnnpack/AlignedAllocator.h",
4309 ] + MICROKERNEL_TEST_HDRS,
4310 deps = MICROKERNEL_TEST_DEPS,
4311)
4312
4313xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004314 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004315 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004316 "test/f32-gavgpool-cw.cc",
4317 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004318 "src/xnnpack/AlignedAllocator.h",
4319 ] + MICROKERNEL_TEST_HDRS,
4320 deps = MICROKERNEL_TEST_DEPS,
4321)
4322
4323xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004324 name = "f32_gemm_test",
4325 srcs = [
4326 "test/f32-gemm.cc",
4327 "test/gemm-microkernel-tester.h",
4328 "src/xnnpack/AlignedAllocator.h",
4329 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004330 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004331)
4332
4333xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004334 name = "f32_gemm_relu_test",
4335 srcs = [
4336 "test/f32-gemm-relu.cc",
4337 "test/gemm-microkernel-tester.h",
4338 "src/xnnpack/AlignedAllocator.h",
4339 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004340 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004341)
4342
4343xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004344 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004346 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004347 "test/gemm-microkernel-tester.h",
4348 "src/xnnpack/AlignedAllocator.h",
4349 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004350 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004351)
4352
4353xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004354 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004356 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357 "test/gemm-microkernel-tester.h",
4358 "src/xnnpack/AlignedAllocator.h",
4359 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004360 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004361)
4362
4363xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004364 name = "f16_hswish_test",
4365 srcs = [
4366 "test/f16-hswish.cc",
4367 "test/hswish-microkernel-tester.h",
4368 ] + MICROKERNEL_TEST_HDRS,
4369 deps = MICROKERNEL_TEST_DEPS,
4370)
4371
4372xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004373 name = "f32_hswish_test",
4374 srcs = [
4375 "test/f32-hswish.cc",
4376 "test/hswish-microkernel-tester.h",
4377 ] + MICROKERNEL_TEST_HDRS,
4378 deps = MICROKERNEL_TEST_DEPS,
4379)
4380
4381xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004382 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004384 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004385 "test/maxpool-microkernel-tester.h",
4386 ] + MICROKERNEL_TEST_HDRS,
4387 deps = MICROKERNEL_TEST_DEPS,
4388)
4389
4390xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004391 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004392 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004393 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394 "test/avgpool-microkernel-tester.h",
4395 "src/xnnpack/AlignedAllocator.h",
4396 ] + MICROKERNEL_TEST_HDRS,
4397 deps = MICROKERNEL_TEST_DEPS,
4398)
4399
4400xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004401 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004402 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004403 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004404 "test/gemm-microkernel-tester.h",
4405 "src/xnnpack/AlignedAllocator.h",
4406 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004407 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004408)
4409
4410xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004411 name = "f16_prelu_test",
4412 srcs = [
4413 "test/f16-prelu.cc",
4414 "test/prelu-microkernel-tester.h",
4415 "src/xnnpack/AlignedAllocator.h",
4416 ] + MICROKERNEL_TEST_HDRS,
4417 deps = MICROKERNEL_TEST_DEPS,
4418)
4419
4420xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004421 name = "f32_prelu_test",
4422 srcs = [
4423 "test/f32-prelu.cc",
4424 "test/prelu-microkernel-tester.h",
4425 "src/xnnpack/AlignedAllocator.h",
4426 ] + MICROKERNEL_TEST_HDRS,
4427 deps = MICROKERNEL_TEST_DEPS,
4428)
4429
4430xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004431 name = "f32_raddexpminusmax_test",
4432 srcs = [
4433 "test/f32-raddexpminusmax.cc",
4434 "test/raddexpminusmax-microkernel-tester.h",
4435 ] + MICROKERNEL_TEST_HDRS,
4436 deps = MICROKERNEL_TEST_DEPS,
4437)
4438
4439xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004440 name = "f32_raddextexp_test",
4441 srcs = [
4442 "test/f32-raddextexp.cc",
4443 "test/raddextexp-microkernel-tester.h",
4444 ] + MICROKERNEL_TEST_HDRS,
4445 deps = MICROKERNEL_TEST_DEPS,
4446)
4447
4448xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004449 name = "f32_raddstoreexpminusmax_test",
4450 srcs = [
4451 "test/f32-raddstoreexpminusmax.cc",
4452 "test/raddstoreexpminusmax-microkernel-tester.h",
4453 ] + MICROKERNEL_TEST_HDRS,
4454 deps = MICROKERNEL_TEST_DEPS,
4455)
4456
4457xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458 name = "f32_rmax_test",
4459 srcs = [
4460 "test/f32-rmax.cc",
4461 "test/rmax-microkernel-tester.h",
4462 ] + MICROKERNEL_TEST_HDRS,
4463 deps = MICROKERNEL_TEST_DEPS,
4464)
4465
4466xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004467 name = "f32_sigmoid_test",
4468 srcs = [
4469 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004470 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004471 ] + MICROKERNEL_TEST_HDRS,
4472 deps = MICROKERNEL_TEST_DEPS,
4473)
4474
4475xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004476 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004477 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004478 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479 "test/spmm-microkernel-tester.h",
4480 "src/xnnpack/AlignedAllocator.h",
4481 ] + MICROKERNEL_TEST_HDRS,
4482 deps = MICROKERNEL_TEST_DEPS,
4483)
4484
4485xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004486 name = "f16_vadd_minmax_test",
4487 srcs = [
4488 "test/f16-vadd-minmax.cc",
4489 "test/vbinary-microkernel-tester.h",
4490 ] + MICROKERNEL_TEST_HDRS,
4491 deps = MICROKERNEL_TEST_DEPS,
4492)
4493
4494xnnpack_unit_test(
4495 name = "f16_vaddc_minmax_test",
4496 srcs = [
4497 "test/f16-vaddc-minmax.cc",
4498 "test/vbinaryc-microkernel-tester.h",
4499 ] + MICROKERNEL_TEST_HDRS,
4500 deps = MICROKERNEL_TEST_DEPS,
4501)
4502
4503xnnpack_unit_test(
4504 name = "f16_vdiv_minmax_test",
4505 srcs = [
4506 "test/f16-vdiv-minmax.cc",
4507 "test/vbinary-microkernel-tester.h",
4508 ] + MICROKERNEL_TEST_HDRS,
4509 deps = MICROKERNEL_TEST_DEPS,
4510)
4511
4512xnnpack_unit_test(
4513 name = "f16_vdivc_minmax_test",
4514 srcs = [
4515 "test/f16-vdivc-minmax.cc",
4516 "test/vbinaryc-microkernel-tester.h",
4517 ] + MICROKERNEL_TEST_HDRS,
4518 deps = MICROKERNEL_TEST_DEPS,
4519)
4520
4521xnnpack_unit_test(
4522 name = "f16_vrdivc_minmax_test",
4523 srcs = [
4524 "test/f16-vrdivc-minmax.cc",
4525 "test/vbinaryc-microkernel-tester.h",
4526 ] + MICROKERNEL_TEST_HDRS,
4527 deps = MICROKERNEL_TEST_DEPS,
4528)
4529
4530xnnpack_unit_test(
4531 name = "f16_vmax_test",
4532 srcs = [
4533 "test/f16-vmax.cc",
4534 "test/vbinary-microkernel-tester.h",
4535 ] + MICROKERNEL_TEST_HDRS,
4536 deps = MICROKERNEL_TEST_DEPS,
4537)
4538
4539xnnpack_unit_test(
4540 name = "f16_vmaxc_test",
4541 srcs = [
4542 "test/f16-vmaxc.cc",
4543 "test/vbinaryc-microkernel-tester.h",
4544 ] + MICROKERNEL_TEST_HDRS,
4545 deps = MICROKERNEL_TEST_DEPS,
4546)
4547
4548xnnpack_unit_test(
4549 name = "f16_vmin_test",
4550 srcs = [
4551 "test/f16-vmin.cc",
4552 "test/vbinary-microkernel-tester.h",
4553 ] + MICROKERNEL_TEST_HDRS,
4554 deps = MICROKERNEL_TEST_DEPS,
4555)
4556
4557xnnpack_unit_test(
4558 name = "f16_vminc_test",
4559 srcs = [
4560 "test/f16-vminc.cc",
4561 "test/vbinaryc-microkernel-tester.h",
4562 ] + MICROKERNEL_TEST_HDRS,
4563 deps = MICROKERNEL_TEST_DEPS,
4564)
4565
4566xnnpack_unit_test(
4567 name = "f16_vmul_minmax_test",
4568 srcs = [
4569 "test/f16-vmul-minmax.cc",
4570 "test/vbinary-microkernel-tester.h",
4571 ] + MICROKERNEL_TEST_HDRS,
4572 deps = MICROKERNEL_TEST_DEPS,
4573)
4574
4575xnnpack_unit_test(
4576 name = "f16_vmulc_minmax_test",
4577 srcs = [
4578 "test/f16-vmulc-minmax.cc",
4579 "test/vbinaryc-microkernel-tester.h",
4580 ] + MICROKERNEL_TEST_HDRS,
4581 deps = MICROKERNEL_TEST_DEPS,
4582)
4583
4584xnnpack_unit_test(
4585 name = "f16_vsub_minmax_test",
4586 srcs = [
4587 "test/f16-vsub-minmax.cc",
4588 "test/vbinary-microkernel-tester.h",
4589 ] + MICROKERNEL_TEST_HDRS,
4590 deps = MICROKERNEL_TEST_DEPS,
4591)
4592
4593xnnpack_unit_test(
4594 name = "f16_vsubc_minmax_test",
4595 srcs = [
4596 "test/f16-vsubc-minmax.cc",
4597 "test/vbinaryc-microkernel-tester.h",
4598 ] + MICROKERNEL_TEST_HDRS,
4599 deps = MICROKERNEL_TEST_DEPS,
4600)
4601
4602xnnpack_unit_test(
4603 name = "f16_vrsubc_minmax_test",
4604 srcs = [
4605 "test/f16-vrsubc-minmax.cc",
4606 "test/vbinaryc-microkernel-tester.h",
4607 ] + MICROKERNEL_TEST_HDRS,
4608 deps = MICROKERNEL_TEST_DEPS,
4609)
4610
4611xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004612 name = "f32_vabs_test",
4613 srcs = [
4614 "test/f32-vabs.cc",
4615 "test/vunary-microkernel-tester.h",
4616 ] + MICROKERNEL_TEST_HDRS,
4617 deps = MICROKERNEL_TEST_DEPS,
4618)
4619
4620xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004621 name = "f32_vadd_test",
4622 srcs = [
4623 "test/f32-vadd.cc",
4624 "test/vbinary-microkernel-tester.h",
4625 ] + MICROKERNEL_TEST_HDRS,
4626 deps = MICROKERNEL_TEST_DEPS,
4627)
4628
4629xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004630 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004631 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004632 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004633 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004634 ] + MICROKERNEL_TEST_HDRS,
4635 deps = MICROKERNEL_TEST_DEPS,
4636)
4637
4638xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004639 name = "f32_vaddc_test",
4640 srcs = [
4641 "test/f32-vaddc.cc",
4642 "test/vbinaryc-microkernel-tester.h",
4643 ] + MICROKERNEL_TEST_HDRS,
4644 deps = MICROKERNEL_TEST_DEPS,
4645)
4646
4647xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004648 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004649 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004650 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004651 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004652 ] + MICROKERNEL_TEST_HDRS,
4653 deps = MICROKERNEL_TEST_DEPS,
4654)
4655
4656xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004657 name = "f32_vdiv_test",
4658 srcs = [
4659 "test/f32-vdiv.cc",
4660 "test/vbinary-microkernel-tester.h",
4661 ] + MICROKERNEL_TEST_HDRS,
4662 deps = MICROKERNEL_TEST_DEPS,
4663)
4664
4665xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004666 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004667 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004668 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004669 "test/vbinary-microkernel-tester.h",
4670 ] + MICROKERNEL_TEST_HDRS,
4671 deps = MICROKERNEL_TEST_DEPS,
4672)
4673
4674xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004675 name = "f32_vdivc_test",
4676 srcs = [
4677 "test/f32-vdivc.cc",
4678 "test/vbinaryc-microkernel-tester.h",
4679 ] + MICROKERNEL_TEST_HDRS,
4680 deps = MICROKERNEL_TEST_DEPS,
4681)
4682
4683xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004684 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004685 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004686 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004687 "test/vbinaryc-microkernel-tester.h",
4688 ] + MICROKERNEL_TEST_HDRS,
4689 deps = MICROKERNEL_TEST_DEPS,
4690)
4691
4692xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004693 name = "f32_vrdivc_test",
4694 srcs = [
4695 "test/f32-vrdivc.cc",
4696 "test/vbinaryc-microkernel-tester.h",
4697 ] + MICROKERNEL_TEST_HDRS,
4698 deps = MICROKERNEL_TEST_DEPS,
4699)
4700
4701xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004702 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004703 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004704 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004705 "test/vbinaryc-microkernel-tester.h",
4706 ] + MICROKERNEL_TEST_HDRS,
4707 deps = MICROKERNEL_TEST_DEPS,
4708)
4709
4710xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004711 name = "f32_vmax_test",
4712 srcs = [
4713 "test/f32-vmax.cc",
4714 "test/vbinary-microkernel-tester.h",
4715 ] + MICROKERNEL_TEST_HDRS,
4716 deps = MICROKERNEL_TEST_DEPS,
4717)
4718
4719xnnpack_unit_test(
4720 name = "f32_vmaxc_test",
4721 srcs = [
4722 "test/f32-vmaxc.cc",
4723 "test/vbinaryc-microkernel-tester.h",
4724 ] + MICROKERNEL_TEST_HDRS,
4725 deps = MICROKERNEL_TEST_DEPS,
4726)
4727
4728xnnpack_unit_test(
4729 name = "f32_vmin_test",
4730 srcs = [
4731 "test/f32-vmin.cc",
4732 "test/vbinary-microkernel-tester.h",
4733 ] + MICROKERNEL_TEST_HDRS,
4734 deps = MICROKERNEL_TEST_DEPS,
4735)
4736
4737xnnpack_unit_test(
4738 name = "f32_vminc_test",
4739 srcs = [
4740 "test/f32-vminc.cc",
4741 "test/vbinaryc-microkernel-tester.h",
4742 ] + MICROKERNEL_TEST_HDRS,
4743 deps = MICROKERNEL_TEST_DEPS,
4744)
4745
4746xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004747 name = "f32_vmul_test",
4748 srcs = [
4749 "test/f32-vmul.cc",
4750 "test/vbinary-microkernel-tester.h",
4751 ] + MICROKERNEL_TEST_HDRS,
4752 deps = MICROKERNEL_TEST_DEPS,
4753)
4754
4755xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004756 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004757 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004758 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004759 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004760 ] + MICROKERNEL_TEST_HDRS,
4761 deps = MICROKERNEL_TEST_DEPS,
4762)
4763
4764xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004765 name = "f32_vmulc_test",
4766 srcs = [
4767 "test/f32-vmulc.cc",
4768 "test/vbinaryc-microkernel-tester.h",
4769 ] + MICROKERNEL_TEST_HDRS,
4770 deps = MICROKERNEL_TEST_DEPS,
4771)
4772
4773xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004774 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004775 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004776 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004777 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004778 ] + MICROKERNEL_TEST_HDRS,
4779 deps = MICROKERNEL_TEST_DEPS,
4780)
4781
4782xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004783 name = "f16_vmulcaddc_minmax_test",
4784 srcs = [
4785 "test/f16-vmulcaddc-minmax.cc",
4786 "test/vmulcaddc-microkernel-tester.h",
4787 "src/xnnpack/AlignedAllocator.h",
4788 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004789 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004790)
4791
4792xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004793 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004794 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004795 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004796 "test/vmulcaddc-microkernel-tester.h",
4797 "src/xnnpack/AlignedAllocator.h",
4798 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004799 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004800)
4801
4802xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004803 name = "f32_vlrelu_test",
4804 srcs = [
4805 "test/f32-vlrelu.cc",
4806 "test/vunary-microkernel-tester.h",
4807 ] + MICROKERNEL_TEST_HDRS,
4808 deps = MICROKERNEL_TEST_DEPS,
4809)
4810
4811xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004812 name = "f32_vneg_test",
4813 srcs = [
4814 "test/f32-vneg.cc",
4815 "test/vunary-microkernel-tester.h",
4816 ] + MICROKERNEL_TEST_HDRS,
4817 deps = MICROKERNEL_TEST_DEPS,
4818)
4819
4820xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004821 name = "f32_vrndne_test",
4822 srcs = [
4823 "test/f32-vrndne.cc",
4824 "test/vunary-microkernel-tester.h",
4825 ] + MICROKERNEL_TEST_HDRS,
4826 deps = MICROKERNEL_TEST_DEPS,
4827)
4828
4829xnnpack_unit_test(
4830 name = "f32_vrndz_test",
4831 srcs = [
4832 "test/f32-vrndz.cc",
4833 "test/vunary-microkernel-tester.h",
4834 ] + MICROKERNEL_TEST_HDRS,
4835 deps = MICROKERNEL_TEST_DEPS,
4836)
4837
4838xnnpack_unit_test(
4839 name = "f32_vrndu_test",
4840 srcs = [
4841 "test/f32-vrndu.cc",
4842 "test/vunary-microkernel-tester.h",
4843 ] + MICROKERNEL_TEST_HDRS,
4844 deps = MICROKERNEL_TEST_DEPS,
4845)
4846
4847xnnpack_unit_test(
4848 name = "f32_vrndd_test",
4849 srcs = [
4850 "test/f32-vrndd.cc",
4851 "test/vunary-microkernel-tester.h",
4852 ] + MICROKERNEL_TEST_HDRS,
4853 deps = MICROKERNEL_TEST_DEPS,
4854)
4855
4856xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004857 name = "f32_vscale_test",
4858 srcs = [
4859 "test/f32-vscale.cc",
4860 "test/vscale-microkernel-tester.h",
4861 ] + MICROKERNEL_TEST_HDRS,
4862 deps = MICROKERNEL_TEST_DEPS,
4863)
4864
4865xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004866 name = "f32_vscaleexpminusmax_test",
4867 srcs = [
4868 "test/f32-vscaleexpminusmax.cc",
4869 "test/vscaleexpminusmax-microkernel-tester.h",
4870 ] + MICROKERNEL_TEST_HDRS,
4871 deps = MICROKERNEL_TEST_DEPS,
4872)
4873
4874xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004875 name = "f32_vscaleextexp_test",
4876 srcs = [
4877 "test/f32-vscaleextexp.cc",
4878 "test/vscaleextexp-microkernel-tester.h",
4879 ] + MICROKERNEL_TEST_HDRS,
4880 deps = MICROKERNEL_TEST_DEPS,
4881)
4882
4883xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004884 name = "f32_vsqr_test",
4885 srcs = [
4886 "test/f32-vsqr.cc",
4887 "test/vunary-microkernel-tester.h",
4888 ] + MICROKERNEL_TEST_HDRS,
4889 deps = MICROKERNEL_TEST_DEPS,
4890)
4891
4892xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004893 name = "f32_vsqrdiff_test",
4894 srcs = [
4895 "test/f32-vsqrdiff.cc",
4896 "test/vbinary-microkernel-tester.h",
4897 ] + MICROKERNEL_TEST_HDRS,
4898 deps = MICROKERNEL_TEST_DEPS,
4899)
4900
4901xnnpack_unit_test(
4902 name = "f32_vsqrdiffc_test",
4903 srcs = [
4904 "test/f32-vsqrdiffc.cc",
4905 "test/vbinaryc-microkernel-tester.h",
4906 ] + MICROKERNEL_TEST_HDRS,
4907 deps = MICROKERNEL_TEST_DEPS,
4908)
4909
4910xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004911 name = "f32_vsqrt_test",
4912 srcs = [
4913 "test/f32-vsqrt.cc",
4914 "test/vunary-microkernel-tester.h",
4915 ] + MICROKERNEL_TEST_HDRS,
4916 deps = MICROKERNEL_TEST_DEPS,
4917)
4918
4919xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004920 name = "f32_vsub_test",
4921 srcs = [
4922 "test/f32-vsub.cc",
4923 "test/vbinary-microkernel-tester.h",
4924 ] + MICROKERNEL_TEST_HDRS,
4925 deps = MICROKERNEL_TEST_DEPS,
4926)
4927
4928xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004929 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004930 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004931 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004932 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004933 ] + MICROKERNEL_TEST_HDRS,
4934 deps = MICROKERNEL_TEST_DEPS,
4935)
4936
4937xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004938 name = "f32_vsubc_test",
4939 srcs = [
4940 "test/f32-vsubc.cc",
4941 "test/vbinaryc-microkernel-tester.h",
4942 ] + MICROKERNEL_TEST_HDRS,
4943 deps = MICROKERNEL_TEST_DEPS,
4944)
4945
4946xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004947 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004948 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004949 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004950 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004951 ] + MICROKERNEL_TEST_HDRS,
4952 deps = MICROKERNEL_TEST_DEPS,
4953)
4954
4955xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004956 name = "f32_vrsubc_test",
4957 srcs = [
4958 "test/f32-vrsubc.cc",
4959 "test/vbinaryc-microkernel-tester.h",
4960 ] + MICROKERNEL_TEST_HDRS,
4961 deps = MICROKERNEL_TEST_DEPS,
4962)
4963
4964xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004965 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004966 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004967 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004968 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004969 ] + MICROKERNEL_TEST_HDRS,
4970 deps = MICROKERNEL_TEST_DEPS,
4971)
4972
4973xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004974 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004975 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004976 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004977 "test/avgpool-microkernel-tester.h",
4978 "src/xnnpack/AlignedAllocator.h",
4979 ] + MICROKERNEL_TEST_HDRS,
4980 deps = MICROKERNEL_TEST_DEPS,
4981)
4982
4983xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004984 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004985 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004986 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987 "test/gemm-microkernel-tester.h",
4988 "src/xnnpack/AlignedAllocator.h",
4989 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004990 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004991)
4992
4993xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004994 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004995 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004996 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004997 "test/dwconv-microkernel-tester.h",
4998 "src/xnnpack/AlignedAllocator.h",
4999 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005000 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005001)
5002
5003xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005004 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005006 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005007 "test/gavgpool-microkernel-tester.h",
5008 "src/xnnpack/AlignedAllocator.h",
5009 ] + MICROKERNEL_TEST_HDRS,
5010 deps = MICROKERNEL_TEST_DEPS,
5011)
5012
5013xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005014 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005016 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005017 "test/gemm-microkernel-tester.h",
5018 "src/xnnpack/AlignedAllocator.h",
5019 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005020 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005021)
5022
5023xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005024 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005025 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005026 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005027 "test/vadd-microkernel-tester.h",
5028 ] + MICROKERNEL_TEST_HDRS,
5029 deps = MICROKERNEL_TEST_DEPS,
5030)
5031
5032xnnpack_unit_test(
5033 name = "u8_clamp_test",
5034 srcs = [
5035 "test/u8-clamp.cc",
5036 "test/clamp-microkernel-tester.h",
5037 ] + MICROKERNEL_TEST_HDRS,
5038 deps = MICROKERNEL_TEST_DEPS,
5039)
5040
5041xnnpack_unit_test(
5042 name = "u8_lut32norm_test",
5043 srcs = [
5044 "test/u8-lut32norm.cc",
5045 "test/lut-norm-microkernel-tester.h",
5046 ] + MICROKERNEL_TEST_HDRS,
5047 deps = MICROKERNEL_TEST_DEPS,
5048)
5049
5050xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005051 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005052 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005053 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054 "test/maxpool-microkernel-tester.h",
5055 ] + MICROKERNEL_TEST_HDRS,
5056 deps = MICROKERNEL_TEST_DEPS,
5057)
5058
5059xnnpack_unit_test(
5060 name = "u8_rmax_test",
5061 srcs = [
5062 "test/u8-rmax.cc",
5063 "test/rmax-microkernel-tester.h",
5064 ] + MICROKERNEL_TEST_HDRS,
5065 deps = MICROKERNEL_TEST_DEPS,
5066)
5067
5068xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005069 name = "x32_fill_test",
5070 srcs = [
5071 "test/x32-fill.cc",
5072 "test/fill-microkernel-tester.h",
5073 ] + MICROKERNEL_TEST_HDRS,
5074 deps = MICROKERNEL_TEST_DEPS,
5075)
5076
5077xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005078 name = "x32_packx_test",
5079 srcs = [
5080 "test/x32-packx.cc",
5081 "test/pack-microkernel-tester.h",
5082 "src/xnnpack/AlignedAllocator.h",
5083 ] + MICROKERNEL_TEST_HDRS,
5084 deps = MICROKERNEL_TEST_DEPS,
5085)
5086
5087xnnpack_unit_test(
5088 name = "x32_pad_test",
5089 srcs = [
5090 "test/x32-pad.cc",
5091 "test/pad-microkernel-tester.h",
5092 ] + MICROKERNEL_TEST_HDRS,
5093 deps = MICROKERNEL_TEST_DEPS,
5094)
5095
5096xnnpack_unit_test(
5097 name = "x32_unpool_test",
5098 srcs = [
5099 "test/x32-unpool.cc",
5100 "test/unpool-microkernel-tester.h",
5101 ] + MICROKERNEL_TEST_HDRS,
5102 deps = MICROKERNEL_TEST_DEPS,
5103)
5104
5105xnnpack_unit_test(
5106 name = "x32_zip_test",
5107 srcs = [
5108 "test/x32-zip.cc",
5109 "test/zip-microkernel-tester.h",
5110 ] + MICROKERNEL_TEST_HDRS,
5111 deps = MICROKERNEL_TEST_DEPS,
5112)
5113
5114xnnpack_unit_test(
5115 name = "x8_lut_test",
5116 srcs = [
5117 "test/x8-lut.cc",
5118 "test/lut-microkernel-tester.h",
5119 ] + MICROKERNEL_TEST_HDRS,
5120 deps = MICROKERNEL_TEST_DEPS,
5121)
5122
5123xnnpack_unit_test(
5124 name = "x8_zip_test",
5125 srcs = [
5126 "test/x8-zip.cc",
5127 "test/zip-microkernel-tester.h",
5128 ] + MICROKERNEL_TEST_HDRS,
5129 deps = MICROKERNEL_TEST_DEPS,
5130)
5131
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005132xnnpack_unit_test(
5133 name = "requantization_test",
5134 srcs = [
5135 "src/xnnpack/requantization-stubs.h",
5136 "test/requantization.cc",
5137 "test/requantization-tester.h",
5138 ] + MICROKERNEL_TEST_HDRS,
5139 deps = MICROKERNEL_TEST_DEPS,
5140)
5141
Marat Dukhan20c3b922020-03-10 03:45:06 -07005142########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005143
5144xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005145 name = "operator_size_test",
5146 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005147 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005148)
5149
Marat Dukhan20c3b922020-03-10 03:45:06 -07005150xnnpack_binary(
5151 name = "subgraph_size_test",
5152 srcs = ["test/subgraph-size.c"],
5153 deps = [":XNNPACK"],
5154)
5155
5156########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005157
5158xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005159 name = "abs_nc_test",
5160 srcs = [
5161 "test/abs-nc.cc",
5162 "test/abs-operator-tester.h",
5163 ],
5164 deps = OPERATOR_TEST_DEPS,
5165)
5166
5167xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005168 name = "add_nd_test",
5169 srcs = [
5170 "test/add-nd.cc",
5171 "test/binary-elementwise-operator-tester.h",
5172 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005173 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005174)
5175
5176xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005177 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005178 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005179 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005180 "test/argmax-pooling-operator-tester.h",
5181 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005182 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183)
5184
5185xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005186 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005188 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005189 "test/average-pooling-operator-tester.h",
5190 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005191 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005192)
5193
5194xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005195 name = "bankers_rounding_nc_test",
5196 srcs = [
5197 "test/bankers-rounding-nc.cc",
5198 "test/bankers-rounding-operator-tester.h",
5199 ],
5200 deps = OPERATOR_TEST_DEPS,
5201)
5202
5203xnnpack_unit_test(
5204 name = "ceiling_nc_test",
5205 srcs = [
5206 "test/ceiling-nc.cc",
5207 "test/ceiling-operator-tester.h",
5208 ],
5209 deps = OPERATOR_TEST_DEPS,
5210)
5211
5212xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005213 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005214 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005215 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005216 "test/channel-shuffle-operator-tester.h",
5217 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005218 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219)
5220
5221xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005222 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005224 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005225 "test/clamp-operator-tester.h",
5226 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005227 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005228)
5229
5230xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005231 name = "constant_pad_nd_test",
5232 srcs = [
5233 "test/constant-pad-nd.cc",
5234 "test/constant-pad-operator-tester.h",
5235 ],
5236 deps = OPERATOR_TEST_DEPS,
5237)
5238
5239xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005240 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005241 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005242 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005243 "test/convolution-operator-tester.h",
5244 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005245 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246)
5247
5248xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005249 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005250 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005251 "test/convolution-nchw.cc",
5252 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005254 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005255)
5256
5257xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005258 name = "copy_nc_test",
5259 srcs = [
5260 "test/copy-nc.cc",
5261 "test/copy-operator-tester.h",
5262 ],
5263 deps = OPERATOR_TEST_DEPS,
5264)
5265
5266xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005267 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005268 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005269 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005270 "test/deconvolution-operator-tester.h",
5271 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005272 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005273)
5274
5275xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005276 name = "divide_nd_test",
5277 srcs = [
5278 "test/binary-elementwise-operator-tester.h",
5279 "test/divide-nd.cc",
5280 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005281 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005282)
5283
5284xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005285 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005286 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005287 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288 "test/fully-connected-operator-tester.h",
5289 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005290 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291)
5292
5293xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005294 name = "floor_nc_test",
5295 srcs = [
5296 "test/floor-nc.cc",
5297 "test/floor-operator-tester.h",
5298 ],
5299 deps = OPERATOR_TEST_DEPS,
5300)
5301
5302xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005303 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005304 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005305 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005307 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005308 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005309)
5310
5311xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005312 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005313 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005314 "test/global-average-pooling-ncw.cc",
5315 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005317 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318)
5319
5320xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005321 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005323 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005324 "test/hardswish-operator-tester.h",
5325 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005326 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327)
5328
5329xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005330 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005331 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005332 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333 "test/leaky-relu-operator-tester.h",
5334 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005335 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336)
5337
5338xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005339 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005340 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005341 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005342 "test/max-pooling-operator-tester.h",
5343 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005344 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345)
5346
5347xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005348 name = "maximum_nd_test",
5349 srcs = [
5350 "test/binary-elementwise-operator-tester.h",
5351 "test/maximum-nd.cc",
5352 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005353 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005354)
5355
5356xnnpack_unit_test(
5357 name = "minimum_nd_test",
5358 srcs = [
5359 "test/binary-elementwise-operator-tester.h",
5360 "test/minimum-nd.cc",
5361 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005362 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005363)
5364
5365xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005366 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005367 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005368 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005369 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005370 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005371 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005372)
5373
5374xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005375 name = "negate_nc_test",
5376 srcs = [
5377 "test/negate-nc.cc",
5378 "test/negate-operator-tester.h",
5379 ],
5380 deps = OPERATOR_TEST_DEPS,
5381)
5382
5383xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005384 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005385 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005386 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005387 "test/prelu-operator-tester.h",
5388 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005389 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005390)
5391
5392xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005393 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005394 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005395 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005396 "test/resize-bilinear-operator-tester.h",
5397 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005398 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005399)
5400
5401xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005402 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005404 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005405 "test/sigmoid-operator-tester.h",
5406 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005407 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005408)
5409
5410xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005411 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005413 "test/softmax-nc.cc",
5414 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005416 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417)
5418
5419xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005420 name = "square_nc_test",
5421 srcs = [
5422 "test/square-nc.cc",
5423 "test/square-operator-tester.h",
5424 ],
5425 deps = OPERATOR_TEST_DEPS,
5426)
5427
5428xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005429 name = "square_root_nc_test",
5430 srcs = [
5431 "test/square-root-nc.cc",
5432 "test/square-root-operator-tester.h",
5433 ],
5434 deps = OPERATOR_TEST_DEPS,
5435)
5436
5437xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005438 name = "squared_difference_nd_test",
5439 srcs = [
5440 "test/binary-elementwise-operator-tester.h",
5441 "test/squared-difference-nd.cc",
5442 ],
5443 deps = OPERATOR_TEST_DEPS,
5444)
5445
5446xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005447 name = "subtract_nd_test",
5448 srcs = [
5449 "test/binary-elementwise-operator-tester.h",
5450 "test/subtract-nd.cc",
5451 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005452 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005453)
5454
5455xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005456 name = "truncation_nc_test",
5457 srcs = [
5458 "test/truncation-nc.cc",
5459 "test/truncation-operator-tester.h",
5460 ],
5461 deps = OPERATOR_TEST_DEPS,
5462)
5463
5464xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005465 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005466 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005467 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005468 "test/unpooling-operator-tester.h",
5469 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005470 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005471)
5472
Chao Mei6ddfc602020-05-13 22:29:36 -07005473############################### Misc unit tests ###############################
5474
5475xnnpack_unit_test(
5476 name = "memory_planner_test",
5477 srcs = [
5478 "test/memory-planner-test.cc",
5479 ],
5480 deps = [
5481 ":XNNPACK",
5482 ":memory_planner",
5483 ],
5484)
5485
Marat Dukhan08c4a432019-10-03 09:29:21 -07005486############################# Build configurations #############################
5487
Marat Dukhanb8642352019-10-30 15:43:02 -07005488# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005489config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005490 name = "xnn_enable_assembly_explicit_true",
5491 define_values = {"xnn_enable_assembly": "true"},
5492)
5493
5494# Disables usage of assembly kernels.
5495config_setting(
5496 name = "xnn_enable_assembly_explicit_false",
5497 define_values = {"xnn_enable_assembly": "false"},
5498)
5499
Marat Dukhan9de90e02020-06-18 16:04:12 -07005500# Enables usage of sparse inference.
5501config_setting(
5502 name = "xnn_enable_sparse_explicit_true",
5503 define_values = {"xnn_enable_sparse": "true"},
5504)
5505
5506# Disables usage of sparse inference.
5507config_setting(
5508 name = "xnn_enable_sparse_explicit_false",
5509 define_values = {"xnn_enable_sparse": "false"},
5510)
5511
Marat Dukhan05702cf2020-03-26 15:41:33 -07005512# Disables usage of HMP-aware optimizations.
5513config_setting(
5514 name = "xnn_enable_hmp_explicit_false",
5515 define_values = {"xnn_enable_hmp": "false"},
5516)
5517
Chao Mei6ddfc602020-05-13 22:29:36 -07005518# Enable usage of optimized memory allocation
5519config_setting(
5520 name = "xnn_enable_memopt_explicit_true",
5521 define_values = {"xnnpack_enable_memopt": "true"},
5522)
5523
5524# Disable usage of optimized memory allocation
5525config_setting(
5526 name = "xnn_enable_memopt_explicit_false",
5527 define_values = {"xnnpack_enable_memopt": "false"},
5528)
5529
Marat Dukhanb8642352019-10-30 15:43:02 -07005530# Builds with -c dbg
5531config_setting(
5532 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005533 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005534 "compilation_mode": "dbg",
5535 },
5536)
5537
5538# Builds with -c opt
5539config_setting(
5540 name = "optimized_build",
5541 values = {
5542 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005543 },
5544)
5545
5546config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005547 name = "linux_k8",
5548 values = {"cpu": "k8"},
5549)
5550
5551config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005552 name = "linux_arm",
5553 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005554)
5555
5556config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005557 name = "linux_armeabi",
5558 values = {"cpu": "armeabi"},
5559)
5560
5561config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005562 name = "linux_armhf",
5563 values = {"cpu": "armhf"},
5564)
5565
5566config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005567 name = "linux_armv7a",
5568 values = {"cpu": "armv7a"},
5569)
5570
5571config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005572 name = "linux_aarch64",
5573 values = {"cpu": "aarch64"},
5574)
5575
5576config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005577 name = "android",
5578 values = {"crosstool_top": "//external:android/crosstool"},
5579)
5580
5581config_setting(
5582 name = "android_armv7",
5583 values = {
5584 "crosstool_top": "//external:android/crosstool",
5585 "cpu": "armeabi-v7a",
5586 },
5587)
5588
5589config_setting(
5590 name = "android_arm64",
5591 values = {
5592 "crosstool_top": "//external:android/crosstool",
5593 "cpu": "arm64-v8a",
5594 },
5595)
5596
5597config_setting(
5598 name = "android_x86",
5599 values = {
5600 "crosstool_top": "//external:android/crosstool",
5601 "cpu": "x86",
5602 },
5603)
5604
5605config_setting(
5606 name = "android_x86_64",
5607 values = {
5608 "crosstool_top": "//external:android/crosstool",
5609 "cpu": "x86_64",
5610 },
5611)
5612
5613config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005614 name = "windows_x86_64",
5615 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005616)
5617
5618config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005619 name = "windows_x86_64_clang",
5620 values = {
5621 "compiler": "clang-cl",
5622 "cpu": "x64_windows",
5623 },
5624)
5625
5626config_setting(
5627 name = "windows_x86_64_mingw",
5628 values = {
5629 "compiler": "mingw-gcc",
5630 "cpu": "x64_windows",
5631 },
5632)
5633
5634config_setting(
5635 name = "windows_x86_64_msys",
5636 values = {
5637 "compiler": "msys-gcc",
5638 "cpu": "x64_windows",
5639 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005640)
5641
5642config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005643 name = "macos_x86_64",
5644 values = {
5645 "apple_platform_type": "macos",
5646 "cpu": "darwin",
5647 },
5648)
5649
5650config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005651 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005652 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005653)
5654
5655config_setting(
5656 name = "emscripten_wasm",
5657 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005658 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005659 "cpu": "wasm",
5660 },
5661)
5662
5663config_setting(
5664 name = "emscripten_wasmsimd",
5665 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005666 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005667 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005668 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005669 },
5670)
5671
5672config_setting(
5673 name = "emscripten_asmjs",
5674 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005675 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005676 "cpu": "asmjs",
5677 },
5678)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005679
5680config_setting(
5681 name = "ios_armv7",
5682 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005683 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005684 "cpu": "ios_armv7",
5685 },
5686)
5687
5688config_setting(
5689 name = "ios_arm64",
5690 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005691 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005692 "cpu": "ios_arm64",
5693 },
5694)
5695
5696config_setting(
5697 name = "ios_arm64e",
5698 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005699 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005700 "cpu": "ios_arm64e",
5701 },
5702)
5703
5704config_setting(
5705 name = "ios_x86",
5706 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005707 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005708 "cpu": "ios_i386",
5709 },
5710)
5711
5712config_setting(
5713 name = "ios_x86_64",
5714 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005715 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005716 "cpu": "ios_x86_64",
5717 },
5718)
5719
5720config_setting(
5721 name = "watchos_armv7k",
5722 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005723 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005724 "cpu": "watchos_armv7k",
5725 },
5726)
5727
5728config_setting(
5729 name = "watchos_arm64_32",
5730 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005731 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005732 "cpu": "watchos_arm64_32",
5733 },
5734)
5735
5736config_setting(
5737 name = "watchos_x86",
5738 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005739 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005740 "cpu": "watchos_i386",
5741 },
5742)
5743
5744config_setting(
5745 name = "watchos_x86_64",
5746 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005747 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005748 "cpu": "watchos_x86_64",
5749 },
5750)
5751
5752config_setting(
5753 name = "tvos_arm64",
5754 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005755 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005756 "cpu": "tvos_arm64",
5757 },
5758)
5759
5760config_setting(
5761 name = "tvos_x86_64",
5762 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005763 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005764 "cpu": "tvos_x86_64",
5765 },
5766)