blob: 39b1d4c6896b264fdf75f5e7a7b24ad40a8d1f3e [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 Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700102 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700103 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700104 "src/subgraph/squared-difference.c",
105 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700106 "src/subgraph/unpooling-2d.c",
107]
108
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800109TABLE_SRCS = [
110 "src/tables/exp2-k-over-64.c",
111 "src/tables/exp2-k-over-2048.c",
112]
113
Marat Dukhan08c4a432019-10-03 09:29:21 -0700114SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800115 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800116 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800117 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700118 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
119 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700120 "src/f32-clamp/gen/scalar-x1.c",
121 "src/f32-clamp/gen/scalar-x2.c",
122 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800123 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700125 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
126 "src/f32-dwconv-chw/3x3p1-scalar.c",
127 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
128 "src/f32-dwconv-chw/5x5p2-scalar.c",
129 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700130 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
131 "src/f32-dwconv/gen/up1x4-scalar.c",
132 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
133 "src/f32-dwconv/gen/up1x9-scalar.c",
134 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
135 "src/f32-dwconv/gen/up1x25-scalar.c",
136 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
137 "src/f32-dwconv/gen/up2x4-scalar.c",
138 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
139 "src/f32-dwconv/gen/up2x9-scalar.c",
140 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
141 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700142 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
144 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700148 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
150 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700152 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700154 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700155 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
156 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700157 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
158 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700160 "src/f32-gemm/gen/1x4-scalar.c",
161 "src/f32-gemm/gen/2x4-scalar.c",
162 "src/f32-gemm/gen/4x2-scalar.c",
163 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700164 "src/f32-gemm/gen/1x4-relu-scalar.c",
165 "src/f32-gemm/gen/2x4-relu-scalar.c",
166 "src/f32-gemm/gen/4x2-relu-scalar.c",
167 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700168 "src/f32-gemm/gen/1x4-minmax-scalar.c",
169 "src/f32-gemm/gen/2x4-minmax-scalar.c",
170 "src/f32-gemm/gen/4x2-minmax-scalar.c",
171 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800172 "src/f32-hswish/gen/scalar-x1.c",
173 "src/f32-hswish/gen/scalar-x2.c",
174 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700175 "src/f32-ibilinear/gen/scalar-c1.c",
176 "src/f32-ibilinear/gen/scalar-c2.c",
177 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700178 "src/f32-igemm/gen/1x4-scalar.c",
179 "src/f32-igemm/gen/2x4-scalar.c",
180 "src/f32-igemm/gen/4x2-scalar.c",
181 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700182 "src/f32-igemm/gen/1x4-relu-scalar.c",
183 "src/f32-igemm/gen/2x4-relu-scalar.c",
184 "src/f32-igemm/gen/4x2-relu-scalar.c",
185 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700186 "src/f32-igemm/gen/1x4-minmax-scalar.c",
187 "src/f32-igemm/gen/2x4-minmax-scalar.c",
188 "src/f32-igemm/gen/4x2-minmax-scalar.c",
189 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700190 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
191 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700193 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
194 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
195 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800197 "src/f32-prelu/gen/scalar-2x1.c",
198 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800199 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700211 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800212 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
213 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
214 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
215 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
216 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
217 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
218 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
219 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
220 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700221 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
222 "src/f32-spmm/gen/1x1-minmax-scalar.c",
223 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
224 "src/f32-spmm/gen/2x1-minmax-scalar.c",
225 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/4x1-minmax-scalar.c",
227 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
228 "src/f32-spmm/gen/8x1-minmax-scalar.c",
229 "src/f32-spmm/gen/8x2-minmax-scalar.c",
230 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700231 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
232 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
233 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700234 "src/f32-vbinary/gen/vadd-scalar-x1.c",
235 "src/f32-vbinary/gen/vadd-scalar-x2.c",
236 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700237 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
238 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
239 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700240 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
241 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
242 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700243 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
244 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
245 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700246 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
247 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
248 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700249 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
250 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
251 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700252 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
253 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
254 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800255 "src/f32-vbinary/gen/vmax-scalar-x1.c",
256 "src/f32-vbinary/gen/vmax-scalar-x2.c",
257 "src/f32-vbinary/gen/vmax-scalar-x4.c",
258 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
259 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
260 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
261 "src/f32-vbinary/gen/vmin-scalar-x1.c",
262 "src/f32-vbinary/gen/vmin-scalar-x2.c",
263 "src/f32-vbinary/gen/vmin-scalar-x4.c",
264 "src/f32-vbinary/gen/vminc-scalar-x1.c",
265 "src/f32-vbinary/gen/vminc-scalar-x2.c",
266 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700267 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
268 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
269 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700270 "src/f32-vbinary/gen/vmul-scalar-x1.c",
271 "src/f32-vbinary/gen/vmul-scalar-x2.c",
272 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700273 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
274 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
275 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700276 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
277 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
278 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700279 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
280 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
281 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700282 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
283 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
284 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700285 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
286 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
287 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700288 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
289 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
290 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700291 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
292 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
293 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
294 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
295 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
296 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700297 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
298 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
299 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700300 "src/f32-vbinary/gen/vsub-scalar-x1.c",
301 "src/f32-vbinary/gen/vsub-scalar-x2.c",
302 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700306 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
307 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
308 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700309 "src/f32-vlrelu/gen/scalar-x1.c",
310 "src/f32-vlrelu/gen/scalar-x2.c",
311 "src/f32-vlrelu/gen/scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700312 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
313 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
314 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700315 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
316 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
317 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
318 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
319 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
320 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
321 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
322 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
323 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
324 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
325 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
326 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700327 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
328 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
329 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700330 "src/f32-vunary/gen/vabs-scalar-x1.c",
331 "src/f32-vunary/gen/vabs-scalar-x2.c",
332 "src/f32-vunary/gen/vabs-scalar-x4.c",
333 "src/f32-vunary/gen/vneg-scalar-x1.c",
334 "src/f32-vunary/gen/vneg-scalar-x2.c",
335 "src/f32-vunary/gen/vneg-scalar-x4.c",
336 "src/f32-vunary/gen/vsqr-scalar-x1.c",
337 "src/f32-vunary/gen/vsqr-scalar-x2.c",
338 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800339 "src/math/expminus-scalar-lut2048-p1.c",
340 "src/math/expminus-scalar-lut64-p2.c",
341 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700342 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700343 "src/math/roundne-scalar-nearbyint.c",
344 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700345 "src/math/roundd-scalar-addsub.c",
346 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700348 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700349 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700350 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700351 "src/math/roundz-scalar-addsub.c",
352 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800354 "src/math/sigmoid-scalar-lut2048-p1-div.c",
355 "src/math/sigmoid-scalar-lut64-p2-div.c",
356 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700357 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
358 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
359 "src/qu8-dwconv/up1x9-minmax-scalar.c",
360 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
361 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
362 "src/qu8-gemm/2x2-minmax-scalar.c",
363 "src/qu8-igemm/2x2-minmax-scalar.c",
364 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700365 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700366 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700367 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700368 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700369 "src/x32-fill/scalar-float.c",
370 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700371 "src/x32-packx/x2-scalar.c",
372 "src/x32-packx/x3-scalar.c",
373 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700374 "src/x32-pad/scalar-int.c",
375 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700376 "src/x32-unpool/scalar.c",
377 "src/x32-zip/x2-scalar.c",
378 "src/x32-zip/x3-scalar.c",
379 "src/x32-zip/x4-scalar.c",
380 "src/x32-zip/xm-scalar.c",
381 "src/x8-lut/scalar.c",
382 "src/x8-zip/x2-scalar.c",
383 "src/x8-zip/x3-scalar.c",
384 "src/x8-zip/x4-scalar.c",
385 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700386 "src/requantization/precise-scalar.c",
387 "src/requantization/fp32-scalar.c",
388 "src/requantization/q31-scalar.c",
389 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700390]
391
Marat Dukhan436ebe62019-12-04 15:10:12 -0800392WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700393 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
394 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700395 "src/f32-clamp/gen/wasm-x1.c",
396 "src/f32-clamp/gen/wasm-x2.c",
397 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700398 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
399 "src/f32-dwconv/gen/up1x4-wasm.c",
400 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
401 "src/f32-dwconv/gen/up1x9-wasm.c",
402 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
403 "src/f32-dwconv/gen/up1x25-wasm.c",
404 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
405 "src/f32-dwconv/gen/up2x4-wasm.c",
406 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
407 "src/f32-dwconv/gen/up2x9-wasm.c",
408 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
409 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700410 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
412 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700414 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700416 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
418 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700420 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700422 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
423 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700424 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
425 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
426 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700427 "src/f32-gemm/gen/1x4-wasm.c",
428 "src/f32-gemm/gen/2x4-wasm.c",
429 "src/f32-gemm/gen/4x2-wasm.c",
430 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700431 "src/f32-gemm/gen/1x4-relu-wasm.c",
432 "src/f32-gemm/gen/2x4-relu-wasm.c",
433 "src/f32-gemm/gen/4x2-relu-wasm.c",
434 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700435 "src/f32-gemm/gen/1x4-minmax-wasm.c",
436 "src/f32-gemm/gen/2x4-minmax-wasm.c",
437 "src/f32-gemm/gen/4x2-minmax-wasm.c",
438 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800439 "src/f32-hswish/gen/wasm-x1.c",
440 "src/f32-hswish/gen/wasm-x2.c",
441 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700442 "src/f32-igemm/gen/1x4-wasm.c",
443 "src/f32-igemm/gen/2x4-wasm.c",
444 "src/f32-igemm/gen/4x2-wasm.c",
445 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700446 "src/f32-igemm/gen/1x4-relu-wasm.c",
447 "src/f32-igemm/gen/2x4-relu-wasm.c",
448 "src/f32-igemm/gen/4x2-relu-wasm.c",
449 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700450 "src/f32-igemm/gen/1x4-minmax-wasm.c",
451 "src/f32-igemm/gen/2x4-minmax-wasm.c",
452 "src/f32-igemm/gen/4x2-minmax-wasm.c",
453 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700454 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
455 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
456 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700457 "src/f32-prelu/gen/wasm-2x1.c",
458 "src/f32-prelu/gen/wasm-2x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700459 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
460 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
461 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
462 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
463 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
464 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
465 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
466 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
467 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
468 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
469 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
470 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800471 "src/f32-vbinary/gen/vmax-wasm-x1.c",
472 "src/f32-vbinary/gen/vmax-wasm-x2.c",
473 "src/f32-vbinary/gen/vmax-wasm-x4.c",
474 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
475 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
476 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
477 "src/f32-vbinary/gen/vmin-wasm-x1.c",
478 "src/f32-vbinary/gen/vmin-wasm-x2.c",
479 "src/f32-vbinary/gen/vmin-wasm-x4.c",
480 "src/f32-vbinary/gen/vminc-wasm-x1.c",
481 "src/f32-vbinary/gen/vminc-wasm-x2.c",
482 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700483 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
484 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
485 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
486 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
487 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
488 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
489 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
490 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
491 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
492 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
493 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
494 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
495 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
496 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
497 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
498 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700501 "src/f32-vlrelu/gen/wasm-x1.c",
502 "src/f32-vlrelu/gen/wasm-x2.c",
503 "src/f32-vlrelu/gen/wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700504 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
505 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
506 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800507]
508
Marat Dukhan290055c2020-06-09 12:24:29 -0700509WASMSIMD_UKERNELS = [
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700510 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
511 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
512 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
513 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700514 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
515 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
516 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
517 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
518 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
519 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700520 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
521 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
522 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
523 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
524 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
525 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
526 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
527 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
528 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
529 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
530 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
531 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
532 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
533 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
534 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
535 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
536 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
537 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
538 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
539 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
540 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
541 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
542 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
543 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700544 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
545 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
546 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700547 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700548 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
549 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
550 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
551 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700552 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
553 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
554 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
555 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
556 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
557 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
558 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
559 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
560 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
561 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
562 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
563 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
564 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
565 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
566 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
567 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
568 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
569 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
570 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
571 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
572 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
573 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
574 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
575 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
576 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
577 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
578 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
579 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
580 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
581 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700582 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
583 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700584 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
585 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
586 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
587 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
588 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
589 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
590 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
591 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
592 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
593 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
594 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
595 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
596 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
597 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
598 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
599 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
600 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
601 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
602 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
603 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
604 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
605 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
606 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
607 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
608 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
609 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
610 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
611 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
612 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
613 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700614 "src/f32-ibilinear/gen/wasmsimd-c4.c",
615 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700616 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
617 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
618 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700619 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700620 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
621 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
622 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
623 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700624 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
625 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
626 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
627 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
628 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
629 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
630 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
631 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
632 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
633 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
634 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
635 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
636 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
637 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
638 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
639 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
640 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
641 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
642 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
643 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
644 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
645 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
646 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
647 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
648 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
649 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
650 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
651 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
652 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
653 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700654 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
655 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan9df9dc62020-07-10 20:08:49 -0700656 "src/f32-hswish/gen/wasmsimd-x4.c",
657 "src/f32-hswish/gen/wasmsimd-x8.c",
658 "src/f32-hswish/gen/wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700659 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
660 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700661 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
662 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
663 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
664 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700665 "src/f32-rmax/wasmsimd-arm.c",
666 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700667 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
668 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
669 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
670 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700671 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
672 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700673 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
674 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
675 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
676 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700677 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
678 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700679 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
680 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
681 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
682 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700683 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
684 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700685 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
686 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
687 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
688 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700689 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
690 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700691 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
692 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
693 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
694 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700695 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
696 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
697 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
698 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700699 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
700 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
701 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
702 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700703 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
704 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
705 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
706 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700707 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
708 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
709 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
710 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700711 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
712 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700713 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
714 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
715 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
716 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700717 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
718 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700719 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
720 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
721 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
722 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700723 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
724 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700725 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
726 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
727 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
728 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700729 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
730 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700731 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
732 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700733 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
734 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700735 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
736 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
737 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
738 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700739 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
740 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700741 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700747 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
748 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
749 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
750 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhan72b399a2020-06-28 22:41:38 -0700751 "src/f32-vrnd/gen/vrndne-wasmsimd-x4.c",
752 "src/f32-vrnd/gen/vrndne-wasmsimd-x8.c",
753 "src/f32-vrnd/gen/vrndz-wasmsimd-x4.c",
754 "src/f32-vrnd/gen/vrndz-wasmsimd-x8.c",
755 "src/f32-vrnd/gen/vrndu-wasmsimd-x4.c",
756 "src/f32-vrnd/gen/vrndu-wasmsimd-x8.c",
757 "src/f32-vrnd/gen/vrndd-wasmsimd-x4.c",
758 "src/f32-vrnd/gen/vrndd-wasmsimd-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700759 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
760 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700761 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
762 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
763 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
764 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
765 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
766 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700767 "src/math/roundne-wasmsimd-addsub.c",
768 "src/math/roundd-wasmsimd-addsub.c",
769 "src/math/roundu-wasmsimd-addsub.c",
770 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700771 "src/requantization/fp32-wasmsimd.c",
772]
773
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800774PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800775 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800776 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800777 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700778 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
779 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700780 "src/f32-clamp/gen/psimd-x4.c",
781 "src/f32-clamp/gen/psimd-x8.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700782 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700783 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700784 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700785 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700786 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700787 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
788 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
789 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
790 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
791 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
792 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
793 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
794 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
795 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
796 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
797 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
798 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700799 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700800 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
801 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700802 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
803 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
804 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
805 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
806 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
807 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
808 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
809 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
810 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
811 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
812 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
813 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
814 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
815 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
816 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
817 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
818 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
819 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
820 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700821 "src/f32-ibilinear/gen/psimd-c4.c",
822 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700823 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
824 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
825 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
826 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
827 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
828 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
829 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
830 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
831 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
832 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700833 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700834 "src/f32-spmm/gen/4x1-minmax-psimd.c",
835 "src/f32-spmm/gen/8x1-minmax-psimd.c",
836 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700837 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
838 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700839 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800840 "src/f32-prelu/gen/psimd-2x4.c",
841 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800842 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700843 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
844 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
845 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
846 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
847 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
848 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
849 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
850 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800851 "src/f32-vbinary/gen/vmax-psimd-x4.c",
852 "src/f32-vbinary/gen/vmax-psimd-x8.c",
853 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
854 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
855 "src/f32-vbinary/gen/vmin-psimd-x4.c",
856 "src/f32-vbinary/gen/vmin-psimd-x8.c",
857 "src/f32-vbinary/gen/vminc-psimd-x4.c",
858 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700859 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
860 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
861 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
862 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
863 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
864 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
865 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
866 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700867 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
868 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
869 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
870 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700871 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
872 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
873 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
874 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700875 "src/f32-vlrelu/gen/psimd-x4.c",
876 "src/f32-vlrelu/gen/psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700877 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
878 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700879 "src/f32-vunary/gen/vabs-psimd-x4.c",
880 "src/f32-vunary/gen/vabs-psimd-x8.c",
881 "src/f32-vunary/gen/vneg-psimd-x4.c",
882 "src/f32-vunary/gen/vneg-psimd-x8.c",
883 "src/f32-vunary/gen/vsqr-psimd-x4.c",
884 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700885 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700886 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700887 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700888 "src/x32-unpool/psimd.c",
889 "src/x32-zip/x2-psimd.c",
890 "src/x32-zip/x3-psimd.c",
891 "src/x32-zip/x4-psimd.c",
892 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700893 "src/requantization/precise-psimd.c",
894 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700895]
896
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800897PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800898 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
899 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
900 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
901 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
902 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
903 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
904 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
905 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
906 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
907 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
908 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
909 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800910 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
911 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
912 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
913 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
914 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
915 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700916 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
917 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
918 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
919 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
920 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
921 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
922 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
923 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700924 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700925 "src/math/roundd-psimd-addsub.c",
926 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700927 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800928 "src/math/sigmoid-psimd-p5-div.c",
929]
930
Marat Dukhan08c4a432019-10-03 09:29:21 -0700931# ISA-specific micro-kernels
932NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700933 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
934 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700935 "src/f32-clamp/gen/neon-x4.c",
936 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700937 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
938 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700939 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
940 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700941 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
942 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700943 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
944 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700945 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
946 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
947 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
948 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700949 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
950 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
951 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
952 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700953 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
954 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
955 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
956 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700957 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700958 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
959 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700960 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
961 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
962 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
963 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
964 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
965 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
966 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
967 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
968 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
969 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
970 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
971 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
972 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
973 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
974 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
975 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
976 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
977 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
978 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
979 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
980 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
981 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
982 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
983 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
984 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
985 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
986 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
987 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
988 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
989 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
990 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800991 "src/f32-hswish/gen/neon-x4.c",
992 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan55dde5b2020-07-10 22:48:54 -0700993 "src/f32-hswish/gen/neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700994 "src/f32-ibilinear/gen/neon-c4.c",
995 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700996 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
997 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
998 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
999 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1000 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1001 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1002 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1003 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1004 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1005 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1006 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1007 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1008 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1009 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1010 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1011 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001012 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1013 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1014 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001015 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1016 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001017 "src/f32-prelu/gen/neon-2x4.c",
1018 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001019 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1020 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1021 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1022 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1023 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1024 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1025 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1026 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1027 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1028 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1029 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1030 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1031 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1032 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1033 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1034 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1035 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1036 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1037 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1038 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1039 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1040 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1041 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1042 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001043 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001044 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001045 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1046 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1047 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1048 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1049 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1050 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1051 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1052 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1053 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1054 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1055 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1056 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1057 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1058 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1059 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1060 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1061 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1062 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001063 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1064 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1065 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1066 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001067 "src/f32-vbinary/gen/vmax-neon-x4.c",
1068 "src/f32-vbinary/gen/vmax-neon-x8.c",
1069 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1070 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1071 "src/f32-vbinary/gen/vmin-neon-x4.c",
1072 "src/f32-vbinary/gen/vmin-neon-x8.c",
1073 "src/f32-vbinary/gen/vminc-neon-x4.c",
1074 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001075 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1076 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1077 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1078 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1079 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1080 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001081 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1082 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1083 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1084 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001085 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1086 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1087 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1088 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001089 "src/f32-vlrelu/gen/neon-x4.c",
1090 "src/f32-vlrelu/gen/neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001091 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1092 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001093 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1094 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1095 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1096 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1097 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1098 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1099 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1100 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001101 "src/f32-vunary/gen/vabs-neon-x4.c",
1102 "src/f32-vunary/gen/vabs-neon-x8.c",
1103 "src/f32-vunary/gen/vneg-neon-x4.c",
1104 "src/f32-vunary/gen/vneg-neon-x8.c",
1105 "src/f32-vunary/gen/vsqr-neon-x4.c",
1106 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001107 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1108 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1109 "src/qu8-dwconv/up8x9-minmax-neon.c",
1110 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1111 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1112 "src/qu8-gemm/4x8-minmax-neon.c",
1113 "src/qu8-gemm/8x8-minmax-neon.c",
1114 "src/qu8-igemm/4x8-minmax-neon.c",
1115 "src/qu8-igemm/8x8-minmax-neon.c",
1116 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001117 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001118 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001119 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001120 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001122 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001123 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001124 "src/x32-zip/x2-neon.c",
1125 "src/x32-zip/x3-neon.c",
1126 "src/x32-zip/x4-neon.c",
1127 "src/x32-zip/xm-neon.c",
1128 "src/x8-zip/x2-neon.c",
1129 "src/x8-zip/x3-neon.c",
1130 "src/x8-zip/x4-neon.c",
1131 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001132 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001133 "src/math/roundd-neon-addsub.c",
1134 "src/math/roundd-neon-cvt.c",
1135 "src/math/roundu-neon-addsub.c",
1136 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001137 "src/math/roundz-neon-addsub.c",
1138 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001139 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001140 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1141 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1142 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1143 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1144 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1145 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001146 "src/math/sqrt-neon-nr1rsqrts.c",
1147 "src/math/sqrt-neon-nr2rsqrts.c",
1148 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001149 "src/requantization/precise-neon.c",
1150 "src/requantization/fp32-neon.c",
1151 "src/requantization/q31-neon.c",
1152 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001153]
1154
1155NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001156 "src/f32-ibilinear/gen/neonfma-c4.c",
1157 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001158 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1159 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1160 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1161 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1162 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1163 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1164 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1165 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1166 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001167 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1168 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1169 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1170 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001171 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1172 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1173 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1174 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001175 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1176 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1177 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1178 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001179 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1180 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1181 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1182 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1183 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1184 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1185 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1186 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1187 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1188 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1189 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1190 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1191 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1192 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1193 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1194 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1195 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1196 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001197 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1198 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001199 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1200 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1201 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1202 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1203 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1204 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1205 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1206 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1207 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1208 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1209 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1210 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1211 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1212 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1213 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1214 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1215 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1216 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1217 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1218 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1219 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1220 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1221 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001223 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1260 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1261 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1262 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1263 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1264 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1265 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1269 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1270 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1271 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1272 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1273 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1275 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1276 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001277 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1278 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001279 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1280 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1281 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1282 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1283 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1284 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1285 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1286 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1287 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1288 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1289 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1290 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1291 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1292 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1293 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1294 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1295 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1296 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1297 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1298 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001299 "src/math/exp-neonfma-lut64-p2.c",
1300 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001301 "src/math/expminus-neonfma-lut2048-p1.c",
1302 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001303 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001304 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1305 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1306 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1307 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1308 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1309 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1310 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1311 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1312 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1313 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1314 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1315 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1316 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1317 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1318 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1319 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1320 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1321 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001322 "src/math/sqrt-neonfma-nr1fma.c",
1323 "src/math/sqrt-neonfma-nr2fma.c",
1324 "src/math/sqrt-neonfma-nr3fma.c",
1325 "src/math/sqrt-neonfma-nr2fma1adj.c",
1326 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001327]
1328
1329AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001330 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1331 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1332 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1333 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1334 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1335 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001336 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1337 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1338 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1339 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1340 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1341 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1342 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1343 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1344 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1345 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1346 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1347 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1348 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1349 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1350 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1351 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1352 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1353 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1354 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1355 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001356 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1357 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001358 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1359 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001360 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1361 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001362 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1363 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001364 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1365 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1366 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1367 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1368 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001369 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1370 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1380 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1381 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1382 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1383 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1384 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1385 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1386 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001387 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1388 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1389 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1390 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1391 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1392 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1393 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1394 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1395 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1396 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1397 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1398 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1399 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1400 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1401 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1402 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1403 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1404 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001405 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1406 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001407 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1408 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1409 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1410 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1411 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1412 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001413]
1414
Marat Dukhan8853b822020-05-07 12:19:01 -07001415NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001416 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1417 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1418 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1419 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1420 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1421 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1422 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1423 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001424 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001425 "src/math/roundd-neonv8.c",
1426 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001427 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001428]
1429
Marat Dukhan08c4a432019-10-03 09:29:21 -07001430AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001431 "src/f16-clamp/gen/neonfp16arith-x16.c",
1432 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001433 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1434 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1435 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1436 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1437 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1438 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1439 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1440 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1441 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1442 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1443 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1444 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001445 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1446 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001447 "src/f16-hswish/gen/neonfp16arith-x16.c",
1448 "src/f16-hswish/gen/neonfp16arith-x8.c",
1449 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1450 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001451 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1452 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1453 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1454 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1455 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1456 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1457 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1458 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1459 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1460 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1461 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1462 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1463 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1464 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1465 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1466 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1467 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1468 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1469 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1470 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1471 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1472 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1473 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1474 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1475 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1476 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1477 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1478 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001479 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1480 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001481 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1482 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1483 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1484 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1485 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1486 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1487 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1488 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001489 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1490 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1491 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1492 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001493 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1494 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1495 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1496 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1497 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1498 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1499 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1500 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1501 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1502 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1503 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1504 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001505 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1506 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1507 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1508 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1509 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1510 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1511 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1512 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001513]
1514
1515SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001516 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1517 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001518 "src/f32-clamp/gen/sse-x4.c",
1519 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001520 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1521 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001522 "src/f32-dwconv-chw/3x3p1-sse.c",
1523 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001524 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1525 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1526 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1527 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1528 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1529 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1530 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1531 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1532 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1533 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1534 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1535 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001536 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001537 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1538 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001539 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1540 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1541 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1542 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1543 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1544 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1545 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1546 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1547 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1548 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1549 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1550 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1551 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001552 "src/f32-hswish/gen/sse-x4.c",
1553 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001554 "src/f32-ibilinear/gen/sse-c4.c",
1555 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001556 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1557 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1558 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1559 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1560 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1561 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1562 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001563 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1564 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1565 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001566 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001567 "src/f32-prelu/gen/sse-2x4.c",
1568 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001569 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001570 "src/f32-spmm/gen/4x1-minmax-sse.c",
1571 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001572 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001573 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1574 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1575 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1576 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1577 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1578 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1579 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1580 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001581 "src/f32-vbinary/gen/vmax-sse-x4.c",
1582 "src/f32-vbinary/gen/vmax-sse-x8.c",
1583 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1584 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1585 "src/f32-vbinary/gen/vmin-sse-x4.c",
1586 "src/f32-vbinary/gen/vmin-sse-x8.c",
1587 "src/f32-vbinary/gen/vminc-sse-x4.c",
1588 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001589 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1590 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1591 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1592 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1593 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1594 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1595 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1596 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001597 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1598 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1599 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1600 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001601 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1602 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1603 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1604 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001605 "src/f32-vlrelu/gen/sse-x4.c",
1606 "src/f32-vlrelu/gen/sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001607 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1608 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001609 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1610 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001611 "src/f32-vunary/gen/vabs-sse-x4.c",
1612 "src/f32-vunary/gen/vabs-sse-x8.c",
1613 "src/f32-vunary/gen/vneg-sse-x4.c",
1614 "src/f32-vunary/gen/vneg-sse-x8.c",
1615 "src/f32-vunary/gen/vsqr-sse-x4.c",
1616 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001617 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001618 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001619 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001620 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001621 "src/math/roundd-sse-addsub.c",
1622 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001623 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001624 "src/math/sqrt-sse-hh1mac.c",
1625 "src/math/sqrt-sse-nr1mac.c",
1626 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001627]
1628
1629SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001630 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1631 "src/f32-argmaxpool/4x-sse2-c4.c",
1632 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001633 "src/f32-prelu/gen/sse2-2x4.c",
1634 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001635 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1636 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1637 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1638 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1639 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1640 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1641 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1642 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1643 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1644 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1645 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1646 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001647 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001648 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001649 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001650 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001651 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1652 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001653 "src/f32-vlrelu/gen/sse2-x4.c",
1654 "src/f32-vlrelu/gen/sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001655 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1656 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1657 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1658 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1659 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1660 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1661 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1662 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001663 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1664 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1665 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1666 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1667 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1668 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1669 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1670 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1671 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001672 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001673 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001674 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001675 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001676 "src/x32-zip/x2-sse2.c",
1677 "src/x32-zip/x3-sse2.c",
1678 "src/x32-zip/x4-sse2.c",
1679 "src/x32-zip/xm-sse2.c",
1680 "src/x8-zip/x2-sse2.c",
1681 "src/x8-zip/x3-sse2.c",
1682 "src/x8-zip/x4-sse2.c",
1683 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001684 "src/math/exp-sse2-p5.c",
1685 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001686 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001687 "src/math/roundd-sse2-cvt.c",
1688 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001689 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001690 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001691 "src/requantization/precise-sse2.c",
1692 "src/requantization/fp32-sse2.c",
1693 "src/requantization/q31-sse2.c",
1694 "src/requantization/gemmlowp-sse2.c",
1695]
1696
1697SSSE3_UKERNELS = [
1698 "src/requantization/precise-ssse3.c",
1699 "src/requantization/q31-ssse3.c",
1700 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001701]
1702
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001703SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001704 "src/f32-prelu/gen/sse41-2x4.c",
1705 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001706 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001707 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001708 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001709 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001710 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1711 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001712 "src/f32-vlrelu/gen/sse41-x4.c",
1713 "src/f32-vlrelu/gen/sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001714 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1715 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1716 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1717 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1718 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1719 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1720 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1721 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001722 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001723 "src/math/roundd-sse41.c",
1724 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001725 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001726 "src/requantization/precise-sse4.c",
1727 "src/requantization/q31-sse4.c",
1728 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001729]
1730
Marat Dukhan08c4a432019-10-03 09:29:21 -07001731AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001732 "src/f32-clamp/gen/avx-x8.c",
1733 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001734 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1735 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1736 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1737 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1738 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1739 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1740 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1741 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1742 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1743 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1744 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1745 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1746 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1747 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1748 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1749 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1750 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1751 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1752 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1753 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1754 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1755 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1756 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1757 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1758 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1759 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1760 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1761 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1762 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1763 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001764 "src/f32-hswish/gen/avx-x8.c",
1765 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001766 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1767 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1768 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1769 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1770 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1771 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1772 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1773 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1774 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001775 "src/f32-prelu/gen/avx-2x8.c",
1776 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001778 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1779 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1780 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1781 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1782 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1783 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1784 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1785 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001786 "src/f32-vbinary/gen/vmax-avx-x8.c",
1787 "src/f32-vbinary/gen/vmax-avx-x16.c",
1788 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1789 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1790 "src/f32-vbinary/gen/vmin-avx-x8.c",
1791 "src/f32-vbinary/gen/vmin-avx-x16.c",
1792 "src/f32-vbinary/gen/vminc-avx-x8.c",
1793 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001794 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1795 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1796 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1797 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1798 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1799 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1800 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1801 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001802 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1803 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1804 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1805 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001806 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1807 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1808 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1809 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001810 "src/f32-vlrelu/gen/avx-x8.c",
1811 "src/f32-vlrelu/gen/avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001812 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001813 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1814 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1815 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1816 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1817 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1818 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1819 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1820 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001821 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1822 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001823 "src/f32-vunary/gen/vabs-avx-x8.c",
1824 "src/f32-vunary/gen/vabs-avx-x16.c",
1825 "src/f32-vunary/gen/vneg-avx-x8.c",
1826 "src/f32-vunary/gen/vneg-avx-x16.c",
1827 "src/f32-vunary/gen/vsqr-avx-x8.c",
1828 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001829]
1830
Marat Dukhanfda12b82019-11-21 12:27:59 -08001831FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001832 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1833 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1834 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1835 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1836 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1837 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1838 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1839 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1840 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1841 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1842 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1843 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1844 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1845 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1846 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1847 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1848 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1849 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1850 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1851 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1852 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1853 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1854 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1855 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1856 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1857 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1858 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1859 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1860 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1861 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1862 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1863 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1864 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1865 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1866 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1867 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1868 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1869 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1870 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1871 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001872 "src/f32-hswish/gen/fma3-x8.c",
1873 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001874 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1875 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1876 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1877 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1878 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1879 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1880 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1881 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1882 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1883 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1884 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1885 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1886 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1887 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001888 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1889 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1890 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1891 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1892 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1893 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1894 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1895 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001896 "src/math/sqrt-fma3-nr1fma.c",
1897 "src/math/sqrt-fma3-nr2fma.c",
1898 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001899]
1900
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001901AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001902 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1903 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1904 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1905 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1906 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1907 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1908 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1909 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1910 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1911 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1912 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1913 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1914 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1915 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1916 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1917 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1918 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1919 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1920 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1921 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1922 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1923 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1924 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1925 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1926 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1927 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1928 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1929 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1930 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1931 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1932 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1933 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1934 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1935 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1936 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1937 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001938 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1939 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1940 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1941 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1942 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1943 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1944 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1945 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1946 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1947 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1948 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1949 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1950 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1951 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1952 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1953 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1954 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1955 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1956 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1957 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1958 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1959 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1960 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1961 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1962 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1963 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1964 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1965 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1966 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1967 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001968 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1969 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1970 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1971 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1972 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1973 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1974 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1975 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1976 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1977 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1978 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1979 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1980 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1981 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1982 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1983 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1984 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1985 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1986 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1987 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1988 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1989 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1990 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1991 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001992 "src/math/exp-avx2-p5.c",
1993 "src/math/exp-avx2-perm-p3.c",
1994 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001995 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001996 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001997 "src/math/sigmoid-avx2-rr2-p5-div.c",
1998 "src/math/sigmoid-avx2-rr1-p5-div.c",
1999 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2000 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2001 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2002 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002003]
2004
Marat Dukhan08c4a432019-10-03 09:29:21 -07002005AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002006 "src/f32-clamp/gen/avx512f-x16.c",
2007 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002008 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2009 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2010 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2011 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2012 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2013 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2014 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2015 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2016 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2017 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2018 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2019 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2020 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2021 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2022 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2023 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2024 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2025 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2026 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2027 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2028 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2029 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2030 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2031 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08002032 "src/f32-hswish/gen/avx512f-x16.c",
2033 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002034 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2035 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2036 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2037 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2038 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2039 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002040 "src/f32-prelu/gen/avx512f-2x16.c",
2041 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002042 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2043 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2044 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2045 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2046 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2047 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2048 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2049 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2050 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2051 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2052 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2053 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2054 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2055 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2056 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2057 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2058 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2059 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2060 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2061 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2062 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2063 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2064 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2065 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2066 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2067 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2068 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2069 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2070 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2071 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2072 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2073 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2074 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2075 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2076 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2077 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002078 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002079 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2080 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2081 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2082 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2083 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2084 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2085 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2086 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002087 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2088 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2089 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2090 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2091 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2092 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2093 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2094 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002095 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2096 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2097 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2098 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2099 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2100 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2101 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2102 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002103 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2104 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2105 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2106 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002107 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2108 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2109 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2110 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07002111 "src/f32-vlrelu/gen/avx512f-x16.c",
2112 "src/f32-vlrelu/gen/avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002113 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002114 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2115 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2116 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2117 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2118 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2119 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2120 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2121 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2122 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2123 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2124 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2125 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2126 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2127 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2128 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2129 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2130 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2131 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2132 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2133 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2134 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2135 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2136 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2137 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002138 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2139 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2140 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2141 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2142 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2143 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2144 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2145 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002146 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2147 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2148 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2149 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2150 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2151 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2152 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2153 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002154 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2155 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2156 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2157 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2158 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2159 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002160 "src/math/exp-avx512f-p5-scalef.c",
2161 "src/math/exp-avx512f-p5.c",
2162 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002163 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002164 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002165 "src/math/sqrt-avx512f-nr1fma.c",
2166 "src/math/sqrt-avx512f-nr2fma.c",
2167 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002168]
2169
2170AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002171 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002172 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002173 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2174 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2175 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2176 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002177 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2178 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2179 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2180 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002181 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2182 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2183 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2184 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002185]
2186
2187AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002188 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002189 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2190 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002191 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002192 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2193 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002194 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2195 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2196 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002197 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002198 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2199 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2200 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002201 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002202 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2203 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002204 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002205 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2206 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2207 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2208 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2209 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2210 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2211 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2212 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2213 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2214 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2215 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2216 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2217 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2218 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2219 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2220 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2221 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2222 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002223 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2224 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002225 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002226 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2227 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2228 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2229 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2230 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2231 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2232 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2233 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2234 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2235 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2236 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2237 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2238 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2239 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2240 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2241 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2242 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2243 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002244 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2245 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2246 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2247 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2248 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2249 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2250 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2251 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2252 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2253 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2254 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2255 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2256 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2257 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2258 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2259 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2260 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2261 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002262]
2263
Marat Dukhan1b354632020-03-23 12:50:22 -07002264INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002265 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002266 "src/xnnpack/argmaxpool.h",
2267 "src/xnnpack/avgpool.h",
2268 "src/xnnpack/clamp.h",
2269 "src/xnnpack/common.h",
2270 "src/xnnpack/conv.h",
2271 "src/xnnpack/dwconv.h",
2272 "src/xnnpack/gavgpool.h",
2273 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002274 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002275 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002276 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002277 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002278 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002279 "src/xnnpack/lut.h",
2280 "src/xnnpack/math.h",
2281 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002282 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002283 "src/xnnpack/packx.h",
2284 "src/xnnpack/pad.h",
2285 "src/xnnpack/params.h",
2286 "src/xnnpack/pavgpool.h",
2287 "src/xnnpack/ppmm.h",
2288 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002289 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002290 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002291 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002292 "src/xnnpack/rmax.h",
2293 "src/xnnpack/scalar-utils.h",
2294 "src/xnnpack/spmm.h",
2295 "src/xnnpack/unpool.h",
2296 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002297 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002298 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002299 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002300 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002301 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002302 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002303 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002304]
2305
2306INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002307 "include/xnnpack.h",
2308 "src/xnnpack/allocator.h",
2309 "src/xnnpack/compute.h",
2310 "src/xnnpack/im2col.h",
2311 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002312 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002313 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002314 "src/xnnpack/operator.h",
2315 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002316 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002317 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002318 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002319 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002320]
2321
Marat Dukhan1b354632020-03-23 12:50:22 -07002322ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002323 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002324]
2325
Marat Dukhan1b354632020-03-23 12:50:22 -07002326MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002327 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002328 "include/xnnpack.h",
2329]
2330
Marat Dukhan1b354632020-03-23 12:50:22 -07002331MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002332 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002333 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002334 "src/xnnpack/requantization.h",
2335 "include/xnnpack.h",
2336]
2337
2338OPERATOR_TEST_PARAMS_HDRS = [
2339 "src/xnnpack/params.h",
2340 "src/xnnpack/common.h",
2341]
2342
2343WEIGHTS_PACK_HDRS = [
2344 "src/xnnpack/pack.h",
2345 "src/xnnpack/operator.h",
2346 "src/xnnpack/compute.h",
2347]
2348
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002349LOGGING_COPTS = select({
2350 # No logging in optimized mode
2351 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2352 # Full logging in debug mode
2353 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2354 # Error-only logging in default (fastbuild) mode
2355 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2356})
2357
Marat Dukhan3b59de22020-06-03 20:15:19 -07002358LOGGING_SRCS = select({
2359 # No logging in optimized mode
2360 ":optimized_build": [],
2361 "//conditions:default": [
2362 "src/operator-strings.c",
2363 "src/subgraph-strings.c",
2364 ],
2365})
2366
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002367LOGGING_HDRS = [
2368 "src/xnnpack/log.h",
2369]
2370
Marat Dukhan08c4a432019-10-03 09:29:21 -07002371xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002372 name = "tables",
2373 srcs = TABLE_SRCS,
2374 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002375 gcc_copts = xnnpack_gcc_std_copts(),
2376 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002377)
2378
2379xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002380 name = "scalar_ukernels",
2381 srcs = SCALAR_UKERNELS,
2382 hdrs = INTERNAL_HDRS,
2383 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002384 gcc_copts = xnnpack_gcc_std_copts(),
2385 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002386 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002387 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002388 "@FP16",
2389 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002390 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391 ],
2392)
2393
2394xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002395 name = "scalar_ukernels_test_mode",
2396 srcs = SCALAR_UKERNELS,
2397 hdrs = INTERNAL_HDRS,
2398 aarch32_copts = ["-marm"],
2399 copts = [
2400 "-UNDEBUG",
2401 "-DXNN_TEST_MODE=1",
2402 ],
2403 gcc_copts = xnnpack_gcc_std_copts(),
2404 msvc_copts = xnnpack_msvc_std_copts(),
2405 deps = [
2406 ":tables",
2407 "@FP16",
2408 "@FXdiv",
2409 "@pthreadpool",
2410 ],
2411)
2412
2413xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002414 name = "wasm_ukernels",
2415 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002416 gcc_copts = xnnpack_gcc_std_copts(),
2417 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002418 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002419 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002420 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002421 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002422 "@FP16",
2423 "@FXdiv",
2424 "@pthreadpool",
2425 ],
2426)
2427
2428xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002429 name = "wasm_ukernels_test_mode",
2430 hdrs = INTERNAL_HDRS,
2431 copts = [
2432 "-UNDEBUG",
2433 "-DXNN_TEST_MODE=1",
2434 ],
2435 gcc_copts = xnnpack_gcc_std_copts(),
2436 msvc_copts = xnnpack_msvc_std_copts(),
2437 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002438 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002439 deps = [
2440 ":tables",
2441 "@FP16",
2442 "@FXdiv",
2443 "@pthreadpool",
2444 ],
2445)
2446
2447xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002448 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002449 hdrs = INTERNAL_HDRS,
2450 aarch32_copts = [
2451 "-marm",
2452 "-mfpu=neon",
2453 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002454 gcc_copts = xnnpack_gcc_std_copts(),
2455 gcc_x86_copts = ["-msse2"],
2456 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002457 optimized_copts = [
2458 "-O3",
2459 "-ffast-math",
2460 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002461 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002462 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002463 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002464 "@FP16",
2465 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002466 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002467 ],
2468)
2469
2470xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002471 name = "psimd_fastmath_ukernels_test_mode",
2472 hdrs = INTERNAL_HDRS,
2473 aarch32_copts = [
2474 "-marm",
2475 "-mfpu=neon",
2476 ],
2477 copts = [
2478 "-UNDEBUG",
2479 "-DXNN_TEST_MODE=1",
2480 ],
2481 gcc_copts = xnnpack_gcc_std_copts(),
2482 gcc_x86_copts = ["-msse2"],
2483 msvc_copts = xnnpack_msvc_std_copts(),
2484 optimized_copts = [
2485 "-O3",
2486 "-ffast-math",
2487 ],
2488 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2489 deps = [
2490 ":tables",
2491 "@FP16",
2492 "@psimd",
2493 "@pthreadpool",
2494 ],
2495)
2496
2497xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002498 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002499 hdrs = INTERNAL_HDRS,
2500 aarch32_copts = [
2501 "-marm",
2502 "-mfpu=neon",
2503 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002504 gcc_copts = xnnpack_gcc_std_copts(),
2505 gcc_x86_copts = ["-msse2"],
2506 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002507 optimized_copts = [
2508 "-O3",
2509 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002510 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002511 deps = [
2512 ":tables",
2513 "@FP16",
2514 "@psimd",
2515 "@pthreadpool",
2516 ],
2517)
2518
2519xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002520 name = "psimd_accmath_ukernels_test_mode",
2521 hdrs = INTERNAL_HDRS,
2522 aarch32_copts = [
2523 "-marm",
2524 "-mfpu=neon",
2525 ],
2526 copts = [
2527 "-UNDEBUG",
2528 "-DXNN_TEST_MODE=1",
2529 ],
2530 gcc_copts = xnnpack_gcc_std_copts(),
2531 gcc_x86_copts = ["-msse2"],
2532 msvc_copts = xnnpack_msvc_std_copts(),
2533 optimized_copts = [
2534 "-O3",
2535 ],
2536 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2537 deps = [
2538 ":tables",
2539 "@FP16",
2540 "@psimd",
2541 "@pthreadpool",
2542 ],
2543)
2544
2545xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002546 name = "neon_ukernels",
2547 hdrs = INTERNAL_HDRS,
2548 aarch32_copts = [
2549 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002550 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002551 "-mfpu=neon",
2552 ],
2553 aarch32_srcs = NEON_UKERNELS,
2554 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002555 gcc_copts = xnnpack_gcc_std_copts(),
2556 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002557 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002558 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002559 "@FP16",
2560 "@pthreadpool",
2561 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562)
2563
2564xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002565 name = "neon_ukernels_test_mode",
2566 hdrs = INTERNAL_HDRS,
2567 aarch32_copts = [
2568 "-marm",
2569 "-march=armv7-a",
2570 "-mfpu=neon",
2571 ],
2572 aarch32_srcs = NEON_UKERNELS,
2573 aarch64_srcs = NEON_UKERNELS,
2574 copts = [
2575 "-UNDEBUG",
2576 "-DXNN_TEST_MODE=1",
2577 ],
2578 gcc_copts = xnnpack_gcc_std_copts(),
2579 msvc_copts = xnnpack_msvc_std_copts(),
2580 deps = [
2581 ":tables",
2582 "@FP16",
2583 "@pthreadpool",
2584 ],
2585)
2586
2587xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002588 name = "neonfma_ukernels",
2589 hdrs = INTERNAL_HDRS,
2590 aarch32_copts = [
2591 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002592 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002593 "-mfpu=neon-vfpv4",
2594 ],
2595 aarch32_srcs = NEONFMA_UKERNELS,
2596 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002597 apple_aarch32_copts = [
2598 "-mcpu=swift",
2599 "-mtune=generic",
2600 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002601 gcc_copts = xnnpack_gcc_std_copts(),
2602 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002603 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002604 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002605 "@FP16",
2606 "@pthreadpool",
2607 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002608)
2609
2610xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002611 name = "neonfma_ukernels_test_mode",
2612 hdrs = INTERNAL_HDRS,
2613 aarch32_copts = [
2614 "-marm",
2615 "-march=armv7-a",
2616 "-mfpu=neon-vfpv4",
2617 ],
2618 aarch32_srcs = NEONFMA_UKERNELS,
2619 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002620 apple_aarch32_copts = [
2621 "-mcpu=swift",
2622 "-mtune=generic",
2623 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002624 copts = [
2625 "-UNDEBUG",
2626 "-DXNN_TEST_MODE=1",
2627 ],
2628 gcc_copts = xnnpack_gcc_std_copts(),
2629 msvc_copts = xnnpack_msvc_std_copts(),
2630 deps = [
2631 ":tables",
2632 "@FP16",
2633 "@pthreadpool",
2634 ],
2635)
2636
2637xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002638 name = "neonv8_ukernels",
2639 hdrs = INTERNAL_HDRS,
2640 aarch32_copts = [
2641 "-marm",
2642 "-march=armv8-a",
2643 "-mfpu=neon-fp-armv8",
2644 ],
2645 aarch32_srcs = NEONV8_UKERNELS,
2646 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002647 apple_aarch32_copts = [
2648 "-mcpu=cyclone",
2649 "-mtune=generic",
2650 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002651 gcc_copts = xnnpack_gcc_std_copts(),
2652 msvc_copts = xnnpack_msvc_std_copts(),
2653 deps = [
2654 ":tables",
2655 "@FP16",
2656 "@pthreadpool",
2657 ],
2658)
2659
2660xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002661 name = "neonv8_ukernels_test_mode",
2662 hdrs = INTERNAL_HDRS,
2663 aarch32_copts = [
2664 "-marm",
2665 "-march=armv8-a",
2666 "-mfpu=neon-fp-armv8",
2667 ],
2668 aarch32_srcs = NEONV8_UKERNELS,
2669 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002670 apple_aarch32_copts = [
2671 "-mcpu=cyclone",
2672 "-mtune=generic",
2673 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002674 copts = [
2675 "-UNDEBUG",
2676 "-DXNN_TEST_MODE=1",
2677 ],
2678 gcc_copts = xnnpack_gcc_std_copts(),
2679 msvc_copts = xnnpack_msvc_std_copts(),
2680 deps = [
2681 ":tables",
2682 "@FP16",
2683 "@pthreadpool",
2684 ],
2685)
2686
2687xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002688 name = "neonfp16arith_ukernels",
2689 hdrs = INTERNAL_HDRS,
2690 aarch64_copts = ["-march=armv8.2-a+fp16"],
2691 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002692 gcc_copts = xnnpack_gcc_std_copts(),
2693 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002694 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002695 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002696 "@FP16",
2697 "@pthreadpool",
2698 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002699)
2700
2701xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002702 name = "neonfp16arith_ukernels_test_mode",
2703 hdrs = INTERNAL_HDRS,
2704 aarch64_copts = ["-march=armv8.2-a+fp16"],
2705 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2706 copts = [
2707 "-UNDEBUG",
2708 "-DXNN_TEST_MODE=1",
2709 ],
2710 gcc_copts = xnnpack_gcc_std_copts(),
2711 msvc_copts = xnnpack_msvc_std_copts(),
2712 deps = [
2713 ":tables",
2714 "@FP16",
2715 "@pthreadpool",
2716 ],
2717)
2718
2719xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002720 name = "sse2_ukernels",
2721 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002722 gcc_copts = xnnpack_gcc_std_copts(),
2723 gcc_x86_copts = ["-msse2"],
2724 msvc_copts = xnnpack_msvc_std_copts(),
2725 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002726 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002727 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002728 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002729 "@FP16",
2730 "@pthreadpool",
2731 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732)
2733
2734xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002735 name = "sse2_ukernels_test_mode",
2736 hdrs = INTERNAL_HDRS,
2737 copts = [
2738 "-UNDEBUG",
2739 "-DXNN_TEST_MODE=1",
2740 ],
2741 gcc_copts = xnnpack_gcc_std_copts(),
2742 gcc_x86_copts = ["-msse2"],
2743 msvc_copts = xnnpack_msvc_std_copts(),
2744 msvc_x86_32_copts = ["/arch:SSE2"],
2745 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2746 deps = [
2747 ":tables",
2748 "@FP16",
2749 "@pthreadpool",
2750 ],
2751)
2752
2753xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002754 name = "ssse3_ukernels",
2755 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002756 gcc_copts = xnnpack_gcc_std_copts(),
2757 gcc_x86_copts = ["-mssse3"],
2758 msvc_copts = xnnpack_msvc_std_copts(),
2759 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002760 x86_srcs = SSSE3_UKERNELS,
2761 deps = [
2762 ":tables",
2763 "@FP16",
2764 "@pthreadpool",
2765 ],
2766)
2767
2768xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002769 name = "ssse3_ukernels_test_mode",
2770 hdrs = INTERNAL_HDRS,
2771 copts = [
2772 "-UNDEBUG",
2773 "-DXNN_TEST_MODE=1",
2774 ],
2775 gcc_copts = xnnpack_gcc_std_copts(),
2776 gcc_x86_copts = ["-mssse3"],
2777 msvc_copts = xnnpack_msvc_std_copts(),
2778 msvc_x86_32_copts = ["/arch:SSE2"],
2779 x86_srcs = SSSE3_UKERNELS,
2780 deps = [
2781 ":tables",
2782 "@FP16",
2783 "@pthreadpool",
2784 ],
2785)
2786
2787xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002788 name = "sse41_ukernels",
2789 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002790 gcc_copts = xnnpack_gcc_std_copts(),
2791 gcc_x86_copts = ["-msse4.1"],
2792 msvc_copts = xnnpack_msvc_std_copts(),
2793 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002794 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002795 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002796 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002797 "@FP16",
2798 "@pthreadpool",
2799 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002800)
2801
2802xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002803 name = "sse41_ukernels_test_mode",
2804 hdrs = INTERNAL_HDRS,
2805 copts = [
2806 "-UNDEBUG",
2807 "-DXNN_TEST_MODE=1",
2808 ],
2809 gcc_copts = xnnpack_gcc_std_copts(),
2810 gcc_x86_copts = ["-msse4.1"],
2811 msvc_copts = xnnpack_msvc_std_copts(),
2812 msvc_x86_32_copts = ["/arch:SSE2"],
2813 x86_srcs = SSE41_UKERNELS,
2814 deps = [
2815 ":tables",
2816 "@FP16",
2817 "@pthreadpool",
2818 ],
2819)
2820
2821xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002822 name = "avx_ukernels",
2823 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002824 gcc_copts = xnnpack_gcc_std_copts(),
2825 gcc_x86_copts = ["-mavx"],
2826 msvc_copts = xnnpack_msvc_std_copts(),
2827 msvc_x86_32_copts = ["/arch:AVX"],
2828 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002829 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002830 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002831 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002832 "@FP16",
2833 "@pthreadpool",
2834 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002835)
2836
2837xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002838 name = "avx_ukernels_test_mode",
2839 hdrs = INTERNAL_HDRS,
2840 copts = [
2841 "-UNDEBUG",
2842 "-DXNN_TEST_MODE=1",
2843 ],
2844 gcc_copts = xnnpack_gcc_std_copts(),
2845 gcc_x86_copts = ["-mavx"],
2846 msvc_copts = xnnpack_msvc_std_copts(),
2847 msvc_x86_32_copts = ["/arch:AVX"],
2848 msvc_x86_64_copts = ["/arch:AVX"],
2849 x86_srcs = AVX_UKERNELS,
2850 deps = [
2851 ":tables",
2852 "@FP16",
2853 "@pthreadpool",
2854 ],
2855)
2856
2857xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002858 name = "fma3_ukernels",
2859 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002860 gcc_copts = xnnpack_gcc_std_copts(),
2861 gcc_x86_copts = ["-mfma"],
2862 msvc_copts = xnnpack_msvc_std_copts(),
2863 msvc_x86_32_copts = ["/arch:AVX"],
2864 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002865 x86_srcs = FMA3_UKERNELS,
2866 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002867 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002868 "@FP16",
2869 "@pthreadpool",
2870 ],
2871)
2872
2873xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002874 name = "fma3_ukernels_test_mode",
2875 hdrs = INTERNAL_HDRS,
2876 copts = [
2877 "-UNDEBUG",
2878 "-DXNN_TEST_MODE=1",
2879 ],
2880 gcc_copts = xnnpack_gcc_std_copts(),
2881 gcc_x86_copts = ["-mfma"],
2882 msvc_copts = xnnpack_msvc_std_copts(),
2883 msvc_x86_32_copts = ["/arch:AVX"],
2884 msvc_x86_64_copts = ["/arch:AVX"],
2885 x86_srcs = FMA3_UKERNELS,
2886 deps = [
2887 ":tables",
2888 "@FP16",
2889 "@pthreadpool",
2890 ],
2891)
2892
2893xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002894 name = "avx2_ukernels",
2895 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002896 gcc_copts = xnnpack_gcc_std_copts(),
2897 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002898 "-mfma",
2899 "-mavx2",
2900 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002901 msvc_copts = xnnpack_msvc_std_copts(),
2902 msvc_x86_32_copts = ["/arch:AVX2"],
2903 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002904 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002905 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002906 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002907 "@FP16",
2908 "@pthreadpool",
2909 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002910)
2911
2912xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002913 name = "avx2_ukernels_test_mode",
2914 hdrs = INTERNAL_HDRS,
2915 copts = [
2916 "-UNDEBUG",
2917 "-DXNN_TEST_MODE=1",
2918 ],
2919 gcc_copts = xnnpack_gcc_std_copts(),
2920 gcc_x86_copts = [
2921 "-mfma",
2922 "-mavx2",
2923 ],
2924 msvc_copts = xnnpack_msvc_std_copts(),
2925 msvc_x86_32_copts = ["/arch:AVX2"],
2926 msvc_x86_64_copts = ["/arch:AVX2"],
2927 x86_srcs = AVX2_UKERNELS,
2928 deps = [
2929 ":tables",
2930 "@FP16",
2931 "@pthreadpool",
2932 ],
2933)
2934
2935xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 name = "avx512f_ukernels",
2937 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002938 gcc_copts = xnnpack_gcc_std_copts(),
2939 gcc_x86_copts = ["-mavx512f"],
2940 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2941 msvc_copts = xnnpack_msvc_std_copts(),
2942 msvc_x86_32_copts = ["/arch:AVX512"],
2943 msvc_x86_64_copts = ["/arch:AVX512"],
2944 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002945 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002946 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002947 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002948 "@FP16",
2949 "@pthreadpool",
2950 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002951)
2952
2953xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002954 name = "avx512f_ukernels_test_mode",
2955 hdrs = INTERNAL_HDRS,
2956 copts = [
2957 "-UNDEBUG",
2958 "-DXNN_TEST_MODE=1",
2959 ],
2960 gcc_copts = xnnpack_gcc_std_copts(),
2961 gcc_x86_copts = ["-mavx512f"],
2962 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2963 msvc_copts = xnnpack_msvc_std_copts(),
2964 msvc_x86_32_copts = ["/arch:AVX512"],
2965 msvc_x86_64_copts = ["/arch:AVX512"],
2966 msys_copts = ["-fno-asynchronous-unwind-tables"],
2967 x86_srcs = AVX512F_UKERNELS,
2968 deps = [
2969 ":tables",
2970 "@FP16",
2971 "@pthreadpool",
2972 ],
2973)
2974
2975xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976 name = "asm_ukernels",
2977 hdrs = ["src/xnnpack/assembly.h"],
2978 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002979 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002980 aarch64_srcs = AARCH64_ASM_UKERNELS,
2981)
2982
Marat Dukhan3b59de22020-06-03 20:15:19 -07002983xnnpack_cc_library(
2984 name = "logging_utils",
2985 srcs = LOGGING_SRCS,
2986 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2987 copts = LOGGING_COPTS + [
2988 "-Isrc",
2989 "-Iinclude",
2990 ] + select({
2991 ":debug_build": [],
2992 "//conditions:default": xnnpack_min_size_copts(),
2993 }),
2994 gcc_copts = xnnpack_gcc_std_copts(),
2995 msvc_copts = xnnpack_msvc_std_copts(),
2996 visibility = xnnpack_visibility(),
2997 deps = [
2998 "@FP16",
2999 "@clog",
3000 "@pthreadpool",
3001 ],
3002)
3003
Marat Dukhan08c4a432019-10-03 09:29:21 -07003004xnnpack_aggregate_library(
3005 name = "ukernels",
3006 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007 ":neon_ukernels",
3008 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003009 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010 ":asm_ukernels",
3011 ],
3012 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003013 ":neon_ukernels",
3014 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003015 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003016 ":neonfp16arith_ukernels",
3017 ":asm_ukernels",
3018 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003019 generic_deps = [
3020 ":scalar_ukernels",
3021 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003022 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003023 ":psimd_fastmath_ukernels",
3024 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003026 wasm_deps = [
3027 ":wasm_ukernels",
3028 ],
3029 wasmsimd_deps = [
3030 ":wasm_ukernels",
3031 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003032 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003033 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003034 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003035 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003036 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003037 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003038 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039 ":avx512f_ukernels",
3040 ],
3041)
3042
Marat Dukhan33fcf782020-05-24 14:27:15 -07003043xnnpack_aggregate_library(
3044 name = "ukernels_test_mode",
3045 aarch32_deps = [
3046 ":neon_ukernels_test_mode",
3047 ":neonfma_ukernels_test_mode",
3048 ":neonv8_ukernels_test_mode",
3049 ":asm_ukernels",
3050 ],
3051 aarch64_deps = [
3052 ":neon_ukernels_test_mode",
3053 ":neonfma_ukernels_test_mode",
3054 ":neonv8_ukernels_test_mode",
3055 ":neonfp16arith_ukernels_test_mode",
3056 ":asm_ukernels",
3057 ],
3058 generic_deps = [
3059 ":scalar_ukernels_test_mode",
3060 ],
3061 psimd_deps = [
3062 ":psimd_fastmath_ukernels_test_mode",
3063 ":psimd_accmath_ukernels_test_mode",
3064 ],
3065 wasm_deps = [
3066 ":wasm_ukernels_test_mode",
3067 ],
3068 wasmsimd_deps = [
3069 ":wasm_ukernels_test_mode",
3070 ],
3071 x86_deps = [
3072 ":sse2_ukernels_test_mode",
3073 ":ssse3_ukernels_test_mode",
3074 ":sse41_ukernels_test_mode",
3075 ":avx_ukernels_test_mode",
3076 ":fma3_ukernels_test_mode",
3077 ":avx2_ukernels_test_mode",
3078 ":avx512f_ukernels_test_mode",
3079 ],
3080)
3081
Marat Dukhan86df3842020-06-17 14:03:31 -07003082# This target must be compatible with Android NDK r17 and compile with Clang 6.
3083# It is similar to :ukernels target, but excludes microkernels which require
3084# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003085xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003086 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003087 aarch32_deps = [
3088 ":neon_ukernels",
3089 ":neonfma_ukernels",
3090 ":neonv8_ukernels",
3091 ":asm_ukernels",
3092 ],
3093 aarch64_deps = [
3094 ":neon_ukernels",
3095 ":neonfma_ukernels",
3096 ":neonv8_ukernels",
3097 ":asm_ukernels",
3098 ],
3099 generic_deps = [
3100 ":scalar_ukernels",
3101 ],
3102 psimd_deps = [
3103 ":psimd_fastmath_ukernels",
3104 ":psimd_accmath_ukernels",
3105 ],
3106 wasm_deps = [
3107 ":wasm_ukernels",
3108 ],
3109 wasmsimd_deps = [
3110 ":wasm_ukernels",
3111 ],
3112 x86_deps = [
3113 ":sse2_ukernels",
3114 ":ssse3_ukernels",
3115 ":sse41_ukernels",
3116 ":avx_ukernels",
3117 ":fma3_ukernels",
3118 ":avx2_ukernels",
3119 ":avx512f_ukernels",
3120 ],
3121)
3122
Marat Dukhan08c4a432019-10-03 09:29:21 -07003123xnnpack_cc_library(
3124 name = "im2col",
3125 srcs = ["src/im2col.c"],
3126 hdrs = [
3127 "src/xnnpack/common.h",
3128 "src/xnnpack/im2col.h",
3129 ],
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)
3133
3134xnnpack_cc_library(
3135 name = "indirection",
3136 srcs = ["src/indirection.c"],
3137 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003138 gcc_copts = xnnpack_gcc_std_copts(),
3139 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140 deps = [
3141 "@FP16",
3142 "@FXdiv",
3143 "@pthreadpool",
3144 ],
3145)
3146
3147xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003148 name = "indirection_test_mode",
3149 srcs = ["src/indirection.c"],
3150 hdrs = INTERNAL_HDRS,
3151 copts = [
3152 "-UNDEBUG",
3153 "-DXNN_TEST_MODE=1",
3154 ],
3155 gcc_copts = xnnpack_gcc_std_copts(),
3156 msvc_copts = xnnpack_msvc_std_copts(),
3157 deps = [
3158 "@FP16",
3159 "@FXdiv",
3160 "@pthreadpool",
3161 ],
3162)
3163
3164xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003165 name = "packing",
3166 srcs = ["src/packing.c"],
3167 hdrs = INTERNAL_HDRS,
3168 gcc_copts = xnnpack_gcc_std_copts(),
3169 msvc_copts = xnnpack_msvc_std_copts(),
3170 deps = [
3171 "@FP16",
3172 "@FXdiv",
3173 "@pthreadpool",
3174 ],
3175)
3176
3177xnnpack_cc_library(
3178 name = "packing_test_mode",
3179 srcs = ["src/packing.c"],
3180 hdrs = INTERNAL_HDRS,
3181 copts = [
3182 "-UNDEBUG",
3183 "-DXNN_TEST_MODE=1",
3184 ],
3185 gcc_copts = xnnpack_gcc_std_copts(),
3186 msvc_copts = xnnpack_msvc_std_copts(),
3187 deps = [
3188 "@FP16",
3189 "@FXdiv",
3190 "@pthreadpool",
3191 ],
3192)
3193
3194xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003195 name = "operator_run",
3196 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003197 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003198 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003199 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3200 "//conditions:default": [],
3201 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003202 gcc_copts = xnnpack_gcc_std_copts(),
3203 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003204 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003205 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003206 "@FP16",
3207 "@FXdiv",
3208 "@clog",
3209 "@pthreadpool",
3210 ],
3211)
3212
Chao Mei6ddfc602020-05-13 22:29:36 -07003213xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003214 name = "operator_run_test_mode",
3215 srcs = ["src/operator-run.c"],
3216 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3217 copts = LOGGING_COPTS + [
3218 "-UNDEBUG",
3219 "-DXNN_TEST_MODE=1",
3220 ] + select({
3221 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3222 "//conditions:default": [],
3223 }),
3224 gcc_copts = xnnpack_gcc_std_copts(),
3225 msvc_copts = xnnpack_msvc_std_copts(),
3226 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003227 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003228 "@FP16",
3229 "@FXdiv",
3230 "@clog",
3231 "@pthreadpool",
3232 ],
3233)
3234
3235xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003236 name = "memory_planner",
3237 srcs = ["src/memory-planner.c"],
3238 hdrs = INTERNAL_HDRS,
3239 defines = select({
3240 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3241 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3242 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3243 }),
3244 gcc_copts = xnnpack_gcc_std_copts(),
3245 msvc_copts = xnnpack_msvc_std_copts(),
3246 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003247 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003248 "@pthreadpool",
3249 ],
3250)
3251
Marat Dukhan33fcf782020-05-24 14:27:15 -07003252xnnpack_cc_library(
3253 name = "memory_planner_test_mode",
3254 srcs = ["src/memory-planner.c"],
3255 hdrs = INTERNAL_HDRS,
3256 copts = [
3257 "-UNDEBUG",
3258 "-DXNN_TEST_MODE=1",
3259 ],
3260 defines = select({
3261 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3262 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3263 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3264 }),
3265 gcc_copts = xnnpack_gcc_std_copts(),
3266 msvc_copts = xnnpack_msvc_std_copts(),
3267 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003268 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003269 "@pthreadpool",
3270 ],
3271)
3272
Marat Dukhan08c4a432019-10-03 09:29:21 -07003273cc_library(
3274 name = "enable_assembly",
3275 defines = select({
3276 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3277 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003278 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003279 }),
3280)
3281
Marat Dukhan9de90e02020-06-18 16:04:12 -07003282cc_library(
3283 name = "enable_sparse",
3284 defines = select({
3285 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3286 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003287 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003288 }),
3289)
3290
Marat Dukhancf056b22019-10-07 10:26:29 -07003291xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003292 name = "operators",
3293 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003294 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003295 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003296 ],
3297 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003298 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003299 "-Isrc",
3300 "-Iinclude",
3301 ] + select({
3302 ":debug_build": [],
3303 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003304 }) + select({
3305 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3306 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003307 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003308 gcc_copts = xnnpack_gcc_std_copts(),
3309 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003310 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003311 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003312 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003313 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003314 "@FP16",
3315 "@FXdiv",
3316 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003317 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003318 ],
3319)
3320
Marat Dukhan10a38082020-04-17 03:58:35 -07003321xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003322 name = "operators_test_mode",
3323 srcs = OPERATOR_SRCS + [
3324 "src/memory.c",
3325 "src/operator-delete.c",
3326 ],
3327 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3328 copts = LOGGING_COPTS + [
3329 "-Isrc",
3330 "-Iinclude",
3331 "-UNDEBUG",
3332 "-DXNN_TEST_MODE=1",
3333 ] + select({
3334 ":debug_build": [],
3335 "//conditions:default": xnnpack_min_size_copts(),
3336 }) + select({
3337 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3338 "//conditions:default": [],
3339 }),
3340 gcc_copts = xnnpack_gcc_std_copts(),
3341 msvc_copts = xnnpack_msvc_std_copts(),
3342 deps = [
3343 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003344 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003345 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003346 "@FP16",
3347 "@FXdiv",
3348 "@clog",
3349 "@pthreadpool",
3350 ],
3351)
3352
3353xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003354 name = "XNNPACK",
3355 srcs = [
3356 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003357 "src/runtime.c",
3358 "src/subgraph.c",
3359 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003360 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003361 hdrs = ["include/xnnpack.h"],
3362 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003363 "-Isrc",
3364 "-Iinclude",
3365 ] + select({
3366 ":debug_build": [],
3367 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003368 }) + select({
3369 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3370 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003371 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003372 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003373 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003374 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003375 visibility = xnnpack_visibility(),
3376 deps = [
3377 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003378 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003379 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003380 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003381 ":operator_run",
3382 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003383 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003384 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003385 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003386 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003387 ] + select({
3388 ":emscripten": [],
3389 "//conditions:default": ["@cpuinfo"],
3390 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003391)
3392
Marat Dukhan10a38082020-04-17 03:58:35 -07003393xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003394 name = "XNNPACK_test_mode",
3395 srcs = [
3396 "src/init.c",
3397 "src/runtime.c",
3398 "src/subgraph.c",
3399 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003400 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003401 hdrs = ["include/xnnpack.h"],
3402 copts = LOGGING_COPTS + [
3403 "-Isrc",
3404 "-Iinclude",
3405 "-UNDEBUG",
3406 "-DXNN_TEST_MODE=1",
3407 ] + select({
3408 ":debug_build": [],
3409 "//conditions:default": xnnpack_min_size_copts(),
3410 }) + select({
3411 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3412 "//conditions:default": [],
3413 }),
3414 gcc_copts = xnnpack_gcc_std_copts(),
3415 includes = ["include"],
3416 msvc_copts = xnnpack_msvc_std_copts(),
3417 visibility = xnnpack_visibility(),
3418 deps = [
3419 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003420 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003421 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003422 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003423 ":operator_run_test_mode",
3424 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003425 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003426 "@clog",
3427 "@FP16",
3428 "@pthreadpool",
3429 ] + select({
3430 ":emscripten": [],
3431 "//conditions:default": ["@cpuinfo"],
3432 }),
3433)
3434
3435xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003436 name = "xnnpack_f32",
3437 srcs = [
3438 "src/init.c",
3439 "src/runtime.c",
3440 "src/subgraph.c",
3441 "src/tensor.c",
3442 ] + SUBGRAPH_SRCS,
3443 hdrs = ["include/xnnpack.h"],
3444 copts = LOGGING_COPTS + [
3445 "-Isrc",
3446 "-Iinclude",
3447 ] + select({
3448 ":debug_build": [],
3449 "//conditions:default": xnnpack_min_size_copts(),
3450 }) + select({
3451 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3452 "//conditions:default": [],
3453 }),
3454 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003455 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003456 "XNN_NO_U8_OPERATORS",
3457 "XNN_NO_X8_OPERATORS",
3458 "XNN_NO_F16_OPERATORS",
3459 "XNN_NO_X16_OPERATORS",
3460 ],
3461 gcc_copts = xnnpack_gcc_std_copts(),
3462 includes = ["include"],
3463 msvc_copts = xnnpack_msvc_std_copts(),
3464 visibility = xnnpack_visibility(),
3465 deps = [
3466 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003467 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003468 ":logging_utils",
3469 ":memory_planner",
3470 ":operator_run",
3471 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003472 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003473 "@clog",
3474 "@FP16",
3475 "@pthreadpool",
3476 ] + select({
3477 ":emscripten": [],
3478 "//conditions:default": ["@cpuinfo"],
3479 }),
3480)
3481
3482xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003483 name = "xnnpack_operators_nhwc_f32",
3484 srcs = [
3485 "src/init.c",
3486 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003487 hdrs = ["include/xnnpack.h"],
3488 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003489 "-Isrc",
3490 "-Iinclude",
3491 ] + select({
3492 ":debug_build": [],
3493 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003494 }) + select({
3495 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3496 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003497 }),
3498 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003499 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003500 "XNN_NO_U8_OPERATORS",
3501 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003502 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003503 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003504 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003506 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003507 visibility = xnnpack_visibility(),
3508 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003509 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003510 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511 ":operator_run",
3512 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003513 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003514 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003516 ] + select({
3517 ":emscripten": [],
3518 "//conditions:default": ["@cpuinfo"],
3519 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003520)
3521
Marat Dukhancf056b22019-10-07 10:26:29 -07003522xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003523 name = "bench_utils",
3524 srcs = ["bench/utils.cc"],
3525 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003526 deps = [
3527 "@com_google_benchmark//:benchmark",
3528 "@cpuinfo",
3529 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530)
3531
Frank Barchard7e955972019-10-11 10:34:25 -07003532######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003533
3534xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003535 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003536 srcs = [
3537 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003538 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539 "src/xnnpack/AlignedAllocator.h",
3540 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003541 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003542 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003543)
3544
3545xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003546 name = "f16_igemm_bench",
3547 srcs = [
3548 "bench/f16-igemm.cc",
3549 "bench/conv.h",
3550 "bench/google/conv.h",
3551 "src/xnnpack/AlignedAllocator.h",
3552 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003553 deps = MICROKERNEL_BENCHMARK_DEPS + [
3554 ":indirection",
3555 ":packing",
3556 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003557)
3558
3559xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003560 name = "f16_gemm_bench",
3561 srcs = [
3562 "bench/f16-gemm.cc",
3563 "bench/gemm.h",
3564 "src/xnnpack/AlignedAllocator.h",
3565 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003566 deps = MICROKERNEL_BENCHMARK_DEPS + [
3567 ":packing",
3568 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003569)
3570
3571xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003572 name = "f16_spmm_bench",
3573 srcs = [
3574 "bench/f16-spmm.cc",
3575 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003576 "src/xnnpack/AlignedAllocator.h",
3577 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003578 deps = MICROKERNEL_BENCHMARK_DEPS,
3579)
3580
3581xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 name = "f32_igemm_bench",
3583 srcs = [
3584 "bench/f32-igemm.cc",
3585 "bench/conv.h",
3586 "src/xnnpack/AlignedAllocator.h",
3587 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003588 deps = MICROKERNEL_BENCHMARK_DEPS + [
3589 ":indirection",
3590 ":packing",
3591 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003592)
3593
3594xnnpack_benchmark(
3595 name = "f32_conv_hwc_bench",
3596 srcs = [
3597 "bench/f32-conv-hwc.cc",
3598 "bench/dconv.h",
3599 "src/xnnpack/AlignedAllocator.h",
3600 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003601 deps = MICROKERNEL_BENCHMARK_DEPS + [
3602 ":packing",
3603 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003604)
3605
3606xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003607 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003608 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003609 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003610 "bench/dconv.h",
3611 "src/xnnpack/AlignedAllocator.h",
3612 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003613 deps = MICROKERNEL_BENCHMARK_DEPS + [
3614 ":packing",
3615 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003616)
3617
3618xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003619 name = "f16_dwconv_bench",
3620 srcs = [
3621 "bench/f16-dwconv.cc",
3622 "bench/dwconv.h",
3623 "bench/google/dwconv.h",
3624 "src/xnnpack/AlignedAllocator.h",
3625 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003626 deps = MICROKERNEL_BENCHMARK_DEPS + [
3627 ":indirection",
3628 ":packing",
3629 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003630)
3631
3632xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003633 name = "f32_dwconv_bench",
3634 srcs = [
3635 "bench/f32-dwconv.cc",
3636 "bench/dwconv.h",
3637 "src/xnnpack/AlignedAllocator.h",
3638 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003639 deps = MICROKERNEL_BENCHMARK_DEPS + [
3640 ":indirection",
3641 ":packing",
3642 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643)
3644
3645xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003646 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003647 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003648 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003649 "bench/dwconv.h",
3650 "src/xnnpack/AlignedAllocator.h",
3651 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003652 deps = MICROKERNEL_BENCHMARK_DEPS + [
3653 ":indirection",
3654 ":packing",
3655 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656)
3657
3658xnnpack_benchmark(
3659 name = "f32_gemm_bench",
3660 srcs = [
3661 "bench/f32-gemm.cc",
3662 "bench/gemm.h",
3663 "src/xnnpack/AlignedAllocator.h",
3664 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003665 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003666 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003667)
3668
3669xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003670 name = "f32_hswish_bench",
3671 srcs = [
3672 "bench/f32-hswish.cc",
3673 "src/xnnpack/AlignedAllocator.h",
3674 ] + MICROKERNEL_BENCHMARK_HDRS,
3675 deps = MICROKERNEL_BENCHMARK_DEPS,
3676)
3677
3678xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003679 name = "f32_raddexpminusmax_bench",
3680 srcs = [
3681 "bench/f32-raddexpminusmax.cc",
3682 "src/xnnpack/AlignedAllocator.h",
3683 ] + MICROKERNEL_BENCHMARK_HDRS,
3684 deps = MICROKERNEL_BENCHMARK_DEPS,
3685)
3686
3687xnnpack_benchmark(
3688 name = "f32_raddextexp_bench",
3689 srcs = [
3690 "bench/f32-raddextexp.cc",
3691 "src/xnnpack/AlignedAllocator.h",
3692 ] + MICROKERNEL_BENCHMARK_HDRS,
3693 deps = MICROKERNEL_BENCHMARK_DEPS,
3694)
3695
3696xnnpack_benchmark(
3697 name = "f32_raddstoreexpminusmax_bench",
3698 srcs = [
3699 "bench/f32-raddstoreexpminusmax.cc",
3700 "src/xnnpack/AlignedAllocator.h",
3701 ] + MICROKERNEL_BENCHMARK_HDRS,
3702 deps = MICROKERNEL_BENCHMARK_DEPS,
3703)
3704
3705xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003706 name = "f32_rmax_bench",
3707 srcs = [
3708 "bench/f32-rmax.cc",
3709 "src/xnnpack/AlignedAllocator.h",
3710 ] + MICROKERNEL_BENCHMARK_HDRS,
3711 deps = MICROKERNEL_BENCHMARK_DEPS,
3712)
3713
3714xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003715 name = "f32_sigmoid_bench",
3716 srcs = [
3717 "bench/f32-sigmoid.cc",
3718 "src/xnnpack/AlignedAllocator.h",
3719 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003720 deps = MICROKERNEL_BENCHMARK_DEPS,
3721)
3722
3723xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003724 name = "f32_spmm_bench",
3725 srcs = [
3726 "bench/f32-spmm.cc",
3727 "bench/gemm.h",
3728 "src/xnnpack/AlignedAllocator.h",
3729 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003730 deps = MICROKERNEL_BENCHMARK_DEPS,
3731)
3732
3733xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003734 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003735 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003736 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003737 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003738 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003739 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003740)
3741
3742xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003743 name = "f32_vscaleexpminusmax_bench",
3744 srcs = [
3745 "bench/f32-vscaleexpminusmax.cc",
3746 "src/xnnpack/AlignedAllocator.h",
3747 ] + MICROKERNEL_BENCHMARK_HDRS,
3748 deps = MICROKERNEL_BENCHMARK_DEPS,
3749)
3750
3751xnnpack_benchmark(
3752 name = "f32_vscaleextexp_bench",
3753 srcs = [
3754 "bench/f32-vscaleextexp.cc",
3755 "src/xnnpack/AlignedAllocator.h",
3756 ] + MICROKERNEL_BENCHMARK_HDRS,
3757 deps = MICROKERNEL_BENCHMARK_DEPS,
3758)
3759
3760xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003761 name = "f32_vsqrt_bench",
3762 srcs = [
3763 "bench/f32-vsqrt.cc",
3764 "src/xnnpack/AlignedAllocator.h",
3765 ] + MICROKERNEL_BENCHMARK_HDRS,
3766 deps = MICROKERNEL_BENCHMARK_DEPS,
3767)
3768
3769xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003770 name = "f32_im2col_gemm_bench",
3771 srcs = [
3772 "bench/f32-im2col-gemm.cc",
3773 "bench/conv.h",
3774 "src/xnnpack/AlignedAllocator.h",
3775 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003776 deps = MICROKERNEL_BENCHMARK_DEPS + [
3777 ":im2col",
3778 ":packing",
3779 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003780)
3781
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003782xnnpack_benchmark(
3783 name = "requantization_bench",
3784 srcs = [
3785 "bench/requantization.cc",
3786 "src/xnnpack/requantization-stubs.h",
3787 "src/xnnpack/AlignedAllocator.h",
3788 ] + MICROKERNEL_BENCHMARK_HDRS,
3789 deps = MICROKERNEL_BENCHMARK_DEPS,
3790)
3791
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003792xnnpack_benchmark(
3793 name = "rounding_bench",
3794 srcs = [
3795 "bench/rounding.cc",
3796 "src/xnnpack/math-stubs.h",
3797 "src/xnnpack/AlignedAllocator.h",
3798 ] + MICROKERNEL_BENCHMARK_HDRS,
3799 deps = MICROKERNEL_BENCHMARK_DEPS,
3800)
3801
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802########################### Benchmarks for operators ###########################
3803
3804xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805 name = "average_pooling_bench",
3806 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003807 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003808 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003809 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003810)
3811
3812xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003813 name = "bankers_rounding_bench",
3814 srcs = ["bench/bankers-rounding.cc"],
3815 copts = xnnpack_optional_tflite_copts(),
3816 tags = ["nowin32"],
3817 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3818)
3819
3820xnnpack_benchmark(
3821 name = "ceiling_bench",
3822 srcs = ["bench/ceiling.cc"],
3823 copts = xnnpack_optional_tflite_copts(),
3824 tags = ["nowin32"],
3825 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3826)
3827
3828xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003829 name = "channel_shuffle_bench",
3830 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003831 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003832)
3833
3834xnnpack_benchmark(
3835 name = "convolution_bench",
3836 srcs = ["bench/convolution.cc"],
3837 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_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() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003840)
3841
3842xnnpack_benchmark(
3843 name = "deconvolution_bench",
3844 srcs = ["bench/deconvolution.cc"],
3845 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003846 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003847 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003848)
3849
3850xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003851 name = "floor_bench",
3852 srcs = ["bench/floor.cc"],
3853 copts = xnnpack_optional_tflite_copts(),
3854 tags = ["nowin32"],
3855 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3856)
3857
3858xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003859 name = "global_average_pooling_bench",
3860 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003861 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003862)
3863
3864xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003865 name = "hardswish_bench",
3866 srcs = ["bench/hardswish.cc"],
3867 copts = xnnpack_optional_tflite_copts(),
3868 tags = ["nowin32"],
3869 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3870)
3871
3872xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003873 name = "max_pooling_bench",
3874 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003875 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003876)
3877
3878xnnpack_benchmark(
3879 name = "sigmoid_bench",
3880 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003881 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003882 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003883 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003884)
3885
3886xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003887 name = "prelu_bench",
3888 srcs = ["bench/prelu.cc"],
3889 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003890 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003891 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003892)
3893
3894xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003895 name = "softmax_bench",
3896 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003897 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003898 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003899 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003900)
3901
Marat Dukhan87727142020-06-24 15:24:10 -07003902xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003903 name = "square_root_bench",
3904 srcs = ["bench/square-root.cc"],
3905 copts = xnnpack_optional_tflite_copts(),
3906 tags = ["nowin32"],
3907 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3908)
3909
3910xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003911 name = "truncation_bench",
3912 srcs = ["bench/truncation.cc"],
3913 deps = OPERATOR_BENCHMARK_DEPS,
3914)
3915
Marat Dukhanc068bb62019-10-04 13:24:39 -07003916############################# End-to-end benchmarks ############################
3917
3918cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003919 name = "fp32_mobilenet_v1",
3920 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003921 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003922 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003923 linkstatic = True,
3924 deps = [
3925 ":XNNPACK",
3926 "@pthreadpool",
3927 ],
3928)
3929
3930cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003931 name = "fp16_mobilenet_v1",
3932 srcs = ["models/fp16-mobilenet-v1.cc"],
3933 hdrs = ["models/models.h"],
3934 copts = xnnpack_std_cxxopts(),
3935 linkstatic = True,
3936 deps = [
3937 ":XNNPACK",
3938 "@FP16",
3939 "@pthreadpool",
3940 ],
3941)
3942
3943cc_library(
3944 name = "fp32_mobilenet_v2",
3945 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003946 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003947 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003948 linkstatic = True,
3949 deps = [
3950 ":XNNPACK",
3951 "@pthreadpool",
3952 ],
3953)
3954
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003955cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003956 name = "fp16_mobilenet_v2",
3957 srcs = ["models/fp16-mobilenet-v2.cc"],
3958 hdrs = ["models/models.h"],
3959 copts = xnnpack_std_cxxopts(),
3960 linkstatic = True,
3961 deps = [
3962 ":XNNPACK",
3963 "@FP16",
3964 "@pthreadpool",
3965 ],
3966)
3967
3968cc_library(
3969 name = "fp32_mobilenet_v3_large",
3970 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003971 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003972 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003973 linkstatic = True,
3974 deps = [
3975 ":XNNPACK",
3976 "@pthreadpool",
3977 ],
3978)
3979
3980cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003981 name = "fp32_mobilenet_v3_small",
3982 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003983 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003984 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003985 linkstatic = True,
3986 deps = [
3987 ":XNNPACK",
3988 "@pthreadpool",
3989 ],
3990)
3991
Marat Dukhanc068bb62019-10-04 13:24:39 -07003992xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003993 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003994 srcs = [
3995 "bench/f32-dwconv-e2e.cc",
3996 "bench/end2end.h",
3997 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003998 deps = MICROKERNEL_BENCHMARK_DEPS + [
3999 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004000 ":fp32_mobilenet_v1",
4001 ":fp32_mobilenet_v2",
4002 ":fp32_mobilenet_v3_large",
4003 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004004 ],
4005)
4006
4007xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004008 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004009 srcs = [
4010 "bench/f32-gemm-e2e.cc",
4011 "bench/end2end.h",
4012 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004013 deps = MICROKERNEL_BENCHMARK_DEPS + [
4014 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004015 ":fp32_mobilenet_v1",
4016 ":fp32_mobilenet_v2",
4017 ":fp32_mobilenet_v3_large",
4018 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004019 ],
4020)
4021
4022xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004023 name = "end2end_bench",
4024 srcs = ["bench/end2end.cc"],
4025 deps = [
4026 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004027 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004028 ":fp16_mobilenet_v1",
4029 ":fp16_mobilenet_v2",
4030 ":fp32_mobilenet_v1",
4031 ":fp32_mobilenet_v2",
4032 ":fp32_mobilenet_v3_large",
4033 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004034 "@pthreadpool",
4035 ],
4036)
4037
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004038#################### Accuracy evaluation for math functions ####################
4039
4040xnnpack_benchmark(
4041 name = "f32_exp_eval",
4042 srcs = [
4043 "eval/f32-exp.cc",
4044 "src/xnnpack/AlignedAllocator.h",
4045 ] + ACCURACY_EVAL_HDRS,
4046 deps = ACCURACY_EVAL_DEPS,
4047)
4048
Marat Dukhan515c9772019-10-17 18:07:57 -07004049xnnpack_benchmark(
4050 name = "f32_expminus_eval",
4051 srcs = [
4052 "eval/f32-expminus.cc",
4053 "src/xnnpack/AlignedAllocator.h",
4054 ] + ACCURACY_EVAL_HDRS,
4055 deps = ACCURACY_EVAL_DEPS,
4056)
4057
Marat Dukhan98ba4412019-10-23 02:14:28 -07004058xnnpack_benchmark(
4059 name = "f32_extexp_eval",
4060 srcs = [
4061 "eval/f32-extexp.cc",
4062 "src/xnnpack/AlignedAllocator.h",
4063 ] + ACCURACY_EVAL_HDRS,
4064 deps = ACCURACY_EVAL_DEPS,
4065)
4066
Marat Dukhan8853b822020-05-07 12:19:01 -07004067xnnpack_unit_test(
4068 name = "f32_roundne_eval",
4069 srcs = [
4070 "eval/f32-roundne.cc",
4071 "src/xnnpack/AlignedAllocator.h",
4072 "src/xnnpack/math-stubs.h",
4073 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004074 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004075 deps = MICROKERNEL_TEST_DEPS,
4076)
4077
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004078xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004079 name = "f32_roundd_eval",
4080 srcs = [
4081 "eval/f32-roundd.cc",
4082 "src/xnnpack/AlignedAllocator.h",
4083 "src/xnnpack/math-stubs.h",
4084 ] + MICROKERNEL_TEST_HDRS,
4085 automatic = False,
4086 deps = MICROKERNEL_TEST_DEPS,
4087)
4088
4089xnnpack_unit_test(
4090 name = "f32_roundu_eval",
4091 srcs = [
4092 "eval/f32-roundu.cc",
4093 "src/xnnpack/AlignedAllocator.h",
4094 "src/xnnpack/math-stubs.h",
4095 ] + MICROKERNEL_TEST_HDRS,
4096 automatic = False,
4097 deps = MICROKERNEL_TEST_DEPS,
4098)
4099
4100xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004101 name = "f32_roundz_eval",
4102 srcs = [
4103 "eval/f32-roundz.cc",
4104 "src/xnnpack/AlignedAllocator.h",
4105 "src/xnnpack/math-stubs.h",
4106 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004107 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004108 deps = MICROKERNEL_TEST_DEPS,
4109)
4110
Marat Dukhan346a9e52019-11-15 09:06:30 -08004111xnnpack_benchmark(
4112 name = "f32_sigmoid_eval",
4113 srcs = [
4114 "eval/f32-sigmoid.cc",
4115 "src/xnnpack/AlignedAllocator.h",
4116 ] + ACCURACY_EVAL_HDRS,
4117 deps = ACCURACY_EVAL_DEPS,
4118)
4119
Marat Dukhan84000762020-06-29 18:38:43 -07004120xnnpack_benchmark(
4121 name = "f32_sqrt_eval",
4122 srcs = [
4123 "eval/f32-sqrt.cc",
4124 "src/xnnpack/AlignedAllocator.h",
4125 ] + ACCURACY_EVAL_HDRS,
4126 deps = ACCURACY_EVAL_DEPS + [
4127 ":bench_utils",
4128 ],
4129)
4130
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131######################### Unit tests for micro-kernels #########################
4132
4133xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004134 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004135 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004136 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004137 "test/gemm-microkernel-tester.h",
4138 "src/xnnpack/AlignedAllocator.h",
4139 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004140 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004141)
4142
4143xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004144 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004145 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004146 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004147 "test/spmm-microkernel-tester.h",
4148 "src/xnnpack/AlignedAllocator.h",
4149 ] + MICROKERNEL_TEST_HDRS,
4150 deps = MICROKERNEL_TEST_DEPS,
4151)
4152
4153xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004154 name = "f32_argmaxpool_test",
4155 srcs = [
4156 "test/f32-argmaxpool.cc",
4157 "test/argmaxpool-microkernel-tester.h",
4158 "src/xnnpack/AlignedAllocator.h",
4159 ] + MICROKERNEL_TEST_HDRS,
4160 deps = MICROKERNEL_TEST_DEPS,
4161)
4162
4163xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004164 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004165 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004166 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004167 "test/avgpool-microkernel-tester.h",
4168 "src/xnnpack/AlignedAllocator.h",
4169 ] + MICROKERNEL_TEST_HDRS,
4170 deps = MICROKERNEL_TEST_DEPS,
4171)
4172
4173xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004174 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004175 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004176 "test/f32-ibilinear.cc",
4177 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004178 "src/xnnpack/AlignedAllocator.h",
4179 ] + MICROKERNEL_TEST_HDRS,
4180 deps = MICROKERNEL_TEST_DEPS,
4181)
4182
4183xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004184 name = "f16_clamp_test",
4185 srcs = [
4186 "test/f16-clamp.cc",
4187 "test/clamp-microkernel-tester.h",
4188 ] + MICROKERNEL_TEST_HDRS,
4189 deps = MICROKERNEL_TEST_DEPS,
4190)
4191
4192xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004193 name = "f32_clamp_test",
4194 srcs = [
4195 "test/f32-clamp.cc",
4196 "test/clamp-microkernel-tester.h",
4197 ] + MICROKERNEL_TEST_HDRS,
4198 deps = MICROKERNEL_TEST_DEPS,
4199)
4200
4201xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004202 name = "f32_igemm_test",
4203 srcs = [
4204 "test/f32-igemm.cc",
4205 "test/gemm-microkernel-tester.h",
4206 "src/xnnpack/AlignedAllocator.h",
4207 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004208 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004209)
4210
4211xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004212 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004214 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004215 "test/gemm-microkernel-tester.h",
4216 "src/xnnpack/AlignedAllocator.h",
4217 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004218 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004219)
4220
4221xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004222 name = "f32_igemm_minmax_test",
4223 srcs = [
4224 "test/f32-igemm-minmax.cc",
4225 "test/gemm-microkernel-tester.h",
4226 "src/xnnpack/AlignedAllocator.h",
4227 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004228 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004229)
4230
4231xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004232 name = "f16_igemm_minmax_test",
4233 srcs = [
4234 "test/f16-igemm-minmax.cc",
4235 "test/gemm-microkernel-tester.h",
4236 "src/xnnpack/AlignedAllocator.h",
4237 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004238 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004239)
4240
4241xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004242 name = "f32_conv_hwc_test",
4243 srcs = [
4244 "test/f32-conv-hwc.cc",
4245 "test/conv-hwc-microkernel-tester.h",
4246 "src/xnnpack/AlignedAllocator.h",
4247 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004248 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004249)
4250
4251xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004252 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004254 "test/f32-conv-hwc2chw.cc",
4255 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004256 "src/xnnpack/AlignedAllocator.h",
4257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004258 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004259)
4260
4261xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004262 name = "f32_dwconv_test",
4263 srcs = [
4264 "test/f32-dwconv.cc",
4265 "test/dwconv-microkernel-tester.h",
4266 "src/xnnpack/AlignedAllocator.h",
4267 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004268 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004269)
4270
4271xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004272 name = "f16_dwconv_minmax_test",
4273 srcs = [
4274 "test/f16-dwconv-minmax.cc",
4275 "test/dwconv-microkernel-tester.h",
4276 "src/xnnpack/AlignedAllocator.h",
4277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004278 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004279)
4280
4281xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004282 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004284 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285 "test/dwconv-microkernel-tester.h",
4286 "src/xnnpack/AlignedAllocator.h",
4287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004288 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289)
4290
4291xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004292 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004293 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004294 "test/f32-dwconv-chw.cc",
4295 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004296 "src/xnnpack/AlignedAllocator.h",
4297 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004298 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299)
4300
4301xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004302 name = "f16_gavgpool_minmax_test",
4303 srcs = [
4304 "test/f16-gavgpool-minmax.cc",
4305 "test/gavgpool-microkernel-tester.h",
4306 "src/xnnpack/AlignedAllocator.h",
4307 ] + MICROKERNEL_TEST_HDRS,
4308 deps = MICROKERNEL_TEST_DEPS,
4309)
4310
4311xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004312 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004313 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004314 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004315 "test/gavgpool-microkernel-tester.h",
4316 "src/xnnpack/AlignedAllocator.h",
4317 ] + MICROKERNEL_TEST_HDRS,
4318 deps = MICROKERNEL_TEST_DEPS,
4319)
4320
4321xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004322 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004323 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004324 "test/f32-gavgpool-cw.cc",
4325 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326 "src/xnnpack/AlignedAllocator.h",
4327 ] + MICROKERNEL_TEST_HDRS,
4328 deps = MICROKERNEL_TEST_DEPS,
4329)
4330
4331xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004332 name = "f32_gemm_test",
4333 srcs = [
4334 "test/f32-gemm.cc",
4335 "test/gemm-microkernel-tester.h",
4336 "src/xnnpack/AlignedAllocator.h",
4337 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004338 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004339)
4340
4341xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004342 name = "f32_gemm_relu_test",
4343 srcs = [
4344 "test/f32-gemm-relu.cc",
4345 "test/gemm-microkernel-tester.h",
4346 "src/xnnpack/AlignedAllocator.h",
4347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004348 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004349)
4350
4351xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004352 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004353 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004354 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 "test/gemm-microkernel-tester.h",
4356 "src/xnnpack/AlignedAllocator.h",
4357 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004358 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004359)
4360
4361xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004362 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004363 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004364 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 "test/gemm-microkernel-tester.h",
4366 "src/xnnpack/AlignedAllocator.h",
4367 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004368 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004369)
4370
4371xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004372 name = "f16_hswish_test",
4373 srcs = [
4374 "test/f16-hswish.cc",
4375 "test/hswish-microkernel-tester.h",
4376 ] + MICROKERNEL_TEST_HDRS,
4377 deps = MICROKERNEL_TEST_DEPS,
4378)
4379
4380xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004381 name = "f32_hswish_test",
4382 srcs = [
4383 "test/f32-hswish.cc",
4384 "test/hswish-microkernel-tester.h",
4385 ] + MICROKERNEL_TEST_HDRS,
4386 deps = MICROKERNEL_TEST_DEPS,
4387)
4388
4389xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004390 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004392 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393 "test/maxpool-microkernel-tester.h",
4394 ] + MICROKERNEL_TEST_HDRS,
4395 deps = MICROKERNEL_TEST_DEPS,
4396)
4397
4398xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004399 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004400 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004401 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004402 "test/avgpool-microkernel-tester.h",
4403 "src/xnnpack/AlignedAllocator.h",
4404 ] + MICROKERNEL_TEST_HDRS,
4405 deps = MICROKERNEL_TEST_DEPS,
4406)
4407
4408xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004409 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004410 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004411 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004412 "test/gemm-microkernel-tester.h",
4413 "src/xnnpack/AlignedAllocator.h",
4414 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004415 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416)
4417
4418xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004419 name = "f16_prelu_test",
4420 srcs = [
4421 "test/f16-prelu.cc",
4422 "test/prelu-microkernel-tester.h",
4423 "src/xnnpack/AlignedAllocator.h",
4424 ] + MICROKERNEL_TEST_HDRS,
4425 deps = MICROKERNEL_TEST_DEPS,
4426)
4427
4428xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004429 name = "f32_prelu_test",
4430 srcs = [
4431 "test/f32-prelu.cc",
4432 "test/prelu-microkernel-tester.h",
4433 "src/xnnpack/AlignedAllocator.h",
4434 ] + MICROKERNEL_TEST_HDRS,
4435 deps = MICROKERNEL_TEST_DEPS,
4436)
4437
4438xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004439 name = "f32_raddexpminusmax_test",
4440 srcs = [
4441 "test/f32-raddexpminusmax.cc",
4442 "test/raddexpminusmax-microkernel-tester.h",
4443 ] + MICROKERNEL_TEST_HDRS,
4444 deps = MICROKERNEL_TEST_DEPS,
4445)
4446
4447xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004448 name = "f32_raddextexp_test",
4449 srcs = [
4450 "test/f32-raddextexp.cc",
4451 "test/raddextexp-microkernel-tester.h",
4452 ] + MICROKERNEL_TEST_HDRS,
4453 deps = MICROKERNEL_TEST_DEPS,
4454)
4455
4456xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004457 name = "f32_raddstoreexpminusmax_test",
4458 srcs = [
4459 "test/f32-raddstoreexpminusmax.cc",
4460 "test/raddstoreexpminusmax-microkernel-tester.h",
4461 ] + MICROKERNEL_TEST_HDRS,
4462 deps = MICROKERNEL_TEST_DEPS,
4463)
4464
4465xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466 name = "f32_rmax_test",
4467 srcs = [
4468 "test/f32-rmax.cc",
4469 "test/rmax-microkernel-tester.h",
4470 ] + MICROKERNEL_TEST_HDRS,
4471 deps = MICROKERNEL_TEST_DEPS,
4472)
4473
4474xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004475 name = "f32_sigmoid_test",
4476 srcs = [
4477 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004478 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004479 ] + MICROKERNEL_TEST_HDRS,
4480 deps = MICROKERNEL_TEST_DEPS,
4481)
4482
4483xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004484 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004485 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004486 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004487 "test/spmm-microkernel-tester.h",
4488 "src/xnnpack/AlignedAllocator.h",
4489 ] + MICROKERNEL_TEST_HDRS,
4490 deps = MICROKERNEL_TEST_DEPS,
4491)
4492
4493xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004494 name = "f16_vadd_minmax_test",
4495 srcs = [
4496 "test/f16-vadd-minmax.cc",
4497 "test/vbinary-microkernel-tester.h",
4498 ] + MICROKERNEL_TEST_HDRS,
4499 deps = MICROKERNEL_TEST_DEPS,
4500)
4501
4502xnnpack_unit_test(
4503 name = "f16_vaddc_minmax_test",
4504 srcs = [
4505 "test/f16-vaddc-minmax.cc",
4506 "test/vbinaryc-microkernel-tester.h",
4507 ] + MICROKERNEL_TEST_HDRS,
4508 deps = MICROKERNEL_TEST_DEPS,
4509)
4510
4511xnnpack_unit_test(
4512 name = "f16_vdiv_minmax_test",
4513 srcs = [
4514 "test/f16-vdiv-minmax.cc",
4515 "test/vbinary-microkernel-tester.h",
4516 ] + MICROKERNEL_TEST_HDRS,
4517 deps = MICROKERNEL_TEST_DEPS,
4518)
4519
4520xnnpack_unit_test(
4521 name = "f16_vdivc_minmax_test",
4522 srcs = [
4523 "test/f16-vdivc-minmax.cc",
4524 "test/vbinaryc-microkernel-tester.h",
4525 ] + MICROKERNEL_TEST_HDRS,
4526 deps = MICROKERNEL_TEST_DEPS,
4527)
4528
4529xnnpack_unit_test(
4530 name = "f16_vrdivc_minmax_test",
4531 srcs = [
4532 "test/f16-vrdivc-minmax.cc",
4533 "test/vbinaryc-microkernel-tester.h",
4534 ] + MICROKERNEL_TEST_HDRS,
4535 deps = MICROKERNEL_TEST_DEPS,
4536)
4537
4538xnnpack_unit_test(
4539 name = "f16_vmax_test",
4540 srcs = [
4541 "test/f16-vmax.cc",
4542 "test/vbinary-microkernel-tester.h",
4543 ] + MICROKERNEL_TEST_HDRS,
4544 deps = MICROKERNEL_TEST_DEPS,
4545)
4546
4547xnnpack_unit_test(
4548 name = "f16_vmaxc_test",
4549 srcs = [
4550 "test/f16-vmaxc.cc",
4551 "test/vbinaryc-microkernel-tester.h",
4552 ] + MICROKERNEL_TEST_HDRS,
4553 deps = MICROKERNEL_TEST_DEPS,
4554)
4555
4556xnnpack_unit_test(
4557 name = "f16_vmin_test",
4558 srcs = [
4559 "test/f16-vmin.cc",
4560 "test/vbinary-microkernel-tester.h",
4561 ] + MICROKERNEL_TEST_HDRS,
4562 deps = MICROKERNEL_TEST_DEPS,
4563)
4564
4565xnnpack_unit_test(
4566 name = "f16_vminc_test",
4567 srcs = [
4568 "test/f16-vminc.cc",
4569 "test/vbinaryc-microkernel-tester.h",
4570 ] + MICROKERNEL_TEST_HDRS,
4571 deps = MICROKERNEL_TEST_DEPS,
4572)
4573
4574xnnpack_unit_test(
4575 name = "f16_vmul_minmax_test",
4576 srcs = [
4577 "test/f16-vmul-minmax.cc",
4578 "test/vbinary-microkernel-tester.h",
4579 ] + MICROKERNEL_TEST_HDRS,
4580 deps = MICROKERNEL_TEST_DEPS,
4581)
4582
4583xnnpack_unit_test(
4584 name = "f16_vmulc_minmax_test",
4585 srcs = [
4586 "test/f16-vmulc-minmax.cc",
4587 "test/vbinaryc-microkernel-tester.h",
4588 ] + MICROKERNEL_TEST_HDRS,
4589 deps = MICROKERNEL_TEST_DEPS,
4590)
4591
4592xnnpack_unit_test(
4593 name = "f16_vsub_minmax_test",
4594 srcs = [
4595 "test/f16-vsub-minmax.cc",
4596 "test/vbinary-microkernel-tester.h",
4597 ] + MICROKERNEL_TEST_HDRS,
4598 deps = MICROKERNEL_TEST_DEPS,
4599)
4600
4601xnnpack_unit_test(
4602 name = "f16_vsubc_minmax_test",
4603 srcs = [
4604 "test/f16-vsubc-minmax.cc",
4605 "test/vbinaryc-microkernel-tester.h",
4606 ] + MICROKERNEL_TEST_HDRS,
4607 deps = MICROKERNEL_TEST_DEPS,
4608)
4609
4610xnnpack_unit_test(
4611 name = "f16_vrsubc_minmax_test",
4612 srcs = [
4613 "test/f16-vrsubc-minmax.cc",
4614 "test/vbinaryc-microkernel-tester.h",
4615 ] + MICROKERNEL_TEST_HDRS,
4616 deps = MICROKERNEL_TEST_DEPS,
4617)
4618
4619xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004620 name = "f32_vabs_test",
4621 srcs = [
4622 "test/f32-vabs.cc",
4623 "test/vunary-microkernel-tester.h",
4624 ] + MICROKERNEL_TEST_HDRS,
4625 deps = MICROKERNEL_TEST_DEPS,
4626)
4627
4628xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004629 name = "f32_vadd_test",
4630 srcs = [
4631 "test/f32-vadd.cc",
4632 "test/vbinary-microkernel-tester.h",
4633 ] + MICROKERNEL_TEST_HDRS,
4634 deps = MICROKERNEL_TEST_DEPS,
4635)
4636
4637xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004638 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004639 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004640 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004641 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004642 ] + MICROKERNEL_TEST_HDRS,
4643 deps = MICROKERNEL_TEST_DEPS,
4644)
4645
4646xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004647 name = "f32_vaddc_test",
4648 srcs = [
4649 "test/f32-vaddc.cc",
4650 "test/vbinaryc-microkernel-tester.h",
4651 ] + MICROKERNEL_TEST_HDRS,
4652 deps = MICROKERNEL_TEST_DEPS,
4653)
4654
4655xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004656 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004657 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004658 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004659 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004660 ] + MICROKERNEL_TEST_HDRS,
4661 deps = MICROKERNEL_TEST_DEPS,
4662)
4663
4664xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004665 name = "f32_vdiv_test",
4666 srcs = [
4667 "test/f32-vdiv.cc",
4668 "test/vbinary-microkernel-tester.h",
4669 ] + MICROKERNEL_TEST_HDRS,
4670 deps = MICROKERNEL_TEST_DEPS,
4671)
4672
4673xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004674 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004675 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004676 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004677 "test/vbinary-microkernel-tester.h",
4678 ] + MICROKERNEL_TEST_HDRS,
4679 deps = MICROKERNEL_TEST_DEPS,
4680)
4681
4682xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004683 name = "f32_vdivc_test",
4684 srcs = [
4685 "test/f32-vdivc.cc",
4686 "test/vbinaryc-microkernel-tester.h",
4687 ] + MICROKERNEL_TEST_HDRS,
4688 deps = MICROKERNEL_TEST_DEPS,
4689)
4690
4691xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004692 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004693 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004694 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004695 "test/vbinaryc-microkernel-tester.h",
4696 ] + MICROKERNEL_TEST_HDRS,
4697 deps = MICROKERNEL_TEST_DEPS,
4698)
4699
4700xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004701 name = "f32_vrdivc_test",
4702 srcs = [
4703 "test/f32-vrdivc.cc",
4704 "test/vbinaryc-microkernel-tester.h",
4705 ] + MICROKERNEL_TEST_HDRS,
4706 deps = MICROKERNEL_TEST_DEPS,
4707)
4708
4709xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004710 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004711 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004712 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004713 "test/vbinaryc-microkernel-tester.h",
4714 ] + MICROKERNEL_TEST_HDRS,
4715 deps = MICROKERNEL_TEST_DEPS,
4716)
4717
4718xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004719 name = "f32_vmax_test",
4720 srcs = [
4721 "test/f32-vmax.cc",
4722 "test/vbinary-microkernel-tester.h",
4723 ] + MICROKERNEL_TEST_HDRS,
4724 deps = MICROKERNEL_TEST_DEPS,
4725)
4726
4727xnnpack_unit_test(
4728 name = "f32_vmaxc_test",
4729 srcs = [
4730 "test/f32-vmaxc.cc",
4731 "test/vbinaryc-microkernel-tester.h",
4732 ] + MICROKERNEL_TEST_HDRS,
4733 deps = MICROKERNEL_TEST_DEPS,
4734)
4735
4736xnnpack_unit_test(
4737 name = "f32_vmin_test",
4738 srcs = [
4739 "test/f32-vmin.cc",
4740 "test/vbinary-microkernel-tester.h",
4741 ] + MICROKERNEL_TEST_HDRS,
4742 deps = MICROKERNEL_TEST_DEPS,
4743)
4744
4745xnnpack_unit_test(
4746 name = "f32_vminc_test",
4747 srcs = [
4748 "test/f32-vminc.cc",
4749 "test/vbinaryc-microkernel-tester.h",
4750 ] + MICROKERNEL_TEST_HDRS,
4751 deps = MICROKERNEL_TEST_DEPS,
4752)
4753
4754xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004755 name = "f32_vmul_test",
4756 srcs = [
4757 "test/f32-vmul.cc",
4758 "test/vbinary-microkernel-tester.h",
4759 ] + MICROKERNEL_TEST_HDRS,
4760 deps = MICROKERNEL_TEST_DEPS,
4761)
4762
4763xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004764 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004765 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004766 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004767 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004768 ] + MICROKERNEL_TEST_HDRS,
4769 deps = MICROKERNEL_TEST_DEPS,
4770)
4771
4772xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004773 name = "f32_vmulc_test",
4774 srcs = [
4775 "test/f32-vmulc.cc",
4776 "test/vbinaryc-microkernel-tester.h",
4777 ] + MICROKERNEL_TEST_HDRS,
4778 deps = MICROKERNEL_TEST_DEPS,
4779)
4780
4781xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004782 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004783 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004784 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004785 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004786 ] + MICROKERNEL_TEST_HDRS,
4787 deps = MICROKERNEL_TEST_DEPS,
4788)
4789
4790xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004791 name = "f16_vmulcaddc_minmax_test",
4792 srcs = [
4793 "test/f16-vmulcaddc-minmax.cc",
4794 "test/vmulcaddc-microkernel-tester.h",
4795 "src/xnnpack/AlignedAllocator.h",
4796 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004797 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004798)
4799
4800xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004801 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004802 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004803 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004804 "test/vmulcaddc-microkernel-tester.h",
4805 "src/xnnpack/AlignedAllocator.h",
4806 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004807 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004808)
4809
4810xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004811 name = "f32_vlrelu_test",
4812 srcs = [
4813 "test/f32-vlrelu.cc",
4814 "test/vunary-microkernel-tester.h",
4815 ] + MICROKERNEL_TEST_HDRS,
4816 deps = MICROKERNEL_TEST_DEPS,
4817)
4818
4819xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004820 name = "f32_vneg_test",
4821 srcs = [
4822 "test/f32-vneg.cc",
4823 "test/vunary-microkernel-tester.h",
4824 ] + MICROKERNEL_TEST_HDRS,
4825 deps = MICROKERNEL_TEST_DEPS,
4826)
4827
4828xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004829 name = "f32_vrndne_test",
4830 srcs = [
4831 "test/f32-vrndne.cc",
4832 "test/vunary-microkernel-tester.h",
4833 ] + MICROKERNEL_TEST_HDRS,
4834 deps = MICROKERNEL_TEST_DEPS,
4835)
4836
4837xnnpack_unit_test(
4838 name = "f32_vrndz_test",
4839 srcs = [
4840 "test/f32-vrndz.cc",
4841 "test/vunary-microkernel-tester.h",
4842 ] + MICROKERNEL_TEST_HDRS,
4843 deps = MICROKERNEL_TEST_DEPS,
4844)
4845
4846xnnpack_unit_test(
4847 name = "f32_vrndu_test",
4848 srcs = [
4849 "test/f32-vrndu.cc",
4850 "test/vunary-microkernel-tester.h",
4851 ] + MICROKERNEL_TEST_HDRS,
4852 deps = MICROKERNEL_TEST_DEPS,
4853)
4854
4855xnnpack_unit_test(
4856 name = "f32_vrndd_test",
4857 srcs = [
4858 "test/f32-vrndd.cc",
4859 "test/vunary-microkernel-tester.h",
4860 ] + MICROKERNEL_TEST_HDRS,
4861 deps = MICROKERNEL_TEST_DEPS,
4862)
4863
4864xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004865 name = "f32_vscale_test",
4866 srcs = [
4867 "test/f32-vscale.cc",
4868 "test/vscale-microkernel-tester.h",
4869 ] + MICROKERNEL_TEST_HDRS,
4870 deps = MICROKERNEL_TEST_DEPS,
4871)
4872
4873xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004874 name = "f32_vscaleexpminusmax_test",
4875 srcs = [
4876 "test/f32-vscaleexpminusmax.cc",
4877 "test/vscaleexpminusmax-microkernel-tester.h",
4878 ] + MICROKERNEL_TEST_HDRS,
4879 deps = MICROKERNEL_TEST_DEPS,
4880)
4881
4882xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004883 name = "f32_vscaleextexp_test",
4884 srcs = [
4885 "test/f32-vscaleextexp.cc",
4886 "test/vscaleextexp-microkernel-tester.h",
4887 ] + MICROKERNEL_TEST_HDRS,
4888 deps = MICROKERNEL_TEST_DEPS,
4889)
4890
4891xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004892 name = "f32_vsqr_test",
4893 srcs = [
4894 "test/f32-vsqr.cc",
4895 "test/vunary-microkernel-tester.h",
4896 ] + MICROKERNEL_TEST_HDRS,
4897 deps = MICROKERNEL_TEST_DEPS,
4898)
4899
4900xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004901 name = "f32_vsqrdiff_test",
4902 srcs = [
4903 "test/f32-vsqrdiff.cc",
4904 "test/vbinary-microkernel-tester.h",
4905 ] + MICROKERNEL_TEST_HDRS,
4906 deps = MICROKERNEL_TEST_DEPS,
4907)
4908
4909xnnpack_unit_test(
4910 name = "f32_vsqrdiffc_test",
4911 srcs = [
4912 "test/f32-vsqrdiffc.cc",
4913 "test/vbinaryc-microkernel-tester.h",
4914 ] + MICROKERNEL_TEST_HDRS,
4915 deps = MICROKERNEL_TEST_DEPS,
4916)
4917
4918xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004919 name = "f32_vsqrt_test",
4920 srcs = [
4921 "test/f32-vsqrt.cc",
4922 "test/vunary-microkernel-tester.h",
4923 ] + MICROKERNEL_TEST_HDRS,
4924 deps = MICROKERNEL_TEST_DEPS,
4925)
4926
4927xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004928 name = "f32_vsub_test",
4929 srcs = [
4930 "test/f32-vsub.cc",
4931 "test/vbinary-microkernel-tester.h",
4932 ] + MICROKERNEL_TEST_HDRS,
4933 deps = MICROKERNEL_TEST_DEPS,
4934)
4935
4936xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004937 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004938 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004939 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004940 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004941 ] + MICROKERNEL_TEST_HDRS,
4942 deps = MICROKERNEL_TEST_DEPS,
4943)
4944
4945xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004946 name = "f32_vsubc_test",
4947 srcs = [
4948 "test/f32-vsubc.cc",
4949 "test/vbinaryc-microkernel-tester.h",
4950 ] + MICROKERNEL_TEST_HDRS,
4951 deps = MICROKERNEL_TEST_DEPS,
4952)
4953
4954xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004955 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004956 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004957 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004958 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004959 ] + MICROKERNEL_TEST_HDRS,
4960 deps = MICROKERNEL_TEST_DEPS,
4961)
4962
4963xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004964 name = "f32_vrsubc_test",
4965 srcs = [
4966 "test/f32-vrsubc.cc",
4967 "test/vbinaryc-microkernel-tester.h",
4968 ] + MICROKERNEL_TEST_HDRS,
4969 deps = MICROKERNEL_TEST_DEPS,
4970)
4971
4972xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004973 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004974 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004975 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004976 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004977 ] + MICROKERNEL_TEST_HDRS,
4978 deps = MICROKERNEL_TEST_DEPS,
4979)
4980
4981xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004982 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004984 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004985 "test/avgpool-microkernel-tester.h",
4986 "src/xnnpack/AlignedAllocator.h",
4987 ] + MICROKERNEL_TEST_HDRS,
4988 deps = MICROKERNEL_TEST_DEPS,
4989)
4990
4991xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004992 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004994 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004995 "test/gemm-microkernel-tester.h",
4996 "src/xnnpack/AlignedAllocator.h",
4997 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004998 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999)
5000
5001xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005002 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005003 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005004 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005 "test/dwconv-microkernel-tester.h",
5006 "src/xnnpack/AlignedAllocator.h",
5007 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005008 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009)
5010
5011xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005012 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005013 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005014 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015 "test/gavgpool-microkernel-tester.h",
5016 "src/xnnpack/AlignedAllocator.h",
5017 ] + MICROKERNEL_TEST_HDRS,
5018 deps = MICROKERNEL_TEST_DEPS,
5019)
5020
5021xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005022 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005023 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005024 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005025 "test/gemm-microkernel-tester.h",
5026 "src/xnnpack/AlignedAllocator.h",
5027 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005028 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005029)
5030
5031xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005032 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005033 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005034 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005035 "test/vadd-microkernel-tester.h",
5036 ] + MICROKERNEL_TEST_HDRS,
5037 deps = MICROKERNEL_TEST_DEPS,
5038)
5039
5040xnnpack_unit_test(
5041 name = "u8_clamp_test",
5042 srcs = [
5043 "test/u8-clamp.cc",
5044 "test/clamp-microkernel-tester.h",
5045 ] + MICROKERNEL_TEST_HDRS,
5046 deps = MICROKERNEL_TEST_DEPS,
5047)
5048
5049xnnpack_unit_test(
5050 name = "u8_lut32norm_test",
5051 srcs = [
5052 "test/u8-lut32norm.cc",
5053 "test/lut-norm-microkernel-tester.h",
5054 ] + MICROKERNEL_TEST_HDRS,
5055 deps = MICROKERNEL_TEST_DEPS,
5056)
5057
5058xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005059 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005060 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005061 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005062 "test/maxpool-microkernel-tester.h",
5063 ] + MICROKERNEL_TEST_HDRS,
5064 deps = MICROKERNEL_TEST_DEPS,
5065)
5066
5067xnnpack_unit_test(
5068 name = "u8_rmax_test",
5069 srcs = [
5070 "test/u8-rmax.cc",
5071 "test/rmax-microkernel-tester.h",
5072 ] + MICROKERNEL_TEST_HDRS,
5073 deps = MICROKERNEL_TEST_DEPS,
5074)
5075
5076xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005077 name = "x32_fill_test",
5078 srcs = [
5079 "test/x32-fill.cc",
5080 "test/fill-microkernel-tester.h",
5081 ] + MICROKERNEL_TEST_HDRS,
5082 deps = MICROKERNEL_TEST_DEPS,
5083)
5084
5085xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005086 name = "x32_packx_test",
5087 srcs = [
5088 "test/x32-packx.cc",
5089 "test/pack-microkernel-tester.h",
5090 "src/xnnpack/AlignedAllocator.h",
5091 ] + MICROKERNEL_TEST_HDRS,
5092 deps = MICROKERNEL_TEST_DEPS,
5093)
5094
5095xnnpack_unit_test(
5096 name = "x32_pad_test",
5097 srcs = [
5098 "test/x32-pad.cc",
5099 "test/pad-microkernel-tester.h",
5100 ] + MICROKERNEL_TEST_HDRS,
5101 deps = MICROKERNEL_TEST_DEPS,
5102)
5103
5104xnnpack_unit_test(
5105 name = "x32_unpool_test",
5106 srcs = [
5107 "test/x32-unpool.cc",
5108 "test/unpool-microkernel-tester.h",
5109 ] + MICROKERNEL_TEST_HDRS,
5110 deps = MICROKERNEL_TEST_DEPS,
5111)
5112
5113xnnpack_unit_test(
5114 name = "x32_zip_test",
5115 srcs = [
5116 "test/x32-zip.cc",
5117 "test/zip-microkernel-tester.h",
5118 ] + MICROKERNEL_TEST_HDRS,
5119 deps = MICROKERNEL_TEST_DEPS,
5120)
5121
5122xnnpack_unit_test(
5123 name = "x8_lut_test",
5124 srcs = [
5125 "test/x8-lut.cc",
5126 "test/lut-microkernel-tester.h",
5127 ] + MICROKERNEL_TEST_HDRS,
5128 deps = MICROKERNEL_TEST_DEPS,
5129)
5130
5131xnnpack_unit_test(
5132 name = "x8_zip_test",
5133 srcs = [
5134 "test/x8-zip.cc",
5135 "test/zip-microkernel-tester.h",
5136 ] + MICROKERNEL_TEST_HDRS,
5137 deps = MICROKERNEL_TEST_DEPS,
5138)
5139
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005140xnnpack_unit_test(
5141 name = "requantization_test",
5142 srcs = [
5143 "src/xnnpack/requantization-stubs.h",
5144 "test/requantization.cc",
5145 "test/requantization-tester.h",
5146 ] + MICROKERNEL_TEST_HDRS,
5147 deps = MICROKERNEL_TEST_DEPS,
5148)
5149
Marat Dukhan20c3b922020-03-10 03:45:06 -07005150########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005151
5152xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005153 name = "operator_size_test",
5154 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005155 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005156)
5157
Marat Dukhan20c3b922020-03-10 03:45:06 -07005158xnnpack_binary(
5159 name = "subgraph_size_test",
5160 srcs = ["test/subgraph-size.c"],
5161 deps = [":XNNPACK"],
5162)
5163
5164########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165
5166xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005167 name = "abs_nc_test",
5168 srcs = [
5169 "test/abs-nc.cc",
5170 "test/abs-operator-tester.h",
5171 ],
5172 deps = OPERATOR_TEST_DEPS,
5173)
5174
5175xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005176 name = "add_nd_test",
5177 srcs = [
5178 "test/add-nd.cc",
5179 "test/binary-elementwise-operator-tester.h",
5180 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005181 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005182)
5183
5184xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005185 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005186 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005187 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005188 "test/argmax-pooling-operator-tester.h",
5189 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005190 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005191)
5192
5193xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005194 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005195 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005196 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197 "test/average-pooling-operator-tester.h",
5198 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005199 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200)
5201
5202xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005203 name = "bankers_rounding_nc_test",
5204 srcs = [
5205 "test/bankers-rounding-nc.cc",
5206 "test/bankers-rounding-operator-tester.h",
5207 ],
5208 deps = OPERATOR_TEST_DEPS,
5209)
5210
5211xnnpack_unit_test(
5212 name = "ceiling_nc_test",
5213 srcs = [
5214 "test/ceiling-nc.cc",
5215 "test/ceiling-operator-tester.h",
5216 ],
5217 deps = OPERATOR_TEST_DEPS,
5218)
5219
5220xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005221 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005222 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005223 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224 "test/channel-shuffle-operator-tester.h",
5225 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005226 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005227)
5228
5229xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005230 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005231 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005232 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005233 "test/clamp-operator-tester.h",
5234 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005235 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005236)
5237
5238xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005239 name = "constant_pad_nd_test",
5240 srcs = [
5241 "test/constant-pad-nd.cc",
5242 "test/constant-pad-operator-tester.h",
5243 ],
5244 deps = OPERATOR_TEST_DEPS,
5245)
5246
5247xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005248 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005249 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005250 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251 "test/convolution-operator-tester.h",
5252 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005253 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005254)
5255
5256xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005257 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005259 "test/convolution-nchw.cc",
5260 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005261 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005262 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263)
5264
5265xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005266 name = "copy_nc_test",
5267 srcs = [
5268 "test/copy-nc.cc",
5269 "test/copy-operator-tester.h",
5270 ],
5271 deps = OPERATOR_TEST_DEPS,
5272)
5273
5274xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005275 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005276 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005277 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005278 "test/deconvolution-operator-tester.h",
5279 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005280 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005281)
5282
5283xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005284 name = "divide_nd_test",
5285 srcs = [
5286 "test/binary-elementwise-operator-tester.h",
5287 "test/divide-nd.cc",
5288 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005289 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005290)
5291
5292xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005293 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005295 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005296 "test/fully-connected-operator-tester.h",
5297 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005299)
5300
5301xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005302 name = "floor_nc_test",
5303 srcs = [
5304 "test/floor-nc.cc",
5305 "test/floor-operator-tester.h",
5306 ],
5307 deps = OPERATOR_TEST_DEPS,
5308)
5309
5310xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005311 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005313 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005314 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005315 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005316 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317)
5318
5319xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005320 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005322 "test/global-average-pooling-ncw.cc",
5323 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005324 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005325 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005326)
5327
5328xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005329 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005330 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005331 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005332 "test/hardswish-operator-tester.h",
5333 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005334 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005335)
5336
5337xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005338 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005340 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005341 "test/leaky-relu-operator-tester.h",
5342 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005343 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005344)
5345
5346xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005347 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005348 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005349 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005350 "test/max-pooling-operator-tester.h",
5351 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005352 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005353)
5354
5355xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005356 name = "maximum_nd_test",
5357 srcs = [
5358 "test/binary-elementwise-operator-tester.h",
5359 "test/maximum-nd.cc",
5360 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005361 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005362)
5363
5364xnnpack_unit_test(
5365 name = "minimum_nd_test",
5366 srcs = [
5367 "test/binary-elementwise-operator-tester.h",
5368 "test/minimum-nd.cc",
5369 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005371)
5372
5373xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005374 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005375 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005376 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005377 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005378 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005379 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005380)
5381
5382xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005383 name = "negate_nc_test",
5384 srcs = [
5385 "test/negate-nc.cc",
5386 "test/negate-operator-tester.h",
5387 ],
5388 deps = OPERATOR_TEST_DEPS,
5389)
5390
5391xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005392 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005393 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005394 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005395 "test/prelu-operator-tester.h",
5396 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005397 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005398)
5399
5400xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005401 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005402 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005403 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005404 "test/resize-bilinear-operator-tester.h",
5405 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005406 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005407)
5408
5409xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005410 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005412 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 "test/sigmoid-operator-tester.h",
5414 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005415 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005416)
5417
5418xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005419 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005420 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005421 "test/softmax-nc.cc",
5422 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005423 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005424 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005425)
5426
5427xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005428 name = "square_nc_test",
5429 srcs = [
5430 "test/square-nc.cc",
5431 "test/square-operator-tester.h",
5432 ],
5433 deps = OPERATOR_TEST_DEPS,
5434)
5435
5436xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005437 name = "square_root_nc_test",
5438 srcs = [
5439 "test/square-root-nc.cc",
5440 "test/square-root-operator-tester.h",
5441 ],
5442 deps = OPERATOR_TEST_DEPS,
5443)
5444
5445xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005446 name = "squared_difference_nd_test",
5447 srcs = [
5448 "test/binary-elementwise-operator-tester.h",
5449 "test/squared-difference-nd.cc",
5450 ],
5451 deps = OPERATOR_TEST_DEPS,
5452)
5453
5454xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005455 name = "subtract_nd_test",
5456 srcs = [
5457 "test/binary-elementwise-operator-tester.h",
5458 "test/subtract-nd.cc",
5459 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005460 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005461)
5462
5463xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005464 name = "truncation_nc_test",
5465 srcs = [
5466 "test/truncation-nc.cc",
5467 "test/truncation-operator-tester.h",
5468 ],
5469 deps = OPERATOR_TEST_DEPS,
5470)
5471
5472xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005473 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005474 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005475 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005476 "test/unpooling-operator-tester.h",
5477 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005478 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005479)
5480
Chao Mei6ddfc602020-05-13 22:29:36 -07005481############################### Misc unit tests ###############################
5482
5483xnnpack_unit_test(
5484 name = "memory_planner_test",
5485 srcs = [
5486 "test/memory-planner-test.cc",
5487 ],
5488 deps = [
5489 ":XNNPACK",
5490 ":memory_planner",
5491 ],
5492)
5493
Marat Dukhan08c4a432019-10-03 09:29:21 -07005494############################# Build configurations #############################
5495
Marat Dukhanb8642352019-10-30 15:43:02 -07005496# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005497config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005498 name = "xnn_enable_assembly_explicit_true",
5499 define_values = {"xnn_enable_assembly": "true"},
5500)
5501
5502# Disables usage of assembly kernels.
5503config_setting(
5504 name = "xnn_enable_assembly_explicit_false",
5505 define_values = {"xnn_enable_assembly": "false"},
5506)
5507
Marat Dukhan9de90e02020-06-18 16:04:12 -07005508# Enables usage of sparse inference.
5509config_setting(
5510 name = "xnn_enable_sparse_explicit_true",
5511 define_values = {"xnn_enable_sparse": "true"},
5512)
5513
5514# Disables usage of sparse inference.
5515config_setting(
5516 name = "xnn_enable_sparse_explicit_false",
5517 define_values = {"xnn_enable_sparse": "false"},
5518)
5519
Marat Dukhan05702cf2020-03-26 15:41:33 -07005520# Disables usage of HMP-aware optimizations.
5521config_setting(
5522 name = "xnn_enable_hmp_explicit_false",
5523 define_values = {"xnn_enable_hmp": "false"},
5524)
5525
Chao Mei6ddfc602020-05-13 22:29:36 -07005526# Enable usage of optimized memory allocation
5527config_setting(
5528 name = "xnn_enable_memopt_explicit_true",
5529 define_values = {"xnnpack_enable_memopt": "true"},
5530)
5531
5532# Disable usage of optimized memory allocation
5533config_setting(
5534 name = "xnn_enable_memopt_explicit_false",
5535 define_values = {"xnnpack_enable_memopt": "false"},
5536)
5537
Marat Dukhanb8642352019-10-30 15:43:02 -07005538# Builds with -c dbg
5539config_setting(
5540 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005541 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005542 "compilation_mode": "dbg",
5543 },
5544)
5545
5546# Builds with -c opt
5547config_setting(
5548 name = "optimized_build",
5549 values = {
5550 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005551 },
5552)
5553
5554config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005555 name = "linux_k8",
5556 values = {"cpu": "k8"},
5557)
5558
5559config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005560 name = "linux_arm",
5561 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005562)
5563
5564config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005565 name = "linux_armeabi",
5566 values = {"cpu": "armeabi"},
5567)
5568
5569config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005570 name = "linux_armhf",
5571 values = {"cpu": "armhf"},
5572)
5573
5574config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005575 name = "linux_armv7a",
5576 values = {"cpu": "armv7a"},
5577)
5578
5579config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005580 name = "linux_aarch64",
5581 values = {"cpu": "aarch64"},
5582)
5583
5584config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005585 name = "android",
5586 values = {"crosstool_top": "//external:android/crosstool"},
5587)
5588
5589config_setting(
5590 name = "android_armv7",
5591 values = {
5592 "crosstool_top": "//external:android/crosstool",
5593 "cpu": "armeabi-v7a",
5594 },
5595)
5596
5597config_setting(
5598 name = "android_arm64",
5599 values = {
5600 "crosstool_top": "//external:android/crosstool",
5601 "cpu": "arm64-v8a",
5602 },
5603)
5604
5605config_setting(
5606 name = "android_x86",
5607 values = {
5608 "crosstool_top": "//external:android/crosstool",
5609 "cpu": "x86",
5610 },
5611)
5612
5613config_setting(
5614 name = "android_x86_64",
5615 values = {
5616 "crosstool_top": "//external:android/crosstool",
5617 "cpu": "x86_64",
5618 },
5619)
5620
5621config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005622 name = "windows_x86_64",
5623 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005624)
5625
5626config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005627 name = "windows_x86_64_clang",
5628 values = {
5629 "compiler": "clang-cl",
5630 "cpu": "x64_windows",
5631 },
5632)
5633
5634config_setting(
5635 name = "windows_x86_64_mingw",
5636 values = {
5637 "compiler": "mingw-gcc",
5638 "cpu": "x64_windows",
5639 },
5640)
5641
5642config_setting(
5643 name = "windows_x86_64_msys",
5644 values = {
5645 "compiler": "msys-gcc",
5646 "cpu": "x64_windows",
5647 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005648)
5649
5650config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005651 name = "macos_x86_64",
5652 values = {
5653 "apple_platform_type": "macos",
5654 "cpu": "darwin",
5655 },
5656)
5657
5658config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005659 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005660 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005661)
5662
5663config_setting(
5664 name = "emscripten_wasm",
5665 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005666 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005667 "cpu": "wasm",
5668 },
5669)
5670
5671config_setting(
5672 name = "emscripten_wasmsimd",
5673 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005674 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005675 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005676 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005677 },
5678)
5679
5680config_setting(
5681 name = "emscripten_asmjs",
5682 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005683 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005684 "cpu": "asmjs",
5685 },
5686)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005687
5688config_setting(
5689 name = "ios_armv7",
5690 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005691 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005692 "cpu": "ios_armv7",
5693 },
5694)
5695
5696config_setting(
5697 name = "ios_arm64",
5698 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005699 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005700 "cpu": "ios_arm64",
5701 },
5702)
5703
5704config_setting(
5705 name = "ios_arm64e",
5706 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005707 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005708 "cpu": "ios_arm64e",
5709 },
5710)
5711
5712config_setting(
5713 name = "ios_x86",
5714 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005715 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005716 "cpu": "ios_i386",
5717 },
5718)
5719
5720config_setting(
5721 name = "ios_x86_64",
5722 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005723 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005724 "cpu": "ios_x86_64",
5725 },
5726)
5727
5728config_setting(
5729 name = "watchos_armv7k",
5730 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005731 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005732 "cpu": "watchos_armv7k",
5733 },
5734)
5735
5736config_setting(
5737 name = "watchos_arm64_32",
5738 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005739 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005740 "cpu": "watchos_arm64_32",
5741 },
5742)
5743
5744config_setting(
5745 name = "watchos_x86",
5746 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005747 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005748 "cpu": "watchos_i386",
5749 },
5750)
5751
5752config_setting(
5753 name = "watchos_x86_64",
5754 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005755 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005756 "cpu": "watchos_x86_64",
5757 },
5758)
5759
5760config_setting(
5761 name = "tvos_arm64",
5762 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005763 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005764 "cpu": "tvos_arm64",
5765 },
5766)
5767
5768config_setting(
5769 name = "tvos_x86_64",
5770 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005771 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005772 "cpu": "tvos_x86_64",
5773 },
5774)