blob: 8bae09ea97a814067372005a8543dc9920e60a6d [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 Dukhan99936602020-04-11 16:47:01 -0700357 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
358 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700359 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700360 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
361 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700362 "src/q8-gemm/2x2-minmax-scalar.c",
363 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700364 "src/q8-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 Dukhan802808c2020-06-16 11:01:17 -0700548 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-arm.c",
549 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat-arm.c",
550 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-x86.c",
551 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700552 "src/f32-gemm/gen/4x2c4-relu-wasmsimd-arm.c",
553 "src/f32-gemm/gen/4x2c4-relu-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700554 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
555 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
556 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
557 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
558 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
559 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
560 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
561 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
562 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
563 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
564 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
565 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
566 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
567 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
568 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
569 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
570 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
571 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
572 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
573 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
574 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
575 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
576 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
577 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
578 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
579 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
580 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
581 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
582 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
583 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700584 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
585 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700586 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
587 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
588 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
589 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
590 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
591 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
592 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
593 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
594 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
595 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
596 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
597 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
598 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
599 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
600 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
601 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
602 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
603 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
604 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
605 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
606 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
607 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
608 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
609 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
610 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
611 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
612 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
613 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
614 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
615 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700616 "src/f32-ibilinear/gen/wasmsimd-c4.c",
617 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700618 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
619 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
620 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700621 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700622 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-arm.c",
623 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat-arm.c",
624 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-x86.c",
625 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700626 "src/f32-igemm/gen/4x2c4-relu-wasmsimd-arm.c",
627 "src/f32-igemm/gen/4x2c4-relu-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700628 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
629 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
630 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
631 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
632 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
633 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
634 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
635 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
636 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
637 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
638 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
639 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
640 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
641 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
642 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
643 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
644 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
645 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
646 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
647 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
648 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
649 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
650 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
651 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
652 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
653 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
654 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
655 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
656 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
657 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700658 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
659 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700660 "src/f32-hswish/gen/wasmsimd-arm-x4.c",
661 "src/f32-hswish/gen/wasmsimd-arm-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700662 "src/f32-hswish/gen/wasmsimd-arm-x16.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700663 "src/f32-hswish/gen/wasmsimd-x86-x4.c",
664 "src/f32-hswish/gen/wasmsimd-x86-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700665 "src/f32-hswish/gen/wasmsimd-x86-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700666 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
667 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700668 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
669 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
670 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
671 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700672 "src/f32-rmax/wasmsimd-arm.c",
673 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700674 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
675 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
676 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
677 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700678 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
679 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700680 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
681 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
682 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
683 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700684 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
685 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700686 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
687 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
688 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
689 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700690 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
691 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700692 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
693 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
694 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
695 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700696 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
697 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700698 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
699 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
700 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
701 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700702 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
703 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
704 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
705 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700706 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
707 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
708 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
709 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700710 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
711 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
712 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
713 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700714 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
715 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
716 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
717 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700718 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
719 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700720 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
721 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
722 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
723 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700724 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
725 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700726 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
727 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
728 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
729 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700730 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
731 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700732 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
733 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
734 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
735 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700736 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
737 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700738 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
739 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700740 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
741 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700742 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
743 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
744 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
745 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700746 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
747 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700748 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
749 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
750 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
751 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700752 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
753 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan72b399a2020-06-28 22:41:38 -0700754 "src/f32-vrnd/gen/vrndne-wasmsimd-x4.c",
755 "src/f32-vrnd/gen/vrndne-wasmsimd-x8.c",
756 "src/f32-vrnd/gen/vrndz-wasmsimd-x4.c",
757 "src/f32-vrnd/gen/vrndz-wasmsimd-x8.c",
758 "src/f32-vrnd/gen/vrndu-wasmsimd-x4.c",
759 "src/f32-vrnd/gen/vrndu-wasmsimd-x8.c",
760 "src/f32-vrnd/gen/vrndd-wasmsimd-x4.c",
761 "src/f32-vrnd/gen/vrndd-wasmsimd-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700762 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
763 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700764 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
765 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
766 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
767 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
768 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
769 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700770 "src/math/roundne-wasmsimd-addsub.c",
771 "src/math/roundd-wasmsimd-addsub.c",
772 "src/math/roundu-wasmsimd-addsub.c",
773 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700774 "src/requantization/fp32-wasmsimd.c",
775]
776
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800777PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800778 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800779 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800780 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700781 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
782 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700783 "src/f32-clamp/gen/psimd-x4.c",
784 "src/f32-clamp/gen/psimd-x8.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700785 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700786 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700787 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700788 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700789 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700790 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
791 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
792 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
793 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
794 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
795 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
796 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
797 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
798 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
799 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
800 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
801 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700802 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700803 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
804 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700805 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
806 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
807 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
808 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
809 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
810 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
811 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
812 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
813 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
814 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
815 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
816 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
817 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
818 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
819 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
820 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
821 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
822 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
823 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800824 "src/f32-hswish/gen/psimd-x4.c",
825 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700826 "src/f32-ibilinear/gen/psimd-c4.c",
827 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700828 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
829 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
830 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
831 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
832 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
833 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
834 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
835 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
836 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
837 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700838 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700839 "src/f32-spmm/gen/4x1-minmax-psimd.c",
840 "src/f32-spmm/gen/8x1-minmax-psimd.c",
841 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700842 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
843 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700844 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800845 "src/f32-prelu/gen/psimd-2x4.c",
846 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800847 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700848 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
849 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
850 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
851 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
852 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
853 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
854 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
855 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800856 "src/f32-vbinary/gen/vmax-psimd-x4.c",
857 "src/f32-vbinary/gen/vmax-psimd-x8.c",
858 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
859 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
860 "src/f32-vbinary/gen/vmin-psimd-x4.c",
861 "src/f32-vbinary/gen/vmin-psimd-x8.c",
862 "src/f32-vbinary/gen/vminc-psimd-x4.c",
863 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700864 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
865 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
866 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
867 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
868 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
869 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
870 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
871 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700872 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
873 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
874 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
875 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700876 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
877 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
878 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
879 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700880 "src/f32-vlrelu/gen/psimd-x4.c",
881 "src/f32-vlrelu/gen/psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700882 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
883 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700884 "src/f32-vunary/gen/vabs-psimd-x4.c",
885 "src/f32-vunary/gen/vabs-psimd-x8.c",
886 "src/f32-vunary/gen/vneg-psimd-x4.c",
887 "src/f32-vunary/gen/vneg-psimd-x8.c",
888 "src/f32-vunary/gen/vsqr-psimd-x4.c",
889 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700890 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700891 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700892 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700893 "src/x32-unpool/psimd.c",
894 "src/x32-zip/x2-psimd.c",
895 "src/x32-zip/x3-psimd.c",
896 "src/x32-zip/x4-psimd.c",
897 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700898 "src/requantization/precise-psimd.c",
899 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700900]
901
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800902PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800903 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
904 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
905 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
906 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
907 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
908 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
909 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
910 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
911 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
912 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
913 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
914 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800915 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
916 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
917 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
918 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
919 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
920 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700921 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
922 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
923 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
924 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
925 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
926 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
927 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
928 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700929 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700930 "src/math/roundd-psimd-addsub.c",
931 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700932 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800933 "src/math/sigmoid-psimd-p5-div.c",
934]
935
Marat Dukhan08c4a432019-10-03 09:29:21 -0700936# ISA-specific micro-kernels
937NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700938 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
939 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700940 "src/f32-clamp/gen/neon-x4.c",
941 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700942 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
943 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700944 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
945 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700946 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
947 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700948 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
949 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700950 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
951 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
952 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
953 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700954 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
955 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
956 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
957 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700958 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
959 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
960 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
961 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700962 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700963 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
964 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700965 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
966 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
967 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
968 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
969 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
970 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
971 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
972 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
973 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
974 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
975 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
976 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
977 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
978 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
979 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
980 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
981 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
982 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
983 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
984 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
985 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
986 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
987 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
988 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
989 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
990 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
991 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
992 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
993 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
994 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
995 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800996 "src/f32-hswish/gen/neon-x4.c",
997 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700998 "src/f32-ibilinear/gen/neon-c4.c",
999 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001000 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1001 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1002 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1003 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1004 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1005 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1006 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1007 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1008 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1009 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1010 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1011 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1012 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1013 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1014 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1015 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001016 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1017 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1018 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001019 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1020 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001021 "src/f32-prelu/gen/neon-2x4.c",
1022 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001023 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1024 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1025 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1026 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1027 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1028 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1029 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1030 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1031 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1032 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1033 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1034 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1035 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1036 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1037 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1038 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1039 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1040 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1041 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1042 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1043 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1044 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1045 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1046 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001047 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001048 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001049 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1050 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1051 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1052 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1053 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1054 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1055 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1056 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1057 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1058 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1059 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1060 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1061 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1062 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1063 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1064 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1065 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1066 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001067 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1068 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1069 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1070 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001071 "src/f32-vbinary/gen/vmax-neon-x4.c",
1072 "src/f32-vbinary/gen/vmax-neon-x8.c",
1073 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1074 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1075 "src/f32-vbinary/gen/vmin-neon-x4.c",
1076 "src/f32-vbinary/gen/vmin-neon-x8.c",
1077 "src/f32-vbinary/gen/vminc-neon-x4.c",
1078 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001079 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1080 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1081 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1082 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1083 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1084 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001085 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1086 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1087 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1088 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001089 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1090 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1091 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1092 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001093 "src/f32-vlrelu/gen/neon-x4.c",
1094 "src/f32-vlrelu/gen/neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001095 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1096 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001097 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1098 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1099 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1100 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1101 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1102 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1103 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1104 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001105 "src/f32-vunary/gen/vabs-neon-x4.c",
1106 "src/f32-vunary/gen/vabs-neon-x8.c",
1107 "src/f32-vunary/gen/vneg-neon-x4.c",
1108 "src/f32-vunary/gen/vneg-neon-x8.c",
1109 "src/f32-vunary/gen/vsqr-neon-x4.c",
1110 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001111 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
1112 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001113 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001114 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
1115 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001116 "src/q8-gemm/4x8-minmax-neon.c",
1117 "src/q8-gemm/8x8-minmax-neon.c",
1118 "src/q8-igemm/4x8-minmax-neon.c",
1119 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001120 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001121 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001122 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001123 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001124 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001125 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001126 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001127 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001128 "src/x32-zip/x2-neon.c",
1129 "src/x32-zip/x3-neon.c",
1130 "src/x32-zip/x4-neon.c",
1131 "src/x32-zip/xm-neon.c",
1132 "src/x8-zip/x2-neon.c",
1133 "src/x8-zip/x3-neon.c",
1134 "src/x8-zip/x4-neon.c",
1135 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001136 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001137 "src/math/roundd-neon-addsub.c",
1138 "src/math/roundd-neon-cvt.c",
1139 "src/math/roundu-neon-addsub.c",
1140 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001141 "src/math/roundz-neon-addsub.c",
1142 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001143 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001144 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1145 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1146 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1147 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1148 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1149 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001150 "src/math/sqrt-neon-nr1rsqrts.c",
1151 "src/math/sqrt-neon-nr2rsqrts.c",
1152 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001153 "src/requantization/precise-neon.c",
1154 "src/requantization/fp32-neon.c",
1155 "src/requantization/q31-neon.c",
1156 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001157]
1158
1159NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001160 "src/f32-ibilinear/gen/neonfma-c4.c",
1161 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001162 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1163 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1164 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1165 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1166 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1167 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1168 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1169 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1170 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001171 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1172 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1173 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1174 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001175 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1176 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1177 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1178 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001179 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1180 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1181 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1182 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001183 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1184 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1185 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1186 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1187 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1188 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1189 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1190 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1191 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1192 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1193 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1194 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1195 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1196 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1197 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1198 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1199 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1200 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001201 "src/f32-hswish/gen/neonfma-x4.c",
1202 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001203 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1204 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001205 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1206 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1207 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1208 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1209 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1210 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1211 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1212 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1213 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1214 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1215 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1216 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1217 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1218 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1219 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1220 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1221 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1223 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1224 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1225 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1226 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1227 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1228 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001229 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1260 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1261 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1262 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1263 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1264 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1265 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1269 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1270 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1271 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1272 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1273 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1275 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1276 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1277 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1278 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1279 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1280 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1281 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1282 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001283 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1284 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001285 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1286 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1287 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1288 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1289 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1290 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1291 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1292 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1293 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1294 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1295 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1296 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1297 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1298 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1299 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1300 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1301 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1302 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1303 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1304 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001305 "src/math/exp-neonfma-lut64-p2.c",
1306 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001307 "src/math/expminus-neonfma-lut2048-p1.c",
1308 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001309 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001310 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1311 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1312 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1313 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1314 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1315 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1316 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1317 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1318 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1319 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1320 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1321 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1322 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1323 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1324 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1325 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1326 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1327 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001328 "src/math/sqrt-neonfma-nr1fma.c",
1329 "src/math/sqrt-neonfma-nr2fma.c",
1330 "src/math/sqrt-neonfma-nr3fma.c",
1331 "src/math/sqrt-neonfma-nr2fma1adj.c",
1332 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001333]
1334
1335AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001336 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1337 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1338 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1339 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1340 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1341 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001342 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1343 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1344 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1345 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1346 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1347 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1348 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1349 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1350 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1351 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1352 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1353 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1354 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1355 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1356 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1357 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1358 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1359 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1360 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1361 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001362 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1363 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001364 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1365 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001366 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1367 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001368 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1369 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001370 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1371 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1372 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1373 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1374 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001375 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1380 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1381 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1382 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1383 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1384 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1385 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1386 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1387 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1388 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1389 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1390 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1391 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1392 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001393 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1394 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1395 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1396 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1397 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1398 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1399 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1400 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1401 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1402 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1403 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1404 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1405 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1406 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1407 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1408 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1409 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1410 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001411 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1412 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001413 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1414 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1415 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1416 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1417 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1418 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001419]
1420
Marat Dukhan8853b822020-05-07 12:19:01 -07001421NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001422 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1423 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1424 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1425 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1426 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1427 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1428 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1429 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001430 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001431 "src/math/roundd-neonv8.c",
1432 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001433 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001434]
1435
Marat Dukhan08c4a432019-10-03 09:29:21 -07001436AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001437 "src/f16-clamp/gen/neonfp16arith-x16.c",
1438 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001439 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1440 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1441 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1442 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1443 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1444 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1445 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1446 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1447 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1448 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1449 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1450 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001451 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1452 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001453 "src/f16-hswish/gen/neonfp16arith-x16.c",
1454 "src/f16-hswish/gen/neonfp16arith-x8.c",
1455 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1456 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001457 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1458 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1459 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1460 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1461 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1462 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1463 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1464 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1465 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1466 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1467 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1468 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1469 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1470 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1471 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1472 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1473 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1474 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1475 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1476 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1477 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1478 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1479 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1480 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1481 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1482 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1483 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1484 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001485 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1486 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001487 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1488 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1489 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1490 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1491 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1492 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1493 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1494 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001495 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1496 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1497 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1498 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001499 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1500 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1501 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1502 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1503 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1504 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1505 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1506 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1507 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1508 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1509 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1510 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001511 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1512 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1513 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1514 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1515 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1516 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1517 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1518 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001519]
1520
1521SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001522 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1523 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001524 "src/f32-clamp/gen/sse-x4.c",
1525 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001526 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1527 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001528 "src/f32-dwconv-chw/3x3p1-sse.c",
1529 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001530 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1531 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1532 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1533 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1534 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1535 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1536 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1537 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1538 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1539 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1540 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1541 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001542 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001543 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1544 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1546 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1547 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1548 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1549 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1550 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1551 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1552 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1553 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1554 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1555 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1556 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1557 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001558 "src/f32-hswish/gen/sse-x4.c",
1559 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001560 "src/f32-ibilinear/gen/sse-c4.c",
1561 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001562 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1563 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1564 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1565 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1566 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1567 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1568 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001569 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1570 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1571 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001572 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001573 "src/f32-prelu/gen/sse-2x4.c",
1574 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001575 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001576 "src/f32-spmm/gen/4x1-minmax-sse.c",
1577 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001578 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001579 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1580 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1581 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1582 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1583 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1584 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1585 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1586 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001587 "src/f32-vbinary/gen/vmax-sse-x4.c",
1588 "src/f32-vbinary/gen/vmax-sse-x8.c",
1589 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1590 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1591 "src/f32-vbinary/gen/vmin-sse-x4.c",
1592 "src/f32-vbinary/gen/vmin-sse-x8.c",
1593 "src/f32-vbinary/gen/vminc-sse-x4.c",
1594 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001595 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1596 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1597 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1598 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1599 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1600 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1601 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1602 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001603 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1604 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1605 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1606 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001607 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1608 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1609 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1610 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001611 "src/f32-vlrelu/gen/sse-x4.c",
1612 "src/f32-vlrelu/gen/sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001613 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1614 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001615 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1616 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001617 "src/f32-vunary/gen/vabs-sse-x4.c",
1618 "src/f32-vunary/gen/vabs-sse-x8.c",
1619 "src/f32-vunary/gen/vneg-sse-x4.c",
1620 "src/f32-vunary/gen/vneg-sse-x8.c",
1621 "src/f32-vunary/gen/vsqr-sse-x4.c",
1622 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001623 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001624 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001625 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001626 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001627 "src/math/roundd-sse-addsub.c",
1628 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001629 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001630 "src/math/sqrt-sse-hh1mac.c",
1631 "src/math/sqrt-sse-nr1mac.c",
1632 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001633]
1634
1635SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001636 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1637 "src/f32-argmaxpool/4x-sse2-c4.c",
1638 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001639 "src/f32-prelu/gen/sse2-2x4.c",
1640 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001641 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1642 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1643 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1644 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1645 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1646 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1647 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1648 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1649 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1650 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1651 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1652 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001653 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001654 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001655 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001656 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001657 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1658 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001659 "src/f32-vlrelu/gen/sse2-x4.c",
1660 "src/f32-vlrelu/gen/sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001661 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1662 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1663 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1664 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1665 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1666 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1667 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1668 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001669 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1670 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001671 "src/q8-igemm/4x4c2-minmax-sse2.c",
1672 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001673 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1674 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001675 "src/q8-gemm/2x4c8-minmax-sse2.c",
1676 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001677 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001678 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001679 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001680 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001681 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001682 "src/x32-zip/x2-sse2.c",
1683 "src/x32-zip/x3-sse2.c",
1684 "src/x32-zip/x4-sse2.c",
1685 "src/x32-zip/xm-sse2.c",
1686 "src/x8-zip/x2-sse2.c",
1687 "src/x8-zip/x3-sse2.c",
1688 "src/x8-zip/x4-sse2.c",
1689 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001690 "src/math/exp-sse2-p5.c",
1691 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001692 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001693 "src/math/roundd-sse2-cvt.c",
1694 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001695 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001696 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001697 "src/requantization/precise-sse2.c",
1698 "src/requantization/fp32-sse2.c",
1699 "src/requantization/q31-sse2.c",
1700 "src/requantization/gemmlowp-sse2.c",
1701]
1702
1703SSSE3_UKERNELS = [
1704 "src/requantization/precise-ssse3.c",
1705 "src/requantization/q31-ssse3.c",
1706 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001707]
1708
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001709SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001710 "src/f32-prelu/gen/sse41-2x4.c",
1711 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001712 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001713 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001714 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001715 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001716 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1717 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001718 "src/f32-vlrelu/gen/sse41-x4.c",
1719 "src/f32-vlrelu/gen/sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001720 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1721 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1722 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1723 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1724 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1725 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1726 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1727 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001728 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001729 "src/math/roundd-sse41.c",
1730 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001731 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001732 "src/requantization/precise-sse4.c",
1733 "src/requantization/q31-sse4.c",
1734 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001735]
1736
Marat Dukhan08c4a432019-10-03 09:29:21 -07001737AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001738 "src/f32-clamp/gen/avx-x8.c",
1739 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001740 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1741 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1742 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1743 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1744 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1745 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1746 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1747 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1748 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1749 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1750 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1751 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1752 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1753 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1754 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1755 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1756 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1757 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1758 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1759 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1760 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1761 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1762 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1763 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1764 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1765 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1766 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1767 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1768 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1769 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001770 "src/f32-hswish/gen/avx-x8.c",
1771 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001772 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1773 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1774 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1775 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1776 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1777 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1778 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1779 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1780 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001781 "src/f32-prelu/gen/avx-2x8.c",
1782 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001783 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001784 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1785 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1786 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1787 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1788 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1789 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1790 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1791 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001792 "src/f32-vbinary/gen/vmax-avx-x8.c",
1793 "src/f32-vbinary/gen/vmax-avx-x16.c",
1794 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1795 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1796 "src/f32-vbinary/gen/vmin-avx-x8.c",
1797 "src/f32-vbinary/gen/vmin-avx-x16.c",
1798 "src/f32-vbinary/gen/vminc-avx-x8.c",
1799 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001800 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1801 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1802 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1803 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1804 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1805 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1806 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1807 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001808 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1809 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1810 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1811 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001812 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1813 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1814 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1815 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001816 "src/f32-vlrelu/gen/avx-x8.c",
1817 "src/f32-vlrelu/gen/avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001818 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001819 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1820 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1821 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1822 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1823 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1824 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1825 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1826 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001827 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1828 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001829 "src/f32-vunary/gen/vabs-avx-x8.c",
1830 "src/f32-vunary/gen/vabs-avx-x16.c",
1831 "src/f32-vunary/gen/vneg-avx-x8.c",
1832 "src/f32-vunary/gen/vneg-avx-x16.c",
1833 "src/f32-vunary/gen/vsqr-avx-x8.c",
1834 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835]
1836
Marat Dukhanfda12b82019-11-21 12:27:59 -08001837FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001838 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1839 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1840 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1841 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1842 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1843 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1844 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1845 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1846 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1847 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1848 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1849 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1850 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1851 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1852 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1853 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1854 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1855 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1856 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1857 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1858 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1859 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1860 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1861 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1862 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1863 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1864 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1865 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1866 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1867 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1868 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1869 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1870 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1871 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1872 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1873 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1874 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1875 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1876 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1877 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001878 "src/f32-hswish/gen/fma3-x8.c",
1879 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001880 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1881 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1882 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1883 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1884 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1885 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1886 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1887 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1888 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1889 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1890 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1891 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1892 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1893 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001894 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1895 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1896 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1897 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1898 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1899 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1900 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1901 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001902 "src/math/sqrt-fma3-nr1fma.c",
1903 "src/math/sqrt-fma3-nr2fma.c",
1904 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001905]
1906
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001907AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001908 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1909 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1910 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1911 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1912 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1913 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1914 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1915 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1916 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1917 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1918 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1919 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1920 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1921 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1922 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1923 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1924 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1925 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1926 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1927 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1928 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1929 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1930 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1931 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1932 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1933 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1934 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1935 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1936 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1937 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1938 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1939 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1940 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1941 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1942 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1943 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001944 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1945 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1946 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1947 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1948 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1949 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1950 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1951 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1952 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1953 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1954 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1955 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1956 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1957 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1958 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1959 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1960 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1961 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1962 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1963 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1964 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1965 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1966 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1967 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1968 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1969 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1970 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1971 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1972 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1973 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001974 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1975 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1976 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1977 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1978 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1979 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1980 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1981 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1982 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1983 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1984 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1985 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1986 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1987 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1988 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1989 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1990 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1991 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1992 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1993 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1994 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1995 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1996 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1997 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001998 "src/math/exp-avx2-p5.c",
1999 "src/math/exp-avx2-perm-p3.c",
2000 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002001 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002002 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002003 "src/math/sigmoid-avx2-rr2-p5-div.c",
2004 "src/math/sigmoid-avx2-rr1-p5-div.c",
2005 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2006 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2007 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2008 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002009]
2010
Marat Dukhan08c4a432019-10-03 09:29:21 -07002011AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002012 "src/f32-clamp/gen/avx512f-x16.c",
2013 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002014 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2015 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2016 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2017 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2018 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2019 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2020 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2021 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2022 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2023 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2024 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2025 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2026 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2027 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2028 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2029 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2030 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2031 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2032 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2033 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2034 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2035 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2036 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2037 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08002038 "src/f32-hswish/gen/avx512f-x16.c",
2039 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002040 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2041 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2042 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2043 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2044 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2045 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002046 "src/f32-prelu/gen/avx512f-2x16.c",
2047 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002048 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2049 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2050 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2051 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2052 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2053 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2054 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2055 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2056 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2057 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2058 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2059 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2060 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2061 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2062 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2063 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2064 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2065 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2066 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2067 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2068 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2069 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2070 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2071 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2072 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2073 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2074 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2075 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2076 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2077 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2078 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2079 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2080 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2081 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2082 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2083 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002084 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002085 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2086 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2087 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2088 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2089 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2090 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2091 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2092 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002093 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2094 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2095 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2096 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2097 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2098 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2099 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2100 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002101 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2102 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2103 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2104 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2105 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2106 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2107 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2108 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002109 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2110 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2111 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2112 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002113 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2114 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2115 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2116 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07002117 "src/f32-vlrelu/gen/avx512f-x16.c",
2118 "src/f32-vlrelu/gen/avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002119 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002120 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2121 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2122 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2123 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2124 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2125 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2126 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2127 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2128 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2129 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2130 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2131 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2132 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2133 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2134 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2135 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2136 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2137 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2138 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2139 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2140 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2141 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2142 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2143 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002144 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2145 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2146 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2147 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2148 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2149 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2150 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2151 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002152 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2153 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2154 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2155 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2156 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2157 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2158 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2159 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002160 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2161 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2162 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2163 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2164 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2165 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002166 "src/math/exp-avx512f-p5-scalef.c",
2167 "src/math/exp-avx512f-p5.c",
2168 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002169 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002170 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002171 "src/math/sqrt-avx512f-nr1fma.c",
2172 "src/math/sqrt-avx512f-nr2fma.c",
2173 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002174]
2175
2176AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002177 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002178 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002179 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2180 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2181 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2182 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002183 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2184 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2185 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2186 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002187 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2188 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2189 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2190 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002191]
2192
2193AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002194 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002195 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2196 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002197 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002198 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2199 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002200 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2201 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2202 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002203 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002204 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2205 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2206 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002207 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002208 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2209 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002210 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002211 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2212 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2213 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2214 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2215 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2216 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2217 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2218 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2219 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2220 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2221 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2222 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2223 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2224 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2225 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2226 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2227 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2228 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002229 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2230 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002231 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002232 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2233 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2234 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2235 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2236 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2237 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2238 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2239 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2240 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2241 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2242 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2243 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2244 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2245 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2246 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2247 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2248 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2249 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002250 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2251 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2252 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2253 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2254 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2255 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2256 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2257 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2258 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2259 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2260 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2261 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2262 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2263 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2264 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2265 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2266 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2267 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268]
2269
Marat Dukhan1b354632020-03-23 12:50:22 -07002270INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002271 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002272 "src/xnnpack/argmaxpool.h",
2273 "src/xnnpack/avgpool.h",
2274 "src/xnnpack/clamp.h",
2275 "src/xnnpack/common.h",
2276 "src/xnnpack/conv.h",
2277 "src/xnnpack/dwconv.h",
2278 "src/xnnpack/gavgpool.h",
2279 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002280 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002281 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002282 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002283 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002284 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002285 "src/xnnpack/lut.h",
2286 "src/xnnpack/math.h",
2287 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002288 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289 "src/xnnpack/packx.h",
2290 "src/xnnpack/pad.h",
2291 "src/xnnpack/params.h",
2292 "src/xnnpack/pavgpool.h",
2293 "src/xnnpack/ppmm.h",
2294 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002295 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002296 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002297 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002298 "src/xnnpack/rmax.h",
2299 "src/xnnpack/scalar-utils.h",
2300 "src/xnnpack/spmm.h",
2301 "src/xnnpack/unpool.h",
2302 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002303 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002304 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002305 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002306 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002307 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002308 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002309 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002310]
2311
2312INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002313 "include/xnnpack.h",
2314 "src/xnnpack/allocator.h",
2315 "src/xnnpack/compute.h",
2316 "src/xnnpack/im2col.h",
2317 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002318 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002319 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002320 "src/xnnpack/operator.h",
2321 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002322 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002323 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002324 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002325 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002326]
2327
Marat Dukhan1b354632020-03-23 12:50:22 -07002328ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002329 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002330]
2331
Marat Dukhan1b354632020-03-23 12:50:22 -07002332MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002333 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002334 "include/xnnpack.h",
2335]
2336
Marat Dukhan1b354632020-03-23 12:50:22 -07002337MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002338 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002339 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340 "src/xnnpack/requantization.h",
2341 "include/xnnpack.h",
2342]
2343
2344OPERATOR_TEST_PARAMS_HDRS = [
2345 "src/xnnpack/params.h",
2346 "src/xnnpack/common.h",
2347]
2348
2349WEIGHTS_PACK_HDRS = [
2350 "src/xnnpack/pack.h",
2351 "src/xnnpack/operator.h",
2352 "src/xnnpack/compute.h",
2353]
2354
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002355LOGGING_COPTS = select({
2356 # No logging in optimized mode
2357 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2358 # Full logging in debug mode
2359 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2360 # Error-only logging in default (fastbuild) mode
2361 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2362})
2363
Marat Dukhan3b59de22020-06-03 20:15:19 -07002364LOGGING_SRCS = select({
2365 # No logging in optimized mode
2366 ":optimized_build": [],
2367 "//conditions:default": [
2368 "src/operator-strings.c",
2369 "src/subgraph-strings.c",
2370 ],
2371})
2372
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002373LOGGING_HDRS = [
2374 "src/xnnpack/log.h",
2375]
2376
Marat Dukhan08c4a432019-10-03 09:29:21 -07002377xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002378 name = "tables",
2379 srcs = TABLE_SRCS,
2380 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002381 gcc_copts = xnnpack_gcc_std_copts(),
2382 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002383)
2384
2385xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002386 name = "scalar_ukernels",
2387 srcs = SCALAR_UKERNELS,
2388 hdrs = INTERNAL_HDRS,
2389 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002390 gcc_copts = xnnpack_gcc_std_copts(),
2391 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002392 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002393 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002394 "@FP16",
2395 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002396 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002397 ],
2398)
2399
2400xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002401 name = "scalar_ukernels_test_mode",
2402 srcs = SCALAR_UKERNELS,
2403 hdrs = INTERNAL_HDRS,
2404 aarch32_copts = ["-marm"],
2405 copts = [
2406 "-UNDEBUG",
2407 "-DXNN_TEST_MODE=1",
2408 ],
2409 gcc_copts = xnnpack_gcc_std_copts(),
2410 msvc_copts = xnnpack_msvc_std_copts(),
2411 deps = [
2412 ":tables",
2413 "@FP16",
2414 "@FXdiv",
2415 "@pthreadpool",
2416 ],
2417)
2418
2419xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002420 name = "wasm_ukernels",
2421 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002422 gcc_copts = xnnpack_gcc_std_copts(),
2423 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002424 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002425 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002426 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002427 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002428 "@FP16",
2429 "@FXdiv",
2430 "@pthreadpool",
2431 ],
2432)
2433
2434xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002435 name = "wasm_ukernels_test_mode",
2436 hdrs = INTERNAL_HDRS,
2437 copts = [
2438 "-UNDEBUG",
2439 "-DXNN_TEST_MODE=1",
2440 ],
2441 gcc_copts = xnnpack_gcc_std_copts(),
2442 msvc_copts = xnnpack_msvc_std_copts(),
2443 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002444 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002445 deps = [
2446 ":tables",
2447 "@FP16",
2448 "@FXdiv",
2449 "@pthreadpool",
2450 ],
2451)
2452
2453xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002454 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002455 hdrs = INTERNAL_HDRS,
2456 aarch32_copts = [
2457 "-marm",
2458 "-mfpu=neon",
2459 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002460 gcc_copts = xnnpack_gcc_std_copts(),
2461 gcc_x86_copts = ["-msse2"],
2462 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002463 optimized_copts = [
2464 "-O3",
2465 "-ffast-math",
2466 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002467 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002468 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002469 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002470 "@FP16",
2471 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002472 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002473 ],
2474)
2475
2476xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002477 name = "psimd_fastmath_ukernels_test_mode",
2478 hdrs = INTERNAL_HDRS,
2479 aarch32_copts = [
2480 "-marm",
2481 "-mfpu=neon",
2482 ],
2483 copts = [
2484 "-UNDEBUG",
2485 "-DXNN_TEST_MODE=1",
2486 ],
2487 gcc_copts = xnnpack_gcc_std_copts(),
2488 gcc_x86_copts = ["-msse2"],
2489 msvc_copts = xnnpack_msvc_std_copts(),
2490 optimized_copts = [
2491 "-O3",
2492 "-ffast-math",
2493 ],
2494 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2495 deps = [
2496 ":tables",
2497 "@FP16",
2498 "@psimd",
2499 "@pthreadpool",
2500 ],
2501)
2502
2503xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002504 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002505 hdrs = INTERNAL_HDRS,
2506 aarch32_copts = [
2507 "-marm",
2508 "-mfpu=neon",
2509 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002510 gcc_copts = xnnpack_gcc_std_copts(),
2511 gcc_x86_copts = ["-msse2"],
2512 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002513 optimized_copts = [
2514 "-O3",
2515 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002516 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002517 deps = [
2518 ":tables",
2519 "@FP16",
2520 "@psimd",
2521 "@pthreadpool",
2522 ],
2523)
2524
2525xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002526 name = "psimd_accmath_ukernels_test_mode",
2527 hdrs = INTERNAL_HDRS,
2528 aarch32_copts = [
2529 "-marm",
2530 "-mfpu=neon",
2531 ],
2532 copts = [
2533 "-UNDEBUG",
2534 "-DXNN_TEST_MODE=1",
2535 ],
2536 gcc_copts = xnnpack_gcc_std_copts(),
2537 gcc_x86_copts = ["-msse2"],
2538 msvc_copts = xnnpack_msvc_std_copts(),
2539 optimized_copts = [
2540 "-O3",
2541 ],
2542 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2543 deps = [
2544 ":tables",
2545 "@FP16",
2546 "@psimd",
2547 "@pthreadpool",
2548 ],
2549)
2550
2551xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002552 name = "neon_ukernels",
2553 hdrs = INTERNAL_HDRS,
2554 aarch32_copts = [
2555 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002556 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002557 "-mfpu=neon",
2558 ],
2559 aarch32_srcs = NEON_UKERNELS,
2560 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002561 gcc_copts = xnnpack_gcc_std_copts(),
2562 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002563 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002564 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002565 "@FP16",
2566 "@pthreadpool",
2567 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002568)
2569
2570xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002571 name = "neon_ukernels_test_mode",
2572 hdrs = INTERNAL_HDRS,
2573 aarch32_copts = [
2574 "-marm",
2575 "-march=armv7-a",
2576 "-mfpu=neon",
2577 ],
2578 aarch32_srcs = NEON_UKERNELS,
2579 aarch64_srcs = NEON_UKERNELS,
2580 copts = [
2581 "-UNDEBUG",
2582 "-DXNN_TEST_MODE=1",
2583 ],
2584 gcc_copts = xnnpack_gcc_std_copts(),
2585 msvc_copts = xnnpack_msvc_std_copts(),
2586 deps = [
2587 ":tables",
2588 "@FP16",
2589 "@pthreadpool",
2590 ],
2591)
2592
2593xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002594 name = "neonfma_ukernels",
2595 hdrs = INTERNAL_HDRS,
2596 aarch32_copts = [
2597 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002598 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002599 "-mfpu=neon-vfpv4",
2600 ],
2601 aarch32_srcs = NEONFMA_UKERNELS,
2602 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002603 apple_aarch32_copts = [
2604 "-mcpu=swift",
2605 "-mtune=generic",
2606 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002607 gcc_copts = xnnpack_gcc_std_copts(),
2608 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002609 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002610 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002611 "@FP16",
2612 "@pthreadpool",
2613 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002614)
2615
2616xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002617 name = "neonfma_ukernels_test_mode",
2618 hdrs = INTERNAL_HDRS,
2619 aarch32_copts = [
2620 "-marm",
2621 "-march=armv7-a",
2622 "-mfpu=neon-vfpv4",
2623 ],
2624 aarch32_srcs = NEONFMA_UKERNELS,
2625 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002626 apple_aarch32_copts = [
2627 "-mcpu=swift",
2628 "-mtune=generic",
2629 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002630 copts = [
2631 "-UNDEBUG",
2632 "-DXNN_TEST_MODE=1",
2633 ],
2634 gcc_copts = xnnpack_gcc_std_copts(),
2635 msvc_copts = xnnpack_msvc_std_copts(),
2636 deps = [
2637 ":tables",
2638 "@FP16",
2639 "@pthreadpool",
2640 ],
2641)
2642
2643xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002644 name = "neonv8_ukernels",
2645 hdrs = INTERNAL_HDRS,
2646 aarch32_copts = [
2647 "-marm",
2648 "-march=armv8-a",
2649 "-mfpu=neon-fp-armv8",
2650 ],
2651 aarch32_srcs = NEONV8_UKERNELS,
2652 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002653 apple_aarch32_copts = [
2654 "-mcpu=cyclone",
2655 "-mtune=generic",
2656 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002657 gcc_copts = xnnpack_gcc_std_copts(),
2658 msvc_copts = xnnpack_msvc_std_copts(),
2659 deps = [
2660 ":tables",
2661 "@FP16",
2662 "@pthreadpool",
2663 ],
2664)
2665
2666xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002667 name = "neonv8_ukernels_test_mode",
2668 hdrs = INTERNAL_HDRS,
2669 aarch32_copts = [
2670 "-marm",
2671 "-march=armv8-a",
2672 "-mfpu=neon-fp-armv8",
2673 ],
2674 aarch32_srcs = NEONV8_UKERNELS,
2675 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002676 apple_aarch32_copts = [
2677 "-mcpu=cyclone",
2678 "-mtune=generic",
2679 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002680 copts = [
2681 "-UNDEBUG",
2682 "-DXNN_TEST_MODE=1",
2683 ],
2684 gcc_copts = xnnpack_gcc_std_copts(),
2685 msvc_copts = xnnpack_msvc_std_copts(),
2686 deps = [
2687 ":tables",
2688 "@FP16",
2689 "@pthreadpool",
2690 ],
2691)
2692
2693xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002694 name = "neonfp16arith_ukernels",
2695 hdrs = INTERNAL_HDRS,
2696 aarch64_copts = ["-march=armv8.2-a+fp16"],
2697 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002698 gcc_copts = xnnpack_gcc_std_copts(),
2699 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002700 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002701 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002702 "@FP16",
2703 "@pthreadpool",
2704 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002705)
2706
2707xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002708 name = "neonfp16arith_ukernels_test_mode",
2709 hdrs = INTERNAL_HDRS,
2710 aarch64_copts = ["-march=armv8.2-a+fp16"],
2711 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2712 copts = [
2713 "-UNDEBUG",
2714 "-DXNN_TEST_MODE=1",
2715 ],
2716 gcc_copts = xnnpack_gcc_std_copts(),
2717 msvc_copts = xnnpack_msvc_std_copts(),
2718 deps = [
2719 ":tables",
2720 "@FP16",
2721 "@pthreadpool",
2722 ],
2723)
2724
2725xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002726 name = "sse2_ukernels",
2727 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002728 gcc_copts = xnnpack_gcc_std_copts(),
2729 gcc_x86_copts = ["-msse2"],
2730 msvc_copts = xnnpack_msvc_std_copts(),
2731 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002733 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002734 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002735 "@FP16",
2736 "@pthreadpool",
2737 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002738)
2739
2740xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002741 name = "sse2_ukernels_test_mode",
2742 hdrs = INTERNAL_HDRS,
2743 copts = [
2744 "-UNDEBUG",
2745 "-DXNN_TEST_MODE=1",
2746 ],
2747 gcc_copts = xnnpack_gcc_std_copts(),
2748 gcc_x86_copts = ["-msse2"],
2749 msvc_copts = xnnpack_msvc_std_copts(),
2750 msvc_x86_32_copts = ["/arch:SSE2"],
2751 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2752 deps = [
2753 ":tables",
2754 "@FP16",
2755 "@pthreadpool",
2756 ],
2757)
2758
2759xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002760 name = "ssse3_ukernels",
2761 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002762 gcc_copts = xnnpack_gcc_std_copts(),
2763 gcc_x86_copts = ["-mssse3"],
2764 msvc_copts = xnnpack_msvc_std_copts(),
2765 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002766 x86_srcs = SSSE3_UKERNELS,
2767 deps = [
2768 ":tables",
2769 "@FP16",
2770 "@pthreadpool",
2771 ],
2772)
2773
2774xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002775 name = "ssse3_ukernels_test_mode",
2776 hdrs = INTERNAL_HDRS,
2777 copts = [
2778 "-UNDEBUG",
2779 "-DXNN_TEST_MODE=1",
2780 ],
2781 gcc_copts = xnnpack_gcc_std_copts(),
2782 gcc_x86_copts = ["-mssse3"],
2783 msvc_copts = xnnpack_msvc_std_copts(),
2784 msvc_x86_32_copts = ["/arch:SSE2"],
2785 x86_srcs = SSSE3_UKERNELS,
2786 deps = [
2787 ":tables",
2788 "@FP16",
2789 "@pthreadpool",
2790 ],
2791)
2792
2793xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002794 name = "sse41_ukernels",
2795 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002796 gcc_copts = xnnpack_gcc_std_copts(),
2797 gcc_x86_copts = ["-msse4.1"],
2798 msvc_copts = xnnpack_msvc_std_copts(),
2799 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002800 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002801 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002802 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002803 "@FP16",
2804 "@pthreadpool",
2805 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002806)
2807
2808xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002809 name = "sse41_ukernels_test_mode",
2810 hdrs = INTERNAL_HDRS,
2811 copts = [
2812 "-UNDEBUG",
2813 "-DXNN_TEST_MODE=1",
2814 ],
2815 gcc_copts = xnnpack_gcc_std_copts(),
2816 gcc_x86_copts = ["-msse4.1"],
2817 msvc_copts = xnnpack_msvc_std_copts(),
2818 msvc_x86_32_copts = ["/arch:SSE2"],
2819 x86_srcs = SSE41_UKERNELS,
2820 deps = [
2821 ":tables",
2822 "@FP16",
2823 "@pthreadpool",
2824 ],
2825)
2826
2827xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002828 name = "avx_ukernels",
2829 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002830 gcc_copts = xnnpack_gcc_std_copts(),
2831 gcc_x86_copts = ["-mavx"],
2832 msvc_copts = xnnpack_msvc_std_copts(),
2833 msvc_x86_32_copts = ["/arch:AVX"],
2834 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002835 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002836 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002837 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002838 "@FP16",
2839 "@pthreadpool",
2840 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002841)
2842
2843xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002844 name = "avx_ukernels_test_mode",
2845 hdrs = INTERNAL_HDRS,
2846 copts = [
2847 "-UNDEBUG",
2848 "-DXNN_TEST_MODE=1",
2849 ],
2850 gcc_copts = xnnpack_gcc_std_copts(),
2851 gcc_x86_copts = ["-mavx"],
2852 msvc_copts = xnnpack_msvc_std_copts(),
2853 msvc_x86_32_copts = ["/arch:AVX"],
2854 msvc_x86_64_copts = ["/arch:AVX"],
2855 x86_srcs = AVX_UKERNELS,
2856 deps = [
2857 ":tables",
2858 "@FP16",
2859 "@pthreadpool",
2860 ],
2861)
2862
2863xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002864 name = "fma3_ukernels",
2865 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002866 gcc_copts = xnnpack_gcc_std_copts(),
2867 gcc_x86_copts = ["-mfma"],
2868 msvc_copts = xnnpack_msvc_std_copts(),
2869 msvc_x86_32_copts = ["/arch:AVX"],
2870 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002871 x86_srcs = FMA3_UKERNELS,
2872 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002873 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002874 "@FP16",
2875 "@pthreadpool",
2876 ],
2877)
2878
2879xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002880 name = "fma3_ukernels_test_mode",
2881 hdrs = INTERNAL_HDRS,
2882 copts = [
2883 "-UNDEBUG",
2884 "-DXNN_TEST_MODE=1",
2885 ],
2886 gcc_copts = xnnpack_gcc_std_copts(),
2887 gcc_x86_copts = ["-mfma"],
2888 msvc_copts = xnnpack_msvc_std_copts(),
2889 msvc_x86_32_copts = ["/arch:AVX"],
2890 msvc_x86_64_copts = ["/arch:AVX"],
2891 x86_srcs = FMA3_UKERNELS,
2892 deps = [
2893 ":tables",
2894 "@FP16",
2895 "@pthreadpool",
2896 ],
2897)
2898
2899xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002900 name = "avx2_ukernels",
2901 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002902 gcc_copts = xnnpack_gcc_std_copts(),
2903 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002904 "-mfma",
2905 "-mavx2",
2906 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002907 msvc_copts = xnnpack_msvc_std_copts(),
2908 msvc_x86_32_copts = ["/arch:AVX2"],
2909 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002910 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002911 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002912 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002913 "@FP16",
2914 "@pthreadpool",
2915 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002916)
2917
2918xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002919 name = "avx2_ukernels_test_mode",
2920 hdrs = INTERNAL_HDRS,
2921 copts = [
2922 "-UNDEBUG",
2923 "-DXNN_TEST_MODE=1",
2924 ],
2925 gcc_copts = xnnpack_gcc_std_copts(),
2926 gcc_x86_copts = [
2927 "-mfma",
2928 "-mavx2",
2929 ],
2930 msvc_copts = xnnpack_msvc_std_copts(),
2931 msvc_x86_32_copts = ["/arch:AVX2"],
2932 msvc_x86_64_copts = ["/arch:AVX2"],
2933 x86_srcs = AVX2_UKERNELS,
2934 deps = [
2935 ":tables",
2936 "@FP16",
2937 "@pthreadpool",
2938 ],
2939)
2940
2941xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002942 name = "avx512f_ukernels",
2943 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002944 gcc_copts = xnnpack_gcc_std_copts(),
2945 gcc_x86_copts = ["-mavx512f"],
2946 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2947 msvc_copts = xnnpack_msvc_std_copts(),
2948 msvc_x86_32_copts = ["/arch:AVX512"],
2949 msvc_x86_64_copts = ["/arch:AVX512"],
2950 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002951 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002952 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002953 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002954 "@FP16",
2955 "@pthreadpool",
2956 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002957)
2958
2959xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002960 name = "avx512f_ukernels_test_mode",
2961 hdrs = INTERNAL_HDRS,
2962 copts = [
2963 "-UNDEBUG",
2964 "-DXNN_TEST_MODE=1",
2965 ],
2966 gcc_copts = xnnpack_gcc_std_copts(),
2967 gcc_x86_copts = ["-mavx512f"],
2968 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2969 msvc_copts = xnnpack_msvc_std_copts(),
2970 msvc_x86_32_copts = ["/arch:AVX512"],
2971 msvc_x86_64_copts = ["/arch:AVX512"],
2972 msys_copts = ["-fno-asynchronous-unwind-tables"],
2973 x86_srcs = AVX512F_UKERNELS,
2974 deps = [
2975 ":tables",
2976 "@FP16",
2977 "@pthreadpool",
2978 ],
2979)
2980
2981xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 name = "asm_ukernels",
2983 hdrs = ["src/xnnpack/assembly.h"],
2984 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002985 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002986 aarch64_srcs = AARCH64_ASM_UKERNELS,
2987)
2988
Marat Dukhan3b59de22020-06-03 20:15:19 -07002989xnnpack_cc_library(
2990 name = "logging_utils",
2991 srcs = LOGGING_SRCS,
2992 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2993 copts = LOGGING_COPTS + [
2994 "-Isrc",
2995 "-Iinclude",
2996 ] + select({
2997 ":debug_build": [],
2998 "//conditions:default": xnnpack_min_size_copts(),
2999 }),
3000 gcc_copts = xnnpack_gcc_std_copts(),
3001 msvc_copts = xnnpack_msvc_std_copts(),
3002 visibility = xnnpack_visibility(),
3003 deps = [
3004 "@FP16",
3005 "@clog",
3006 "@pthreadpool",
3007 ],
3008)
3009
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010xnnpack_aggregate_library(
3011 name = "ukernels",
3012 aarch32_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 ":asm_ukernels",
3017 ],
3018 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003019 ":neon_ukernels",
3020 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003021 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 ":neonfp16arith_ukernels",
3023 ":asm_ukernels",
3024 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003025 generic_deps = [
3026 ":scalar_ukernels",
3027 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003028 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003029 ":psimd_fastmath_ukernels",
3030 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003031 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003032 wasm_deps = [
3033 ":wasm_ukernels",
3034 ],
3035 wasmsimd_deps = [
3036 ":wasm_ukernels",
3037 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003038 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003040 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003041 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003042 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003043 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003044 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003045 ":avx512f_ukernels",
3046 ],
3047)
3048
Marat Dukhan33fcf782020-05-24 14:27:15 -07003049xnnpack_aggregate_library(
3050 name = "ukernels_test_mode",
3051 aarch32_deps = [
3052 ":neon_ukernels_test_mode",
3053 ":neonfma_ukernels_test_mode",
3054 ":neonv8_ukernels_test_mode",
3055 ":asm_ukernels",
3056 ],
3057 aarch64_deps = [
3058 ":neon_ukernels_test_mode",
3059 ":neonfma_ukernels_test_mode",
3060 ":neonv8_ukernels_test_mode",
3061 ":neonfp16arith_ukernels_test_mode",
3062 ":asm_ukernels",
3063 ],
3064 generic_deps = [
3065 ":scalar_ukernels_test_mode",
3066 ],
3067 psimd_deps = [
3068 ":psimd_fastmath_ukernels_test_mode",
3069 ":psimd_accmath_ukernels_test_mode",
3070 ],
3071 wasm_deps = [
3072 ":wasm_ukernels_test_mode",
3073 ],
3074 wasmsimd_deps = [
3075 ":wasm_ukernels_test_mode",
3076 ],
3077 x86_deps = [
3078 ":sse2_ukernels_test_mode",
3079 ":ssse3_ukernels_test_mode",
3080 ":sse41_ukernels_test_mode",
3081 ":avx_ukernels_test_mode",
3082 ":fma3_ukernels_test_mode",
3083 ":avx2_ukernels_test_mode",
3084 ":avx512f_ukernels_test_mode",
3085 ],
3086)
3087
Marat Dukhan86df3842020-06-17 14:03:31 -07003088# This target must be compatible with Android NDK r17 and compile with Clang 6.
3089# It is similar to :ukernels target, but excludes microkernels which require
3090# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003091xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003092 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003093 aarch32_deps = [
3094 ":neon_ukernels",
3095 ":neonfma_ukernels",
3096 ":neonv8_ukernels",
3097 ":asm_ukernels",
3098 ],
3099 aarch64_deps = [
3100 ":neon_ukernels",
3101 ":neonfma_ukernels",
3102 ":neonv8_ukernels",
3103 ":asm_ukernels",
3104 ],
3105 generic_deps = [
3106 ":scalar_ukernels",
3107 ],
3108 psimd_deps = [
3109 ":psimd_fastmath_ukernels",
3110 ":psimd_accmath_ukernels",
3111 ],
3112 wasm_deps = [
3113 ":wasm_ukernels",
3114 ],
3115 wasmsimd_deps = [
3116 ":wasm_ukernels",
3117 ],
3118 x86_deps = [
3119 ":sse2_ukernels",
3120 ":ssse3_ukernels",
3121 ":sse41_ukernels",
3122 ":avx_ukernels",
3123 ":fma3_ukernels",
3124 ":avx2_ukernels",
3125 ":avx512f_ukernels",
3126 ],
3127)
3128
Marat Dukhan08c4a432019-10-03 09:29:21 -07003129xnnpack_cc_library(
3130 name = "im2col",
3131 srcs = ["src/im2col.c"],
3132 hdrs = [
3133 "src/xnnpack/common.h",
3134 "src/xnnpack/im2col.h",
3135 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003136 gcc_copts = xnnpack_gcc_std_copts(),
3137 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003138)
3139
3140xnnpack_cc_library(
3141 name = "indirection",
3142 srcs = ["src/indirection.c"],
3143 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003144 gcc_copts = xnnpack_gcc_std_copts(),
3145 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003146 deps = [
3147 "@FP16",
3148 "@FXdiv",
3149 "@pthreadpool",
3150 ],
3151)
3152
3153xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003154 name = "indirection_test_mode",
3155 srcs = ["src/indirection.c"],
3156 hdrs = INTERNAL_HDRS,
3157 copts = [
3158 "-UNDEBUG",
3159 "-DXNN_TEST_MODE=1",
3160 ],
3161 gcc_copts = xnnpack_gcc_std_copts(),
3162 msvc_copts = xnnpack_msvc_std_copts(),
3163 deps = [
3164 "@FP16",
3165 "@FXdiv",
3166 "@pthreadpool",
3167 ],
3168)
3169
3170xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003171 name = "packing",
3172 srcs = ["src/packing.c"],
3173 hdrs = INTERNAL_HDRS,
3174 gcc_copts = xnnpack_gcc_std_copts(),
3175 msvc_copts = xnnpack_msvc_std_copts(),
3176 deps = [
3177 "@FP16",
3178 "@FXdiv",
3179 "@pthreadpool",
3180 ],
3181)
3182
3183xnnpack_cc_library(
3184 name = "packing_test_mode",
3185 srcs = ["src/packing.c"],
3186 hdrs = INTERNAL_HDRS,
3187 copts = [
3188 "-UNDEBUG",
3189 "-DXNN_TEST_MODE=1",
3190 ],
3191 gcc_copts = xnnpack_gcc_std_copts(),
3192 msvc_copts = xnnpack_msvc_std_copts(),
3193 deps = [
3194 "@FP16",
3195 "@FXdiv",
3196 "@pthreadpool",
3197 ],
3198)
3199
3200xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003201 name = "operator_run",
3202 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003203 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003204 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003205 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3206 "//conditions:default": [],
3207 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003208 gcc_copts = xnnpack_gcc_std_copts(),
3209 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003211 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212 "@FP16",
3213 "@FXdiv",
3214 "@clog",
3215 "@pthreadpool",
3216 ],
3217)
3218
Chao Mei6ddfc602020-05-13 22:29:36 -07003219xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003220 name = "operator_run_test_mode",
3221 srcs = ["src/operator-run.c"],
3222 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3223 copts = LOGGING_COPTS + [
3224 "-UNDEBUG",
3225 "-DXNN_TEST_MODE=1",
3226 ] + select({
3227 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3228 "//conditions:default": [],
3229 }),
3230 gcc_copts = xnnpack_gcc_std_copts(),
3231 msvc_copts = xnnpack_msvc_std_copts(),
3232 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003233 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003234 "@FP16",
3235 "@FXdiv",
3236 "@clog",
3237 "@pthreadpool",
3238 ],
3239)
3240
3241xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003242 name = "memory_planner",
3243 srcs = ["src/memory-planner.c"],
3244 hdrs = INTERNAL_HDRS,
3245 defines = select({
3246 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3247 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3248 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3249 }),
3250 gcc_copts = xnnpack_gcc_std_copts(),
3251 msvc_copts = xnnpack_msvc_std_copts(),
3252 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003253 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003254 "@pthreadpool",
3255 ],
3256)
3257
Marat Dukhan33fcf782020-05-24 14:27:15 -07003258xnnpack_cc_library(
3259 name = "memory_planner_test_mode",
3260 srcs = ["src/memory-planner.c"],
3261 hdrs = INTERNAL_HDRS,
3262 copts = [
3263 "-UNDEBUG",
3264 "-DXNN_TEST_MODE=1",
3265 ],
3266 defines = select({
3267 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3268 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3269 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3270 }),
3271 gcc_copts = xnnpack_gcc_std_copts(),
3272 msvc_copts = xnnpack_msvc_std_copts(),
3273 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003274 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003275 "@pthreadpool",
3276 ],
3277)
3278
Marat Dukhan08c4a432019-10-03 09:29:21 -07003279cc_library(
3280 name = "enable_assembly",
3281 defines = select({
3282 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3283 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003284 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003285 }),
3286)
3287
Marat Dukhan9de90e02020-06-18 16:04:12 -07003288cc_library(
3289 name = "enable_sparse",
3290 defines = select({
3291 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3292 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003293 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003294 }),
3295)
3296
Marat Dukhancf056b22019-10-07 10:26:29 -07003297xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003298 name = "operators",
3299 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003300 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003301 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003302 ],
3303 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003304 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003305 "-Isrc",
3306 "-Iinclude",
3307 ] + select({
3308 ":debug_build": [],
3309 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003310 }) + select({
3311 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3312 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003313 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003314 gcc_copts = xnnpack_gcc_std_copts(),
3315 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003316 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003317 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003318 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003319 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003320 "@FP16",
3321 "@FXdiv",
3322 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003323 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003324 ],
3325)
3326
Marat Dukhan10a38082020-04-17 03:58:35 -07003327xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003328 name = "operators_test_mode",
3329 srcs = OPERATOR_SRCS + [
3330 "src/memory.c",
3331 "src/operator-delete.c",
3332 ],
3333 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3334 copts = LOGGING_COPTS + [
3335 "-Isrc",
3336 "-Iinclude",
3337 "-UNDEBUG",
3338 "-DXNN_TEST_MODE=1",
3339 ] + select({
3340 ":debug_build": [],
3341 "//conditions:default": xnnpack_min_size_copts(),
3342 }) + select({
3343 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3344 "//conditions:default": [],
3345 }),
3346 gcc_copts = xnnpack_gcc_std_copts(),
3347 msvc_copts = xnnpack_msvc_std_copts(),
3348 deps = [
3349 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003350 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003351 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003352 "@FP16",
3353 "@FXdiv",
3354 "@clog",
3355 "@pthreadpool",
3356 ],
3357)
3358
3359xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003360 name = "XNNPACK",
3361 srcs = [
3362 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003363 "src/runtime.c",
3364 "src/subgraph.c",
3365 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003366 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003367 hdrs = ["include/xnnpack.h"],
3368 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003369 "-Isrc",
3370 "-Iinclude",
3371 ] + select({
3372 ":debug_build": [],
3373 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003374 }) + select({
3375 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3376 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003377 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003378 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003379 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003380 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003381 visibility = xnnpack_visibility(),
3382 deps = [
3383 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003384 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003385 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003386 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003387 ":operator_run",
3388 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003389 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003390 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003391 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003392 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003393 ] + select({
3394 ":emscripten": [],
3395 "//conditions:default": ["@cpuinfo"],
3396 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003397)
3398
Marat Dukhan10a38082020-04-17 03:58:35 -07003399xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003400 name = "XNNPACK_test_mode",
3401 srcs = [
3402 "src/init.c",
3403 "src/runtime.c",
3404 "src/subgraph.c",
3405 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003406 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003407 hdrs = ["include/xnnpack.h"],
3408 copts = LOGGING_COPTS + [
3409 "-Isrc",
3410 "-Iinclude",
3411 "-UNDEBUG",
3412 "-DXNN_TEST_MODE=1",
3413 ] + select({
3414 ":debug_build": [],
3415 "//conditions:default": xnnpack_min_size_copts(),
3416 }) + select({
3417 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3418 "//conditions:default": [],
3419 }),
3420 gcc_copts = xnnpack_gcc_std_copts(),
3421 includes = ["include"],
3422 msvc_copts = xnnpack_msvc_std_copts(),
3423 visibility = xnnpack_visibility(),
3424 deps = [
3425 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003426 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003427 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003428 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003429 ":operator_run_test_mode",
3430 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003431 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003432 "@clog",
3433 "@FP16",
3434 "@pthreadpool",
3435 ] + select({
3436 ":emscripten": [],
3437 "//conditions:default": ["@cpuinfo"],
3438 }),
3439)
3440
3441xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003442 name = "xnnpack_f32",
3443 srcs = [
3444 "src/init.c",
3445 "src/runtime.c",
3446 "src/subgraph.c",
3447 "src/tensor.c",
3448 ] + SUBGRAPH_SRCS,
3449 hdrs = ["include/xnnpack.h"],
3450 copts = LOGGING_COPTS + [
3451 "-Isrc",
3452 "-Iinclude",
3453 ] + select({
3454 ":debug_build": [],
3455 "//conditions:default": xnnpack_min_size_copts(),
3456 }) + select({
3457 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3458 "//conditions:default": [],
3459 }),
3460 defines = [
3461 "XNN_NO_Q8_OPERATORS",
3462 "XNN_NO_U8_OPERATORS",
3463 "XNN_NO_X8_OPERATORS",
3464 "XNN_NO_F16_OPERATORS",
3465 "XNN_NO_X16_OPERATORS",
3466 ],
3467 gcc_copts = xnnpack_gcc_std_copts(),
3468 includes = ["include"],
3469 msvc_copts = xnnpack_msvc_std_copts(),
3470 visibility = xnnpack_visibility(),
3471 deps = [
3472 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003473 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003474 ":logging_utils",
3475 ":memory_planner",
3476 ":operator_run",
3477 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003478 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003479 "@clog",
3480 "@FP16",
3481 "@pthreadpool",
3482 ] + select({
3483 ":emscripten": [],
3484 "//conditions:default": ["@cpuinfo"],
3485 }),
3486)
3487
3488xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003489 name = "xnnpack_operators_nhwc_f32",
3490 srcs = [
3491 "src/init.c",
3492 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003493 hdrs = ["include/xnnpack.h"],
3494 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003495 "-Isrc",
3496 "-Iinclude",
3497 ] + select({
3498 ":debug_build": [],
3499 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003500 }) + select({
3501 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3502 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003503 }),
3504 defines = [
3505 "XNN_NO_Q8_OPERATORS",
3506 "XNN_NO_U8_OPERATORS",
3507 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003508 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003509 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003510 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003512 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513 visibility = xnnpack_visibility(),
3514 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003515 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003516 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517 ":operator_run",
3518 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003519 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003520 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003521 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003522 ] + select({
3523 ":emscripten": [],
3524 "//conditions:default": ["@cpuinfo"],
3525 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003526)
3527
Marat Dukhancf056b22019-10-07 10:26:29 -07003528xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003529 name = "bench_utils",
3530 srcs = ["bench/utils.cc"],
3531 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003532 deps = [
3533 "@com_google_benchmark//:benchmark",
3534 "@cpuinfo",
3535 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003536)
3537
Frank Barchard7e955972019-10-11 10:34:25 -07003538######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539
3540xnnpack_benchmark(
3541 name = "q8_gemm_bench",
3542 srcs = [
3543 "bench/gemm.h",
3544 "bench/q8-gemm.cc",
3545 "src/xnnpack/AlignedAllocator.h",
3546 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003547 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003548 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003549)
3550
3551xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003552 name = "f16_igemm_bench",
3553 srcs = [
3554 "bench/f16-igemm.cc",
3555 "bench/conv.h",
3556 "bench/google/conv.h",
3557 "src/xnnpack/AlignedAllocator.h",
3558 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003559 deps = MICROKERNEL_BENCHMARK_DEPS + [
3560 ":indirection",
3561 ":packing",
3562 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003563)
3564
3565xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003566 name = "f16_gemm_bench",
3567 srcs = [
3568 "bench/f16-gemm.cc",
3569 "bench/gemm.h",
3570 "src/xnnpack/AlignedAllocator.h",
3571 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003572 deps = MICROKERNEL_BENCHMARK_DEPS + [
3573 ":packing",
3574 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003575)
3576
3577xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003578 name = "f16_spmm_bench",
3579 srcs = [
3580 "bench/f16-spmm.cc",
3581 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003582 "src/xnnpack/AlignedAllocator.h",
3583 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003584 deps = MICROKERNEL_BENCHMARK_DEPS,
3585)
3586
3587xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 name = "f32_igemm_bench",
3589 srcs = [
3590 "bench/f32-igemm.cc",
3591 "bench/conv.h",
3592 "src/xnnpack/AlignedAllocator.h",
3593 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003594 deps = MICROKERNEL_BENCHMARK_DEPS + [
3595 ":indirection",
3596 ":packing",
3597 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003598)
3599
3600xnnpack_benchmark(
3601 name = "f32_conv_hwc_bench",
3602 srcs = [
3603 "bench/f32-conv-hwc.cc",
3604 "bench/dconv.h",
3605 "src/xnnpack/AlignedAllocator.h",
3606 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003607 deps = MICROKERNEL_BENCHMARK_DEPS + [
3608 ":packing",
3609 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610)
3611
3612xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003613 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003614 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003615 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003616 "bench/dconv.h",
3617 "src/xnnpack/AlignedAllocator.h",
3618 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003619 deps = MICROKERNEL_BENCHMARK_DEPS + [
3620 ":packing",
3621 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003622)
3623
3624xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003625 name = "f16_dwconv_bench",
3626 srcs = [
3627 "bench/f16-dwconv.cc",
3628 "bench/dwconv.h",
3629 "bench/google/dwconv.h",
3630 "src/xnnpack/AlignedAllocator.h",
3631 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003632 deps = MICROKERNEL_BENCHMARK_DEPS + [
3633 ":indirection",
3634 ":packing",
3635 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003636)
3637
3638xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003639 name = "f32_dwconv_bench",
3640 srcs = [
3641 "bench/f32-dwconv.cc",
3642 "bench/dwconv.h",
3643 "src/xnnpack/AlignedAllocator.h",
3644 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003645 deps = MICROKERNEL_BENCHMARK_DEPS + [
3646 ":indirection",
3647 ":packing",
3648 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003649)
3650
3651xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003652 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003654 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003655 "bench/dwconv.h",
3656 "src/xnnpack/AlignedAllocator.h",
3657 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003658 deps = MICROKERNEL_BENCHMARK_DEPS + [
3659 ":indirection",
3660 ":packing",
3661 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662)
3663
3664xnnpack_benchmark(
3665 name = "f32_gemm_bench",
3666 srcs = [
3667 "bench/f32-gemm.cc",
3668 "bench/gemm.h",
3669 "src/xnnpack/AlignedAllocator.h",
3670 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003671 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003672 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003673)
3674
3675xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003676 name = "f32_hswish_bench",
3677 srcs = [
3678 "bench/f32-hswish.cc",
3679 "src/xnnpack/AlignedAllocator.h",
3680 ] + MICROKERNEL_BENCHMARK_HDRS,
3681 deps = MICROKERNEL_BENCHMARK_DEPS,
3682)
3683
3684xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003685 name = "f32_raddexpminusmax_bench",
3686 srcs = [
3687 "bench/f32-raddexpminusmax.cc",
3688 "src/xnnpack/AlignedAllocator.h",
3689 ] + MICROKERNEL_BENCHMARK_HDRS,
3690 deps = MICROKERNEL_BENCHMARK_DEPS,
3691)
3692
3693xnnpack_benchmark(
3694 name = "f32_raddextexp_bench",
3695 srcs = [
3696 "bench/f32-raddextexp.cc",
3697 "src/xnnpack/AlignedAllocator.h",
3698 ] + MICROKERNEL_BENCHMARK_HDRS,
3699 deps = MICROKERNEL_BENCHMARK_DEPS,
3700)
3701
3702xnnpack_benchmark(
3703 name = "f32_raddstoreexpminusmax_bench",
3704 srcs = [
3705 "bench/f32-raddstoreexpminusmax.cc",
3706 "src/xnnpack/AlignedAllocator.h",
3707 ] + MICROKERNEL_BENCHMARK_HDRS,
3708 deps = MICROKERNEL_BENCHMARK_DEPS,
3709)
3710
3711xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003712 name = "f32_rmax_bench",
3713 srcs = [
3714 "bench/f32-rmax.cc",
3715 "src/xnnpack/AlignedAllocator.h",
3716 ] + MICROKERNEL_BENCHMARK_HDRS,
3717 deps = MICROKERNEL_BENCHMARK_DEPS,
3718)
3719
3720xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003721 name = "f32_sigmoid_bench",
3722 srcs = [
3723 "bench/f32-sigmoid.cc",
3724 "src/xnnpack/AlignedAllocator.h",
3725 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003726 deps = MICROKERNEL_BENCHMARK_DEPS,
3727)
3728
3729xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003730 name = "f32_spmm_bench",
3731 srcs = [
3732 "bench/f32-spmm.cc",
3733 "bench/gemm.h",
3734 "src/xnnpack/AlignedAllocator.h",
3735 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003736 deps = MICROKERNEL_BENCHMARK_DEPS,
3737)
3738
3739xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003740 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003741 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003742 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003743 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003744 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003745 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003746)
3747
3748xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003749 name = "f32_vscaleexpminusmax_bench",
3750 srcs = [
3751 "bench/f32-vscaleexpminusmax.cc",
3752 "src/xnnpack/AlignedAllocator.h",
3753 ] + MICROKERNEL_BENCHMARK_HDRS,
3754 deps = MICROKERNEL_BENCHMARK_DEPS,
3755)
3756
3757xnnpack_benchmark(
3758 name = "f32_vscaleextexp_bench",
3759 srcs = [
3760 "bench/f32-vscaleextexp.cc",
3761 "src/xnnpack/AlignedAllocator.h",
3762 ] + MICROKERNEL_BENCHMARK_HDRS,
3763 deps = MICROKERNEL_BENCHMARK_DEPS,
3764)
3765
3766xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003767 name = "f32_vsqrt_bench",
3768 srcs = [
3769 "bench/f32-vsqrt.cc",
3770 "src/xnnpack/AlignedAllocator.h",
3771 ] + MICROKERNEL_BENCHMARK_HDRS,
3772 deps = MICROKERNEL_BENCHMARK_DEPS,
3773)
3774
3775xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003776 name = "f32_im2col_gemm_bench",
3777 srcs = [
3778 "bench/f32-im2col-gemm.cc",
3779 "bench/conv.h",
3780 "src/xnnpack/AlignedAllocator.h",
3781 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003782 deps = MICROKERNEL_BENCHMARK_DEPS + [
3783 ":im2col",
3784 ":packing",
3785 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003786)
3787
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003788xnnpack_benchmark(
3789 name = "requantization_bench",
3790 srcs = [
3791 "bench/requantization.cc",
3792 "src/xnnpack/requantization-stubs.h",
3793 "src/xnnpack/AlignedAllocator.h",
3794 ] + MICROKERNEL_BENCHMARK_HDRS,
3795 deps = MICROKERNEL_BENCHMARK_DEPS,
3796)
3797
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003798xnnpack_benchmark(
3799 name = "rounding_bench",
3800 srcs = [
3801 "bench/rounding.cc",
3802 "src/xnnpack/math-stubs.h",
3803 "src/xnnpack/AlignedAllocator.h",
3804 ] + MICROKERNEL_BENCHMARK_HDRS,
3805 deps = MICROKERNEL_BENCHMARK_DEPS,
3806)
3807
Marat Dukhan08c4a432019-10-03 09:29:21 -07003808########################### Benchmarks for operators ###########################
3809
3810xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003811 name = "average_pooling_bench",
3812 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003813 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003814 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003815 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816)
3817
3818xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003819 name = "bankers_rounding_bench",
3820 srcs = ["bench/bankers-rounding.cc"],
3821 copts = xnnpack_optional_tflite_copts(),
3822 tags = ["nowin32"],
3823 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3824)
3825
3826xnnpack_benchmark(
3827 name = "ceiling_bench",
3828 srcs = ["bench/ceiling.cc"],
3829 copts = xnnpack_optional_tflite_copts(),
3830 tags = ["nowin32"],
3831 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3832)
3833
3834xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003835 name = "channel_shuffle_bench",
3836 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003837 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003838)
3839
3840xnnpack_benchmark(
3841 name = "convolution_bench",
3842 srcs = ["bench/convolution.cc"],
3843 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003844 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003845 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003846)
3847
3848xnnpack_benchmark(
3849 name = "deconvolution_bench",
3850 srcs = ["bench/deconvolution.cc"],
3851 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003852 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003853 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003854)
3855
3856xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003857 name = "floor_bench",
3858 srcs = ["bench/floor.cc"],
3859 copts = xnnpack_optional_tflite_copts(),
3860 tags = ["nowin32"],
3861 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3862)
3863
3864xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003865 name = "global_average_pooling_bench",
3866 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003867 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003868)
3869
3870xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003871 name = "hardswish_bench",
3872 srcs = ["bench/hardswish.cc"],
3873 copts = xnnpack_optional_tflite_copts(),
3874 tags = ["nowin32"],
3875 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3876)
3877
3878xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003879 name = "max_pooling_bench",
3880 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003881 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003882)
3883
3884xnnpack_benchmark(
3885 name = "sigmoid_bench",
3886 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003887 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003888 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003889 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890)
3891
3892xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003893 name = "prelu_bench",
3894 srcs = ["bench/prelu.cc"],
3895 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003896 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003897 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003898)
3899
3900xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003901 name = "softmax_bench",
3902 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003903 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003904 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003905 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906)
3907
Marat Dukhan87727142020-06-24 15:24:10 -07003908xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003909 name = "square_root_bench",
3910 srcs = ["bench/square-root.cc"],
3911 copts = xnnpack_optional_tflite_copts(),
3912 tags = ["nowin32"],
3913 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3914)
3915
3916xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003917 name = "truncation_bench",
3918 srcs = ["bench/truncation.cc"],
3919 deps = OPERATOR_BENCHMARK_DEPS,
3920)
3921
Marat Dukhanc068bb62019-10-04 13:24:39 -07003922############################# End-to-end benchmarks ############################
3923
3924cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003925 name = "fp32_mobilenet_v1",
3926 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003927 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003928 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003929 linkstatic = True,
3930 deps = [
3931 ":XNNPACK",
3932 "@pthreadpool",
3933 ],
3934)
3935
3936cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003937 name = "fp16_mobilenet_v1",
3938 srcs = ["models/fp16-mobilenet-v1.cc"],
3939 hdrs = ["models/models.h"],
3940 copts = xnnpack_std_cxxopts(),
3941 linkstatic = True,
3942 deps = [
3943 ":XNNPACK",
3944 "@FP16",
3945 "@pthreadpool",
3946 ],
3947)
3948
3949cc_library(
3950 name = "fp32_mobilenet_v2",
3951 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003952 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003953 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003954 linkstatic = True,
3955 deps = [
3956 ":XNNPACK",
3957 "@pthreadpool",
3958 ],
3959)
3960
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003961cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003962 name = "fp16_mobilenet_v2",
3963 srcs = ["models/fp16-mobilenet-v2.cc"],
3964 hdrs = ["models/models.h"],
3965 copts = xnnpack_std_cxxopts(),
3966 linkstatic = True,
3967 deps = [
3968 ":XNNPACK",
3969 "@FP16",
3970 "@pthreadpool",
3971 ],
3972)
3973
3974cc_library(
3975 name = "fp32_mobilenet_v3_large",
3976 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003977 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003978 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003979 linkstatic = True,
3980 deps = [
3981 ":XNNPACK",
3982 "@pthreadpool",
3983 ],
3984)
3985
3986cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003987 name = "fp32_mobilenet_v3_small",
3988 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003989 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003990 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003991 linkstatic = True,
3992 deps = [
3993 ":XNNPACK",
3994 "@pthreadpool",
3995 ],
3996)
3997
Marat Dukhanc068bb62019-10-04 13:24:39 -07003998xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003999 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004000 srcs = [
4001 "bench/f32-dwconv-e2e.cc",
4002 "bench/end2end.h",
4003 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004004 deps = MICROKERNEL_BENCHMARK_DEPS + [
4005 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004006 ":fp32_mobilenet_v1",
4007 ":fp32_mobilenet_v2",
4008 ":fp32_mobilenet_v3_large",
4009 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004010 ],
4011)
4012
4013xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004014 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004015 srcs = [
4016 "bench/f32-gemm-e2e.cc",
4017 "bench/end2end.h",
4018 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004019 deps = MICROKERNEL_BENCHMARK_DEPS + [
4020 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004021 ":fp32_mobilenet_v1",
4022 ":fp32_mobilenet_v2",
4023 ":fp32_mobilenet_v3_large",
4024 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004025 ],
4026)
4027
4028xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004029 name = "end2end_bench",
4030 srcs = ["bench/end2end.cc"],
4031 deps = [
4032 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004033 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004034 ":fp16_mobilenet_v1",
4035 ":fp16_mobilenet_v2",
4036 ":fp32_mobilenet_v1",
4037 ":fp32_mobilenet_v2",
4038 ":fp32_mobilenet_v3_large",
4039 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004040 "@pthreadpool",
4041 ],
4042)
4043
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004044#################### Accuracy evaluation for math functions ####################
4045
4046xnnpack_benchmark(
4047 name = "f32_exp_eval",
4048 srcs = [
4049 "eval/f32-exp.cc",
4050 "src/xnnpack/AlignedAllocator.h",
4051 ] + ACCURACY_EVAL_HDRS,
4052 deps = ACCURACY_EVAL_DEPS,
4053)
4054
Marat Dukhan515c9772019-10-17 18:07:57 -07004055xnnpack_benchmark(
4056 name = "f32_expminus_eval",
4057 srcs = [
4058 "eval/f32-expminus.cc",
4059 "src/xnnpack/AlignedAllocator.h",
4060 ] + ACCURACY_EVAL_HDRS,
4061 deps = ACCURACY_EVAL_DEPS,
4062)
4063
Marat Dukhan98ba4412019-10-23 02:14:28 -07004064xnnpack_benchmark(
4065 name = "f32_extexp_eval",
4066 srcs = [
4067 "eval/f32-extexp.cc",
4068 "src/xnnpack/AlignedAllocator.h",
4069 ] + ACCURACY_EVAL_HDRS,
4070 deps = ACCURACY_EVAL_DEPS,
4071)
4072
Marat Dukhan8853b822020-05-07 12:19:01 -07004073xnnpack_unit_test(
4074 name = "f32_roundne_eval",
4075 srcs = [
4076 "eval/f32-roundne.cc",
4077 "src/xnnpack/AlignedAllocator.h",
4078 "src/xnnpack/math-stubs.h",
4079 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004080 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004081 deps = MICROKERNEL_TEST_DEPS,
4082)
4083
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004084xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004085 name = "f32_roundd_eval",
4086 srcs = [
4087 "eval/f32-roundd.cc",
4088 "src/xnnpack/AlignedAllocator.h",
4089 "src/xnnpack/math-stubs.h",
4090 ] + MICROKERNEL_TEST_HDRS,
4091 automatic = False,
4092 deps = MICROKERNEL_TEST_DEPS,
4093)
4094
4095xnnpack_unit_test(
4096 name = "f32_roundu_eval",
4097 srcs = [
4098 "eval/f32-roundu.cc",
4099 "src/xnnpack/AlignedAllocator.h",
4100 "src/xnnpack/math-stubs.h",
4101 ] + MICROKERNEL_TEST_HDRS,
4102 automatic = False,
4103 deps = MICROKERNEL_TEST_DEPS,
4104)
4105
4106xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004107 name = "f32_roundz_eval",
4108 srcs = [
4109 "eval/f32-roundz.cc",
4110 "src/xnnpack/AlignedAllocator.h",
4111 "src/xnnpack/math-stubs.h",
4112 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004113 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004114 deps = MICROKERNEL_TEST_DEPS,
4115)
4116
Marat Dukhan346a9e52019-11-15 09:06:30 -08004117xnnpack_benchmark(
4118 name = "f32_sigmoid_eval",
4119 srcs = [
4120 "eval/f32-sigmoid.cc",
4121 "src/xnnpack/AlignedAllocator.h",
4122 ] + ACCURACY_EVAL_HDRS,
4123 deps = ACCURACY_EVAL_DEPS,
4124)
4125
Marat Dukhan84000762020-06-29 18:38:43 -07004126xnnpack_benchmark(
4127 name = "f32_sqrt_eval",
4128 srcs = [
4129 "eval/f32-sqrt.cc",
4130 "src/xnnpack/AlignedAllocator.h",
4131 ] + ACCURACY_EVAL_HDRS,
4132 deps = ACCURACY_EVAL_DEPS + [
4133 ":bench_utils",
4134 ],
4135)
4136
Marat Dukhan08c4a432019-10-03 09:29:21 -07004137######################### Unit tests for micro-kernels #########################
4138
4139xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004140 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004141 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004142 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004143 "test/gemm-microkernel-tester.h",
4144 "src/xnnpack/AlignedAllocator.h",
4145 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004146 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004147)
4148
4149xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004150 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004151 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004152 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004153 "test/spmm-microkernel-tester.h",
4154 "src/xnnpack/AlignedAllocator.h",
4155 ] + MICROKERNEL_TEST_HDRS,
4156 deps = MICROKERNEL_TEST_DEPS,
4157)
4158
4159xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004160 name = "f32_argmaxpool_test",
4161 srcs = [
4162 "test/f32-argmaxpool.cc",
4163 "test/argmaxpool-microkernel-tester.h",
4164 "src/xnnpack/AlignedAllocator.h",
4165 ] + MICROKERNEL_TEST_HDRS,
4166 deps = MICROKERNEL_TEST_DEPS,
4167)
4168
4169xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004170 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004171 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004172 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004173 "test/avgpool-microkernel-tester.h",
4174 "src/xnnpack/AlignedAllocator.h",
4175 ] + MICROKERNEL_TEST_HDRS,
4176 deps = MICROKERNEL_TEST_DEPS,
4177)
4178
4179xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004180 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004181 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004182 "test/f32-ibilinear.cc",
4183 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004184 "src/xnnpack/AlignedAllocator.h",
4185 ] + MICROKERNEL_TEST_HDRS,
4186 deps = MICROKERNEL_TEST_DEPS,
4187)
4188
4189xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004190 name = "f16_clamp_test",
4191 srcs = [
4192 "test/f16-clamp.cc",
4193 "test/clamp-microkernel-tester.h",
4194 ] + MICROKERNEL_TEST_HDRS,
4195 deps = MICROKERNEL_TEST_DEPS,
4196)
4197
4198xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004199 name = "f32_clamp_test",
4200 srcs = [
4201 "test/f32-clamp.cc",
4202 "test/clamp-microkernel-tester.h",
4203 ] + MICROKERNEL_TEST_HDRS,
4204 deps = MICROKERNEL_TEST_DEPS,
4205)
4206
4207xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004208 name = "f32_igemm_test",
4209 srcs = [
4210 "test/f32-igemm.cc",
4211 "test/gemm-microkernel-tester.h",
4212 "src/xnnpack/AlignedAllocator.h",
4213 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004214 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004215)
4216
4217xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004218 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004219 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004220 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004221 "test/gemm-microkernel-tester.h",
4222 "src/xnnpack/AlignedAllocator.h",
4223 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004224 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004225)
4226
4227xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004228 name = "f32_igemm_minmax_test",
4229 srcs = [
4230 "test/f32-igemm-minmax.cc",
4231 "test/gemm-microkernel-tester.h",
4232 "src/xnnpack/AlignedAllocator.h",
4233 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004234 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004235)
4236
4237xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004238 name = "f16_igemm_minmax_test",
4239 srcs = [
4240 "test/f16-igemm-minmax.cc",
4241 "test/gemm-microkernel-tester.h",
4242 "src/xnnpack/AlignedAllocator.h",
4243 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004244 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004245)
4246
4247xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004248 name = "f32_conv_hwc_test",
4249 srcs = [
4250 "test/f32-conv-hwc.cc",
4251 "test/conv-hwc-microkernel-tester.h",
4252 "src/xnnpack/AlignedAllocator.h",
4253 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004254 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004255)
4256
4257xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004258 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004259 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004260 "test/f32-conv-hwc2chw.cc",
4261 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004262 "src/xnnpack/AlignedAllocator.h",
4263 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004264 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004265)
4266
4267xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004268 name = "f32_dwconv_test",
4269 srcs = [
4270 "test/f32-dwconv.cc",
4271 "test/dwconv-microkernel-tester.h",
4272 "src/xnnpack/AlignedAllocator.h",
4273 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004274 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004275)
4276
4277xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004278 name = "f16_dwconv_minmax_test",
4279 srcs = [
4280 "test/f16-dwconv-minmax.cc",
4281 "test/dwconv-microkernel-tester.h",
4282 "src/xnnpack/AlignedAllocator.h",
4283 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004284 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004285)
4286
4287xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004288 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004290 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004291 "test/dwconv-microkernel-tester.h",
4292 "src/xnnpack/AlignedAllocator.h",
4293 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004294 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004295)
4296
4297xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004298 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004300 "test/f32-dwconv-chw.cc",
4301 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004302 "src/xnnpack/AlignedAllocator.h",
4303 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004304 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305)
4306
4307xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004308 name = "f16_gavgpool_minmax_test",
4309 srcs = [
4310 "test/f16-gavgpool-minmax.cc",
4311 "test/gavgpool-microkernel-tester.h",
4312 "src/xnnpack/AlignedAllocator.h",
4313 ] + MICROKERNEL_TEST_HDRS,
4314 deps = MICROKERNEL_TEST_DEPS,
4315)
4316
4317xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004318 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004320 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004321 "test/gavgpool-microkernel-tester.h",
4322 "src/xnnpack/AlignedAllocator.h",
4323 ] + MICROKERNEL_TEST_HDRS,
4324 deps = MICROKERNEL_TEST_DEPS,
4325)
4326
4327xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004328 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004330 "test/f32-gavgpool-cw.cc",
4331 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004332 "src/xnnpack/AlignedAllocator.h",
4333 ] + MICROKERNEL_TEST_HDRS,
4334 deps = MICROKERNEL_TEST_DEPS,
4335)
4336
4337xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004338 name = "f32_gemm_test",
4339 srcs = [
4340 "test/f32-gemm.cc",
4341 "test/gemm-microkernel-tester.h",
4342 "src/xnnpack/AlignedAllocator.h",
4343 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004344 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004345)
4346
4347xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004348 name = "f32_gemm_relu_test",
4349 srcs = [
4350 "test/f32-gemm-relu.cc",
4351 "test/gemm-microkernel-tester.h",
4352 "src/xnnpack/AlignedAllocator.h",
4353 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004354 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004355)
4356
4357xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004358 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004359 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004360 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004361 "test/gemm-microkernel-tester.h",
4362 "src/xnnpack/AlignedAllocator.h",
4363 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004364 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365)
4366
4367xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004368 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004369 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004370 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371 "test/gemm-microkernel-tester.h",
4372 "src/xnnpack/AlignedAllocator.h",
4373 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004374 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004375)
4376
4377xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004378 name = "f16_hswish_test",
4379 srcs = [
4380 "test/f16-hswish.cc",
4381 "test/hswish-microkernel-tester.h",
4382 ] + MICROKERNEL_TEST_HDRS,
4383 deps = MICROKERNEL_TEST_DEPS,
4384)
4385
4386xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004387 name = "f32_hswish_test",
4388 srcs = [
4389 "test/f32-hswish.cc",
4390 "test/hswish-microkernel-tester.h",
4391 ] + MICROKERNEL_TEST_HDRS,
4392 deps = MICROKERNEL_TEST_DEPS,
4393)
4394
4395xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004396 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004397 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004398 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004399 "test/maxpool-microkernel-tester.h",
4400 ] + MICROKERNEL_TEST_HDRS,
4401 deps = MICROKERNEL_TEST_DEPS,
4402)
4403
4404xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004405 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004406 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004407 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004408 "test/avgpool-microkernel-tester.h",
4409 "src/xnnpack/AlignedAllocator.h",
4410 ] + MICROKERNEL_TEST_HDRS,
4411 deps = MICROKERNEL_TEST_DEPS,
4412)
4413
4414xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004415 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004417 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004418 "test/gemm-microkernel-tester.h",
4419 "src/xnnpack/AlignedAllocator.h",
4420 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004421 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004422)
4423
4424xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004425 name = "f16_prelu_test",
4426 srcs = [
4427 "test/f16-prelu.cc",
4428 "test/prelu-microkernel-tester.h",
4429 "src/xnnpack/AlignedAllocator.h",
4430 ] + MICROKERNEL_TEST_HDRS,
4431 deps = MICROKERNEL_TEST_DEPS,
4432)
4433
4434xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004435 name = "f32_prelu_test",
4436 srcs = [
4437 "test/f32-prelu.cc",
4438 "test/prelu-microkernel-tester.h",
4439 "src/xnnpack/AlignedAllocator.h",
4440 ] + MICROKERNEL_TEST_HDRS,
4441 deps = MICROKERNEL_TEST_DEPS,
4442)
4443
4444xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004445 name = "f32_raddexpminusmax_test",
4446 srcs = [
4447 "test/f32-raddexpminusmax.cc",
4448 "test/raddexpminusmax-microkernel-tester.h",
4449 ] + MICROKERNEL_TEST_HDRS,
4450 deps = MICROKERNEL_TEST_DEPS,
4451)
4452
4453xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004454 name = "f32_raddextexp_test",
4455 srcs = [
4456 "test/f32-raddextexp.cc",
4457 "test/raddextexp-microkernel-tester.h",
4458 ] + MICROKERNEL_TEST_HDRS,
4459 deps = MICROKERNEL_TEST_DEPS,
4460)
4461
4462xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004463 name = "f32_raddstoreexpminusmax_test",
4464 srcs = [
4465 "test/f32-raddstoreexpminusmax.cc",
4466 "test/raddstoreexpminusmax-microkernel-tester.h",
4467 ] + MICROKERNEL_TEST_HDRS,
4468 deps = MICROKERNEL_TEST_DEPS,
4469)
4470
4471xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004472 name = "f32_rmax_test",
4473 srcs = [
4474 "test/f32-rmax.cc",
4475 "test/rmax-microkernel-tester.h",
4476 ] + MICROKERNEL_TEST_HDRS,
4477 deps = MICROKERNEL_TEST_DEPS,
4478)
4479
4480xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004481 name = "f32_sigmoid_test",
4482 srcs = [
4483 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004484 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004485 ] + MICROKERNEL_TEST_HDRS,
4486 deps = MICROKERNEL_TEST_DEPS,
4487)
4488
4489xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004490 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004491 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004492 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493 "test/spmm-microkernel-tester.h",
4494 "src/xnnpack/AlignedAllocator.h",
4495 ] + MICROKERNEL_TEST_HDRS,
4496 deps = MICROKERNEL_TEST_DEPS,
4497)
4498
4499xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004500 name = "f16_vadd_minmax_test",
4501 srcs = [
4502 "test/f16-vadd-minmax.cc",
4503 "test/vbinary-microkernel-tester.h",
4504 ] + MICROKERNEL_TEST_HDRS,
4505 deps = MICROKERNEL_TEST_DEPS,
4506)
4507
4508xnnpack_unit_test(
4509 name = "f16_vaddc_minmax_test",
4510 srcs = [
4511 "test/f16-vaddc-minmax.cc",
4512 "test/vbinaryc-microkernel-tester.h",
4513 ] + MICROKERNEL_TEST_HDRS,
4514 deps = MICROKERNEL_TEST_DEPS,
4515)
4516
4517xnnpack_unit_test(
4518 name = "f16_vdiv_minmax_test",
4519 srcs = [
4520 "test/f16-vdiv-minmax.cc",
4521 "test/vbinary-microkernel-tester.h",
4522 ] + MICROKERNEL_TEST_HDRS,
4523 deps = MICROKERNEL_TEST_DEPS,
4524)
4525
4526xnnpack_unit_test(
4527 name = "f16_vdivc_minmax_test",
4528 srcs = [
4529 "test/f16-vdivc-minmax.cc",
4530 "test/vbinaryc-microkernel-tester.h",
4531 ] + MICROKERNEL_TEST_HDRS,
4532 deps = MICROKERNEL_TEST_DEPS,
4533)
4534
4535xnnpack_unit_test(
4536 name = "f16_vrdivc_minmax_test",
4537 srcs = [
4538 "test/f16-vrdivc-minmax.cc",
4539 "test/vbinaryc-microkernel-tester.h",
4540 ] + MICROKERNEL_TEST_HDRS,
4541 deps = MICROKERNEL_TEST_DEPS,
4542)
4543
4544xnnpack_unit_test(
4545 name = "f16_vmax_test",
4546 srcs = [
4547 "test/f16-vmax.cc",
4548 "test/vbinary-microkernel-tester.h",
4549 ] + MICROKERNEL_TEST_HDRS,
4550 deps = MICROKERNEL_TEST_DEPS,
4551)
4552
4553xnnpack_unit_test(
4554 name = "f16_vmaxc_test",
4555 srcs = [
4556 "test/f16-vmaxc.cc",
4557 "test/vbinaryc-microkernel-tester.h",
4558 ] + MICROKERNEL_TEST_HDRS,
4559 deps = MICROKERNEL_TEST_DEPS,
4560)
4561
4562xnnpack_unit_test(
4563 name = "f16_vmin_test",
4564 srcs = [
4565 "test/f16-vmin.cc",
4566 "test/vbinary-microkernel-tester.h",
4567 ] + MICROKERNEL_TEST_HDRS,
4568 deps = MICROKERNEL_TEST_DEPS,
4569)
4570
4571xnnpack_unit_test(
4572 name = "f16_vminc_test",
4573 srcs = [
4574 "test/f16-vminc.cc",
4575 "test/vbinaryc-microkernel-tester.h",
4576 ] + MICROKERNEL_TEST_HDRS,
4577 deps = MICROKERNEL_TEST_DEPS,
4578)
4579
4580xnnpack_unit_test(
4581 name = "f16_vmul_minmax_test",
4582 srcs = [
4583 "test/f16-vmul-minmax.cc",
4584 "test/vbinary-microkernel-tester.h",
4585 ] + MICROKERNEL_TEST_HDRS,
4586 deps = MICROKERNEL_TEST_DEPS,
4587)
4588
4589xnnpack_unit_test(
4590 name = "f16_vmulc_minmax_test",
4591 srcs = [
4592 "test/f16-vmulc-minmax.cc",
4593 "test/vbinaryc-microkernel-tester.h",
4594 ] + MICROKERNEL_TEST_HDRS,
4595 deps = MICROKERNEL_TEST_DEPS,
4596)
4597
4598xnnpack_unit_test(
4599 name = "f16_vsub_minmax_test",
4600 srcs = [
4601 "test/f16-vsub-minmax.cc",
4602 "test/vbinary-microkernel-tester.h",
4603 ] + MICROKERNEL_TEST_HDRS,
4604 deps = MICROKERNEL_TEST_DEPS,
4605)
4606
4607xnnpack_unit_test(
4608 name = "f16_vsubc_minmax_test",
4609 srcs = [
4610 "test/f16-vsubc-minmax.cc",
4611 "test/vbinaryc-microkernel-tester.h",
4612 ] + MICROKERNEL_TEST_HDRS,
4613 deps = MICROKERNEL_TEST_DEPS,
4614)
4615
4616xnnpack_unit_test(
4617 name = "f16_vrsubc_minmax_test",
4618 srcs = [
4619 "test/f16-vrsubc-minmax.cc",
4620 "test/vbinaryc-microkernel-tester.h",
4621 ] + MICROKERNEL_TEST_HDRS,
4622 deps = MICROKERNEL_TEST_DEPS,
4623)
4624
4625xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004626 name = "f32_vabs_test",
4627 srcs = [
4628 "test/f32-vabs.cc",
4629 "test/vunary-microkernel-tester.h",
4630 ] + MICROKERNEL_TEST_HDRS,
4631 deps = MICROKERNEL_TEST_DEPS,
4632)
4633
4634xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004635 name = "f32_vadd_test",
4636 srcs = [
4637 "test/f32-vadd.cc",
4638 "test/vbinary-microkernel-tester.h",
4639 ] + MICROKERNEL_TEST_HDRS,
4640 deps = MICROKERNEL_TEST_DEPS,
4641)
4642
4643xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004644 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004645 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004646 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004647 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004648 ] + MICROKERNEL_TEST_HDRS,
4649 deps = MICROKERNEL_TEST_DEPS,
4650)
4651
4652xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004653 name = "f32_vaddc_test",
4654 srcs = [
4655 "test/f32-vaddc.cc",
4656 "test/vbinaryc-microkernel-tester.h",
4657 ] + MICROKERNEL_TEST_HDRS,
4658 deps = MICROKERNEL_TEST_DEPS,
4659)
4660
4661xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004662 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004663 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004664 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004665 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004666 ] + MICROKERNEL_TEST_HDRS,
4667 deps = MICROKERNEL_TEST_DEPS,
4668)
4669
4670xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004671 name = "f32_vdiv_test",
4672 srcs = [
4673 "test/f32-vdiv.cc",
4674 "test/vbinary-microkernel-tester.h",
4675 ] + MICROKERNEL_TEST_HDRS,
4676 deps = MICROKERNEL_TEST_DEPS,
4677)
4678
4679xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004680 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004681 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004682 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004683 "test/vbinary-microkernel-tester.h",
4684 ] + MICROKERNEL_TEST_HDRS,
4685 deps = MICROKERNEL_TEST_DEPS,
4686)
4687
4688xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004689 name = "f32_vdivc_test",
4690 srcs = [
4691 "test/f32-vdivc.cc",
4692 "test/vbinaryc-microkernel-tester.h",
4693 ] + MICROKERNEL_TEST_HDRS,
4694 deps = MICROKERNEL_TEST_DEPS,
4695)
4696
4697xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004698 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004699 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004700 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004701 "test/vbinaryc-microkernel-tester.h",
4702 ] + MICROKERNEL_TEST_HDRS,
4703 deps = MICROKERNEL_TEST_DEPS,
4704)
4705
4706xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004707 name = "f32_vrdivc_test",
4708 srcs = [
4709 "test/f32-vrdivc.cc",
4710 "test/vbinaryc-microkernel-tester.h",
4711 ] + MICROKERNEL_TEST_HDRS,
4712 deps = MICROKERNEL_TEST_DEPS,
4713)
4714
4715xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004716 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004717 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004718 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004719 "test/vbinaryc-microkernel-tester.h",
4720 ] + MICROKERNEL_TEST_HDRS,
4721 deps = MICROKERNEL_TEST_DEPS,
4722)
4723
4724xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004725 name = "f32_vmax_test",
4726 srcs = [
4727 "test/f32-vmax.cc",
4728 "test/vbinary-microkernel-tester.h",
4729 ] + MICROKERNEL_TEST_HDRS,
4730 deps = MICROKERNEL_TEST_DEPS,
4731)
4732
4733xnnpack_unit_test(
4734 name = "f32_vmaxc_test",
4735 srcs = [
4736 "test/f32-vmaxc.cc",
4737 "test/vbinaryc-microkernel-tester.h",
4738 ] + MICROKERNEL_TEST_HDRS,
4739 deps = MICROKERNEL_TEST_DEPS,
4740)
4741
4742xnnpack_unit_test(
4743 name = "f32_vmin_test",
4744 srcs = [
4745 "test/f32-vmin.cc",
4746 "test/vbinary-microkernel-tester.h",
4747 ] + MICROKERNEL_TEST_HDRS,
4748 deps = MICROKERNEL_TEST_DEPS,
4749)
4750
4751xnnpack_unit_test(
4752 name = "f32_vminc_test",
4753 srcs = [
4754 "test/f32-vminc.cc",
4755 "test/vbinaryc-microkernel-tester.h",
4756 ] + MICROKERNEL_TEST_HDRS,
4757 deps = MICROKERNEL_TEST_DEPS,
4758)
4759
4760xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004761 name = "f32_vmul_test",
4762 srcs = [
4763 "test/f32-vmul.cc",
4764 "test/vbinary-microkernel-tester.h",
4765 ] + MICROKERNEL_TEST_HDRS,
4766 deps = MICROKERNEL_TEST_DEPS,
4767)
4768
4769xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004770 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004771 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004772 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004773 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004774 ] + MICROKERNEL_TEST_HDRS,
4775 deps = MICROKERNEL_TEST_DEPS,
4776)
4777
4778xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004779 name = "f32_vmulc_test",
4780 srcs = [
4781 "test/f32-vmulc.cc",
4782 "test/vbinaryc-microkernel-tester.h",
4783 ] + MICROKERNEL_TEST_HDRS,
4784 deps = MICROKERNEL_TEST_DEPS,
4785)
4786
4787xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004788 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004789 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004790 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004791 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004792 ] + MICROKERNEL_TEST_HDRS,
4793 deps = MICROKERNEL_TEST_DEPS,
4794)
4795
4796xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004797 name = "f16_vmulcaddc_minmax_test",
4798 srcs = [
4799 "test/f16-vmulcaddc-minmax.cc",
4800 "test/vmulcaddc-microkernel-tester.h",
4801 "src/xnnpack/AlignedAllocator.h",
4802 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004803 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004804)
4805
4806xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004807 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004808 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004809 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004810 "test/vmulcaddc-microkernel-tester.h",
4811 "src/xnnpack/AlignedAllocator.h",
4812 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004813 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004814)
4815
4816xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004817 name = "f32_vlrelu_test",
4818 srcs = [
4819 "test/f32-vlrelu.cc",
4820 "test/vunary-microkernel-tester.h",
4821 ] + MICROKERNEL_TEST_HDRS,
4822 deps = MICROKERNEL_TEST_DEPS,
4823)
4824
4825xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004826 name = "f32_vneg_test",
4827 srcs = [
4828 "test/f32-vneg.cc",
4829 "test/vunary-microkernel-tester.h",
4830 ] + MICROKERNEL_TEST_HDRS,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004835 name = "f32_vrndne_test",
4836 srcs = [
4837 "test/f32-vrndne.cc",
4838 "test/vunary-microkernel-tester.h",
4839 ] + MICROKERNEL_TEST_HDRS,
4840 deps = MICROKERNEL_TEST_DEPS,
4841)
4842
4843xnnpack_unit_test(
4844 name = "f32_vrndz_test",
4845 srcs = [
4846 "test/f32-vrndz.cc",
4847 "test/vunary-microkernel-tester.h",
4848 ] + MICROKERNEL_TEST_HDRS,
4849 deps = MICROKERNEL_TEST_DEPS,
4850)
4851
4852xnnpack_unit_test(
4853 name = "f32_vrndu_test",
4854 srcs = [
4855 "test/f32-vrndu.cc",
4856 "test/vunary-microkernel-tester.h",
4857 ] + MICROKERNEL_TEST_HDRS,
4858 deps = MICROKERNEL_TEST_DEPS,
4859)
4860
4861xnnpack_unit_test(
4862 name = "f32_vrndd_test",
4863 srcs = [
4864 "test/f32-vrndd.cc",
4865 "test/vunary-microkernel-tester.h",
4866 ] + MICROKERNEL_TEST_HDRS,
4867 deps = MICROKERNEL_TEST_DEPS,
4868)
4869
4870xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004871 name = "f32_vscale_test",
4872 srcs = [
4873 "test/f32-vscale.cc",
4874 "test/vscale-microkernel-tester.h",
4875 ] + MICROKERNEL_TEST_HDRS,
4876 deps = MICROKERNEL_TEST_DEPS,
4877)
4878
4879xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004880 name = "f32_vscaleexpminusmax_test",
4881 srcs = [
4882 "test/f32-vscaleexpminusmax.cc",
4883 "test/vscaleexpminusmax-microkernel-tester.h",
4884 ] + MICROKERNEL_TEST_HDRS,
4885 deps = MICROKERNEL_TEST_DEPS,
4886)
4887
4888xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004889 name = "f32_vscaleextexp_test",
4890 srcs = [
4891 "test/f32-vscaleextexp.cc",
4892 "test/vscaleextexp-microkernel-tester.h",
4893 ] + MICROKERNEL_TEST_HDRS,
4894 deps = MICROKERNEL_TEST_DEPS,
4895)
4896
4897xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004898 name = "f32_vsqr_test",
4899 srcs = [
4900 "test/f32-vsqr.cc",
4901 "test/vunary-microkernel-tester.h",
4902 ] + MICROKERNEL_TEST_HDRS,
4903 deps = MICROKERNEL_TEST_DEPS,
4904)
4905
4906xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004907 name = "f32_vsqrdiff_test",
4908 srcs = [
4909 "test/f32-vsqrdiff.cc",
4910 "test/vbinary-microkernel-tester.h",
4911 ] + MICROKERNEL_TEST_HDRS,
4912 deps = MICROKERNEL_TEST_DEPS,
4913)
4914
4915xnnpack_unit_test(
4916 name = "f32_vsqrdiffc_test",
4917 srcs = [
4918 "test/f32-vsqrdiffc.cc",
4919 "test/vbinaryc-microkernel-tester.h",
4920 ] + MICROKERNEL_TEST_HDRS,
4921 deps = MICROKERNEL_TEST_DEPS,
4922)
4923
4924xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004925 name = "f32_vsqrt_test",
4926 srcs = [
4927 "test/f32-vsqrt.cc",
4928 "test/vunary-microkernel-tester.h",
4929 ] + MICROKERNEL_TEST_HDRS,
4930 deps = MICROKERNEL_TEST_DEPS,
4931)
4932
4933xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004934 name = "f32_vsub_test",
4935 srcs = [
4936 "test/f32-vsub.cc",
4937 "test/vbinary-microkernel-tester.h",
4938 ] + MICROKERNEL_TEST_HDRS,
4939 deps = MICROKERNEL_TEST_DEPS,
4940)
4941
4942xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004943 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004944 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004945 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004946 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004947 ] + MICROKERNEL_TEST_HDRS,
4948 deps = MICROKERNEL_TEST_DEPS,
4949)
4950
4951xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004952 name = "f32_vsubc_test",
4953 srcs = [
4954 "test/f32-vsubc.cc",
4955 "test/vbinaryc-microkernel-tester.h",
4956 ] + MICROKERNEL_TEST_HDRS,
4957 deps = MICROKERNEL_TEST_DEPS,
4958)
4959
4960xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004961 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004962 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004963 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004964 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004965 ] + MICROKERNEL_TEST_HDRS,
4966 deps = MICROKERNEL_TEST_DEPS,
4967)
4968
4969xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004970 name = "f32_vrsubc_test",
4971 srcs = [
4972 "test/f32-vrsubc.cc",
4973 "test/vbinaryc-microkernel-tester.h",
4974 ] + MICROKERNEL_TEST_HDRS,
4975 deps = MICROKERNEL_TEST_DEPS,
4976)
4977
4978xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004979 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004980 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004981 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004982 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004983 ] + MICROKERNEL_TEST_HDRS,
4984 deps = MICROKERNEL_TEST_DEPS,
4985)
4986
4987xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004988 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004990 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004991 "test/avgpool-microkernel-tester.h",
4992 "src/xnnpack/AlignedAllocator.h",
4993 ] + MICROKERNEL_TEST_HDRS,
4994 deps = MICROKERNEL_TEST_DEPS,
4995)
4996
4997xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004998 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005000 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005001 "test/gemm-microkernel-tester.h",
5002 "src/xnnpack/AlignedAllocator.h",
5003 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005004 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005)
5006
5007xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005008 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005010 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005011 "test/dwconv-microkernel-tester.h",
5012 "src/xnnpack/AlignedAllocator.h",
5013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005014 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015)
5016
5017xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005018 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005019 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005020 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005021 "test/gavgpool-microkernel-tester.h",
5022 "src/xnnpack/AlignedAllocator.h",
5023 ] + MICROKERNEL_TEST_HDRS,
5024 deps = MICROKERNEL_TEST_DEPS,
5025)
5026
5027xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005028 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005029 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005030 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005031 "test/gemm-microkernel-tester.h",
5032 "src/xnnpack/AlignedAllocator.h",
5033 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005034 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005035)
5036
5037xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005038 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005039 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005040 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005041 "test/vadd-microkernel-tester.h",
5042 ] + MICROKERNEL_TEST_HDRS,
5043 deps = MICROKERNEL_TEST_DEPS,
5044)
5045
5046xnnpack_unit_test(
5047 name = "u8_clamp_test",
5048 srcs = [
5049 "test/u8-clamp.cc",
5050 "test/clamp-microkernel-tester.h",
5051 ] + MICROKERNEL_TEST_HDRS,
5052 deps = MICROKERNEL_TEST_DEPS,
5053)
5054
5055xnnpack_unit_test(
5056 name = "u8_lut32norm_test",
5057 srcs = [
5058 "test/u8-lut32norm.cc",
5059 "test/lut-norm-microkernel-tester.h",
5060 ] + MICROKERNEL_TEST_HDRS,
5061 deps = MICROKERNEL_TEST_DEPS,
5062)
5063
5064xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005065 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005066 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005067 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005068 "test/maxpool-microkernel-tester.h",
5069 ] + MICROKERNEL_TEST_HDRS,
5070 deps = MICROKERNEL_TEST_DEPS,
5071)
5072
5073xnnpack_unit_test(
5074 name = "u8_rmax_test",
5075 srcs = [
5076 "test/u8-rmax.cc",
5077 "test/rmax-microkernel-tester.h",
5078 ] + MICROKERNEL_TEST_HDRS,
5079 deps = MICROKERNEL_TEST_DEPS,
5080)
5081
5082xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005083 name = "x32_fill_test",
5084 srcs = [
5085 "test/x32-fill.cc",
5086 "test/fill-microkernel-tester.h",
5087 ] + MICROKERNEL_TEST_HDRS,
5088 deps = MICROKERNEL_TEST_DEPS,
5089)
5090
5091xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092 name = "x32_packx_test",
5093 srcs = [
5094 "test/x32-packx.cc",
5095 "test/pack-microkernel-tester.h",
5096 "src/xnnpack/AlignedAllocator.h",
5097 ] + MICROKERNEL_TEST_HDRS,
5098 deps = MICROKERNEL_TEST_DEPS,
5099)
5100
5101xnnpack_unit_test(
5102 name = "x32_pad_test",
5103 srcs = [
5104 "test/x32-pad.cc",
5105 "test/pad-microkernel-tester.h",
5106 ] + MICROKERNEL_TEST_HDRS,
5107 deps = MICROKERNEL_TEST_DEPS,
5108)
5109
5110xnnpack_unit_test(
5111 name = "x32_unpool_test",
5112 srcs = [
5113 "test/x32-unpool.cc",
5114 "test/unpool-microkernel-tester.h",
5115 ] + MICROKERNEL_TEST_HDRS,
5116 deps = MICROKERNEL_TEST_DEPS,
5117)
5118
5119xnnpack_unit_test(
5120 name = "x32_zip_test",
5121 srcs = [
5122 "test/x32-zip.cc",
5123 "test/zip-microkernel-tester.h",
5124 ] + MICROKERNEL_TEST_HDRS,
5125 deps = MICROKERNEL_TEST_DEPS,
5126)
5127
5128xnnpack_unit_test(
5129 name = "x8_lut_test",
5130 srcs = [
5131 "test/x8-lut.cc",
5132 "test/lut-microkernel-tester.h",
5133 ] + MICROKERNEL_TEST_HDRS,
5134 deps = MICROKERNEL_TEST_DEPS,
5135)
5136
5137xnnpack_unit_test(
5138 name = "x8_zip_test",
5139 srcs = [
5140 "test/x8-zip.cc",
5141 "test/zip-microkernel-tester.h",
5142 ] + MICROKERNEL_TEST_HDRS,
5143 deps = MICROKERNEL_TEST_DEPS,
5144)
5145
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005146xnnpack_unit_test(
5147 name = "requantization_test",
5148 srcs = [
5149 "src/xnnpack/requantization-stubs.h",
5150 "test/requantization.cc",
5151 "test/requantization-tester.h",
5152 ] + MICROKERNEL_TEST_HDRS,
5153 deps = MICROKERNEL_TEST_DEPS,
5154)
5155
Marat Dukhan20c3b922020-03-10 03:45:06 -07005156########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005157
5158xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005159 name = "operator_size_test",
5160 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005161 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005162)
5163
Marat Dukhan20c3b922020-03-10 03:45:06 -07005164xnnpack_binary(
5165 name = "subgraph_size_test",
5166 srcs = ["test/subgraph-size.c"],
5167 deps = [":XNNPACK"],
5168)
5169
5170########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005171
5172xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005173 name = "abs_nc_test",
5174 srcs = [
5175 "test/abs-nc.cc",
5176 "test/abs-operator-tester.h",
5177 ],
5178 deps = OPERATOR_TEST_DEPS,
5179)
5180
5181xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005182 name = "add_nd_test",
5183 srcs = [
5184 "test/add-nd.cc",
5185 "test/binary-elementwise-operator-tester.h",
5186 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005187 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005188)
5189
5190xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005191 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005192 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005193 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 "test/argmax-pooling-operator-tester.h",
5195 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005196 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197)
5198
5199xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005200 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005201 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005202 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005203 "test/average-pooling-operator-tester.h",
5204 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005205 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206)
5207
5208xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005209 name = "bankers_rounding_nc_test",
5210 srcs = [
5211 "test/bankers-rounding-nc.cc",
5212 "test/bankers-rounding-operator-tester.h",
5213 ],
5214 deps = OPERATOR_TEST_DEPS,
5215)
5216
5217xnnpack_unit_test(
5218 name = "ceiling_nc_test",
5219 srcs = [
5220 "test/ceiling-nc.cc",
5221 "test/ceiling-operator-tester.h",
5222 ],
5223 deps = OPERATOR_TEST_DEPS,
5224)
5225
5226xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005227 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005228 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005229 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230 "test/channel-shuffle-operator-tester.h",
5231 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005232 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005233)
5234
5235xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005236 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005238 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005239 "test/clamp-operator-tester.h",
5240 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005241 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005242)
5243
5244xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005245 name = "constant_pad_nd_test",
5246 srcs = [
5247 "test/constant-pad-nd.cc",
5248 "test/constant-pad-operator-tester.h",
5249 ],
5250 deps = OPERATOR_TEST_DEPS,
5251)
5252
5253xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005254 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005255 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005256 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005257 "test/convolution-operator-tester.h",
5258 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005259 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260)
5261
5262xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005263 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005265 "test/convolution-nchw.cc",
5266 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005268 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005269)
5270
5271xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005272 name = "copy_nc_test",
5273 srcs = [
5274 "test/copy-nc.cc",
5275 "test/copy-operator-tester.h",
5276 ],
5277 deps = OPERATOR_TEST_DEPS,
5278)
5279
5280xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005281 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005283 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 "test/deconvolution-operator-tester.h",
5285 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005286 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287)
5288
5289xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005290 name = "divide_nd_test",
5291 srcs = [
5292 "test/binary-elementwise-operator-tester.h",
5293 "test/divide-nd.cc",
5294 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005295 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005296)
5297
5298xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005299 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005301 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005302 "test/fully-connected-operator-tester.h",
5303 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005304 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305)
5306
5307xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005308 name = "floor_nc_test",
5309 srcs = [
5310 "test/floor-nc.cc",
5311 "test/floor-operator-tester.h",
5312 ],
5313 deps = OPERATOR_TEST_DEPS,
5314)
5315
5316xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005317 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005319 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005320 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005321 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005322 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323)
5324
5325xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005326 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005328 "test/global-average-pooling-ncw.cc",
5329 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005330 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005331 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005332)
5333
5334xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005335 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005337 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005338 "test/hardswish-operator-tester.h",
5339 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005340 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005341)
5342
5343xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005344 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005346 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005347 "test/leaky-relu-operator-tester.h",
5348 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005349 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005350)
5351
5352xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005353 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005354 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005355 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005356 "test/max-pooling-operator-tester.h",
5357 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005358 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005359)
5360
5361xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005362 name = "maximum_nd_test",
5363 srcs = [
5364 "test/binary-elementwise-operator-tester.h",
5365 "test/maximum-nd.cc",
5366 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005367 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005368)
5369
5370xnnpack_unit_test(
5371 name = "minimum_nd_test",
5372 srcs = [
5373 "test/binary-elementwise-operator-tester.h",
5374 "test/minimum-nd.cc",
5375 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005376 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005377)
5378
5379xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005380 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005381 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005382 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005383 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005384 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005385 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005386)
5387
5388xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005389 name = "negate_nc_test",
5390 srcs = [
5391 "test/negate-nc.cc",
5392 "test/negate-operator-tester.h",
5393 ],
5394 deps = OPERATOR_TEST_DEPS,
5395)
5396
5397xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005398 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005400 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005401 "test/prelu-operator-tester.h",
5402 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005403 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005404)
5405
5406xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005407 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005408 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005409 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005410 "test/resize-bilinear-operator-tester.h",
5411 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005412 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005413)
5414
5415xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005416 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005418 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005419 "test/sigmoid-operator-tester.h",
5420 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005421 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005422)
5423
5424xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005425 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005426 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005427 "test/softmax-nc.cc",
5428 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005429 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005430 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005431)
5432
5433xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005434 name = "square_nc_test",
5435 srcs = [
5436 "test/square-nc.cc",
5437 "test/square-operator-tester.h",
5438 ],
5439 deps = OPERATOR_TEST_DEPS,
5440)
5441
5442xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005443 name = "square_root_nc_test",
5444 srcs = [
5445 "test/square-root-nc.cc",
5446 "test/square-root-operator-tester.h",
5447 ],
5448 deps = OPERATOR_TEST_DEPS,
5449)
5450
5451xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005452 name = "squared_difference_nd_test",
5453 srcs = [
5454 "test/binary-elementwise-operator-tester.h",
5455 "test/squared-difference-nd.cc",
5456 ],
5457 deps = OPERATOR_TEST_DEPS,
5458)
5459
5460xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005461 name = "subtract_nd_test",
5462 srcs = [
5463 "test/binary-elementwise-operator-tester.h",
5464 "test/subtract-nd.cc",
5465 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005466 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005467)
5468
5469xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005470 name = "truncation_nc_test",
5471 srcs = [
5472 "test/truncation-nc.cc",
5473 "test/truncation-operator-tester.h",
5474 ],
5475 deps = OPERATOR_TEST_DEPS,
5476)
5477
5478xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005479 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005480 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005481 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005482 "test/unpooling-operator-tester.h",
5483 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005484 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005485)
5486
Chao Mei6ddfc602020-05-13 22:29:36 -07005487############################### Misc unit tests ###############################
5488
5489xnnpack_unit_test(
5490 name = "memory_planner_test",
5491 srcs = [
5492 "test/memory-planner-test.cc",
5493 ],
5494 deps = [
5495 ":XNNPACK",
5496 ":memory_planner",
5497 ],
5498)
5499
Marat Dukhan08c4a432019-10-03 09:29:21 -07005500############################# Build configurations #############################
5501
Marat Dukhanb8642352019-10-30 15:43:02 -07005502# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005503config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005504 name = "xnn_enable_assembly_explicit_true",
5505 define_values = {"xnn_enable_assembly": "true"},
5506)
5507
5508# Disables usage of assembly kernels.
5509config_setting(
5510 name = "xnn_enable_assembly_explicit_false",
5511 define_values = {"xnn_enable_assembly": "false"},
5512)
5513
Marat Dukhan9de90e02020-06-18 16:04:12 -07005514# Enables usage of sparse inference.
5515config_setting(
5516 name = "xnn_enable_sparse_explicit_true",
5517 define_values = {"xnn_enable_sparse": "true"},
5518)
5519
5520# Disables usage of sparse inference.
5521config_setting(
5522 name = "xnn_enable_sparse_explicit_false",
5523 define_values = {"xnn_enable_sparse": "false"},
5524)
5525
Marat Dukhan05702cf2020-03-26 15:41:33 -07005526# Disables usage of HMP-aware optimizations.
5527config_setting(
5528 name = "xnn_enable_hmp_explicit_false",
5529 define_values = {"xnn_enable_hmp": "false"},
5530)
5531
Chao Mei6ddfc602020-05-13 22:29:36 -07005532# Enable usage of optimized memory allocation
5533config_setting(
5534 name = "xnn_enable_memopt_explicit_true",
5535 define_values = {"xnnpack_enable_memopt": "true"},
5536)
5537
5538# Disable usage of optimized memory allocation
5539config_setting(
5540 name = "xnn_enable_memopt_explicit_false",
5541 define_values = {"xnnpack_enable_memopt": "false"},
5542)
5543
Marat Dukhanb8642352019-10-30 15:43:02 -07005544# Builds with -c dbg
5545config_setting(
5546 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005547 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005548 "compilation_mode": "dbg",
5549 },
5550)
5551
5552# Builds with -c opt
5553config_setting(
5554 name = "optimized_build",
5555 values = {
5556 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005557 },
5558)
5559
5560config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005561 name = "linux_k8",
5562 values = {"cpu": "k8"},
5563)
5564
5565config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005566 name = "linux_arm",
5567 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005568)
5569
5570config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005571 name = "linux_armeabi",
5572 values = {"cpu": "armeabi"},
5573)
5574
5575config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005576 name = "linux_armhf",
5577 values = {"cpu": "armhf"},
5578)
5579
5580config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005581 name = "linux_armv7a",
5582 values = {"cpu": "armv7a"},
5583)
5584
5585config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005586 name = "linux_aarch64",
5587 values = {"cpu": "aarch64"},
5588)
5589
5590config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005591 name = "android",
5592 values = {"crosstool_top": "//external:android/crosstool"},
5593)
5594
5595config_setting(
5596 name = "android_armv7",
5597 values = {
5598 "crosstool_top": "//external:android/crosstool",
5599 "cpu": "armeabi-v7a",
5600 },
5601)
5602
5603config_setting(
5604 name = "android_arm64",
5605 values = {
5606 "crosstool_top": "//external:android/crosstool",
5607 "cpu": "arm64-v8a",
5608 },
5609)
5610
5611config_setting(
5612 name = "android_x86",
5613 values = {
5614 "crosstool_top": "//external:android/crosstool",
5615 "cpu": "x86",
5616 },
5617)
5618
5619config_setting(
5620 name = "android_x86_64",
5621 values = {
5622 "crosstool_top": "//external:android/crosstool",
5623 "cpu": "x86_64",
5624 },
5625)
5626
5627config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005628 name = "windows_x86_64",
5629 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005630)
5631
5632config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005633 name = "windows_x86_64_clang",
5634 values = {
5635 "compiler": "clang-cl",
5636 "cpu": "x64_windows",
5637 },
5638)
5639
5640config_setting(
5641 name = "windows_x86_64_mingw",
5642 values = {
5643 "compiler": "mingw-gcc",
5644 "cpu": "x64_windows",
5645 },
5646)
5647
5648config_setting(
5649 name = "windows_x86_64_msys",
5650 values = {
5651 "compiler": "msys-gcc",
5652 "cpu": "x64_windows",
5653 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005654)
5655
5656config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005657 name = "macos_x86_64",
5658 values = {
5659 "apple_platform_type": "macos",
5660 "cpu": "darwin",
5661 },
5662)
5663
5664config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005665 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005666 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005667)
5668
5669config_setting(
5670 name = "emscripten_wasm",
5671 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005672 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005673 "cpu": "wasm",
5674 },
5675)
5676
5677config_setting(
5678 name = "emscripten_wasmsimd",
5679 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005680 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005681 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005682 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005683 },
5684)
5685
5686config_setting(
5687 name = "emscripten_asmjs",
5688 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005689 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005690 "cpu": "asmjs",
5691 },
5692)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005693
5694config_setting(
5695 name = "ios_armv7",
5696 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005697 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005698 "cpu": "ios_armv7",
5699 },
5700)
5701
5702config_setting(
5703 name = "ios_arm64",
5704 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005705 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005706 "cpu": "ios_arm64",
5707 },
5708)
5709
5710config_setting(
5711 name = "ios_arm64e",
5712 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005713 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005714 "cpu": "ios_arm64e",
5715 },
5716)
5717
5718config_setting(
5719 name = "ios_x86",
5720 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005721 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005722 "cpu": "ios_i386",
5723 },
5724)
5725
5726config_setting(
5727 name = "ios_x86_64",
5728 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005729 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005730 "cpu": "ios_x86_64",
5731 },
5732)
5733
5734config_setting(
5735 name = "watchos_armv7k",
5736 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005737 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005738 "cpu": "watchos_armv7k",
5739 },
5740)
5741
5742config_setting(
5743 name = "watchos_arm64_32",
5744 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005745 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005746 "cpu": "watchos_arm64_32",
5747 },
5748)
5749
5750config_setting(
5751 name = "watchos_x86",
5752 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005753 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005754 "cpu": "watchos_i386",
5755 },
5756)
5757
5758config_setting(
5759 name = "watchos_x86_64",
5760 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005761 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005762 "cpu": "watchos_x86_64",
5763 },
5764)
5765
5766config_setting(
5767 name = "tvos_arm64",
5768 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005769 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005770 "cpu": "tvos_arm64",
5771 },
5772)
5773
5774config_setting(
5775 name = "tvos_x86_64",
5776 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005777 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005778 "cpu": "tvos_x86_64",
5779 },
5780)