blob: e92731ee39067b15e070a8b7e7f8b246a6a26d18 [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 Dukhanefaac272020-07-16 16:58:07 -0700172 "src/f32-hswish/gen/hswish-scalar-x1.c",
173 "src/f32-hswish/gen/hswish-scalar-x2.c",
174 "src/f32-hswish/gen/hswish-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",
Frank Barchardfb158e22020-07-15 16:10:10 -0700211 "src/f32-relu/gen/scalar-x1.c",
212 "src/f32-relu/gen/scalar-x2.c",
213 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700214 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800215 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
218 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
221 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700224 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
225 "src/f32-spmm/gen/1x1-minmax-scalar.c",
226 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar.c",
228 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar.c",
230 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar.c",
232 "src/f32-spmm/gen/8x2-minmax-scalar.c",
233 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700234 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
235 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700237 "src/f32-vbinary/gen/vadd-scalar-x1.c",
238 "src/f32-vbinary/gen/vadd-scalar-x2.c",
239 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700240 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700243 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
244 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700246 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700249 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
250 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700252 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700255 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
256 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800258 "src/f32-vbinary/gen/vmax-scalar-x1.c",
259 "src/f32-vbinary/gen/vmax-scalar-x2.c",
260 "src/f32-vbinary/gen/vmax-scalar-x4.c",
261 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
264 "src/f32-vbinary/gen/vmin-scalar-x1.c",
265 "src/f32-vbinary/gen/vmin-scalar-x2.c",
266 "src/f32-vbinary/gen/vmin-scalar-x4.c",
267 "src/f32-vbinary/gen/vminc-scalar-x1.c",
268 "src/f32-vbinary/gen/vminc-scalar-x2.c",
269 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700270 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
271 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700273 "src/f32-vbinary/gen/vmul-scalar-x1.c",
274 "src/f32-vbinary/gen/vmul-scalar-x2.c",
275 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700276 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700279 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
280 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700282 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700285 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
286 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700288 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
292 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700294 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
295 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
297 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700300 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
301 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vsub-scalar-x1.c",
304 "src/f32-vbinary/gen/vsub-scalar-x2.c",
305 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700306 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700309 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
310 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700312 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
313 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700315 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
316 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700318 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
319 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
321 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
324 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
327 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700330 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
331 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700333 "src/f32-vunary/gen/vabs-scalar-x1.c",
334 "src/f32-vunary/gen/vabs-scalar-x2.c",
335 "src/f32-vunary/gen/vabs-scalar-x4.c",
336 "src/f32-vunary/gen/vneg-scalar-x1.c",
337 "src/f32-vunary/gen/vneg-scalar-x2.c",
338 "src/f32-vunary/gen/vneg-scalar-x4.c",
339 "src/f32-vunary/gen/vsqr-scalar-x1.c",
340 "src/f32-vunary/gen/vsqr-scalar-x2.c",
341 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800342 "src/math/expminus-scalar-lut2048-p1.c",
343 "src/math/expminus-scalar-lut64-p2.c",
344 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700345 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700346 "src/math/roundne-scalar-nearbyint.c",
347 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700348 "src/math/roundd-scalar-addsub.c",
349 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700350 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700351 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700352 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700353 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700354 "src/math/roundz-scalar-addsub.c",
355 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700356 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800357 "src/math/sigmoid-scalar-lut2048-p1-div.c",
358 "src/math/sigmoid-scalar-lut64-p2-div.c",
359 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700360 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
361 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
362 "src/qu8-dwconv/up1x9-minmax-scalar.c",
363 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
364 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
365 "src/qu8-gemm/2x2-minmax-scalar.c",
366 "src/qu8-igemm/2x2-minmax-scalar.c",
367 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700368 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700369 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700370 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700371 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700372 "src/x32-fill/scalar-float.c",
373 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700374 "src/x32-packx/x2-scalar.c",
375 "src/x32-packx/x3-scalar.c",
376 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700377 "src/x32-pad/scalar-int.c",
378 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700379 "src/x32-unpool/scalar.c",
380 "src/x32-zip/x2-scalar.c",
381 "src/x32-zip/x3-scalar.c",
382 "src/x32-zip/x4-scalar.c",
383 "src/x32-zip/xm-scalar.c",
384 "src/x8-lut/scalar.c",
385 "src/x8-zip/x2-scalar.c",
386 "src/x8-zip/x3-scalar.c",
387 "src/x8-zip/x4-scalar.c",
388 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700389 "src/requantization/precise-scalar.c",
390 "src/requantization/fp32-scalar.c",
391 "src/requantization/q31-scalar.c",
392 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700393]
394
Marat Dukhan436ebe62019-12-04 15:10:12 -0800395WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700396 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
397 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700398 "src/f32-clamp/gen/wasm-x1.c",
399 "src/f32-clamp/gen/wasm-x2.c",
400 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700401 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
402 "src/f32-dwconv/gen/up1x4-wasm.c",
403 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
404 "src/f32-dwconv/gen/up1x9-wasm.c",
405 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
406 "src/f32-dwconv/gen/up1x25-wasm.c",
407 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
408 "src/f32-dwconv/gen/up2x4-wasm.c",
409 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
410 "src/f32-dwconv/gen/up2x9-wasm.c",
411 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
412 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700413 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
414 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
415 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
416 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700417 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
418 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700419 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
420 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
421 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
422 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700423 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
424 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700425 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
426 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700427 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
428 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
429 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700430 "src/f32-gemm/gen/1x4-wasm.c",
431 "src/f32-gemm/gen/2x4-wasm.c",
432 "src/f32-gemm/gen/4x2-wasm.c",
433 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700434 "src/f32-gemm/gen/1x4-relu-wasm.c",
435 "src/f32-gemm/gen/2x4-relu-wasm.c",
436 "src/f32-gemm/gen/4x2-relu-wasm.c",
437 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700438 "src/f32-gemm/gen/1x4-minmax-wasm.c",
439 "src/f32-gemm/gen/2x4-minmax-wasm.c",
440 "src/f32-gemm/gen/4x2-minmax-wasm.c",
441 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700442 "src/f32-hswish/gen/hswish-wasm-x1.c",
443 "src/f32-hswish/gen/hswish-wasm-x2.c",
444 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700445 "src/f32-igemm/gen/1x4-wasm.c",
446 "src/f32-igemm/gen/2x4-wasm.c",
447 "src/f32-igemm/gen/4x2-wasm.c",
448 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700449 "src/f32-igemm/gen/1x4-relu-wasm.c",
450 "src/f32-igemm/gen/2x4-relu-wasm.c",
451 "src/f32-igemm/gen/4x2-relu-wasm.c",
452 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700453 "src/f32-igemm/gen/1x4-minmax-wasm.c",
454 "src/f32-igemm/gen/2x4-minmax-wasm.c",
455 "src/f32-igemm/gen/4x2-minmax-wasm.c",
456 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700457 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
458 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
459 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700460 "src/f32-prelu/gen/wasm-2x1.c",
461 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700462 "src/f32-relu/gen/wasm-x1.c",
463 "src/f32-relu/gen/wasm-x2.c",
464 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700465 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
466 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
467 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
468 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
469 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
470 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
471 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
472 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
473 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
474 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800477 "src/f32-vbinary/gen/vmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
481 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
482 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
483 "src/f32-vbinary/gen/vmin-wasm-x1.c",
484 "src/f32-vbinary/gen/vmin-wasm-x2.c",
485 "src/f32-vbinary/gen/vmin-wasm-x4.c",
486 "src/f32-vbinary/gen/vminc-wasm-x1.c",
487 "src/f32-vbinary/gen/vminc-wasm-x2.c",
488 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700489 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
490 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
491 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
492 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
493 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
494 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
495 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
496 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
497 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
498 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700507 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
508 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
509 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700510 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
511 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
512 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800513]
514
Marat Dukhan290055c2020-06-09 12:24:29 -0700515WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700516 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
517 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
518 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700519 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
520 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
521 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
522 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700523 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
524 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
525 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
526 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700527 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
528 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
529 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
530 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
531 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
532 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700533 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
534 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
535 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
536 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
537 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
538 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
539 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
540 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
541 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
542 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
543 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
544 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
545 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
546 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
547 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
548 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
549 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
550 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
551 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
552 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
553 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
554 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
555 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
556 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700557 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
558 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
559 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700560 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700561 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
562 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
563 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
564 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700565 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
566 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
567 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
568 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
569 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
570 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
571 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
572 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
573 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
574 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
581 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
582 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
583 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
584 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
585 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
586 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
587 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
588 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
589 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700595 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
596 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700597 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
598 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
599 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
600 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
601 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
602 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
603 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
604 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
605 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
606 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
613 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
614 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
615 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
616 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
617 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
618 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
619 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
620 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
621 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700627 "src/f32-ibilinear/gen/wasmsimd-c4.c",
628 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700629 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
630 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
631 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700632 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700633 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
634 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
635 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
636 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700637 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
638 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
639 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
640 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
641 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
642 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
643 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
644 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
645 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
646 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
653 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
654 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
655 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
656 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
657 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
658 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
659 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
660 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
661 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700667 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
668 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700669 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
670 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
671 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
672 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700673 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
674 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
675 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700676 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
677 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700678 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700682 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
683 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
684 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
685 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700686 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
687 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
688 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
689 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
690 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
691 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
692 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
693 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
694 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
695 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
696 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
697 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700698 "src/f32-relu/gen/wasmsimd-x4.c",
699 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700700 "src/f32-rmax/wasmsimd-arm.c",
701 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700702 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
703 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
704 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
705 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
706 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
707 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700708 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
709 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
710 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
711 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
712 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
713 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700714 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
715 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
716 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
717 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700718 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
719 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700720 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
721 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
722 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
723 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700724 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
725 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700726 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
727 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
728 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
729 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700730 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
731 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700732 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
733 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
734 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
735 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700736 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
737 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700738 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
739 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
740 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
741 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700742 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
743 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
744 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
745 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700746 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
747 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
748 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
749 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700750 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
751 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
752 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
753 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700754 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
755 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
756 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
757 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700758 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
759 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700760 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
761 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
762 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
763 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700764 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
765 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700766 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
767 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
768 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
769 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700770 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
771 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700772 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
773 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
774 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
775 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700776 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
777 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700778 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
779 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700780 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
781 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700782 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
783 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
784 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
785 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700786 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
787 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700788 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
789 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
790 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
791 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700792 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
793 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700794 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
795 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
796 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
797 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700798 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
799 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
800 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
801 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700802 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
803 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
804 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
805 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
806 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
807 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
808 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
809 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
810 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
811 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
812 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
813 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
814 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
815 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700816 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
817 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700818 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
819 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
820 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
821 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
822 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
823 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700824 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700825 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700826 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700827 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700828 "src/x32-zip/x2-wasmsimd.c",
829 "src/x32-zip/x3-wasmsimd.c",
830 "src/x32-zip/x4-wasmsimd.c",
831 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700832 "src/math/roundne-wasmsimd-addsub.c",
833 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700834 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700835 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700836 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700837 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700838 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700839 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
840 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700841 "src/requantization/fp32-wasmsimd.c",
842]
843
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800844PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800845 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800846 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800847 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700848 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
849 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700850 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700851 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700852 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700853 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700854 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700855 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700856 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
857 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700858 "src/f32-ibilinear/gen/psimd-c4.c",
859 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700860 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700861 "src/f32-spmm/gen/4x1-minmax-psimd.c",
862 "src/f32-spmm/gen/8x1-minmax-psimd.c",
863 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700864 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
865 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700866 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700867 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700868 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700869 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700870 "src/x32-unpool/psimd.c",
871 "src/x32-zip/x2-psimd.c",
872 "src/x32-zip/x3-psimd.c",
873 "src/x32-zip/x4-psimd.c",
874 "src/x32-zip/xm-psimd.c",
875]
876
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800877PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan249cd6a2020-07-17 14:05:30 -0700878 "src/requantization/precise-psimd.c",
879 "src/requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800880]
881
Marat Dukhan08c4a432019-10-03 09:29:21 -0700882# ISA-specific micro-kernels
883NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700884 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
885 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700886 "src/f32-clamp/gen/neon-x4.c",
887 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700888 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
889 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700890 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
891 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700892 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
893 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700894 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
895 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700896 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
897 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
898 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
899 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700900 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
901 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
902 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
903 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700904 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
905 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
906 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
907 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700908 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700909 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
910 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700911 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
912 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
913 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
914 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
915 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
916 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
917 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
918 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
919 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
920 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
921 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
922 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
923 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
924 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
925 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
926 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
927 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
928 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
929 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
930 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
931 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
932 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
933 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
934 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
935 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
936 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
937 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
938 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
939 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
940 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
941 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700942 "src/f32-hswish/gen/hswish-neon-x4.c",
943 "src/f32-hswish/gen/hswish-neon-x8.c",
944 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700945 "src/f32-ibilinear/gen/neon-c4.c",
946 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700947 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
948 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
949 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
950 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
951 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
952 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
953 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
954 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
955 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
956 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
957 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
958 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
959 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
960 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
961 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
962 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700963 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
964 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
965 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700966 "src/f32-ppmm/gen/4x8-minmax-neon.c",
967 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800968 "src/f32-prelu/gen/neon-2x4.c",
969 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800970 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
971 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
972 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
973 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
974 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
975 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
976 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
977 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
978 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
979 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
980 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
981 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
982 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
983 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
984 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
985 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
986 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
987 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
988 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
989 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700994 "src/f32-relu/gen/neon-x4.c",
995 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700996 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800997 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800998 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
999 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1000 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1001 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1002 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1003 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1004 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1005 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1006 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1007 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1008 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1009 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1010 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1011 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1012 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1013 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1014 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1015 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001016 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1017 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1018 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1019 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001020 "src/f32-vbinary/gen/vmax-neon-x4.c",
1021 "src/f32-vbinary/gen/vmax-neon-x8.c",
1022 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1023 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1024 "src/f32-vbinary/gen/vmin-neon-x4.c",
1025 "src/f32-vbinary/gen/vmin-neon-x8.c",
1026 "src/f32-vbinary/gen/vminc-neon-x4.c",
1027 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001028 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1029 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1030 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1031 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1032 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1033 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001034 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1035 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1036 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1037 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001038 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1039 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1040 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1041 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001042 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1043 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001044 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1045 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001046 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1047 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1048 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1049 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1050 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1051 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1052 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1053 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001054 "src/f32-vunary/gen/vabs-neon-x4.c",
1055 "src/f32-vunary/gen/vabs-neon-x8.c",
1056 "src/f32-vunary/gen/vneg-neon-x4.c",
1057 "src/f32-vunary/gen/vneg-neon-x8.c",
1058 "src/f32-vunary/gen/vsqr-neon-x4.c",
1059 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001060 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1061 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1062 "src/qu8-dwconv/up8x9-minmax-neon.c",
1063 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1064 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1065 "src/qu8-gemm/4x8-minmax-neon.c",
1066 "src/qu8-gemm/8x8-minmax-neon.c",
1067 "src/qu8-igemm/4x8-minmax-neon.c",
1068 "src/qu8-igemm/8x8-minmax-neon.c",
1069 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001070 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001071 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001072 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001073 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001074 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001075 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001076 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001077 "src/x32-zip/x2-neon.c",
1078 "src/x32-zip/x3-neon.c",
1079 "src/x32-zip/x4-neon.c",
1080 "src/x32-zip/xm-neon.c",
1081 "src/x8-zip/x2-neon.c",
1082 "src/x8-zip/x3-neon.c",
1083 "src/x8-zip/x4-neon.c",
1084 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001085 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001086 "src/math/roundd-neon-addsub.c",
1087 "src/math/roundd-neon-cvt.c",
1088 "src/math/roundu-neon-addsub.c",
1089 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001090 "src/math/roundz-neon-addsub.c",
1091 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001092 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001093 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1094 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1095 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1096 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1097 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1098 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001099 "src/math/sqrt-neon-nr1rsqrts.c",
1100 "src/math/sqrt-neon-nr2rsqrts.c",
1101 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001102 "src/requantization/precise-neon.c",
1103 "src/requantization/fp32-neon.c",
1104 "src/requantization/q31-neon.c",
1105 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001106]
1107
1108NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001109 "src/f32-ibilinear/gen/neonfma-c4.c",
1110 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001111 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1112 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1113 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1114 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1115 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1116 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1117 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1118 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1119 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001120 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1121 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1122 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1123 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001124 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1125 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1126 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1127 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001128 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1129 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1130 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1131 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001132 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1133 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1134 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1135 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1136 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1137 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1138 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1139 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1140 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1141 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1142 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1143 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1144 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1145 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1146 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1147 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1148 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1149 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001150 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1151 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001152 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1153 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1154 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1155 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1156 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1157 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1158 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1159 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1160 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1161 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1162 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1163 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1164 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1165 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1166 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1167 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1168 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1169 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1170 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1171 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1172 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1173 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001176 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1177 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1178 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1179 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1180 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1181 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1182 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1183 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1184 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1185 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1186 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1187 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001230 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1231 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001232 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1233 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1234 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1235 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1236 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1237 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1238 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1239 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1240 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1241 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1242 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1243 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1244 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1245 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1246 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1247 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1248 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1249 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1250 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1251 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001252 "src/math/exp-neonfma-lut64-p2.c",
1253 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001254 "src/math/expminus-neonfma-lut2048-p1.c",
1255 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001256 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001257 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1258 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1259 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1260 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1261 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1262 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1263 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1264 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1265 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1266 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1267 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1268 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1269 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1270 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1271 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1272 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1273 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1274 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001275 "src/math/sqrt-neonfma-nr1fma.c",
1276 "src/math/sqrt-neonfma-nr2fma.c",
1277 "src/math/sqrt-neonfma-nr3fma.c",
1278 "src/math/sqrt-neonfma-nr2fma1adj.c",
1279 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001280]
1281
1282AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001283 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1284 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1285 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1286 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1287 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1288 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001289 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1290 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1291 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1292 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1293 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1294 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1295 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1296 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1297 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1298 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1299 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1300 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1301 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1302 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1303 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1304 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1305 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1306 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1307 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1308 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001309 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1310 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001311 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1312 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001313 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1314 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001315 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1316 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001317 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1318 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1319 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1320 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1321 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001322 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1323 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1325 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1326 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1327 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001340 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1341 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1342 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1343 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1344 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1345 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1346 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1347 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1348 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1349 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1350 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1351 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1352 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1353 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1354 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1355 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1356 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1357 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001358 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1359 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001360 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1361 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1362 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1363 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1364 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1365 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001366]
1367
Marat Dukhan8853b822020-05-07 12:19:01 -07001368NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001369 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1370 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1371 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1372 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1373 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1374 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1375 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1376 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001377 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001378 "src/math/roundd-neonv8.c",
1379 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001380 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001381]
1382
Marat Dukhan08c4a432019-10-03 09:29:21 -07001383AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001384 "src/f16-clamp/gen/neonfp16arith-x16.c",
1385 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001386 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1387 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1388 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1389 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1390 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1391 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1392 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1393 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1394 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1395 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1396 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1397 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001398 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1399 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001400 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1401 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001402 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1403 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001404 "src/f16-relu/gen/neonfp16arith-x16.c",
1405 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001406 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1407 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1408 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1409 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1410 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1411 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1412 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1413 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1414 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1415 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1416 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1417 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1418 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1419 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1420 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1421 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1422 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1423 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1424 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1425 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1426 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1427 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1428 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1429 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1430 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1431 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1432 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1433 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001434 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1435 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001436 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1437 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1438 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1439 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1440 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1441 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1442 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1443 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001444 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1445 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1446 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1447 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001448 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1449 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1450 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1451 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1452 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1453 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1454 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1455 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1456 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1457 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1458 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1459 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001460 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1461 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1462 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1463 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1464 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1465 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1466 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1467 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001468]
1469
1470SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001471 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1472 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001473 "src/f32-clamp/gen/sse-x4.c",
1474 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001475 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1476 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001477 "src/f32-dwconv-chw/3x3p1-sse.c",
1478 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001479 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1480 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1481 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1482 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1483 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1484 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1485 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1486 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1487 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1488 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1489 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1490 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001491 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001492 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1493 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001494 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1495 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1496 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1497 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1498 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1499 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1500 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1501 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1502 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1503 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1504 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1505 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1506 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001507 "src/f32-hswish/gen/hswish-sse-x4.c",
1508 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001509 "src/f32-ibilinear/gen/sse-c4.c",
1510 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001511 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1512 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1513 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1514 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1515 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1516 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1517 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001518 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1519 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1520 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001521 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001522 "src/f32-prelu/gen/sse-2x4.c",
1523 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001524 "src/f32-relu/gen/sse-x4.c",
1525 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001526 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001527 "src/f32-spmm/gen/4x1-minmax-sse.c",
1528 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001529 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001530 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1531 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1532 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1533 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1534 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1535 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1536 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1537 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001538 "src/f32-vbinary/gen/vmax-sse-x4.c",
1539 "src/f32-vbinary/gen/vmax-sse-x8.c",
1540 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1541 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1542 "src/f32-vbinary/gen/vmin-sse-x4.c",
1543 "src/f32-vbinary/gen/vmin-sse-x8.c",
1544 "src/f32-vbinary/gen/vminc-sse-x4.c",
1545 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001546 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1547 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1548 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1549 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1550 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1551 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1552 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1553 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001554 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1555 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1556 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1557 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001558 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1559 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1560 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1561 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001562 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1563 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001564 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1565 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001566 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1567 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001568 "src/f32-vunary/gen/vabs-sse-x4.c",
1569 "src/f32-vunary/gen/vabs-sse-x8.c",
1570 "src/f32-vunary/gen/vneg-sse-x4.c",
1571 "src/f32-vunary/gen/vneg-sse-x8.c",
1572 "src/f32-vunary/gen/vsqr-sse-x4.c",
1573 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001574 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001575 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001576 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001577 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001578 "src/math/roundd-sse-addsub.c",
1579 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001580 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001581 "src/math/sqrt-sse-hh1mac.c",
1582 "src/math/sqrt-sse-nr1mac.c",
1583 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001584]
1585
1586SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001587 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1588 "src/f32-argmaxpool/4x-sse2-c4.c",
1589 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001590 "src/f32-prelu/gen/sse2-2x4.c",
1591 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001592 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1593 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1594 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1595 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1596 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1597 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1598 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1599 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1600 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1601 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1602 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1603 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001604 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001605 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001606 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001607 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001608 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1609 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001610 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1611 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001612 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1613 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1614 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1615 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1616 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1617 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1618 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1619 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001620 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1621 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1622 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1623 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1624 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1625 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1626 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1627 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1628 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001629 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001630 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001631 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001632 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001633 "src/x32-zip/x2-sse2.c",
1634 "src/x32-zip/x3-sse2.c",
1635 "src/x32-zip/x4-sse2.c",
1636 "src/x32-zip/xm-sse2.c",
1637 "src/x8-zip/x2-sse2.c",
1638 "src/x8-zip/x3-sse2.c",
1639 "src/x8-zip/x4-sse2.c",
1640 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001641 "src/math/exp-sse2-p5.c",
1642 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001643 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001644 "src/math/roundd-sse2-cvt.c",
1645 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001646 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001647 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001648 "src/requantization/precise-sse2.c",
1649 "src/requantization/fp32-sse2.c",
1650 "src/requantization/q31-sse2.c",
1651 "src/requantization/gemmlowp-sse2.c",
1652]
1653
1654SSSE3_UKERNELS = [
1655 "src/requantization/precise-ssse3.c",
1656 "src/requantization/q31-ssse3.c",
1657 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001658]
1659
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001660SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001661 "src/f32-prelu/gen/sse41-2x4.c",
1662 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001663 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001664 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001665 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001666 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001667 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1668 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001669 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1670 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001671 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1672 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1673 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1674 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1675 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1676 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1677 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1678 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001679 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001680 "src/math/roundd-sse41.c",
1681 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001682 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001683 "src/requantization/precise-sse4.c",
1684 "src/requantization/q31-sse4.c",
1685 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001686]
1687
Marat Dukhan08c4a432019-10-03 09:29:21 -07001688AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001689 "src/f32-clamp/gen/avx-x8.c",
1690 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001691 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1692 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1693 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1694 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1695 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1696 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1697 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1698 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1699 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1700 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1701 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1702 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1703 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1704 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1705 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1706 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1707 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1708 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1709 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1710 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1711 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1712 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1713 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1714 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1715 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1716 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1717 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1718 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1719 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1720 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001721 "src/f32-hswish/gen/hswish-avx-x8.c",
1722 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001723 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1724 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1725 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1726 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1727 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1728 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1729 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1730 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1731 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001732 "src/f32-prelu/gen/avx-2x8.c",
1733 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001734 "src/f32-relu/gen/avx-x16.c",
1735 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001736 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001737 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1738 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1739 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1740 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1741 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1742 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1743 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1744 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001745 "src/f32-vbinary/gen/vmax-avx-x8.c",
1746 "src/f32-vbinary/gen/vmax-avx-x16.c",
1747 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1748 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1749 "src/f32-vbinary/gen/vmin-avx-x8.c",
1750 "src/f32-vbinary/gen/vmin-avx-x16.c",
1751 "src/f32-vbinary/gen/vminc-avx-x8.c",
1752 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001753 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1754 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1755 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1756 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1757 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1758 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1759 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1760 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001761 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1762 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1763 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1764 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001765 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1766 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1767 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1768 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001769 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1770 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001771 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001772 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1773 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1774 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1775 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1776 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1777 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1778 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1779 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001780 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1781 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001782 "src/f32-vunary/gen/vabs-avx-x8.c",
1783 "src/f32-vunary/gen/vabs-avx-x16.c",
1784 "src/f32-vunary/gen/vneg-avx-x8.c",
1785 "src/f32-vunary/gen/vneg-avx-x16.c",
1786 "src/f32-vunary/gen/vsqr-avx-x8.c",
1787 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001788]
1789
Marat Dukhanfda12b82019-11-21 12:27:59 -08001790FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001791 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1792 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1793 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1794 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1795 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1796 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1797 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1798 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1799 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1800 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1801 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1802 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1803 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1804 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1805 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1806 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1807 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1808 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1809 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1810 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1811 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1812 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1813 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1814 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1815 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1816 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1817 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1818 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1819 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1820 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1821 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1822 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1823 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1824 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1825 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1826 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1827 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1828 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1829 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1830 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001831 "src/f32-hswish/gen/hswish-fma3-x8.c",
1832 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001833 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1834 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1835 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1836 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1837 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1838 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1839 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1840 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1841 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1842 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1843 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1844 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1845 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1846 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001847 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1848 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1849 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1850 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1851 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1852 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1853 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1854 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001855 "src/math/sqrt-fma3-nr1fma.c",
1856 "src/math/sqrt-fma3-nr2fma.c",
1857 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001858]
1859
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001860AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001861 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1862 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1863 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1864 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1865 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1866 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1867 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1868 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1869 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1870 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1871 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1872 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1873 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1874 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1875 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1876 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1877 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1878 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1879 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1880 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1881 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1882 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1883 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1884 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1885 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1886 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1887 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1888 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1889 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1890 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1891 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1892 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1893 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1894 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1895 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1896 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001897 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1898 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1899 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1900 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1901 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1902 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1903 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1904 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1905 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1906 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1907 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1908 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1909 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1910 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1911 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1912 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1913 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1914 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1915 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1916 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1917 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1918 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1919 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1920 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1921 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1922 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1923 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1924 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1925 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1926 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001927 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1928 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1929 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1930 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1931 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1932 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1933 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1934 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1935 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1936 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1937 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1938 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1939 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1940 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1941 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1942 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1943 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1944 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1945 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1946 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1947 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1948 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1949 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1950 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001951 "src/math/exp-avx2-p5.c",
1952 "src/math/exp-avx2-perm-p3.c",
1953 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001954 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001955 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001956 "src/math/sigmoid-avx2-rr2-p5-div.c",
1957 "src/math/sigmoid-avx2-rr1-p5-div.c",
1958 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1959 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1960 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1961 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001962]
1963
Marat Dukhan08c4a432019-10-03 09:29:21 -07001964AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001965 "src/f32-clamp/gen/avx512f-x16.c",
1966 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001967 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1968 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1969 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1970 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1971 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1972 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1973 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1974 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1975 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1976 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1977 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1978 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1979 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1980 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1981 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1982 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1983 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1984 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1985 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1986 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1987 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1988 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1989 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1990 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001991 "src/f32-hswish/gen/hswish-avx512f-x16.c",
1992 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001993 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1994 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1995 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1996 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1997 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1998 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001999 "src/f32-prelu/gen/avx512f-2x16.c",
2000 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002001 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2002 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2003 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2004 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2005 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2006 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2007 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2008 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2009 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2010 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2011 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2012 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2013 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2014 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2015 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2016 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2017 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2018 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2019 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2020 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2021 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2022 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2023 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2024 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2025 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2026 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2027 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2028 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2029 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2030 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2031 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2032 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2033 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2034 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2035 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2036 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002037 "src/f32-relu/gen/avx512f-x16.c",
2038 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002039 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002040 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2041 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2042 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2043 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2044 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2045 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2046 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2047 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002048 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2049 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2050 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2051 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2052 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2053 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2054 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2055 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002056 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2057 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2058 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2059 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2060 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2061 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2062 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2063 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002064 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2065 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2066 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2067 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002068 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2069 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2070 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2071 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002072 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2073 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002074 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002075 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2076 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2077 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2078 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2079 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2080 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2081 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2082 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2083 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2084 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2085 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2086 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2087 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2088 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2089 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2090 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2091 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2092 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2093 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2094 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2095 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2096 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2097 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2098 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002099 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2100 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2101 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2102 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2103 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2104 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2105 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2106 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002107 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2108 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2109 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2110 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2111 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2112 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2113 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2114 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002115 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2116 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2117 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2118 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2119 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2120 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002121 "src/math/exp-avx512f-p5-scalef.c",
2122 "src/math/exp-avx512f-p5.c",
2123 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002124 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002125 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002126 "src/math/sqrt-avx512f-nr1fma.c",
2127 "src/math/sqrt-avx512f-nr2fma.c",
2128 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002129]
2130
2131AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002132 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002133 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002134 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2135 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002136 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002137 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002138 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002139 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002140 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2141 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002142 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2143 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2144 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2145 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002146]
2147
2148AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002149 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002150 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2151 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002152 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002153 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2154 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002155 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2156 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2157 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002158 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002159 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2160 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2161 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002162 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002163 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2164 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002165 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002166 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2167 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2168 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2169 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2170 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2171 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2172 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2173 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2174 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2175 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2176 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2177 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2178 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2179 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2180 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2181 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2182 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2183 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002184 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2185 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002186 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002187 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2188 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2189 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2190 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2191 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2192 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2193 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2194 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2195 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2196 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2197 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2198 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2199 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2200 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2201 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2202 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2203 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2204 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002205 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2206 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2207 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2208 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2209 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2210 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2211 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2212 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2213 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2214 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2215 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2216 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2217 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2218 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2219 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2220 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2221 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2222 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002223]
2224
Marat Dukhan1b354632020-03-23 12:50:22 -07002225INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002226 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002227 "src/xnnpack/argmaxpool.h",
2228 "src/xnnpack/avgpool.h",
2229 "src/xnnpack/clamp.h",
2230 "src/xnnpack/common.h",
2231 "src/xnnpack/conv.h",
2232 "src/xnnpack/dwconv.h",
2233 "src/xnnpack/gavgpool.h",
2234 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002235 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002236 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002237 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002239 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002240 "src/xnnpack/lut.h",
2241 "src/xnnpack/math.h",
2242 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002243 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244 "src/xnnpack/packx.h",
2245 "src/xnnpack/pad.h",
2246 "src/xnnpack/params.h",
2247 "src/xnnpack/pavgpool.h",
2248 "src/xnnpack/ppmm.h",
2249 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002250 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002251 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002252 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002253 "src/xnnpack/rmax.h",
2254 "src/xnnpack/scalar-utils.h",
2255 "src/xnnpack/spmm.h",
2256 "src/xnnpack/unpool.h",
2257 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002258 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002259 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002260 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002261 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002262 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002263 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002264 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002265]
2266
2267INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268 "include/xnnpack.h",
2269 "src/xnnpack/allocator.h",
2270 "src/xnnpack/compute.h",
2271 "src/xnnpack/im2col.h",
2272 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002273 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002274 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002275 "src/xnnpack/operator.h",
2276 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002277 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002278 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002279 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002280 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002281]
2282
Marat Dukhan1b354632020-03-23 12:50:22 -07002283ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002284 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002285]
2286
Marat Dukhan1b354632020-03-23 12:50:22 -07002287MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002288 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289 "include/xnnpack.h",
2290]
2291
Marat Dukhan1b354632020-03-23 12:50:22 -07002292MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002293 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002294 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002295 "src/xnnpack/requantization.h",
2296 "include/xnnpack.h",
2297]
2298
2299OPERATOR_TEST_PARAMS_HDRS = [
2300 "src/xnnpack/params.h",
2301 "src/xnnpack/common.h",
2302]
2303
2304WEIGHTS_PACK_HDRS = [
2305 "src/xnnpack/pack.h",
2306 "src/xnnpack/operator.h",
2307 "src/xnnpack/compute.h",
2308]
2309
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002310LOGGING_COPTS = select({
2311 # No logging in optimized mode
2312 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2313 # Full logging in debug mode
2314 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2315 # Error-only logging in default (fastbuild) mode
2316 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2317})
2318
Marat Dukhan3b59de22020-06-03 20:15:19 -07002319LOGGING_SRCS = select({
2320 # No logging in optimized mode
2321 ":optimized_build": [],
2322 "//conditions:default": [
2323 "src/operator-strings.c",
2324 "src/subgraph-strings.c",
2325 ],
2326})
2327
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002328LOGGING_HDRS = [
2329 "src/xnnpack/log.h",
2330]
2331
Marat Dukhan08c4a432019-10-03 09:29:21 -07002332xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002333 name = "tables",
2334 srcs = TABLE_SRCS,
2335 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002336 gcc_copts = xnnpack_gcc_std_copts(),
2337 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002338)
2339
2340xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002341 name = "scalar_ukernels",
2342 srcs = SCALAR_UKERNELS,
2343 hdrs = INTERNAL_HDRS,
2344 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002345 gcc_copts = xnnpack_gcc_std_copts(),
2346 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002347 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002348 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002349 "@FP16",
2350 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002351 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352 ],
2353)
2354
2355xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002356 name = "scalar_ukernels_test_mode",
2357 srcs = SCALAR_UKERNELS,
2358 hdrs = INTERNAL_HDRS,
2359 aarch32_copts = ["-marm"],
2360 copts = [
2361 "-UNDEBUG",
2362 "-DXNN_TEST_MODE=1",
2363 ],
2364 gcc_copts = xnnpack_gcc_std_copts(),
2365 msvc_copts = xnnpack_msvc_std_copts(),
2366 deps = [
2367 ":tables",
2368 "@FP16",
2369 "@FXdiv",
2370 "@pthreadpool",
2371 ],
2372)
2373
2374xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002375 name = "wasm_ukernels",
2376 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002377 gcc_copts = xnnpack_gcc_std_copts(),
2378 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002379 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002380 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002381 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002382 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002383 "@FP16",
2384 "@FXdiv",
2385 "@pthreadpool",
2386 ],
2387)
2388
2389xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002390 name = "wasm_ukernels_test_mode",
2391 hdrs = INTERNAL_HDRS,
2392 copts = [
2393 "-UNDEBUG",
2394 "-DXNN_TEST_MODE=1",
2395 ],
2396 gcc_copts = xnnpack_gcc_std_copts(),
2397 msvc_copts = xnnpack_msvc_std_copts(),
2398 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002399 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002400 deps = [
2401 ":tables",
2402 "@FP16",
2403 "@FXdiv",
2404 "@pthreadpool",
2405 ],
2406)
2407
2408xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002409 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002410 hdrs = INTERNAL_HDRS,
2411 aarch32_copts = [
2412 "-marm",
2413 "-mfpu=neon",
2414 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002415 gcc_copts = xnnpack_gcc_std_copts(),
2416 gcc_x86_copts = ["-msse2"],
2417 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002418 optimized_copts = [
2419 "-O3",
2420 "-ffast-math",
2421 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002422 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002423 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002424 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002425 "@FP16",
2426 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002427 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002428 ],
2429)
2430
2431xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002432 name = "psimd_fastmath_ukernels_test_mode",
2433 hdrs = INTERNAL_HDRS,
2434 aarch32_copts = [
2435 "-marm",
2436 "-mfpu=neon",
2437 ],
2438 copts = [
2439 "-UNDEBUG",
2440 "-DXNN_TEST_MODE=1",
2441 ],
2442 gcc_copts = xnnpack_gcc_std_copts(),
2443 gcc_x86_copts = ["-msse2"],
2444 msvc_copts = xnnpack_msvc_std_copts(),
2445 optimized_copts = [
2446 "-O3",
2447 "-ffast-math",
2448 ],
2449 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2450 deps = [
2451 ":tables",
2452 "@FP16",
2453 "@psimd",
2454 "@pthreadpool",
2455 ],
2456)
2457
2458xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002459 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002460 hdrs = INTERNAL_HDRS,
2461 aarch32_copts = [
2462 "-marm",
2463 "-mfpu=neon",
2464 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002465 gcc_copts = xnnpack_gcc_std_copts(),
2466 gcc_x86_copts = ["-msse2"],
2467 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002468 optimized_copts = [
2469 "-O3",
2470 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002471 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002472 deps = [
2473 ":tables",
2474 "@FP16",
2475 "@psimd",
2476 "@pthreadpool",
2477 ],
2478)
2479
2480xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002481 name = "psimd_accmath_ukernels_test_mode",
2482 hdrs = INTERNAL_HDRS,
2483 aarch32_copts = [
2484 "-marm",
2485 "-mfpu=neon",
2486 ],
2487 copts = [
2488 "-UNDEBUG",
2489 "-DXNN_TEST_MODE=1",
2490 ],
2491 gcc_copts = xnnpack_gcc_std_copts(),
2492 gcc_x86_copts = ["-msse2"],
2493 msvc_copts = xnnpack_msvc_std_copts(),
2494 optimized_copts = [
2495 "-O3",
2496 ],
2497 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2498 deps = [
2499 ":tables",
2500 "@FP16",
2501 "@psimd",
2502 "@pthreadpool",
2503 ],
2504)
2505
2506xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002507 name = "neon_ukernels",
2508 hdrs = INTERNAL_HDRS,
2509 aarch32_copts = [
2510 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002511 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002512 "-mfpu=neon",
2513 ],
2514 aarch32_srcs = NEON_UKERNELS,
2515 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002516 gcc_copts = xnnpack_gcc_std_copts(),
2517 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002518 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002519 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002520 "@FP16",
2521 "@pthreadpool",
2522 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002523)
2524
2525xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002526 name = "neon_ukernels_test_mode",
2527 hdrs = INTERNAL_HDRS,
2528 aarch32_copts = [
2529 "-marm",
2530 "-march=armv7-a",
2531 "-mfpu=neon",
2532 ],
2533 aarch32_srcs = NEON_UKERNELS,
2534 aarch64_srcs = NEON_UKERNELS,
2535 copts = [
2536 "-UNDEBUG",
2537 "-DXNN_TEST_MODE=1",
2538 ],
2539 gcc_copts = xnnpack_gcc_std_copts(),
2540 msvc_copts = xnnpack_msvc_std_copts(),
2541 deps = [
2542 ":tables",
2543 "@FP16",
2544 "@pthreadpool",
2545 ],
2546)
2547
2548xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002549 name = "neonfma_ukernels",
2550 hdrs = INTERNAL_HDRS,
2551 aarch32_copts = [
2552 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002553 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002554 "-mfpu=neon-vfpv4",
2555 ],
2556 aarch32_srcs = NEONFMA_UKERNELS,
2557 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002558 apple_aarch32_copts = [
2559 "-mcpu=swift",
2560 "-mtune=generic",
2561 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002562 gcc_copts = xnnpack_gcc_std_copts(),
2563 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002564 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002565 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002566 "@FP16",
2567 "@pthreadpool",
2568 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002569)
2570
2571xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002572 name = "neonfma_ukernels_test_mode",
2573 hdrs = INTERNAL_HDRS,
2574 aarch32_copts = [
2575 "-marm",
2576 "-march=armv7-a",
2577 "-mfpu=neon-vfpv4",
2578 ],
2579 aarch32_srcs = NEONFMA_UKERNELS,
2580 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002581 apple_aarch32_copts = [
2582 "-mcpu=swift",
2583 "-mtune=generic",
2584 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002585 copts = [
2586 "-UNDEBUG",
2587 "-DXNN_TEST_MODE=1",
2588 ],
2589 gcc_copts = xnnpack_gcc_std_copts(),
2590 msvc_copts = xnnpack_msvc_std_copts(),
2591 deps = [
2592 ":tables",
2593 "@FP16",
2594 "@pthreadpool",
2595 ],
2596)
2597
2598xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002599 name = "neonv8_ukernels",
2600 hdrs = INTERNAL_HDRS,
2601 aarch32_copts = [
2602 "-marm",
2603 "-march=armv8-a",
2604 "-mfpu=neon-fp-armv8",
2605 ],
2606 aarch32_srcs = NEONV8_UKERNELS,
2607 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002608 apple_aarch32_copts = [
2609 "-mcpu=cyclone",
2610 "-mtune=generic",
2611 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002612 gcc_copts = xnnpack_gcc_std_copts(),
2613 msvc_copts = xnnpack_msvc_std_copts(),
2614 deps = [
2615 ":tables",
2616 "@FP16",
2617 "@pthreadpool",
2618 ],
2619)
2620
2621xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002622 name = "neonv8_ukernels_test_mode",
2623 hdrs = INTERNAL_HDRS,
2624 aarch32_copts = [
2625 "-marm",
2626 "-march=armv8-a",
2627 "-mfpu=neon-fp-armv8",
2628 ],
2629 aarch32_srcs = NEONV8_UKERNELS,
2630 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002631 apple_aarch32_copts = [
2632 "-mcpu=cyclone",
2633 "-mtune=generic",
2634 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002635 copts = [
2636 "-UNDEBUG",
2637 "-DXNN_TEST_MODE=1",
2638 ],
2639 gcc_copts = xnnpack_gcc_std_copts(),
2640 msvc_copts = xnnpack_msvc_std_copts(),
2641 deps = [
2642 ":tables",
2643 "@FP16",
2644 "@pthreadpool",
2645 ],
2646)
2647
2648xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002649 name = "neonfp16arith_ukernels",
2650 hdrs = INTERNAL_HDRS,
2651 aarch64_copts = ["-march=armv8.2-a+fp16"],
2652 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002653 gcc_copts = xnnpack_gcc_std_copts(),
2654 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002655 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002656 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002657 "@FP16",
2658 "@pthreadpool",
2659 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002660)
2661
2662xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002663 name = "neonfp16arith_ukernels_test_mode",
2664 hdrs = INTERNAL_HDRS,
2665 aarch64_copts = ["-march=armv8.2-a+fp16"],
2666 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2667 copts = [
2668 "-UNDEBUG",
2669 "-DXNN_TEST_MODE=1",
2670 ],
2671 gcc_copts = xnnpack_gcc_std_copts(),
2672 msvc_copts = xnnpack_msvc_std_copts(),
2673 deps = [
2674 ":tables",
2675 "@FP16",
2676 "@pthreadpool",
2677 ],
2678)
2679
2680xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002681 name = "sse2_ukernels",
2682 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002683 gcc_copts = xnnpack_gcc_std_copts(),
2684 gcc_x86_copts = ["-msse2"],
2685 msvc_copts = xnnpack_msvc_std_copts(),
2686 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002687 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002688 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002689 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002690 "@FP16",
2691 "@pthreadpool",
2692 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002693)
2694
2695xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002696 name = "sse2_ukernels_test_mode",
2697 hdrs = INTERNAL_HDRS,
2698 copts = [
2699 "-UNDEBUG",
2700 "-DXNN_TEST_MODE=1",
2701 ],
2702 gcc_copts = xnnpack_gcc_std_copts(),
2703 gcc_x86_copts = ["-msse2"],
2704 msvc_copts = xnnpack_msvc_std_copts(),
2705 msvc_x86_32_copts = ["/arch:SSE2"],
2706 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2707 deps = [
2708 ":tables",
2709 "@FP16",
2710 "@pthreadpool",
2711 ],
2712)
2713
2714xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002715 name = "ssse3_ukernels",
2716 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002717 gcc_copts = xnnpack_gcc_std_copts(),
2718 gcc_x86_copts = ["-mssse3"],
2719 msvc_copts = xnnpack_msvc_std_copts(),
2720 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002721 x86_srcs = SSSE3_UKERNELS,
2722 deps = [
2723 ":tables",
2724 "@FP16",
2725 "@pthreadpool",
2726 ],
2727)
2728
2729xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002730 name = "ssse3_ukernels_test_mode",
2731 hdrs = INTERNAL_HDRS,
2732 copts = [
2733 "-UNDEBUG",
2734 "-DXNN_TEST_MODE=1",
2735 ],
2736 gcc_copts = xnnpack_gcc_std_copts(),
2737 gcc_x86_copts = ["-mssse3"],
2738 msvc_copts = xnnpack_msvc_std_copts(),
2739 msvc_x86_32_copts = ["/arch:SSE2"],
2740 x86_srcs = SSSE3_UKERNELS,
2741 deps = [
2742 ":tables",
2743 "@FP16",
2744 "@pthreadpool",
2745 ],
2746)
2747
2748xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002749 name = "sse41_ukernels",
2750 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002751 gcc_copts = xnnpack_gcc_std_copts(),
2752 gcc_x86_copts = ["-msse4.1"],
2753 msvc_copts = xnnpack_msvc_std_copts(),
2754 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002755 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002756 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002757 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002758 "@FP16",
2759 "@pthreadpool",
2760 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002761)
2762
2763xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002764 name = "sse41_ukernels_test_mode",
2765 hdrs = INTERNAL_HDRS,
2766 copts = [
2767 "-UNDEBUG",
2768 "-DXNN_TEST_MODE=1",
2769 ],
2770 gcc_copts = xnnpack_gcc_std_copts(),
2771 gcc_x86_copts = ["-msse4.1"],
2772 msvc_copts = xnnpack_msvc_std_copts(),
2773 msvc_x86_32_copts = ["/arch:SSE2"],
2774 x86_srcs = SSE41_UKERNELS,
2775 deps = [
2776 ":tables",
2777 "@FP16",
2778 "@pthreadpool",
2779 ],
2780)
2781
2782xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002783 name = "avx_ukernels",
2784 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002785 gcc_copts = xnnpack_gcc_std_copts(),
2786 gcc_x86_copts = ["-mavx"],
2787 msvc_copts = xnnpack_msvc_std_copts(),
2788 msvc_x86_32_copts = ["/arch:AVX"],
2789 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002790 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002791 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002792 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002793 "@FP16",
2794 "@pthreadpool",
2795 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002796)
2797
2798xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002799 name = "avx_ukernels_test_mode",
2800 hdrs = INTERNAL_HDRS,
2801 copts = [
2802 "-UNDEBUG",
2803 "-DXNN_TEST_MODE=1",
2804 ],
2805 gcc_copts = xnnpack_gcc_std_copts(),
2806 gcc_x86_copts = ["-mavx"],
2807 msvc_copts = xnnpack_msvc_std_copts(),
2808 msvc_x86_32_copts = ["/arch:AVX"],
2809 msvc_x86_64_copts = ["/arch:AVX"],
2810 x86_srcs = AVX_UKERNELS,
2811 deps = [
2812 ":tables",
2813 "@FP16",
2814 "@pthreadpool",
2815 ],
2816)
2817
2818xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002819 name = "fma3_ukernels",
2820 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002821 gcc_copts = xnnpack_gcc_std_copts(),
2822 gcc_x86_copts = ["-mfma"],
2823 msvc_copts = xnnpack_msvc_std_copts(),
2824 msvc_x86_32_copts = ["/arch:AVX"],
2825 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002826 x86_srcs = FMA3_UKERNELS,
2827 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002828 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002829 "@FP16",
2830 "@pthreadpool",
2831 ],
2832)
2833
2834xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002835 name = "fma3_ukernels_test_mode",
2836 hdrs = INTERNAL_HDRS,
2837 copts = [
2838 "-UNDEBUG",
2839 "-DXNN_TEST_MODE=1",
2840 ],
2841 gcc_copts = xnnpack_gcc_std_copts(),
2842 gcc_x86_copts = ["-mfma"],
2843 msvc_copts = xnnpack_msvc_std_copts(),
2844 msvc_x86_32_copts = ["/arch:AVX"],
2845 msvc_x86_64_copts = ["/arch:AVX"],
2846 x86_srcs = FMA3_UKERNELS,
2847 deps = [
2848 ":tables",
2849 "@FP16",
2850 "@pthreadpool",
2851 ],
2852)
2853
2854xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002855 name = "avx2_ukernels",
2856 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002857 gcc_copts = xnnpack_gcc_std_copts(),
2858 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002859 "-mfma",
2860 "-mavx2",
2861 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002862 msvc_copts = xnnpack_msvc_std_copts(),
2863 msvc_x86_32_copts = ["/arch:AVX2"],
2864 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002865 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002866 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002867 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002868 "@FP16",
2869 "@pthreadpool",
2870 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002871)
2872
2873xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002874 name = "avx2_ukernels_test_mode",
2875 hdrs = INTERNAL_HDRS,
2876 copts = [
2877 "-UNDEBUG",
2878 "-DXNN_TEST_MODE=1",
2879 ],
2880 gcc_copts = xnnpack_gcc_std_copts(),
2881 gcc_x86_copts = [
2882 "-mfma",
2883 "-mavx2",
2884 ],
2885 msvc_copts = xnnpack_msvc_std_copts(),
2886 msvc_x86_32_copts = ["/arch:AVX2"],
2887 msvc_x86_64_copts = ["/arch:AVX2"],
2888 x86_srcs = AVX2_UKERNELS,
2889 deps = [
2890 ":tables",
2891 "@FP16",
2892 "@pthreadpool",
2893 ],
2894)
2895
2896xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 name = "avx512f_ukernels",
2898 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002899 gcc_copts = xnnpack_gcc_std_copts(),
2900 gcc_x86_copts = ["-mavx512f"],
2901 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2902 msvc_copts = xnnpack_msvc_std_copts(),
2903 msvc_x86_32_copts = ["/arch:AVX512"],
2904 msvc_x86_64_copts = ["/arch:AVX512"],
2905 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002906 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002907 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002908 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002909 "@FP16",
2910 "@pthreadpool",
2911 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002912)
2913
2914xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002915 name = "avx512f_ukernels_test_mode",
2916 hdrs = INTERNAL_HDRS,
2917 copts = [
2918 "-UNDEBUG",
2919 "-DXNN_TEST_MODE=1",
2920 ],
2921 gcc_copts = xnnpack_gcc_std_copts(),
2922 gcc_x86_copts = ["-mavx512f"],
2923 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2924 msvc_copts = xnnpack_msvc_std_copts(),
2925 msvc_x86_32_copts = ["/arch:AVX512"],
2926 msvc_x86_64_copts = ["/arch:AVX512"],
2927 msys_copts = ["-fno-asynchronous-unwind-tables"],
2928 x86_srcs = AVX512F_UKERNELS,
2929 deps = [
2930 ":tables",
2931 "@FP16",
2932 "@pthreadpool",
2933 ],
2934)
2935
2936xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002937 name = "asm_ukernels",
2938 hdrs = ["src/xnnpack/assembly.h"],
2939 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002940 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002941 aarch64_srcs = AARCH64_ASM_UKERNELS,
2942)
2943
Marat Dukhan3b59de22020-06-03 20:15:19 -07002944xnnpack_cc_library(
2945 name = "logging_utils",
2946 srcs = LOGGING_SRCS,
2947 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2948 copts = LOGGING_COPTS + [
2949 "-Isrc",
2950 "-Iinclude",
2951 ] + select({
2952 ":debug_build": [],
2953 "//conditions:default": xnnpack_min_size_copts(),
2954 }),
2955 gcc_copts = xnnpack_gcc_std_copts(),
2956 msvc_copts = xnnpack_msvc_std_copts(),
2957 visibility = xnnpack_visibility(),
2958 deps = [
2959 "@FP16",
2960 "@clog",
2961 "@pthreadpool",
2962 ],
2963)
2964
Marat Dukhan08c4a432019-10-03 09:29:21 -07002965xnnpack_aggregate_library(
2966 name = "ukernels",
2967 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002968 ":neon_ukernels",
2969 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002970 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002971 ":asm_ukernels",
2972 ],
2973 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002974 ":neon_ukernels",
2975 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002976 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977 ":neonfp16arith_ukernels",
2978 ":asm_ukernels",
2979 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002980 generic_deps = [
2981 ":scalar_ukernels",
2982 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002983 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002984 ":psimd_fastmath_ukernels",
2985 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002986 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002987 wasm_deps = [
2988 ":wasm_ukernels",
2989 ],
2990 wasmsimd_deps = [
2991 ":wasm_ukernels",
2992 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002994 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002995 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002996 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002997 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002998 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002999 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000 ":avx512f_ukernels",
3001 ],
3002)
3003
Marat Dukhan33fcf782020-05-24 14:27:15 -07003004xnnpack_aggregate_library(
3005 name = "ukernels_test_mode",
3006 aarch32_deps = [
3007 ":neon_ukernels_test_mode",
3008 ":neonfma_ukernels_test_mode",
3009 ":neonv8_ukernels_test_mode",
3010 ":asm_ukernels",
3011 ],
3012 aarch64_deps = [
3013 ":neon_ukernels_test_mode",
3014 ":neonfma_ukernels_test_mode",
3015 ":neonv8_ukernels_test_mode",
3016 ":neonfp16arith_ukernels_test_mode",
3017 ":asm_ukernels",
3018 ],
3019 generic_deps = [
3020 ":scalar_ukernels_test_mode",
3021 ],
3022 psimd_deps = [
3023 ":psimd_fastmath_ukernels_test_mode",
3024 ":psimd_accmath_ukernels_test_mode",
3025 ],
3026 wasm_deps = [
3027 ":wasm_ukernels_test_mode",
3028 ],
3029 wasmsimd_deps = [
3030 ":wasm_ukernels_test_mode",
3031 ],
3032 x86_deps = [
3033 ":sse2_ukernels_test_mode",
3034 ":ssse3_ukernels_test_mode",
3035 ":sse41_ukernels_test_mode",
3036 ":avx_ukernels_test_mode",
3037 ":fma3_ukernels_test_mode",
3038 ":avx2_ukernels_test_mode",
3039 ":avx512f_ukernels_test_mode",
3040 ],
3041)
3042
Marat Dukhan86df3842020-06-17 14:03:31 -07003043# This target must be compatible with Android NDK r17 and compile with Clang 6.
3044# It is similar to :ukernels target, but excludes microkernels which require
3045# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003046xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003047 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003048 aarch32_deps = [
3049 ":neon_ukernels",
3050 ":neonfma_ukernels",
3051 ":neonv8_ukernels",
3052 ":asm_ukernels",
3053 ],
3054 aarch64_deps = [
3055 ":neon_ukernels",
3056 ":neonfma_ukernels",
3057 ":neonv8_ukernels",
3058 ":asm_ukernels",
3059 ],
3060 generic_deps = [
3061 ":scalar_ukernels",
3062 ],
3063 psimd_deps = [
3064 ":psimd_fastmath_ukernels",
3065 ":psimd_accmath_ukernels",
3066 ],
3067 wasm_deps = [
3068 ":wasm_ukernels",
3069 ],
3070 wasmsimd_deps = [
3071 ":wasm_ukernels",
3072 ],
3073 x86_deps = [
3074 ":sse2_ukernels",
3075 ":ssse3_ukernels",
3076 ":sse41_ukernels",
3077 ":avx_ukernels",
3078 ":fma3_ukernels",
3079 ":avx2_ukernels",
3080 ":avx512f_ukernels",
3081 ],
3082)
3083
Marat Dukhan08c4a432019-10-03 09:29:21 -07003084xnnpack_cc_library(
3085 name = "im2col",
3086 srcs = ["src/im2col.c"],
3087 hdrs = [
3088 "src/xnnpack/common.h",
3089 "src/xnnpack/im2col.h",
3090 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003091 gcc_copts = xnnpack_gcc_std_copts(),
3092 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003093)
3094
3095xnnpack_cc_library(
3096 name = "indirection",
3097 srcs = ["src/indirection.c"],
3098 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003099 gcc_copts = xnnpack_gcc_std_copts(),
3100 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003101 deps = [
3102 "@FP16",
3103 "@FXdiv",
3104 "@pthreadpool",
3105 ],
3106)
3107
3108xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003109 name = "indirection_test_mode",
3110 srcs = ["src/indirection.c"],
3111 hdrs = INTERNAL_HDRS,
3112 copts = [
3113 "-UNDEBUG",
3114 "-DXNN_TEST_MODE=1",
3115 ],
3116 gcc_copts = xnnpack_gcc_std_copts(),
3117 msvc_copts = xnnpack_msvc_std_copts(),
3118 deps = [
3119 "@FP16",
3120 "@FXdiv",
3121 "@pthreadpool",
3122 ],
3123)
3124
3125xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003126 name = "packing",
3127 srcs = ["src/packing.c"],
3128 hdrs = INTERNAL_HDRS,
3129 gcc_copts = xnnpack_gcc_std_copts(),
3130 msvc_copts = xnnpack_msvc_std_copts(),
3131 deps = [
3132 "@FP16",
3133 "@FXdiv",
3134 "@pthreadpool",
3135 ],
3136)
3137
3138xnnpack_cc_library(
3139 name = "packing_test_mode",
3140 srcs = ["src/packing.c"],
3141 hdrs = INTERNAL_HDRS,
3142 copts = [
3143 "-UNDEBUG",
3144 "-DXNN_TEST_MODE=1",
3145 ],
3146 gcc_copts = xnnpack_gcc_std_copts(),
3147 msvc_copts = xnnpack_msvc_std_copts(),
3148 deps = [
3149 "@FP16",
3150 "@FXdiv",
3151 "@pthreadpool",
3152 ],
3153)
3154
3155xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156 name = "operator_run",
3157 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003158 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003159 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003160 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3161 "//conditions:default": [],
3162 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003163 gcc_copts = xnnpack_gcc_std_copts(),
3164 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003166 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 "@FP16",
3168 "@FXdiv",
3169 "@clog",
3170 "@pthreadpool",
3171 ],
3172)
3173
Chao Mei6ddfc602020-05-13 22:29:36 -07003174xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003175 name = "operator_run_test_mode",
3176 srcs = ["src/operator-run.c"],
3177 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3178 copts = LOGGING_COPTS + [
3179 "-UNDEBUG",
3180 "-DXNN_TEST_MODE=1",
3181 ] + select({
3182 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3183 "//conditions:default": [],
3184 }),
3185 gcc_copts = xnnpack_gcc_std_copts(),
3186 msvc_copts = xnnpack_msvc_std_copts(),
3187 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003188 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003189 "@FP16",
3190 "@FXdiv",
3191 "@clog",
3192 "@pthreadpool",
3193 ],
3194)
3195
3196xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003197 name = "memory_planner",
3198 srcs = ["src/memory-planner.c"],
3199 hdrs = INTERNAL_HDRS,
3200 defines = select({
3201 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3202 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3203 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3204 }),
3205 gcc_copts = xnnpack_gcc_std_copts(),
3206 msvc_copts = xnnpack_msvc_std_copts(),
3207 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003208 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003209 "@pthreadpool",
3210 ],
3211)
3212
Marat Dukhan33fcf782020-05-24 14:27:15 -07003213xnnpack_cc_library(
3214 name = "memory_planner_test_mode",
3215 srcs = ["src/memory-planner.c"],
3216 hdrs = INTERNAL_HDRS,
3217 copts = [
3218 "-UNDEBUG",
3219 "-DXNN_TEST_MODE=1",
3220 ],
3221 defines = select({
3222 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3223 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3224 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3225 }),
3226 gcc_copts = xnnpack_gcc_std_copts(),
3227 msvc_copts = xnnpack_msvc_std_copts(),
3228 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003229 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003230 "@pthreadpool",
3231 ],
3232)
3233
Marat Dukhan08c4a432019-10-03 09:29:21 -07003234cc_library(
3235 name = "enable_assembly",
3236 defines = select({
3237 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3238 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003239 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 }),
3241)
3242
Marat Dukhan9de90e02020-06-18 16:04:12 -07003243cc_library(
3244 name = "enable_sparse",
3245 defines = select({
3246 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3247 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003248 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003249 }),
3250)
3251
Marat Dukhancf056b22019-10-07 10:26:29 -07003252xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 name = "operators",
3254 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003255 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003256 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003257 ],
3258 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003259 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003260 "-Isrc",
3261 "-Iinclude",
3262 ] + select({
3263 ":debug_build": [],
3264 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003265 }) + select({
3266 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3267 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003269 gcc_copts = xnnpack_gcc_std_copts(),
3270 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003273 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003274 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003275 "@FP16",
3276 "@FXdiv",
3277 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003279 ],
3280)
3281
Marat Dukhan10a38082020-04-17 03:58:35 -07003282xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003283 name = "operators_test_mode",
3284 srcs = OPERATOR_SRCS + [
3285 "src/memory.c",
3286 "src/operator-delete.c",
3287 ],
3288 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3289 copts = LOGGING_COPTS + [
3290 "-Isrc",
3291 "-Iinclude",
3292 "-UNDEBUG",
3293 "-DXNN_TEST_MODE=1",
3294 ] + select({
3295 ":debug_build": [],
3296 "//conditions:default": xnnpack_min_size_copts(),
3297 }) + select({
3298 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3299 "//conditions:default": [],
3300 }),
3301 gcc_copts = xnnpack_gcc_std_copts(),
3302 msvc_copts = xnnpack_msvc_std_copts(),
3303 deps = [
3304 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003305 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003306 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003307 "@FP16",
3308 "@FXdiv",
3309 "@clog",
3310 "@pthreadpool",
3311 ],
3312)
3313
3314xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003315 name = "XNNPACK",
3316 srcs = [
3317 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003318 "src/runtime.c",
3319 "src/subgraph.c",
3320 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003321 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003322 hdrs = ["include/xnnpack.h"],
3323 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003324 "-Isrc",
3325 "-Iinclude",
3326 ] + select({
3327 ":debug_build": [],
3328 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003329 }) + select({
3330 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3331 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003332 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003333 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003334 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003335 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003336 visibility = xnnpack_visibility(),
3337 deps = [
3338 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003339 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003340 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003341 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003342 ":operator_run",
3343 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003344 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003345 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003346 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003347 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003348 ] + select({
3349 ":emscripten": [],
3350 "//conditions:default": ["@cpuinfo"],
3351 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003352)
3353
Marat Dukhan10a38082020-04-17 03:58:35 -07003354xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003355 name = "XNNPACK_test_mode",
3356 srcs = [
3357 "src/init.c",
3358 "src/runtime.c",
3359 "src/subgraph.c",
3360 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003361 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003362 hdrs = ["include/xnnpack.h"],
3363 copts = LOGGING_COPTS + [
3364 "-Isrc",
3365 "-Iinclude",
3366 "-UNDEBUG",
3367 "-DXNN_TEST_MODE=1",
3368 ] + select({
3369 ":debug_build": [],
3370 "//conditions:default": xnnpack_min_size_copts(),
3371 }) + select({
3372 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3373 "//conditions:default": [],
3374 }),
3375 gcc_copts = xnnpack_gcc_std_copts(),
3376 includes = ["include"],
3377 msvc_copts = xnnpack_msvc_std_copts(),
3378 visibility = xnnpack_visibility(),
3379 deps = [
3380 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003381 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003382 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003383 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003384 ":operator_run_test_mode",
3385 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003386 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003387 "@clog",
3388 "@FP16",
3389 "@pthreadpool",
3390 ] + select({
3391 ":emscripten": [],
3392 "//conditions:default": ["@cpuinfo"],
3393 }),
3394)
3395
3396xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003397 name = "xnnpack_f32",
3398 srcs = [
3399 "src/init.c",
3400 "src/runtime.c",
3401 "src/subgraph.c",
3402 "src/tensor.c",
3403 ] + SUBGRAPH_SRCS,
3404 hdrs = ["include/xnnpack.h"],
3405 copts = LOGGING_COPTS + [
3406 "-Isrc",
3407 "-Iinclude",
3408 ] + select({
3409 ":debug_build": [],
3410 "//conditions:default": xnnpack_min_size_copts(),
3411 }) + select({
3412 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3413 "//conditions:default": [],
3414 }),
3415 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003416 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003417 "XNN_NO_U8_OPERATORS",
3418 "XNN_NO_X8_OPERATORS",
3419 "XNN_NO_F16_OPERATORS",
3420 "XNN_NO_X16_OPERATORS",
3421 ],
3422 gcc_copts = xnnpack_gcc_std_copts(),
3423 includes = ["include"],
3424 msvc_copts = xnnpack_msvc_std_copts(),
3425 visibility = xnnpack_visibility(),
3426 deps = [
3427 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003428 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003429 ":logging_utils",
3430 ":memory_planner",
3431 ":operator_run",
3432 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003433 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003434 "@clog",
3435 "@FP16",
3436 "@pthreadpool",
3437 ] + select({
3438 ":emscripten": [],
3439 "//conditions:default": ["@cpuinfo"],
3440 }),
3441)
3442
3443xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003444 name = "xnnpack_operators_nhwc_f32",
3445 srcs = [
3446 "src/init.c",
3447 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003448 hdrs = ["include/xnnpack.h"],
3449 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003450 "-Isrc",
3451 "-Iinclude",
3452 ] + select({
3453 ":debug_build": [],
3454 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003455 }) + select({
3456 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3457 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003458 }),
3459 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003460 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003461 "XNN_NO_U8_OPERATORS",
3462 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003463 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003464 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003465 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003466 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003467 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003468 visibility = xnnpack_visibility(),
3469 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003470 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003471 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003472 ":operator_run",
3473 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003474 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003475 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003476 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003477 ] + select({
3478 ":emscripten": [],
3479 "//conditions:default": ["@cpuinfo"],
3480 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003481)
3482
Marat Dukhancf056b22019-10-07 10:26:29 -07003483xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484 name = "bench_utils",
3485 srcs = ["bench/utils.cc"],
3486 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003487 deps = [
3488 "@com_google_benchmark//:benchmark",
3489 "@cpuinfo",
3490 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003491)
3492
Frank Barchard7e955972019-10-11 10:34:25 -07003493######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003494
3495xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003496 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003497 srcs = [
3498 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003499 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003500 "src/xnnpack/AlignedAllocator.h",
3501 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003502 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003503 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003504)
3505
3506xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003507 name = "f16_igemm_bench",
3508 srcs = [
3509 "bench/f16-igemm.cc",
3510 "bench/conv.h",
3511 "bench/google/conv.h",
3512 "src/xnnpack/AlignedAllocator.h",
3513 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003514 deps = MICROKERNEL_BENCHMARK_DEPS + [
3515 ":indirection",
3516 ":packing",
3517 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003518)
3519
3520xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003521 name = "f16_gemm_bench",
3522 srcs = [
3523 "bench/f16-gemm.cc",
3524 "bench/gemm.h",
3525 "src/xnnpack/AlignedAllocator.h",
3526 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003527 deps = MICROKERNEL_BENCHMARK_DEPS + [
3528 ":packing",
3529 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530)
3531
3532xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003533 name = "f16_spmm_bench",
3534 srcs = [
3535 "bench/f16-spmm.cc",
3536 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003537 "src/xnnpack/AlignedAllocator.h",
3538 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003539 deps = MICROKERNEL_BENCHMARK_DEPS,
3540)
3541
3542xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003543 name = "f32_igemm_bench",
3544 srcs = [
3545 "bench/f32-igemm.cc",
3546 "bench/conv.h",
3547 "src/xnnpack/AlignedAllocator.h",
3548 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003549 deps = MICROKERNEL_BENCHMARK_DEPS + [
3550 ":indirection",
3551 ":packing",
3552 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553)
3554
3555xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003556 name = "f16_relu_bench",
3557 srcs = [
3558 "bench/f16-relu.cc",
3559 "src/xnnpack/AlignedAllocator.h",
3560 ] + MICROKERNEL_BENCHMARK_HDRS,
3561 deps = MICROKERNEL_BENCHMARK_DEPS,
3562)
3563
3564xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003565 name = "f32_conv_hwc_bench",
3566 srcs = [
3567 "bench/f32-conv-hwc.cc",
3568 "bench/dconv.h",
3569 "src/xnnpack/AlignedAllocator.h",
3570 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003571 deps = MICROKERNEL_BENCHMARK_DEPS + [
3572 ":packing",
3573 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574)
3575
3576xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003577 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003578 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003579 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003580 "bench/dconv.h",
3581 "src/xnnpack/AlignedAllocator.h",
3582 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003583 deps = MICROKERNEL_BENCHMARK_DEPS + [
3584 ":packing",
3585 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003586)
3587
3588xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003589 name = "f16_dwconv_bench",
3590 srcs = [
3591 "bench/f16-dwconv.cc",
3592 "bench/dwconv.h",
3593 "bench/google/dwconv.h",
3594 "src/xnnpack/AlignedAllocator.h",
3595 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003596 deps = MICROKERNEL_BENCHMARK_DEPS + [
3597 ":indirection",
3598 ":packing",
3599 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003600)
3601
3602xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603 name = "f32_dwconv_bench",
3604 srcs = [
3605 "bench/f32-dwconv.cc",
3606 "bench/dwconv.h",
3607 "src/xnnpack/AlignedAllocator.h",
3608 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003609 deps = MICROKERNEL_BENCHMARK_DEPS + [
3610 ":indirection",
3611 ":packing",
3612 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003613)
3614
3615xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003616 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003618 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003619 "bench/dwconv.h",
3620 "src/xnnpack/AlignedAllocator.h",
3621 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003622 deps = MICROKERNEL_BENCHMARK_DEPS + [
3623 ":indirection",
3624 ":packing",
3625 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003626)
3627
3628xnnpack_benchmark(
3629 name = "f32_gemm_bench",
3630 srcs = [
3631 "bench/f32-gemm.cc",
3632 "bench/gemm.h",
3633 "src/xnnpack/AlignedAllocator.h",
3634 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003635 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003636 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637)
3638
3639xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003640 name = "f32_hswish_bench",
3641 srcs = [
3642 "bench/f32-hswish.cc",
3643 "src/xnnpack/AlignedAllocator.h",
3644 ] + MICROKERNEL_BENCHMARK_HDRS,
3645 deps = MICROKERNEL_BENCHMARK_DEPS,
3646)
3647
3648xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003649 name = "f32_raddexpminusmax_bench",
3650 srcs = [
3651 "bench/f32-raddexpminusmax.cc",
3652 "src/xnnpack/AlignedAllocator.h",
3653 ] + MICROKERNEL_BENCHMARK_HDRS,
3654 deps = MICROKERNEL_BENCHMARK_DEPS,
3655)
3656
3657xnnpack_benchmark(
3658 name = "f32_raddextexp_bench",
3659 srcs = [
3660 "bench/f32-raddextexp.cc",
3661 "src/xnnpack/AlignedAllocator.h",
3662 ] + MICROKERNEL_BENCHMARK_HDRS,
3663 deps = MICROKERNEL_BENCHMARK_DEPS,
3664)
3665
3666xnnpack_benchmark(
3667 name = "f32_raddstoreexpminusmax_bench",
3668 srcs = [
3669 "bench/f32-raddstoreexpminusmax.cc",
3670 "src/xnnpack/AlignedAllocator.h",
3671 ] + MICROKERNEL_BENCHMARK_HDRS,
3672 deps = MICROKERNEL_BENCHMARK_DEPS,
3673)
3674
3675xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003676 name = "f32_relu_bench",
3677 srcs = [
3678 "bench/f32-relu.cc",
3679 "src/xnnpack/AlignedAllocator.h",
3680 ] + MICROKERNEL_BENCHMARK_HDRS,
3681 deps = MICROKERNEL_BENCHMARK_DEPS,
3682)
3683
3684xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003685 name = "f32_rmax_bench",
3686 srcs = [
3687 "bench/f32-rmax.cc",
3688 "src/xnnpack/AlignedAllocator.h",
3689 ] + MICROKERNEL_BENCHMARK_HDRS,
3690 deps = MICROKERNEL_BENCHMARK_DEPS,
3691)
3692
3693xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003694 name = "f32_sigmoid_bench",
3695 srcs = [
3696 "bench/f32-sigmoid.cc",
3697 "src/xnnpack/AlignedAllocator.h",
3698 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003699 deps = MICROKERNEL_BENCHMARK_DEPS,
3700)
3701
3702xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703 name = "f32_spmm_bench",
3704 srcs = [
3705 "bench/f32-spmm.cc",
3706 "bench/gemm.h",
3707 "src/xnnpack/AlignedAllocator.h",
3708 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709 deps = MICROKERNEL_BENCHMARK_DEPS,
3710)
3711
3712xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003713 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003714 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003715 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003716 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003717 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003718 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003719)
3720
3721xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003722 name = "f32_vscaleexpminusmax_bench",
3723 srcs = [
3724 "bench/f32-vscaleexpminusmax.cc",
3725 "src/xnnpack/AlignedAllocator.h",
3726 ] + MICROKERNEL_BENCHMARK_HDRS,
3727 deps = MICROKERNEL_BENCHMARK_DEPS,
3728)
3729
3730xnnpack_benchmark(
3731 name = "f32_vscaleextexp_bench",
3732 srcs = [
3733 "bench/f32-vscaleextexp.cc",
3734 "src/xnnpack/AlignedAllocator.h",
3735 ] + MICROKERNEL_BENCHMARK_HDRS,
3736 deps = MICROKERNEL_BENCHMARK_DEPS,
3737)
3738
3739xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003740 name = "f32_vsqrt_bench",
3741 srcs = [
3742 "bench/f32-vsqrt.cc",
3743 "src/xnnpack/AlignedAllocator.h",
3744 ] + MICROKERNEL_BENCHMARK_HDRS,
3745 deps = MICROKERNEL_BENCHMARK_DEPS,
3746)
3747
3748xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003749 name = "f32_im2col_gemm_bench",
3750 srcs = [
3751 "bench/f32-im2col-gemm.cc",
3752 "bench/conv.h",
3753 "src/xnnpack/AlignedAllocator.h",
3754 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003755 deps = MICROKERNEL_BENCHMARK_DEPS + [
3756 ":im2col",
3757 ":packing",
3758 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759)
3760
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003761xnnpack_benchmark(
3762 name = "requantization_bench",
3763 srcs = [
3764 "bench/requantization.cc",
3765 "src/xnnpack/requantization-stubs.h",
3766 "src/xnnpack/AlignedAllocator.h",
3767 ] + MICROKERNEL_BENCHMARK_HDRS,
3768 deps = MICROKERNEL_BENCHMARK_DEPS,
3769)
3770
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003771xnnpack_benchmark(
3772 name = "rounding_bench",
3773 srcs = [
3774 "bench/rounding.cc",
3775 "src/xnnpack/math-stubs.h",
3776 "src/xnnpack/AlignedAllocator.h",
3777 ] + MICROKERNEL_BENCHMARK_HDRS,
3778 deps = MICROKERNEL_BENCHMARK_DEPS,
3779)
3780
Marat Dukhan08c4a432019-10-03 09:29:21 -07003781########################### Benchmarks for operators ###########################
3782
3783xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003784 name = "average_pooling_bench",
3785 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003786 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003787 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003788 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003789)
3790
3791xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003792 name = "bankers_rounding_bench",
3793 srcs = ["bench/bankers-rounding.cc"],
3794 copts = xnnpack_optional_tflite_copts(),
3795 tags = ["nowin32"],
3796 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3797)
3798
3799xnnpack_benchmark(
3800 name = "ceiling_bench",
3801 srcs = ["bench/ceiling.cc"],
3802 copts = xnnpack_optional_tflite_copts(),
3803 tags = ["nowin32"],
3804 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3805)
3806
3807xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003808 name = "channel_shuffle_bench",
3809 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003810 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003811)
3812
3813xnnpack_benchmark(
3814 name = "convolution_bench",
3815 srcs = ["bench/convolution.cc"],
3816 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003817 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003818 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819)
3820
3821xnnpack_benchmark(
3822 name = "deconvolution_bench",
3823 srcs = ["bench/deconvolution.cc"],
3824 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003825 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003826 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003827)
3828
3829xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003830 name = "floor_bench",
3831 srcs = ["bench/floor.cc"],
3832 copts = xnnpack_optional_tflite_copts(),
3833 tags = ["nowin32"],
3834 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3835)
3836
3837xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003838 name = "global_average_pooling_bench",
3839 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003840 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003841)
3842
3843xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003844 name = "hardswish_bench",
3845 srcs = ["bench/hardswish.cc"],
3846 copts = xnnpack_optional_tflite_copts(),
3847 tags = ["nowin32"],
3848 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3849)
3850
3851xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003852 name = "max_pooling_bench",
3853 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003854 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003855)
3856
3857xnnpack_benchmark(
3858 name = "sigmoid_bench",
3859 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003860 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003861 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003862 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003863)
3864
3865xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003866 name = "prelu_bench",
3867 srcs = ["bench/prelu.cc"],
3868 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003869 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003870 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003871)
3872
3873xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003874 name = "softmax_bench",
3875 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003876 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003877 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003878 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003879)
3880
Marat Dukhan87727142020-06-24 15:24:10 -07003881xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003882 name = "square_root_bench",
3883 srcs = ["bench/square-root.cc"],
3884 copts = xnnpack_optional_tflite_copts(),
3885 tags = ["nowin32"],
3886 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3887)
3888
3889xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003890 name = "truncation_bench",
3891 srcs = ["bench/truncation.cc"],
3892 deps = OPERATOR_BENCHMARK_DEPS,
3893)
3894
Marat Dukhanc068bb62019-10-04 13:24:39 -07003895############################# End-to-end benchmarks ############################
3896
3897cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003898 name = "fp32_mobilenet_v1",
3899 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003900 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003901 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003902 linkstatic = True,
3903 deps = [
3904 ":XNNPACK",
3905 "@pthreadpool",
3906 ],
3907)
3908
3909cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003910 name = "fp16_mobilenet_v1",
3911 srcs = ["models/fp16-mobilenet-v1.cc"],
3912 hdrs = ["models/models.h"],
3913 copts = xnnpack_std_cxxopts(),
3914 linkstatic = True,
3915 deps = [
3916 ":XNNPACK",
3917 "@FP16",
3918 "@pthreadpool",
3919 ],
3920)
3921
3922cc_library(
3923 name = "fp32_mobilenet_v2",
3924 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003925 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003926 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003927 linkstatic = True,
3928 deps = [
3929 ":XNNPACK",
3930 "@pthreadpool",
3931 ],
3932)
3933
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003934cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003935 name = "fp16_mobilenet_v2",
3936 srcs = ["models/fp16-mobilenet-v2.cc"],
3937 hdrs = ["models/models.h"],
3938 copts = xnnpack_std_cxxopts(),
3939 linkstatic = True,
3940 deps = [
3941 ":XNNPACK",
3942 "@FP16",
3943 "@pthreadpool",
3944 ],
3945)
3946
3947cc_library(
3948 name = "fp32_mobilenet_v3_large",
3949 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003950 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003951 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003952 linkstatic = True,
3953 deps = [
3954 ":XNNPACK",
3955 "@pthreadpool",
3956 ],
3957)
3958
3959cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003960 name = "fp32_mobilenet_v3_small",
3961 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003962 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003963 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003964 linkstatic = True,
3965 deps = [
3966 ":XNNPACK",
3967 "@pthreadpool",
3968 ],
3969)
3970
Marat Dukhanc068bb62019-10-04 13:24:39 -07003971xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003972 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003973 srcs = [
3974 "bench/f32-dwconv-e2e.cc",
3975 "bench/end2end.h",
3976 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003977 deps = MICROKERNEL_BENCHMARK_DEPS + [
3978 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003979 ":fp32_mobilenet_v1",
3980 ":fp32_mobilenet_v2",
3981 ":fp32_mobilenet_v3_large",
3982 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003983 ],
3984)
3985
3986xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003987 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003988 srcs = [
3989 "bench/f32-gemm-e2e.cc",
3990 "bench/end2end.h",
3991 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003992 deps = MICROKERNEL_BENCHMARK_DEPS + [
3993 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003994 ":fp32_mobilenet_v1",
3995 ":fp32_mobilenet_v2",
3996 ":fp32_mobilenet_v3_large",
3997 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003998 ],
3999)
4000
4001xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004002 name = "end2end_bench",
4003 srcs = ["bench/end2end.cc"],
4004 deps = [
4005 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004006 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004007 ":fp16_mobilenet_v1",
4008 ":fp16_mobilenet_v2",
4009 ":fp32_mobilenet_v1",
4010 ":fp32_mobilenet_v2",
4011 ":fp32_mobilenet_v3_large",
4012 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004013 "@pthreadpool",
4014 ],
4015)
4016
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004017#################### Accuracy evaluation for math functions ####################
4018
4019xnnpack_benchmark(
4020 name = "f32_exp_eval",
4021 srcs = [
4022 "eval/f32-exp.cc",
4023 "src/xnnpack/AlignedAllocator.h",
4024 ] + ACCURACY_EVAL_HDRS,
4025 deps = ACCURACY_EVAL_DEPS,
4026)
4027
Marat Dukhan515c9772019-10-17 18:07:57 -07004028xnnpack_benchmark(
4029 name = "f32_expminus_eval",
4030 srcs = [
4031 "eval/f32-expminus.cc",
4032 "src/xnnpack/AlignedAllocator.h",
4033 ] + ACCURACY_EVAL_HDRS,
4034 deps = ACCURACY_EVAL_DEPS,
4035)
4036
Marat Dukhan98ba4412019-10-23 02:14:28 -07004037xnnpack_benchmark(
4038 name = "f32_extexp_eval",
4039 srcs = [
4040 "eval/f32-extexp.cc",
4041 "src/xnnpack/AlignedAllocator.h",
4042 ] + ACCURACY_EVAL_HDRS,
4043 deps = ACCURACY_EVAL_DEPS,
4044)
4045
Marat Dukhan8853b822020-05-07 12:19:01 -07004046xnnpack_unit_test(
4047 name = "f32_roundne_eval",
4048 srcs = [
4049 "eval/f32-roundne.cc",
4050 "src/xnnpack/AlignedAllocator.h",
4051 "src/xnnpack/math-stubs.h",
4052 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004053 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004054 deps = MICROKERNEL_TEST_DEPS,
4055)
4056
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004057xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004058 name = "f32_roundd_eval",
4059 srcs = [
4060 "eval/f32-roundd.cc",
4061 "src/xnnpack/AlignedAllocator.h",
4062 "src/xnnpack/math-stubs.h",
4063 ] + MICROKERNEL_TEST_HDRS,
4064 automatic = False,
4065 deps = MICROKERNEL_TEST_DEPS,
4066)
4067
4068xnnpack_unit_test(
4069 name = "f32_roundu_eval",
4070 srcs = [
4071 "eval/f32-roundu.cc",
4072 "src/xnnpack/AlignedAllocator.h",
4073 "src/xnnpack/math-stubs.h",
4074 ] + MICROKERNEL_TEST_HDRS,
4075 automatic = False,
4076 deps = MICROKERNEL_TEST_DEPS,
4077)
4078
4079xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004080 name = "f32_roundz_eval",
4081 srcs = [
4082 "eval/f32-roundz.cc",
4083 "src/xnnpack/AlignedAllocator.h",
4084 "src/xnnpack/math-stubs.h",
4085 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004086 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004087 deps = MICROKERNEL_TEST_DEPS,
4088)
4089
Marat Dukhan346a9e52019-11-15 09:06:30 -08004090xnnpack_benchmark(
4091 name = "f32_sigmoid_eval",
4092 srcs = [
4093 "eval/f32-sigmoid.cc",
4094 "src/xnnpack/AlignedAllocator.h",
4095 ] + ACCURACY_EVAL_HDRS,
4096 deps = ACCURACY_EVAL_DEPS,
4097)
4098
Marat Dukhan84000762020-06-29 18:38:43 -07004099xnnpack_benchmark(
4100 name = "f32_sqrt_eval",
4101 srcs = [
4102 "eval/f32-sqrt.cc",
4103 "src/xnnpack/AlignedAllocator.h",
4104 ] + ACCURACY_EVAL_HDRS,
4105 deps = ACCURACY_EVAL_DEPS + [
4106 ":bench_utils",
4107 ],
4108)
4109
Marat Dukhan08c4a432019-10-03 09:29:21 -07004110######################### Unit tests for micro-kernels #########################
4111
4112xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004113 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004114 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004115 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004116 "test/gemm-microkernel-tester.h",
4117 "src/xnnpack/AlignedAllocator.h",
4118 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004119 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004120)
4121
4122xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004123 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004124 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004125 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004126 "test/spmm-microkernel-tester.h",
4127 "src/xnnpack/AlignedAllocator.h",
4128 ] + MICROKERNEL_TEST_HDRS,
4129 deps = MICROKERNEL_TEST_DEPS,
4130)
4131
4132xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004133 name = "f32_argmaxpool_test",
4134 srcs = [
4135 "test/f32-argmaxpool.cc",
4136 "test/argmaxpool-microkernel-tester.h",
4137 "src/xnnpack/AlignedAllocator.h",
4138 ] + MICROKERNEL_TEST_HDRS,
4139 deps = MICROKERNEL_TEST_DEPS,
4140)
4141
4142xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004143 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004144 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004145 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004146 "test/avgpool-microkernel-tester.h",
4147 "src/xnnpack/AlignedAllocator.h",
4148 ] + MICROKERNEL_TEST_HDRS,
4149 deps = MICROKERNEL_TEST_DEPS,
4150)
4151
4152xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004153 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004154 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004155 "test/f32-ibilinear.cc",
4156 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004157 "src/xnnpack/AlignedAllocator.h",
4158 ] + MICROKERNEL_TEST_HDRS,
4159 deps = MICROKERNEL_TEST_DEPS,
4160)
4161
4162xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004163 name = "f16_clamp_test",
4164 srcs = [
4165 "test/f16-clamp.cc",
4166 "test/clamp-microkernel-tester.h",
4167 ] + MICROKERNEL_TEST_HDRS,
4168 deps = MICROKERNEL_TEST_DEPS,
4169)
4170
4171xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004172 name = "f32_clamp_test",
4173 srcs = [
4174 "test/f32-clamp.cc",
4175 "test/clamp-microkernel-tester.h",
4176 ] + MICROKERNEL_TEST_HDRS,
4177 deps = MICROKERNEL_TEST_DEPS,
4178)
4179
4180xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004181 name = "f32_igemm_test",
4182 srcs = [
4183 "test/f32-igemm.cc",
4184 "test/gemm-microkernel-tester.h",
4185 "src/xnnpack/AlignedAllocator.h",
4186 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004187 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004188)
4189
4190xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004191 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004192 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004193 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194 "test/gemm-microkernel-tester.h",
4195 "src/xnnpack/AlignedAllocator.h",
4196 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004197 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198)
4199
4200xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004201 name = "f32_igemm_minmax_test",
4202 srcs = [
4203 "test/f32-igemm-minmax.cc",
4204 "test/gemm-microkernel-tester.h",
4205 "src/xnnpack/AlignedAllocator.h",
4206 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004207 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004208)
4209
4210xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004211 name = "f16_igemm_minmax_test",
4212 srcs = [
4213 "test/f16-igemm-minmax.cc",
4214 "test/gemm-microkernel-tester.h",
4215 "src/xnnpack/AlignedAllocator.h",
4216 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004217 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004218)
4219
4220xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004221 name = "f32_conv_hwc_test",
4222 srcs = [
4223 "test/f32-conv-hwc.cc",
4224 "test/conv-hwc-microkernel-tester.h",
4225 "src/xnnpack/AlignedAllocator.h",
4226 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004227 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228)
4229
4230xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004231 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004232 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004233 "test/f32-conv-hwc2chw.cc",
4234 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004235 "src/xnnpack/AlignedAllocator.h",
4236 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004237 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004238)
4239
4240xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004241 name = "f32_dwconv_test",
4242 srcs = [
4243 "test/f32-dwconv.cc",
4244 "test/dwconv-microkernel-tester.h",
4245 "src/xnnpack/AlignedAllocator.h",
4246 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004247 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004248)
4249
4250xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004251 name = "f16_dwconv_minmax_test",
4252 srcs = [
4253 "test/f16-dwconv-minmax.cc",
4254 "test/dwconv-microkernel-tester.h",
4255 "src/xnnpack/AlignedAllocator.h",
4256 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004257 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004258)
4259
4260xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004261 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004262 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004263 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004264 "test/dwconv-microkernel-tester.h",
4265 "src/xnnpack/AlignedAllocator.h",
4266 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004267 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004268)
4269
4270xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004271 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004272 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004273 "test/f32-dwconv-chw.cc",
4274 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275 "src/xnnpack/AlignedAllocator.h",
4276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004277 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278)
4279
4280xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004281 name = "f16_gavgpool_minmax_test",
4282 srcs = [
4283 "test/f16-gavgpool-minmax.cc",
4284 "test/gavgpool-microkernel-tester.h",
4285 "src/xnnpack/AlignedAllocator.h",
4286 ] + MICROKERNEL_TEST_HDRS,
4287 deps = MICROKERNEL_TEST_DEPS,
4288)
4289
4290xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004291 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004292 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004293 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004294 "test/gavgpool-microkernel-tester.h",
4295 "src/xnnpack/AlignedAllocator.h",
4296 ] + MICROKERNEL_TEST_HDRS,
4297 deps = MICROKERNEL_TEST_DEPS,
4298)
4299
4300xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004301 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004302 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004303 "test/f32-gavgpool-cw.cc",
4304 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305 "src/xnnpack/AlignedAllocator.h",
4306 ] + MICROKERNEL_TEST_HDRS,
4307 deps = MICROKERNEL_TEST_DEPS,
4308)
4309
4310xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004311 name = "f32_gemm_test",
4312 srcs = [
4313 "test/f32-gemm.cc",
4314 "test/gemm-microkernel-tester.h",
4315 "src/xnnpack/AlignedAllocator.h",
4316 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004317 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004318)
4319
4320xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004321 name = "f32_gemm_relu_test",
4322 srcs = [
4323 "test/f32-gemm-relu.cc",
4324 "test/gemm-microkernel-tester.h",
4325 "src/xnnpack/AlignedAllocator.h",
4326 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004327 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004328)
4329
4330xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004331 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004332 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004333 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004334 "test/gemm-microkernel-tester.h",
4335 "src/xnnpack/AlignedAllocator.h",
4336 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004337 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004338)
4339
4340xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004341 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004342 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004343 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004344 "test/gemm-microkernel-tester.h",
4345 "src/xnnpack/AlignedAllocator.h",
4346 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004347 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348)
4349
4350xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004351 name = "f16_hswish_test",
4352 srcs = [
4353 "test/f16-hswish.cc",
4354 "test/hswish-microkernel-tester.h",
4355 ] + MICROKERNEL_TEST_HDRS,
4356 deps = MICROKERNEL_TEST_DEPS,
4357)
4358
4359xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004360 name = "f32_hswish_test",
4361 srcs = [
4362 "test/f32-hswish.cc",
4363 "test/hswish-microkernel-tester.h",
4364 ] + MICROKERNEL_TEST_HDRS,
4365 deps = MICROKERNEL_TEST_DEPS,
4366)
4367
4368xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004369 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004370 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004371 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004372 "test/maxpool-microkernel-tester.h",
4373 ] + MICROKERNEL_TEST_HDRS,
4374 deps = MICROKERNEL_TEST_DEPS,
4375)
4376
4377xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004378 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004379 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004380 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004381 "test/avgpool-microkernel-tester.h",
4382 "src/xnnpack/AlignedAllocator.h",
4383 ] + MICROKERNEL_TEST_HDRS,
4384 deps = MICROKERNEL_TEST_DEPS,
4385)
4386
4387xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004388 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004389 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004390 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391 "test/gemm-microkernel-tester.h",
4392 "src/xnnpack/AlignedAllocator.h",
4393 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004394 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004395)
4396
4397xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004398 name = "f16_prelu_test",
4399 srcs = [
4400 "test/f16-prelu.cc",
4401 "test/prelu-microkernel-tester.h",
4402 "src/xnnpack/AlignedAllocator.h",
4403 ] + MICROKERNEL_TEST_HDRS,
4404 deps = MICROKERNEL_TEST_DEPS,
4405)
4406
4407xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004408 name = "f32_prelu_test",
4409 srcs = [
4410 "test/f32-prelu.cc",
4411 "test/prelu-microkernel-tester.h",
4412 "src/xnnpack/AlignedAllocator.h",
4413 ] + MICROKERNEL_TEST_HDRS,
4414 deps = MICROKERNEL_TEST_DEPS,
4415)
4416
4417xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004418 name = "f32_raddexpminusmax_test",
4419 srcs = [
4420 "test/f32-raddexpminusmax.cc",
4421 "test/raddexpminusmax-microkernel-tester.h",
4422 ] + MICROKERNEL_TEST_HDRS,
4423 deps = MICROKERNEL_TEST_DEPS,
4424)
4425
4426xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004427 name = "f32_raddextexp_test",
4428 srcs = [
4429 "test/f32-raddextexp.cc",
4430 "test/raddextexp-microkernel-tester.h",
4431 ] + MICROKERNEL_TEST_HDRS,
4432 deps = MICROKERNEL_TEST_DEPS,
4433)
4434
4435xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004436 name = "f32_raddstoreexpminusmax_test",
4437 srcs = [
4438 "test/f32-raddstoreexpminusmax.cc",
4439 "test/raddstoreexpminusmax-microkernel-tester.h",
4440 ] + MICROKERNEL_TEST_HDRS,
4441 deps = MICROKERNEL_TEST_DEPS,
4442)
4443
4444xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004445 name = "f32_relu_test",
4446 srcs = [
4447 "test/f32-relu.cc",
4448 "test/vunary-microkernel-tester.h",
4449 ] + MICROKERNEL_TEST_HDRS,
4450 deps = MICROKERNEL_TEST_DEPS,
4451)
4452
4453xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004454 name = "f32_rmax_test",
4455 srcs = [
4456 "test/f32-rmax.cc",
4457 "test/rmax-microkernel-tester.h",
4458 ] + MICROKERNEL_TEST_HDRS,
4459 deps = MICROKERNEL_TEST_DEPS,
4460)
4461
4462xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004463 name = "f32_sigmoid_test",
4464 srcs = [
4465 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004466 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004467 ] + MICROKERNEL_TEST_HDRS,
4468 deps = MICROKERNEL_TEST_DEPS,
4469)
4470
4471xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004472 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004474 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004475 "test/spmm-microkernel-tester.h",
4476 "src/xnnpack/AlignedAllocator.h",
4477 ] + MICROKERNEL_TEST_HDRS,
4478 deps = MICROKERNEL_TEST_DEPS,
4479)
4480
4481xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004482 name = "f16_vadd_minmax_test",
4483 srcs = [
4484 "test/f16-vadd-minmax.cc",
4485 "test/vbinary-microkernel-tester.h",
4486 ] + MICROKERNEL_TEST_HDRS,
4487 deps = MICROKERNEL_TEST_DEPS,
4488)
4489
4490xnnpack_unit_test(
4491 name = "f16_vaddc_minmax_test",
4492 srcs = [
4493 "test/f16-vaddc-minmax.cc",
4494 "test/vbinaryc-microkernel-tester.h",
4495 ] + MICROKERNEL_TEST_HDRS,
4496 deps = MICROKERNEL_TEST_DEPS,
4497)
4498
4499xnnpack_unit_test(
4500 name = "f16_vdiv_minmax_test",
4501 srcs = [
4502 "test/f16-vdiv-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_vdivc_minmax_test",
4510 srcs = [
4511 "test/f16-vdivc-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_vrdivc_minmax_test",
4519 srcs = [
4520 "test/f16-vrdivc-minmax.cc",
4521 "test/vbinaryc-microkernel-tester.h",
4522 ] + MICROKERNEL_TEST_HDRS,
4523 deps = MICROKERNEL_TEST_DEPS,
4524)
4525
4526xnnpack_unit_test(
4527 name = "f16_vmax_test",
4528 srcs = [
4529 "test/f16-vmax.cc",
4530 "test/vbinary-microkernel-tester.h",
4531 ] + MICROKERNEL_TEST_HDRS,
4532 deps = MICROKERNEL_TEST_DEPS,
4533)
4534
4535xnnpack_unit_test(
4536 name = "f16_vmaxc_test",
4537 srcs = [
4538 "test/f16-vmaxc.cc",
4539 "test/vbinaryc-microkernel-tester.h",
4540 ] + MICROKERNEL_TEST_HDRS,
4541 deps = MICROKERNEL_TEST_DEPS,
4542)
4543
4544xnnpack_unit_test(
4545 name = "f16_vmin_test",
4546 srcs = [
4547 "test/f16-vmin.cc",
4548 "test/vbinary-microkernel-tester.h",
4549 ] + MICROKERNEL_TEST_HDRS,
4550 deps = MICROKERNEL_TEST_DEPS,
4551)
4552
4553xnnpack_unit_test(
4554 name = "f16_vminc_test",
4555 srcs = [
4556 "test/f16-vminc.cc",
4557 "test/vbinaryc-microkernel-tester.h",
4558 ] + MICROKERNEL_TEST_HDRS,
4559 deps = MICROKERNEL_TEST_DEPS,
4560)
4561
4562xnnpack_unit_test(
4563 name = "f16_vmul_minmax_test",
4564 srcs = [
4565 "test/f16-vmul-minmax.cc",
4566 "test/vbinary-microkernel-tester.h",
4567 ] + MICROKERNEL_TEST_HDRS,
4568 deps = MICROKERNEL_TEST_DEPS,
4569)
4570
4571xnnpack_unit_test(
4572 name = "f16_vmulc_minmax_test",
4573 srcs = [
4574 "test/f16-vmulc-minmax.cc",
4575 "test/vbinaryc-microkernel-tester.h",
4576 ] + MICROKERNEL_TEST_HDRS,
4577 deps = MICROKERNEL_TEST_DEPS,
4578)
4579
4580xnnpack_unit_test(
4581 name = "f16_vsub_minmax_test",
4582 srcs = [
4583 "test/f16-vsub-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_vsubc_minmax_test",
4591 srcs = [
4592 "test/f16-vsubc-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_vrsubc_minmax_test",
4600 srcs = [
4601 "test/f16-vrsubc-minmax.cc",
4602 "test/vbinaryc-microkernel-tester.h",
4603 ] + MICROKERNEL_TEST_HDRS,
4604 deps = MICROKERNEL_TEST_DEPS,
4605)
4606
4607xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004608 name = "f32_vabs_test",
4609 srcs = [
4610 "test/f32-vabs.cc",
4611 "test/vunary-microkernel-tester.h",
4612 ] + MICROKERNEL_TEST_HDRS,
4613 deps = MICROKERNEL_TEST_DEPS,
4614)
4615
4616xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004617 name = "f32_vadd_test",
4618 srcs = [
4619 "test/f32-vadd.cc",
4620 "test/vbinary-microkernel-tester.h",
4621 ] + MICROKERNEL_TEST_HDRS,
4622 deps = MICROKERNEL_TEST_DEPS,
4623)
4624
4625xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004626 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004627 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004628 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004629 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004630 ] + MICROKERNEL_TEST_HDRS,
4631 deps = MICROKERNEL_TEST_DEPS,
4632)
4633
4634xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004635 name = "f32_vaddc_test",
4636 srcs = [
4637 "test/f32-vaddc.cc",
4638 "test/vbinaryc-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_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004645 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004646 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004647 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004648 ] + MICROKERNEL_TEST_HDRS,
4649 deps = MICROKERNEL_TEST_DEPS,
4650)
4651
4652xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004653 name = "f32_vdiv_test",
4654 srcs = [
4655 "test/f32-vdiv.cc",
4656 "test/vbinary-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_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004663 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004664 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004665 "test/vbinary-microkernel-tester.h",
4666 ] + MICROKERNEL_TEST_HDRS,
4667 deps = MICROKERNEL_TEST_DEPS,
4668)
4669
4670xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004671 name = "f32_vdivc_test",
4672 srcs = [
4673 "test/f32-vdivc.cc",
4674 "test/vbinaryc-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_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004681 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004682 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004683 "test/vbinaryc-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_vrdivc_test",
4690 srcs = [
4691 "test/f32-vrdivc.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_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004699 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004700 "test/f32-vrdivc-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(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004707 name = "f32_vmax_test",
4708 srcs = [
4709 "test/f32-vmax.cc",
4710 "test/vbinary-microkernel-tester.h",
4711 ] + MICROKERNEL_TEST_HDRS,
4712 deps = MICROKERNEL_TEST_DEPS,
4713)
4714
4715xnnpack_unit_test(
4716 name = "f32_vmaxc_test",
4717 srcs = [
4718 "test/f32-vmaxc.cc",
4719 "test/vbinaryc-microkernel-tester.h",
4720 ] + MICROKERNEL_TEST_HDRS,
4721 deps = MICROKERNEL_TEST_DEPS,
4722)
4723
4724xnnpack_unit_test(
4725 name = "f32_vmin_test",
4726 srcs = [
4727 "test/f32-vmin.cc",
4728 "test/vbinary-microkernel-tester.h",
4729 ] + MICROKERNEL_TEST_HDRS,
4730 deps = MICROKERNEL_TEST_DEPS,
4731)
4732
4733xnnpack_unit_test(
4734 name = "f32_vminc_test",
4735 srcs = [
4736 "test/f32-vminc.cc",
4737 "test/vbinaryc-microkernel-tester.h",
4738 ] + MICROKERNEL_TEST_HDRS,
4739 deps = MICROKERNEL_TEST_DEPS,
4740)
4741
4742xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004743 name = "f32_vmul_test",
4744 srcs = [
4745 "test/f32-vmul.cc",
4746 "test/vbinary-microkernel-tester.h",
4747 ] + MICROKERNEL_TEST_HDRS,
4748 deps = MICROKERNEL_TEST_DEPS,
4749)
4750
4751xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004752 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004753 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004754 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004755 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004756 ] + MICROKERNEL_TEST_HDRS,
4757 deps = MICROKERNEL_TEST_DEPS,
4758)
4759
4760xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004761 name = "f32_vmulc_test",
4762 srcs = [
4763 "test/f32-vmulc.cc",
4764 "test/vbinaryc-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_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004771 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004772 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004773 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004774 ] + MICROKERNEL_TEST_HDRS,
4775 deps = MICROKERNEL_TEST_DEPS,
4776)
4777
4778xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004779 name = "f16_vmulcaddc_minmax_test",
4780 srcs = [
4781 "test/f16-vmulcaddc-minmax.cc",
4782 "test/vmulcaddc-microkernel-tester.h",
4783 "src/xnnpack/AlignedAllocator.h",
4784 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004785 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004786)
4787
4788xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004789 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004790 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004791 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004792 "test/vmulcaddc-microkernel-tester.h",
4793 "src/xnnpack/AlignedAllocator.h",
4794 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004795 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004796)
4797
4798xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004799 name = "f32_vlrelu_test",
4800 srcs = [
4801 "test/f32-vlrelu.cc",
4802 "test/vunary-microkernel-tester.h",
4803 ] + MICROKERNEL_TEST_HDRS,
4804 deps = MICROKERNEL_TEST_DEPS,
4805)
4806
4807xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004808 name = "f32_vneg_test",
4809 srcs = [
4810 "test/f32-vneg.cc",
4811 "test/vunary-microkernel-tester.h",
4812 ] + MICROKERNEL_TEST_HDRS,
4813 deps = MICROKERNEL_TEST_DEPS,
4814)
4815
4816xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004817 name = "f32_vrndne_test",
4818 srcs = [
4819 "test/f32-vrndne.cc",
4820 "test/vunary-microkernel-tester.h",
4821 ] + MICROKERNEL_TEST_HDRS,
4822 deps = MICROKERNEL_TEST_DEPS,
4823)
4824
4825xnnpack_unit_test(
4826 name = "f32_vrndz_test",
4827 srcs = [
4828 "test/f32-vrndz.cc",
4829 "test/vunary-microkernel-tester.h",
4830 ] + MICROKERNEL_TEST_HDRS,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
4835 name = "f32_vrndu_test",
4836 srcs = [
4837 "test/f32-vrndu.cc",
4838 "test/vunary-microkernel-tester.h",
4839 ] + MICROKERNEL_TEST_HDRS,
4840 deps = MICROKERNEL_TEST_DEPS,
4841)
4842
4843xnnpack_unit_test(
4844 name = "f32_vrndd_test",
4845 srcs = [
4846 "test/f32-vrndd.cc",
4847 "test/vunary-microkernel-tester.h",
4848 ] + MICROKERNEL_TEST_HDRS,
4849 deps = MICROKERNEL_TEST_DEPS,
4850)
4851
4852xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004853 name = "f32_vscale_test",
4854 srcs = [
4855 "test/f32-vscale.cc",
4856 "test/vscale-microkernel-tester.h",
4857 ] + MICROKERNEL_TEST_HDRS,
4858 deps = MICROKERNEL_TEST_DEPS,
4859)
4860
4861xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004862 name = "f32_vscaleexpminusmax_test",
4863 srcs = [
4864 "test/f32-vscaleexpminusmax.cc",
4865 "test/vscaleexpminusmax-microkernel-tester.h",
4866 ] + MICROKERNEL_TEST_HDRS,
4867 deps = MICROKERNEL_TEST_DEPS,
4868)
4869
4870xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004871 name = "f32_vscaleextexp_test",
4872 srcs = [
4873 "test/f32-vscaleextexp.cc",
4874 "test/vscaleextexp-microkernel-tester.h",
4875 ] + MICROKERNEL_TEST_HDRS,
4876 deps = MICROKERNEL_TEST_DEPS,
4877)
4878
4879xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004880 name = "f32_vsqr_test",
4881 srcs = [
4882 "test/f32-vsqr.cc",
4883 "test/vunary-microkernel-tester.h",
4884 ] + MICROKERNEL_TEST_HDRS,
4885 deps = MICROKERNEL_TEST_DEPS,
4886)
4887
4888xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004889 name = "f32_vsqrdiff_test",
4890 srcs = [
4891 "test/f32-vsqrdiff.cc",
4892 "test/vbinary-microkernel-tester.h",
4893 ] + MICROKERNEL_TEST_HDRS,
4894 deps = MICROKERNEL_TEST_DEPS,
4895)
4896
4897xnnpack_unit_test(
4898 name = "f32_vsqrdiffc_test",
4899 srcs = [
4900 "test/f32-vsqrdiffc.cc",
4901 "test/vbinaryc-microkernel-tester.h",
4902 ] + MICROKERNEL_TEST_HDRS,
4903 deps = MICROKERNEL_TEST_DEPS,
4904)
4905
4906xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004907 name = "f32_vsqrt_test",
4908 srcs = [
4909 "test/f32-vsqrt.cc",
4910 "test/vunary-microkernel-tester.h",
4911 ] + MICROKERNEL_TEST_HDRS,
4912 deps = MICROKERNEL_TEST_DEPS,
4913)
4914
4915xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004916 name = "f32_vsub_test",
4917 srcs = [
4918 "test/f32-vsub.cc",
4919 "test/vbinary-microkernel-tester.h",
4920 ] + MICROKERNEL_TEST_HDRS,
4921 deps = MICROKERNEL_TEST_DEPS,
4922)
4923
4924xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004925 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004926 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004927 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004928 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004929 ] + MICROKERNEL_TEST_HDRS,
4930 deps = MICROKERNEL_TEST_DEPS,
4931)
4932
4933xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004934 name = "f32_vsubc_test",
4935 srcs = [
4936 "test/f32-vsubc.cc",
4937 "test/vbinaryc-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_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004944 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004945 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004946 "test/vbinaryc-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_vrsubc_test",
4953 srcs = [
4954 "test/f32-vrsubc.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_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004962 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004963 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004964 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004965 ] + MICROKERNEL_TEST_HDRS,
4966 deps = MICROKERNEL_TEST_DEPS,
4967)
4968
4969xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004970 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004971 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004972 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004973 "test/avgpool-microkernel-tester.h",
4974 "src/xnnpack/AlignedAllocator.h",
4975 ] + MICROKERNEL_TEST_HDRS,
4976 deps = MICROKERNEL_TEST_DEPS,
4977)
4978
4979xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004980 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004981 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004982 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983 "test/gemm-microkernel-tester.h",
4984 "src/xnnpack/AlignedAllocator.h",
4985 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004986 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987)
4988
4989xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004990 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004991 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004992 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993 "test/dwconv-microkernel-tester.h",
4994 "src/xnnpack/AlignedAllocator.h",
4995 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004996 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004997)
4998
4999xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005000 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005001 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005002 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005003 "test/gavgpool-microkernel-tester.h",
5004 "src/xnnpack/AlignedAllocator.h",
5005 ] + MICROKERNEL_TEST_HDRS,
5006 deps = MICROKERNEL_TEST_DEPS,
5007)
5008
5009xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005010 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005011 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005012 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005013 "test/gemm-microkernel-tester.h",
5014 "src/xnnpack/AlignedAllocator.h",
5015 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005016 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005017)
5018
5019xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005020 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005021 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005022 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005023 "test/vadd-microkernel-tester.h",
5024 ] + MICROKERNEL_TEST_HDRS,
5025 deps = MICROKERNEL_TEST_DEPS,
5026)
5027
5028xnnpack_unit_test(
5029 name = "u8_clamp_test",
5030 srcs = [
5031 "test/u8-clamp.cc",
5032 "test/clamp-microkernel-tester.h",
5033 ] + MICROKERNEL_TEST_HDRS,
5034 deps = MICROKERNEL_TEST_DEPS,
5035)
5036
5037xnnpack_unit_test(
5038 name = "u8_lut32norm_test",
5039 srcs = [
5040 "test/u8-lut32norm.cc",
5041 "test/lut-norm-microkernel-tester.h",
5042 ] + MICROKERNEL_TEST_HDRS,
5043 deps = MICROKERNEL_TEST_DEPS,
5044)
5045
5046xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005047 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005048 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005049 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050 "test/maxpool-microkernel-tester.h",
5051 ] + MICROKERNEL_TEST_HDRS,
5052 deps = MICROKERNEL_TEST_DEPS,
5053)
5054
5055xnnpack_unit_test(
5056 name = "u8_rmax_test",
5057 srcs = [
5058 "test/u8-rmax.cc",
5059 "test/rmax-microkernel-tester.h",
5060 ] + MICROKERNEL_TEST_HDRS,
5061 deps = MICROKERNEL_TEST_DEPS,
5062)
5063
5064xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005065 name = "x32_fill_test",
5066 srcs = [
5067 "test/x32-fill.cc",
5068 "test/fill-microkernel-tester.h",
5069 ] + MICROKERNEL_TEST_HDRS,
5070 deps = MICROKERNEL_TEST_DEPS,
5071)
5072
5073xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005074 name = "x32_packx_test",
5075 srcs = [
5076 "test/x32-packx.cc",
5077 "test/pack-microkernel-tester.h",
5078 "src/xnnpack/AlignedAllocator.h",
5079 ] + MICROKERNEL_TEST_HDRS,
5080 deps = MICROKERNEL_TEST_DEPS,
5081)
5082
5083xnnpack_unit_test(
5084 name = "x32_pad_test",
5085 srcs = [
5086 "test/x32-pad.cc",
5087 "test/pad-microkernel-tester.h",
5088 ] + MICROKERNEL_TEST_HDRS,
5089 deps = MICROKERNEL_TEST_DEPS,
5090)
5091
5092xnnpack_unit_test(
5093 name = "x32_unpool_test",
5094 srcs = [
5095 "test/x32-unpool.cc",
5096 "test/unpool-microkernel-tester.h",
5097 ] + MICROKERNEL_TEST_HDRS,
5098 deps = MICROKERNEL_TEST_DEPS,
5099)
5100
5101xnnpack_unit_test(
5102 name = "x32_zip_test",
5103 srcs = [
5104 "test/x32-zip.cc",
5105 "test/zip-microkernel-tester.h",
5106 ] + MICROKERNEL_TEST_HDRS,
5107 deps = MICROKERNEL_TEST_DEPS,
5108)
5109
5110xnnpack_unit_test(
5111 name = "x8_lut_test",
5112 srcs = [
5113 "test/x8-lut.cc",
5114 "test/lut-microkernel-tester.h",
5115 ] + MICROKERNEL_TEST_HDRS,
5116 deps = MICROKERNEL_TEST_DEPS,
5117)
5118
5119xnnpack_unit_test(
5120 name = "x8_zip_test",
5121 srcs = [
5122 "test/x8-zip.cc",
5123 "test/zip-microkernel-tester.h",
5124 ] + MICROKERNEL_TEST_HDRS,
5125 deps = MICROKERNEL_TEST_DEPS,
5126)
5127
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005128xnnpack_unit_test(
5129 name = "requantization_test",
5130 srcs = [
5131 "src/xnnpack/requantization-stubs.h",
5132 "test/requantization.cc",
5133 "test/requantization-tester.h",
5134 ] + MICROKERNEL_TEST_HDRS,
5135 deps = MICROKERNEL_TEST_DEPS,
5136)
5137
Marat Dukhan20c3b922020-03-10 03:45:06 -07005138########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005139
5140xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005141 name = "operator_size_test",
5142 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005143 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005144)
5145
Marat Dukhan20c3b922020-03-10 03:45:06 -07005146xnnpack_binary(
5147 name = "subgraph_size_test",
5148 srcs = ["test/subgraph-size.c"],
5149 deps = [":XNNPACK"],
5150)
5151
5152########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005153
5154xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005155 name = "abs_nc_test",
5156 srcs = [
5157 "test/abs-nc.cc",
5158 "test/abs-operator-tester.h",
5159 ],
5160 deps = OPERATOR_TEST_DEPS,
5161)
5162
5163xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005164 name = "add_nd_test",
5165 srcs = [
5166 "test/add-nd.cc",
5167 "test/binary-elementwise-operator-tester.h",
5168 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005169 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005170)
5171
5172xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005173 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005174 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005175 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005176 "test/argmax-pooling-operator-tester.h",
5177 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005178 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005179)
5180
5181xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005182 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005184 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005185 "test/average-pooling-operator-tester.h",
5186 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005187 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005188)
5189
5190xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005191 name = "bankers_rounding_nc_test",
5192 srcs = [
5193 "test/bankers-rounding-nc.cc",
5194 "test/bankers-rounding-operator-tester.h",
5195 ],
5196 deps = OPERATOR_TEST_DEPS,
5197)
5198
5199xnnpack_unit_test(
5200 name = "ceiling_nc_test",
5201 srcs = [
5202 "test/ceiling-nc.cc",
5203 "test/ceiling-operator-tester.h",
5204 ],
5205 deps = OPERATOR_TEST_DEPS,
5206)
5207
5208xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005209 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005211 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005212 "test/channel-shuffle-operator-tester.h",
5213 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005214 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215)
5216
5217xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005218 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005220 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005221 "test/clamp-operator-tester.h",
5222 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005223 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224)
5225
5226xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005227 name = "constant_pad_nd_test",
5228 srcs = [
5229 "test/constant-pad-nd.cc",
5230 "test/constant-pad-operator-tester.h",
5231 ],
5232 deps = OPERATOR_TEST_DEPS,
5233)
5234
5235xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005236 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005238 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005239 "test/convolution-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 Dukhanefc47b82019-11-18 09:25:38 -08005245 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005247 "test/convolution-nchw.cc",
5248 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005249 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005250 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251)
5252
5253xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005254 name = "copy_nc_test",
5255 srcs = [
5256 "test/copy-nc.cc",
5257 "test/copy-operator-tester.h",
5258 ],
5259 deps = OPERATOR_TEST_DEPS,
5260)
5261
5262xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005263 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005265 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005266 "test/deconvolution-operator-tester.h",
5267 ] + OPERATOR_TEST_PARAMS_HDRS,
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 Dukhan69180502019-12-06 15:00:31 -08005272 name = "divide_nd_test",
5273 srcs = [
5274 "test/binary-elementwise-operator-tester.h",
5275 "test/divide-nd.cc",
5276 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005277 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005278)
5279
5280xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005281 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005283 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 "test/fully-connected-operator-tester.h",
5285 ],
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 Dukhan64e52512020-06-09 13:41:16 -07005290 name = "floor_nc_test",
5291 srcs = [
5292 "test/floor-nc.cc",
5293 "test/floor-operator-tester.h",
5294 ],
5295 deps = OPERATOR_TEST_DEPS,
5296)
5297
5298xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005299 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005301 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005302 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005303 ] + OPERATOR_TEST_PARAMS_HDRS,
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 Dukhanefc47b82019-11-18 09:25:38 -08005308 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005309 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005310 "test/global-average-pooling-ncw.cc",
5311 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005313 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005314)
5315
5316xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005317 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005319 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005320 "test/hardswish-operator-tester.h",
5321 ],
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 = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005328 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005329 "test/leaky-relu-operator-tester.h",
5330 ],
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 = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005337 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005338 "test/max-pooling-operator-tester.h",
5339 ] + OPERATOR_TEST_PARAMS_HDRS,
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 Dukhan79e7f842019-12-05 14:35:50 -08005344 name = "maximum_nd_test",
5345 srcs = [
5346 "test/binary-elementwise-operator-tester.h",
5347 "test/maximum-nd.cc",
5348 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005349 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005350)
5351
5352xnnpack_unit_test(
5353 name = "minimum_nd_test",
5354 srcs = [
5355 "test/binary-elementwise-operator-tester.h",
5356 "test/minimum-nd.cc",
5357 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005358 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005359)
5360
5361xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005362 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005363 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005364 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005365 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005366 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005367 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005368)
5369
5370xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005371 name = "negate_nc_test",
5372 srcs = [
5373 "test/negate-nc.cc",
5374 "test/negate-operator-tester.h",
5375 ],
5376 deps = OPERATOR_TEST_DEPS,
5377)
5378
5379xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005380 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005381 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005382 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005383 "test/prelu-operator-tester.h",
5384 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005385 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005386)
5387
5388xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005389 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005390 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005391 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005392 "test/resize-bilinear-operator-tester.h",
5393 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005394 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005395)
5396
5397xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005398 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005400 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005401 "test/sigmoid-operator-tester.h",
5402 ],
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 Dukhanfd8e6892020-01-27 15:25:25 -08005407 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005408 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005409 "test/softmax-nc.cc",
5410 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005412 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413)
5414
5415xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005416 name = "square_nc_test",
5417 srcs = [
5418 "test/square-nc.cc",
5419 "test/square-operator-tester.h",
5420 ],
5421 deps = OPERATOR_TEST_DEPS,
5422)
5423
5424xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005425 name = "square_root_nc_test",
5426 srcs = [
5427 "test/square-root-nc.cc",
5428 "test/square-root-operator-tester.h",
5429 ],
5430 deps = OPERATOR_TEST_DEPS,
5431)
5432
5433xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005434 name = "squared_difference_nd_test",
5435 srcs = [
5436 "test/binary-elementwise-operator-tester.h",
5437 "test/squared-difference-nd.cc",
5438 ],
5439 deps = OPERATOR_TEST_DEPS,
5440)
5441
5442xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005443 name = "subtract_nd_test",
5444 srcs = [
5445 "test/binary-elementwise-operator-tester.h",
5446 "test/subtract-nd.cc",
5447 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005448 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005449)
5450
5451xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005452 name = "truncation_nc_test",
5453 srcs = [
5454 "test/truncation-nc.cc",
5455 "test/truncation-operator-tester.h",
5456 ],
5457 deps = OPERATOR_TEST_DEPS,
5458)
5459
5460xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005461 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005463 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005464 "test/unpooling-operator-tester.h",
5465 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005466 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467)
5468
Chao Mei6ddfc602020-05-13 22:29:36 -07005469############################### Misc unit tests ###############################
5470
5471xnnpack_unit_test(
5472 name = "memory_planner_test",
5473 srcs = [
5474 "test/memory-planner-test.cc",
5475 ],
5476 deps = [
5477 ":XNNPACK",
5478 ":memory_planner",
5479 ],
5480)
5481
Marat Dukhan08c4a432019-10-03 09:29:21 -07005482############################# Build configurations #############################
5483
Marat Dukhanb8642352019-10-30 15:43:02 -07005484# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005485config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005486 name = "xnn_enable_assembly_explicit_true",
5487 define_values = {"xnn_enable_assembly": "true"},
5488)
5489
5490# Disables usage of assembly kernels.
5491config_setting(
5492 name = "xnn_enable_assembly_explicit_false",
5493 define_values = {"xnn_enable_assembly": "false"},
5494)
5495
Marat Dukhan9de90e02020-06-18 16:04:12 -07005496# Enables usage of sparse inference.
5497config_setting(
5498 name = "xnn_enable_sparse_explicit_true",
5499 define_values = {"xnn_enable_sparse": "true"},
5500)
5501
5502# Disables usage of sparse inference.
5503config_setting(
5504 name = "xnn_enable_sparse_explicit_false",
5505 define_values = {"xnn_enable_sparse": "false"},
5506)
5507
Marat Dukhan05702cf2020-03-26 15:41:33 -07005508# Disables usage of HMP-aware optimizations.
5509config_setting(
5510 name = "xnn_enable_hmp_explicit_false",
5511 define_values = {"xnn_enable_hmp": "false"},
5512)
5513
Chao Mei6ddfc602020-05-13 22:29:36 -07005514# Enable usage of optimized memory allocation
5515config_setting(
5516 name = "xnn_enable_memopt_explicit_true",
5517 define_values = {"xnnpack_enable_memopt": "true"},
5518)
5519
5520# Disable usage of optimized memory allocation
5521config_setting(
5522 name = "xnn_enable_memopt_explicit_false",
5523 define_values = {"xnnpack_enable_memopt": "false"},
5524)
5525
Marat Dukhanb8642352019-10-30 15:43:02 -07005526# Builds with -c dbg
5527config_setting(
5528 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005529 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005530 "compilation_mode": "dbg",
5531 },
5532)
5533
5534# Builds with -c opt
5535config_setting(
5536 name = "optimized_build",
5537 values = {
5538 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005539 },
5540)
5541
5542config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005543 name = "linux_k8",
5544 values = {"cpu": "k8"},
5545)
5546
5547config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005548 name = "linux_arm",
5549 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005550)
5551
5552config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005553 name = "linux_armeabi",
5554 values = {"cpu": "armeabi"},
5555)
5556
5557config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005558 name = "linux_armhf",
5559 values = {"cpu": "armhf"},
5560)
5561
5562config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005563 name = "linux_armv7a",
5564 values = {"cpu": "armv7a"},
5565)
5566
5567config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005568 name = "linux_aarch64",
5569 values = {"cpu": "aarch64"},
5570)
5571
5572config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573 name = "android",
5574 values = {"crosstool_top": "//external:android/crosstool"},
5575)
5576
5577config_setting(
5578 name = "android_armv7",
5579 values = {
5580 "crosstool_top": "//external:android/crosstool",
5581 "cpu": "armeabi-v7a",
5582 },
5583)
5584
5585config_setting(
5586 name = "android_arm64",
5587 values = {
5588 "crosstool_top": "//external:android/crosstool",
5589 "cpu": "arm64-v8a",
5590 },
5591)
5592
5593config_setting(
5594 name = "android_x86",
5595 values = {
5596 "crosstool_top": "//external:android/crosstool",
5597 "cpu": "x86",
5598 },
5599)
5600
5601config_setting(
5602 name = "android_x86_64",
5603 values = {
5604 "crosstool_top": "//external:android/crosstool",
5605 "cpu": "x86_64",
5606 },
5607)
5608
5609config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005610 name = "windows_x86_64",
5611 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005612)
5613
5614config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005615 name = "windows_x86_64_clang",
5616 values = {
5617 "compiler": "clang-cl",
5618 "cpu": "x64_windows",
5619 },
5620)
5621
5622config_setting(
5623 name = "windows_x86_64_mingw",
5624 values = {
5625 "compiler": "mingw-gcc",
5626 "cpu": "x64_windows",
5627 },
5628)
5629
5630config_setting(
5631 name = "windows_x86_64_msys",
5632 values = {
5633 "compiler": "msys-gcc",
5634 "cpu": "x64_windows",
5635 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005636)
5637
5638config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005639 name = "macos_x86_64",
5640 values = {
5641 "apple_platform_type": "macos",
5642 "cpu": "darwin",
5643 },
5644)
5645
5646config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005647 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005648 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005649)
5650
5651config_setting(
5652 name = "emscripten_wasm",
5653 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005654 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005655 "cpu": "wasm",
5656 },
5657)
5658
5659config_setting(
5660 name = "emscripten_wasmsimd",
5661 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005662 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005663 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005664 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005665 },
5666)
5667
5668config_setting(
5669 name = "emscripten_asmjs",
5670 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005671 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005672 "cpu": "asmjs",
5673 },
5674)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005675
5676config_setting(
5677 name = "ios_armv7",
5678 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005679 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005680 "cpu": "ios_armv7",
5681 },
5682)
5683
5684config_setting(
5685 name = "ios_arm64",
5686 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005687 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005688 "cpu": "ios_arm64",
5689 },
5690)
5691
5692config_setting(
5693 name = "ios_arm64e",
5694 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005695 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005696 "cpu": "ios_arm64e",
5697 },
5698)
5699
5700config_setting(
5701 name = "ios_x86",
5702 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005703 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005704 "cpu": "ios_i386",
5705 },
5706)
5707
5708config_setting(
5709 name = "ios_x86_64",
5710 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005711 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005712 "cpu": "ios_x86_64",
5713 },
5714)
5715
5716config_setting(
5717 name = "watchos_armv7k",
5718 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005719 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005720 "cpu": "watchos_armv7k",
5721 },
5722)
5723
5724config_setting(
5725 name = "watchos_arm64_32",
5726 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005727 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005728 "cpu": "watchos_arm64_32",
5729 },
5730)
5731
5732config_setting(
5733 name = "watchos_x86",
5734 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005735 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005736 "cpu": "watchos_i386",
5737 },
5738)
5739
5740config_setting(
5741 name = "watchos_x86_64",
5742 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005743 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005744 "cpu": "watchos_x86_64",
5745 },
5746)
5747
5748config_setting(
5749 name = "tvos_arm64",
5750 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005751 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005752 "cpu": "tvos_arm64",
5753 },
5754)
5755
5756config_setting(
5757 name = "tvos_x86_64",
5758 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005759 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005760 "cpu": "tvos_x86_64",
5761 },
5762)