blob: db4129ad26709d6ee26413a9acb43cdca1db66fb [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 Dukhan93d1ba12020-06-26 12:33:35 -0700708 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
709 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
710 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
711 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700712 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
713 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700714 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
715 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
716 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
717 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700718 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
719 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700720 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
721 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
722 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
723 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700724 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
725 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700726 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
727 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
728 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
729 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700730 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
731 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700732 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
733 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
734 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
735 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700736 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
737 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
738 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
739 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700740 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
741 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
742 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
743 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700744 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
745 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
746 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
747 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700748 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
749 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
750 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
751 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700752 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
753 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700754 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
755 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
756 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
757 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700758 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
759 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700760 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
761 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
762 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
763 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700764 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
765 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700766 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
767 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
768 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
769 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700770 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
771 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700772 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
773 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700774 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
775 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700776 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
777 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
778 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
779 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700780 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
781 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700782 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
783 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
784 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
785 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700786 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
787 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700788 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
789 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
790 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
791 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700792 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
793 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
794 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
795 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700796 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
797 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
798 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
799 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
800 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
801 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
802 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
803 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
804 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
805 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
806 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
807 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
808 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
809 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700810 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
811 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700812 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
813 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
814 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
815 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
816 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
817 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700818 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700819 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700820 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700821 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700822 "src/x32-zip/x2-wasmsimd.c",
823 "src/x32-zip/x3-wasmsimd.c",
824 "src/x32-zip/x4-wasmsimd.c",
825 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700826 "src/math/roundne-wasmsimd-addsub.c",
827 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700828 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700829 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700830 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700831 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700832 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan9340b6c2020-07-16 22:59:48 -0700833 "src/math/sigmoid-wasmsimd-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700834 "src/requantization/fp32-wasmsimd.c",
835]
836
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800837PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800838 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800839 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800840 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700841 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
842 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700843 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700844 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700845 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700846 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700847 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700848 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700849 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
850 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700851 "src/f32-ibilinear/gen/psimd-c4.c",
852 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700853 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700854 "src/f32-spmm/gen/4x1-minmax-psimd.c",
855 "src/f32-spmm/gen/8x1-minmax-psimd.c",
856 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700857 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
858 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700859 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700860 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700861 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700862 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700863 "src/x32-unpool/psimd.c",
864 "src/x32-zip/x2-psimd.c",
865 "src/x32-zip/x3-psimd.c",
866 "src/x32-zip/x4-psimd.c",
867 "src/x32-zip/xm-psimd.c",
868]
869
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800870PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan249cd6a2020-07-17 14:05:30 -0700871 "src/requantization/precise-psimd.c",
872 "src/requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800873]
874
Marat Dukhan08c4a432019-10-03 09:29:21 -0700875# ISA-specific micro-kernels
876NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700877 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
878 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700879 "src/f32-clamp/gen/neon-x4.c",
880 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700881 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
882 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700883 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
884 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700885 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
886 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700887 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
888 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700889 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
890 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
891 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
892 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700893 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
894 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
895 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
896 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700897 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
898 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
899 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
900 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700901 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700902 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
903 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700904 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
905 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
906 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
907 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
908 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
909 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
910 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
911 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
912 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
913 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
914 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
915 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
916 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
917 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
918 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
919 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
920 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
921 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
922 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
923 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
924 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
925 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
926 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
927 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
928 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
929 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
930 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
931 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
932 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
933 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
934 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700935 "src/f32-hswish/gen/hswish-neon-x4.c",
936 "src/f32-hswish/gen/hswish-neon-x8.c",
937 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700938 "src/f32-ibilinear/gen/neon-c4.c",
939 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700940 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
941 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
942 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
943 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
944 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
945 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
946 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
947 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
948 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
949 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
950 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
951 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
952 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
953 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
954 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
955 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700956 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
957 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
958 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700959 "src/f32-ppmm/gen/4x8-minmax-neon.c",
960 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800961 "src/f32-prelu/gen/neon-2x4.c",
962 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800963 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
964 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
965 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
966 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
967 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
968 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
969 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
970 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
971 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
972 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
973 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
974 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
975 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
976 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
977 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
978 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
979 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
980 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
981 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
982 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
983 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
984 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
985 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
986 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700987 "src/f32-relu/gen/neon-x4.c",
988 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700989 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800990 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800991 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
992 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
993 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
994 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
995 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
996 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
997 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
998 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
999 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1000 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1001 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1002 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1003 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1004 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1005 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1006 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1007 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1008 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001009 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1010 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1011 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1012 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001013 "src/f32-vbinary/gen/vmax-neon-x4.c",
1014 "src/f32-vbinary/gen/vmax-neon-x8.c",
1015 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1016 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1017 "src/f32-vbinary/gen/vmin-neon-x4.c",
1018 "src/f32-vbinary/gen/vmin-neon-x8.c",
1019 "src/f32-vbinary/gen/vminc-neon-x4.c",
1020 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001021 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1022 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1023 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1024 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1025 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1026 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001027 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1028 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1029 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1030 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001031 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1032 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1033 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1034 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001035 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1036 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001037 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1038 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001039 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1040 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1041 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1042 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1043 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1044 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1045 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1046 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001047 "src/f32-vunary/gen/vabs-neon-x4.c",
1048 "src/f32-vunary/gen/vabs-neon-x8.c",
1049 "src/f32-vunary/gen/vneg-neon-x4.c",
1050 "src/f32-vunary/gen/vneg-neon-x8.c",
1051 "src/f32-vunary/gen/vsqr-neon-x4.c",
1052 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001053 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1054 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1055 "src/qu8-dwconv/up8x9-minmax-neon.c",
1056 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1057 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1058 "src/qu8-gemm/4x8-minmax-neon.c",
1059 "src/qu8-gemm/8x8-minmax-neon.c",
1060 "src/qu8-igemm/4x8-minmax-neon.c",
1061 "src/qu8-igemm/8x8-minmax-neon.c",
1062 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001063 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001064 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001065 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001066 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001067 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001068 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001069 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001070 "src/x32-zip/x2-neon.c",
1071 "src/x32-zip/x3-neon.c",
1072 "src/x32-zip/x4-neon.c",
1073 "src/x32-zip/xm-neon.c",
1074 "src/x8-zip/x2-neon.c",
1075 "src/x8-zip/x3-neon.c",
1076 "src/x8-zip/x4-neon.c",
1077 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001078 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001079 "src/math/roundd-neon-addsub.c",
1080 "src/math/roundd-neon-cvt.c",
1081 "src/math/roundu-neon-addsub.c",
1082 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001083 "src/math/roundz-neon-addsub.c",
1084 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001085 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001086 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1087 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1088 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1089 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1090 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1091 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001092 "src/math/sqrt-neon-nr1rsqrts.c",
1093 "src/math/sqrt-neon-nr2rsqrts.c",
1094 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001095 "src/requantization/precise-neon.c",
1096 "src/requantization/fp32-neon.c",
1097 "src/requantization/q31-neon.c",
1098 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001099]
1100
1101NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001102 "src/f32-ibilinear/gen/neonfma-c4.c",
1103 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001104 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1105 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1106 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1107 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1108 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1109 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1110 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1111 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1112 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001113 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1114 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1115 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1116 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001117 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1118 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1119 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1120 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001121 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1122 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1123 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1124 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001125 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1126 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1127 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1128 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1129 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1130 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1131 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1132 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1133 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1134 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1135 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1136 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1137 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1138 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1139 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1140 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1141 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1142 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001143 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1144 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001145 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1146 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1147 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1148 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1149 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1150 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1151 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1152 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1153 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1154 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1155 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1156 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1157 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1158 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1159 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1160 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1161 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1162 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1163 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1164 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1165 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1166 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1167 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1168 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001169 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1170 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1171 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1172 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1173 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1174 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1175 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1176 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1177 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1178 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1179 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1180 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1181 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1182 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1183 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1184 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1185 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1186 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1187 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001223 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1224 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001225 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1226 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1227 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1228 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1229 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1230 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1231 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1232 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1233 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1234 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1235 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1236 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1237 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1238 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1239 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1240 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1241 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1242 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1243 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1244 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001245 "src/math/exp-neonfma-lut64-p2.c",
1246 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001247 "src/math/expminus-neonfma-lut2048-p1.c",
1248 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001249 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001250 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1251 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1252 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1253 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1254 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1255 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1256 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1257 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1258 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1259 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1260 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1261 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1262 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1263 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1264 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1265 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1266 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1267 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001268 "src/math/sqrt-neonfma-nr1fma.c",
1269 "src/math/sqrt-neonfma-nr2fma.c",
1270 "src/math/sqrt-neonfma-nr3fma.c",
1271 "src/math/sqrt-neonfma-nr2fma1adj.c",
1272 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001273]
1274
1275AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001276 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1277 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1278 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1279 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1280 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1281 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001282 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1283 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1284 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1285 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1286 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1287 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1288 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1289 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1290 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1291 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1292 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1293 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1294 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1295 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1296 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1297 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1298 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1299 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1300 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1301 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001302 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1303 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001304 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1305 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001306 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1307 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001308 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1309 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001310 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1311 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1312 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1313 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1314 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001315 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1316 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1318 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1319 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1320 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1321 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1322 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1323 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1325 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1326 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1327 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001333 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1334 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1335 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1336 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1337 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1338 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1339 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1340 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1341 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1342 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1343 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1344 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1345 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1346 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1347 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1348 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1349 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1350 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001351 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1352 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001353 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1354 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1355 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1356 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1357 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1358 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001359]
1360
Marat Dukhan8853b822020-05-07 12:19:01 -07001361NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001362 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1363 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1364 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1365 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1366 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1367 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1368 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1369 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001370 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001371 "src/math/roundd-neonv8.c",
1372 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001373 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001374]
1375
Marat Dukhan08c4a432019-10-03 09:29:21 -07001376AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001377 "src/f16-clamp/gen/neonfp16arith-x16.c",
1378 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001379 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1380 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1381 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1382 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1383 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1384 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1385 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1386 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1387 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1388 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1389 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1390 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001391 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1392 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001393 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1394 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001395 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1396 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001397 "src/f16-relu/gen/neonfp16arith-x16.c",
1398 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001399 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1400 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1401 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1402 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1403 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1404 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1405 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1406 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1407 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1408 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1409 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1410 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1411 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1412 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1413 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1414 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1415 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1416 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1417 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1418 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1419 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1420 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1421 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1422 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1423 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1424 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1425 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1426 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001427 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1428 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001429 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1430 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1431 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1432 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1433 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1434 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1435 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1436 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001437 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1438 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1439 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1440 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001441 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1442 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1443 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1444 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1445 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1446 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1447 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1448 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1449 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1450 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1451 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1452 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001453 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1454 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1455 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1456 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1457 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1458 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1459 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1460 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001461]
1462
1463SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001464 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1465 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001466 "src/f32-clamp/gen/sse-x4.c",
1467 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001468 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1469 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001470 "src/f32-dwconv-chw/3x3p1-sse.c",
1471 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001472 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1473 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1474 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1475 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1476 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1477 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1478 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1479 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1480 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1481 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1482 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1483 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001484 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001485 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1486 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001487 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1488 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1489 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1490 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1491 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1492 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1493 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1494 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1495 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1496 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1497 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1498 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1499 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001500 "src/f32-hswish/gen/hswish-sse-x4.c",
1501 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001502 "src/f32-ibilinear/gen/sse-c4.c",
1503 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001504 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1505 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1506 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1507 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1508 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1509 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1510 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001511 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1512 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1513 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001514 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001515 "src/f32-prelu/gen/sse-2x4.c",
1516 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001517 "src/f32-relu/gen/sse-x4.c",
1518 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001519 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001520 "src/f32-spmm/gen/4x1-minmax-sse.c",
1521 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001522 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001523 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1524 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1525 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1526 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1527 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1528 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1529 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1530 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001531 "src/f32-vbinary/gen/vmax-sse-x4.c",
1532 "src/f32-vbinary/gen/vmax-sse-x8.c",
1533 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1534 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1535 "src/f32-vbinary/gen/vmin-sse-x4.c",
1536 "src/f32-vbinary/gen/vmin-sse-x8.c",
1537 "src/f32-vbinary/gen/vminc-sse-x4.c",
1538 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001539 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1540 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1541 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1542 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1543 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1544 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1545 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1546 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001547 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1548 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1549 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1550 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001551 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1552 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1553 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1554 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001555 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1556 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001557 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1558 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001559 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1560 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001561 "src/f32-vunary/gen/vabs-sse-x4.c",
1562 "src/f32-vunary/gen/vabs-sse-x8.c",
1563 "src/f32-vunary/gen/vneg-sse-x4.c",
1564 "src/f32-vunary/gen/vneg-sse-x8.c",
1565 "src/f32-vunary/gen/vsqr-sse-x4.c",
1566 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001567 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001568 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001569 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001570 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001571 "src/math/roundd-sse-addsub.c",
1572 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001573 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001574 "src/math/sqrt-sse-hh1mac.c",
1575 "src/math/sqrt-sse-nr1mac.c",
1576 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001577]
1578
1579SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001580 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1581 "src/f32-argmaxpool/4x-sse2-c4.c",
1582 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001583 "src/f32-prelu/gen/sse2-2x4.c",
1584 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001585 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1586 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1587 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1588 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1589 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1590 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1591 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1592 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1593 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1594 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1595 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1596 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001597 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001598 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001599 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001600 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001601 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1602 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001603 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1604 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001605 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1606 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1607 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1608 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1609 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1610 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1611 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1612 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001613 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1614 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1615 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1616 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1617 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1618 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1619 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1620 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1621 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001622 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001623 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001624 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001625 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001626 "src/x32-zip/x2-sse2.c",
1627 "src/x32-zip/x3-sse2.c",
1628 "src/x32-zip/x4-sse2.c",
1629 "src/x32-zip/xm-sse2.c",
1630 "src/x8-zip/x2-sse2.c",
1631 "src/x8-zip/x3-sse2.c",
1632 "src/x8-zip/x4-sse2.c",
1633 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001634 "src/math/exp-sse2-p5.c",
1635 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001636 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001637 "src/math/roundd-sse2-cvt.c",
1638 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001639 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001640 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001641 "src/requantization/precise-sse2.c",
1642 "src/requantization/fp32-sse2.c",
1643 "src/requantization/q31-sse2.c",
1644 "src/requantization/gemmlowp-sse2.c",
1645]
1646
1647SSSE3_UKERNELS = [
1648 "src/requantization/precise-ssse3.c",
1649 "src/requantization/q31-ssse3.c",
1650 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001651]
1652
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001653SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001654 "src/f32-prelu/gen/sse41-2x4.c",
1655 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001656 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001657 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001658 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001659 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001660 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1661 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001662 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1663 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001664 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1665 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1666 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1667 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1668 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1669 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1670 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1671 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001672 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001673 "src/math/roundd-sse41.c",
1674 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001675 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001676 "src/requantization/precise-sse4.c",
1677 "src/requantization/q31-sse4.c",
1678 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001679]
1680
Marat Dukhan08c4a432019-10-03 09:29:21 -07001681AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001682 "src/f32-clamp/gen/avx-x8.c",
1683 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001684 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1685 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1686 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1687 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1688 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1689 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1690 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1691 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1692 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1693 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1694 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1695 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1696 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1697 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1698 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1699 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1700 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1701 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1702 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1703 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1704 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1705 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1706 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1707 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1708 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1709 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1710 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1711 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1712 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1713 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001714 "src/f32-hswish/gen/hswish-avx-x8.c",
1715 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001716 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1717 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1718 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1719 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1720 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1721 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1722 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1723 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1724 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001725 "src/f32-prelu/gen/avx-2x8.c",
1726 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001727 "src/f32-relu/gen/avx-x16.c",
1728 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001729 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001730 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1731 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1732 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1733 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1734 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1735 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1736 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1737 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001738 "src/f32-vbinary/gen/vmax-avx-x8.c",
1739 "src/f32-vbinary/gen/vmax-avx-x16.c",
1740 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1741 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1742 "src/f32-vbinary/gen/vmin-avx-x8.c",
1743 "src/f32-vbinary/gen/vmin-avx-x16.c",
1744 "src/f32-vbinary/gen/vminc-avx-x8.c",
1745 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001746 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1747 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1748 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1749 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1750 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1751 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1752 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1753 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001754 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1755 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1756 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1757 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001758 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1759 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1760 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1761 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001762 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1763 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001764 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001765 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1766 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1767 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1768 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1769 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1770 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1771 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1772 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001773 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1774 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001775 "src/f32-vunary/gen/vabs-avx-x8.c",
1776 "src/f32-vunary/gen/vabs-avx-x16.c",
1777 "src/f32-vunary/gen/vneg-avx-x8.c",
1778 "src/f32-vunary/gen/vneg-avx-x16.c",
1779 "src/f32-vunary/gen/vsqr-avx-x8.c",
1780 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001781]
1782
Marat Dukhanfda12b82019-11-21 12:27:59 -08001783FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001784 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1785 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1786 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1787 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1788 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1789 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1790 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1791 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1792 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1793 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1794 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1795 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1796 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1797 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1798 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1799 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1800 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1801 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1802 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1803 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1804 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1805 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1806 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1807 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1808 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1809 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1810 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1811 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1812 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1813 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1814 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1815 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1816 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1817 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1818 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1819 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1820 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1821 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1822 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1823 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001824 "src/f32-hswish/gen/hswish-fma3-x8.c",
1825 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001826 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1827 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1828 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1829 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1830 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1831 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1832 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1833 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1834 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1835 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1836 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1837 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1838 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1839 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001840 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1841 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1842 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1843 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1844 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1845 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1846 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1847 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001848 "src/math/sqrt-fma3-nr1fma.c",
1849 "src/math/sqrt-fma3-nr2fma.c",
1850 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001851]
1852
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001853AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001854 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1855 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1856 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1857 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1858 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1859 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1860 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1861 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1862 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1863 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1864 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1865 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1866 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1867 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1868 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1869 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1870 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1871 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1872 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1873 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1874 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1875 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1876 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1877 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1878 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1879 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1880 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1881 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1882 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1883 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1884 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1885 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1886 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1887 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1888 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1889 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001890 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1891 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1892 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1893 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1894 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1895 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1896 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1897 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1898 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1899 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1900 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1901 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1902 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1903 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1904 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1905 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1906 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1907 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1908 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1909 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1910 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1911 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1912 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1913 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1914 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1915 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1916 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1917 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1918 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1919 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001920 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1921 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1922 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1923 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1924 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1925 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1926 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1927 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1928 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1929 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1930 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1931 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1932 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1933 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1934 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1935 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1936 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1937 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1938 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1939 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1940 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1941 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1942 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1943 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001944 "src/math/exp-avx2-p5.c",
1945 "src/math/exp-avx2-perm-p3.c",
1946 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001947 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001948 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001949 "src/math/sigmoid-avx2-rr2-p5-div.c",
1950 "src/math/sigmoid-avx2-rr1-p5-div.c",
1951 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1952 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1953 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1954 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001955]
1956
Marat Dukhan08c4a432019-10-03 09:29:21 -07001957AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001958 "src/f32-clamp/gen/avx512f-x16.c",
1959 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001960 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1961 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1962 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1963 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1964 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1965 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1966 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1967 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1968 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1969 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1970 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1971 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1972 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1973 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1974 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1975 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1976 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1977 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1978 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1979 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1980 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1981 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1982 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1983 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001984 "src/f32-hswish/gen/hswish-avx512f-x16.c",
1985 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001986 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1987 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1988 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1989 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1990 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1991 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001992 "src/f32-prelu/gen/avx512f-2x16.c",
1993 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001994 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1995 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1996 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1997 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1998 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1999 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2000 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2001 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2002 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2003 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2004 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2005 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2006 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2007 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2008 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2009 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2010 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2011 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2012 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2013 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2014 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2015 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2016 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2017 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2018 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2019 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2020 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2021 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2022 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2023 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2024 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2025 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2026 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2027 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2028 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2029 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002030 "src/f32-relu/gen/avx512f-x16.c",
2031 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002032 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002033 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2034 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2035 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2036 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2037 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2038 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2039 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2040 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002041 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2042 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2043 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2044 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2045 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2046 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2047 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2048 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002049 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2050 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2051 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2052 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2053 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2054 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2055 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2056 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002057 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2058 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2059 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2060 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002061 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2062 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2063 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2064 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002065 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2066 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002067 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002068 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2069 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2070 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2071 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2072 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2073 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2074 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2075 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2076 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2077 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2078 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2079 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2080 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2081 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2082 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2083 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2084 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2085 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2086 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2087 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2088 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2089 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2090 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2091 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002092 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2093 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2094 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2095 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2096 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2097 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2098 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2099 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002100 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2101 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2102 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2103 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2104 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2105 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2106 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2107 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002108 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2109 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2110 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2111 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2112 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2113 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002114 "src/math/exp-avx512f-p5-scalef.c",
2115 "src/math/exp-avx512f-p5.c",
2116 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002117 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002118 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002119 "src/math/sqrt-avx512f-nr1fma.c",
2120 "src/math/sqrt-avx512f-nr2fma.c",
2121 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002122]
2123
2124AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002125 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002126 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002127 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2128 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002129 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002130 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002131 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002132 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002133 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2134 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002135 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2136 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2137 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2138 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002139]
2140
2141AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002142 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002143 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2144 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002145 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002146 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2147 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002148 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2149 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2150 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002151 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002152 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2153 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2154 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002155 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002156 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2157 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002158 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002159 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2160 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2161 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2162 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2163 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2164 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2165 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2166 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2167 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2168 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2169 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2170 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2171 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2172 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2173 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2174 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2175 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2176 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002177 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2178 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002179 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002180 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2181 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2182 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2183 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2184 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2185 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2186 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2187 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2188 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2189 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2190 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2191 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2192 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2193 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2194 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2195 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2196 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2197 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002198 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2199 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2200 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2201 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2202 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2203 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2204 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2205 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2206 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2207 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2208 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2209 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2210 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2211 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2212 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2213 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2214 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2215 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002216]
2217
Marat Dukhan1b354632020-03-23 12:50:22 -07002218INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002219 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002220 "src/xnnpack/argmaxpool.h",
2221 "src/xnnpack/avgpool.h",
2222 "src/xnnpack/clamp.h",
2223 "src/xnnpack/common.h",
2224 "src/xnnpack/conv.h",
2225 "src/xnnpack/dwconv.h",
2226 "src/xnnpack/gavgpool.h",
2227 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002228 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002229 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002230 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002231 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002232 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002233 "src/xnnpack/lut.h",
2234 "src/xnnpack/math.h",
2235 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002236 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002237 "src/xnnpack/packx.h",
2238 "src/xnnpack/pad.h",
2239 "src/xnnpack/params.h",
2240 "src/xnnpack/pavgpool.h",
2241 "src/xnnpack/ppmm.h",
2242 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002243 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002244 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002245 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002246 "src/xnnpack/rmax.h",
2247 "src/xnnpack/scalar-utils.h",
2248 "src/xnnpack/spmm.h",
2249 "src/xnnpack/unpool.h",
2250 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002251 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002252 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002253 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002254 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002255 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002256 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002257 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002258]
2259
2260INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002261 "include/xnnpack.h",
2262 "src/xnnpack/allocator.h",
2263 "src/xnnpack/compute.h",
2264 "src/xnnpack/im2col.h",
2265 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002266 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002267 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268 "src/xnnpack/operator.h",
2269 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002270 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002271 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002272 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002273 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002274]
2275
Marat Dukhan1b354632020-03-23 12:50:22 -07002276ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002277 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002278]
2279
Marat Dukhan1b354632020-03-23 12:50:22 -07002280MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002281 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002282 "include/xnnpack.h",
2283]
2284
Marat Dukhan1b354632020-03-23 12:50:22 -07002285MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002286 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002287 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002288 "src/xnnpack/requantization.h",
2289 "include/xnnpack.h",
2290]
2291
2292OPERATOR_TEST_PARAMS_HDRS = [
2293 "src/xnnpack/params.h",
2294 "src/xnnpack/common.h",
2295]
2296
2297WEIGHTS_PACK_HDRS = [
2298 "src/xnnpack/pack.h",
2299 "src/xnnpack/operator.h",
2300 "src/xnnpack/compute.h",
2301]
2302
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002303LOGGING_COPTS = select({
2304 # No logging in optimized mode
2305 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2306 # Full logging in debug mode
2307 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2308 # Error-only logging in default (fastbuild) mode
2309 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2310})
2311
Marat Dukhan3b59de22020-06-03 20:15:19 -07002312LOGGING_SRCS = select({
2313 # No logging in optimized mode
2314 ":optimized_build": [],
2315 "//conditions:default": [
2316 "src/operator-strings.c",
2317 "src/subgraph-strings.c",
2318 ],
2319})
2320
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002321LOGGING_HDRS = [
2322 "src/xnnpack/log.h",
2323]
2324
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002326 name = "tables",
2327 srcs = TABLE_SRCS,
2328 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002329 gcc_copts = xnnpack_gcc_std_copts(),
2330 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002331)
2332
2333xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002334 name = "scalar_ukernels",
2335 srcs = SCALAR_UKERNELS,
2336 hdrs = INTERNAL_HDRS,
2337 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002338 gcc_copts = xnnpack_gcc_std_copts(),
2339 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002341 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002342 "@FP16",
2343 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002344 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002345 ],
2346)
2347
2348xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002349 name = "scalar_ukernels_test_mode",
2350 srcs = SCALAR_UKERNELS,
2351 hdrs = INTERNAL_HDRS,
2352 aarch32_copts = ["-marm"],
2353 copts = [
2354 "-UNDEBUG",
2355 "-DXNN_TEST_MODE=1",
2356 ],
2357 gcc_copts = xnnpack_gcc_std_copts(),
2358 msvc_copts = xnnpack_msvc_std_copts(),
2359 deps = [
2360 ":tables",
2361 "@FP16",
2362 "@FXdiv",
2363 "@pthreadpool",
2364 ],
2365)
2366
2367xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002368 name = "wasm_ukernels",
2369 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002370 gcc_copts = xnnpack_gcc_std_copts(),
2371 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002372 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002373 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002374 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002375 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002376 "@FP16",
2377 "@FXdiv",
2378 "@pthreadpool",
2379 ],
2380)
2381
2382xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002383 name = "wasm_ukernels_test_mode",
2384 hdrs = INTERNAL_HDRS,
2385 copts = [
2386 "-UNDEBUG",
2387 "-DXNN_TEST_MODE=1",
2388 ],
2389 gcc_copts = xnnpack_gcc_std_copts(),
2390 msvc_copts = xnnpack_msvc_std_copts(),
2391 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002392 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002393 deps = [
2394 ":tables",
2395 "@FP16",
2396 "@FXdiv",
2397 "@pthreadpool",
2398 ],
2399)
2400
2401xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002402 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002403 hdrs = INTERNAL_HDRS,
2404 aarch32_copts = [
2405 "-marm",
2406 "-mfpu=neon",
2407 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002408 gcc_copts = xnnpack_gcc_std_copts(),
2409 gcc_x86_copts = ["-msse2"],
2410 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002411 optimized_copts = [
2412 "-O3",
2413 "-ffast-math",
2414 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002415 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002416 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002417 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002418 "@FP16",
2419 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002420 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002421 ],
2422)
2423
2424xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002425 name = "psimd_fastmath_ukernels_test_mode",
2426 hdrs = INTERNAL_HDRS,
2427 aarch32_copts = [
2428 "-marm",
2429 "-mfpu=neon",
2430 ],
2431 copts = [
2432 "-UNDEBUG",
2433 "-DXNN_TEST_MODE=1",
2434 ],
2435 gcc_copts = xnnpack_gcc_std_copts(),
2436 gcc_x86_copts = ["-msse2"],
2437 msvc_copts = xnnpack_msvc_std_copts(),
2438 optimized_copts = [
2439 "-O3",
2440 "-ffast-math",
2441 ],
2442 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2443 deps = [
2444 ":tables",
2445 "@FP16",
2446 "@psimd",
2447 "@pthreadpool",
2448 ],
2449)
2450
2451xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002452 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002453 hdrs = INTERNAL_HDRS,
2454 aarch32_copts = [
2455 "-marm",
2456 "-mfpu=neon",
2457 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002458 gcc_copts = xnnpack_gcc_std_copts(),
2459 gcc_x86_copts = ["-msse2"],
2460 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002461 optimized_copts = [
2462 "-O3",
2463 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002464 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002465 deps = [
2466 ":tables",
2467 "@FP16",
2468 "@psimd",
2469 "@pthreadpool",
2470 ],
2471)
2472
2473xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002474 name = "psimd_accmath_ukernels_test_mode",
2475 hdrs = INTERNAL_HDRS,
2476 aarch32_copts = [
2477 "-marm",
2478 "-mfpu=neon",
2479 ],
2480 copts = [
2481 "-UNDEBUG",
2482 "-DXNN_TEST_MODE=1",
2483 ],
2484 gcc_copts = xnnpack_gcc_std_copts(),
2485 gcc_x86_copts = ["-msse2"],
2486 msvc_copts = xnnpack_msvc_std_copts(),
2487 optimized_copts = [
2488 "-O3",
2489 ],
2490 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2491 deps = [
2492 ":tables",
2493 "@FP16",
2494 "@psimd",
2495 "@pthreadpool",
2496 ],
2497)
2498
2499xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002500 name = "neon_ukernels",
2501 hdrs = INTERNAL_HDRS,
2502 aarch32_copts = [
2503 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002504 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002505 "-mfpu=neon",
2506 ],
2507 aarch32_srcs = NEON_UKERNELS,
2508 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002509 gcc_copts = xnnpack_gcc_std_copts(),
2510 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002511 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002512 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002513 "@FP16",
2514 "@pthreadpool",
2515 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002516)
2517
2518xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002519 name = "neon_ukernels_test_mode",
2520 hdrs = INTERNAL_HDRS,
2521 aarch32_copts = [
2522 "-marm",
2523 "-march=armv7-a",
2524 "-mfpu=neon",
2525 ],
2526 aarch32_srcs = NEON_UKERNELS,
2527 aarch64_srcs = NEON_UKERNELS,
2528 copts = [
2529 "-UNDEBUG",
2530 "-DXNN_TEST_MODE=1",
2531 ],
2532 gcc_copts = xnnpack_gcc_std_copts(),
2533 msvc_copts = xnnpack_msvc_std_copts(),
2534 deps = [
2535 ":tables",
2536 "@FP16",
2537 "@pthreadpool",
2538 ],
2539)
2540
2541xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542 name = "neonfma_ukernels",
2543 hdrs = INTERNAL_HDRS,
2544 aarch32_copts = [
2545 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002546 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002547 "-mfpu=neon-vfpv4",
2548 ],
2549 aarch32_srcs = NEONFMA_UKERNELS,
2550 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002551 apple_aarch32_copts = [
2552 "-mcpu=swift",
2553 "-mtune=generic",
2554 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002555 gcc_copts = xnnpack_gcc_std_copts(),
2556 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002557 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002558 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002559 "@FP16",
2560 "@pthreadpool",
2561 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562)
2563
2564xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002565 name = "neonfma_ukernels_test_mode",
2566 hdrs = INTERNAL_HDRS,
2567 aarch32_copts = [
2568 "-marm",
2569 "-march=armv7-a",
2570 "-mfpu=neon-vfpv4",
2571 ],
2572 aarch32_srcs = NEONFMA_UKERNELS,
2573 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002574 apple_aarch32_copts = [
2575 "-mcpu=swift",
2576 "-mtune=generic",
2577 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002578 copts = [
2579 "-UNDEBUG",
2580 "-DXNN_TEST_MODE=1",
2581 ],
2582 gcc_copts = xnnpack_gcc_std_copts(),
2583 msvc_copts = xnnpack_msvc_std_copts(),
2584 deps = [
2585 ":tables",
2586 "@FP16",
2587 "@pthreadpool",
2588 ],
2589)
2590
2591xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002592 name = "neonv8_ukernels",
2593 hdrs = INTERNAL_HDRS,
2594 aarch32_copts = [
2595 "-marm",
2596 "-march=armv8-a",
2597 "-mfpu=neon-fp-armv8",
2598 ],
2599 aarch32_srcs = NEONV8_UKERNELS,
2600 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002601 apple_aarch32_copts = [
2602 "-mcpu=cyclone",
2603 "-mtune=generic",
2604 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002605 gcc_copts = xnnpack_gcc_std_copts(),
2606 msvc_copts = xnnpack_msvc_std_copts(),
2607 deps = [
2608 ":tables",
2609 "@FP16",
2610 "@pthreadpool",
2611 ],
2612)
2613
2614xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002615 name = "neonv8_ukernels_test_mode",
2616 hdrs = INTERNAL_HDRS,
2617 aarch32_copts = [
2618 "-marm",
2619 "-march=armv8-a",
2620 "-mfpu=neon-fp-armv8",
2621 ],
2622 aarch32_srcs = NEONV8_UKERNELS,
2623 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002624 apple_aarch32_copts = [
2625 "-mcpu=cyclone",
2626 "-mtune=generic",
2627 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002628 copts = [
2629 "-UNDEBUG",
2630 "-DXNN_TEST_MODE=1",
2631 ],
2632 gcc_copts = xnnpack_gcc_std_copts(),
2633 msvc_copts = xnnpack_msvc_std_copts(),
2634 deps = [
2635 ":tables",
2636 "@FP16",
2637 "@pthreadpool",
2638 ],
2639)
2640
2641xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002642 name = "neonfp16arith_ukernels",
2643 hdrs = INTERNAL_HDRS,
2644 aarch64_copts = ["-march=armv8.2-a+fp16"],
2645 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002646 gcc_copts = xnnpack_gcc_std_copts(),
2647 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002648 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002649 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002650 "@FP16",
2651 "@pthreadpool",
2652 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002653)
2654
2655xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002656 name = "neonfp16arith_ukernels_test_mode",
2657 hdrs = INTERNAL_HDRS,
2658 aarch64_copts = ["-march=armv8.2-a+fp16"],
2659 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2660 copts = [
2661 "-UNDEBUG",
2662 "-DXNN_TEST_MODE=1",
2663 ],
2664 gcc_copts = xnnpack_gcc_std_copts(),
2665 msvc_copts = xnnpack_msvc_std_copts(),
2666 deps = [
2667 ":tables",
2668 "@FP16",
2669 "@pthreadpool",
2670 ],
2671)
2672
2673xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002674 name = "sse2_ukernels",
2675 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002676 gcc_copts = xnnpack_gcc_std_copts(),
2677 gcc_x86_copts = ["-msse2"],
2678 msvc_copts = xnnpack_msvc_std_copts(),
2679 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002680 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002681 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002682 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002683 "@FP16",
2684 "@pthreadpool",
2685 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686)
2687
2688xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002689 name = "sse2_ukernels_test_mode",
2690 hdrs = INTERNAL_HDRS,
2691 copts = [
2692 "-UNDEBUG",
2693 "-DXNN_TEST_MODE=1",
2694 ],
2695 gcc_copts = xnnpack_gcc_std_copts(),
2696 gcc_x86_copts = ["-msse2"],
2697 msvc_copts = xnnpack_msvc_std_copts(),
2698 msvc_x86_32_copts = ["/arch:SSE2"],
2699 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2700 deps = [
2701 ":tables",
2702 "@FP16",
2703 "@pthreadpool",
2704 ],
2705)
2706
2707xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002708 name = "ssse3_ukernels",
2709 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002710 gcc_copts = xnnpack_gcc_std_copts(),
2711 gcc_x86_copts = ["-mssse3"],
2712 msvc_copts = xnnpack_msvc_std_copts(),
2713 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002714 x86_srcs = SSSE3_UKERNELS,
2715 deps = [
2716 ":tables",
2717 "@FP16",
2718 "@pthreadpool",
2719 ],
2720)
2721
2722xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002723 name = "ssse3_ukernels_test_mode",
2724 hdrs = INTERNAL_HDRS,
2725 copts = [
2726 "-UNDEBUG",
2727 "-DXNN_TEST_MODE=1",
2728 ],
2729 gcc_copts = xnnpack_gcc_std_copts(),
2730 gcc_x86_copts = ["-mssse3"],
2731 msvc_copts = xnnpack_msvc_std_copts(),
2732 msvc_x86_32_copts = ["/arch:SSE2"],
2733 x86_srcs = SSSE3_UKERNELS,
2734 deps = [
2735 ":tables",
2736 "@FP16",
2737 "@pthreadpool",
2738 ],
2739)
2740
2741xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002742 name = "sse41_ukernels",
2743 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002744 gcc_copts = xnnpack_gcc_std_copts(),
2745 gcc_x86_copts = ["-msse4.1"],
2746 msvc_copts = xnnpack_msvc_std_copts(),
2747 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002748 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002749 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002750 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002751 "@FP16",
2752 "@pthreadpool",
2753 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002754)
2755
2756xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002757 name = "sse41_ukernels_test_mode",
2758 hdrs = INTERNAL_HDRS,
2759 copts = [
2760 "-UNDEBUG",
2761 "-DXNN_TEST_MODE=1",
2762 ],
2763 gcc_copts = xnnpack_gcc_std_copts(),
2764 gcc_x86_copts = ["-msse4.1"],
2765 msvc_copts = xnnpack_msvc_std_copts(),
2766 msvc_x86_32_copts = ["/arch:SSE2"],
2767 x86_srcs = SSE41_UKERNELS,
2768 deps = [
2769 ":tables",
2770 "@FP16",
2771 "@pthreadpool",
2772 ],
2773)
2774
2775xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002776 name = "avx_ukernels",
2777 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002778 gcc_copts = xnnpack_gcc_std_copts(),
2779 gcc_x86_copts = ["-mavx"],
2780 msvc_copts = xnnpack_msvc_std_copts(),
2781 msvc_x86_32_copts = ["/arch:AVX"],
2782 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002783 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002784 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002785 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002786 "@FP16",
2787 "@pthreadpool",
2788 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002789)
2790
2791xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002792 name = "avx_ukernels_test_mode",
2793 hdrs = INTERNAL_HDRS,
2794 copts = [
2795 "-UNDEBUG",
2796 "-DXNN_TEST_MODE=1",
2797 ],
2798 gcc_copts = xnnpack_gcc_std_copts(),
2799 gcc_x86_copts = ["-mavx"],
2800 msvc_copts = xnnpack_msvc_std_copts(),
2801 msvc_x86_32_copts = ["/arch:AVX"],
2802 msvc_x86_64_copts = ["/arch:AVX"],
2803 x86_srcs = AVX_UKERNELS,
2804 deps = [
2805 ":tables",
2806 "@FP16",
2807 "@pthreadpool",
2808 ],
2809)
2810
2811xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002812 name = "fma3_ukernels",
2813 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002814 gcc_copts = xnnpack_gcc_std_copts(),
2815 gcc_x86_copts = ["-mfma"],
2816 msvc_copts = xnnpack_msvc_std_copts(),
2817 msvc_x86_32_copts = ["/arch:AVX"],
2818 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002819 x86_srcs = FMA3_UKERNELS,
2820 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002821 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002822 "@FP16",
2823 "@pthreadpool",
2824 ],
2825)
2826
2827xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002828 name = "fma3_ukernels_test_mode",
2829 hdrs = INTERNAL_HDRS,
2830 copts = [
2831 "-UNDEBUG",
2832 "-DXNN_TEST_MODE=1",
2833 ],
2834 gcc_copts = xnnpack_gcc_std_copts(),
2835 gcc_x86_copts = ["-mfma"],
2836 msvc_copts = xnnpack_msvc_std_copts(),
2837 msvc_x86_32_copts = ["/arch:AVX"],
2838 msvc_x86_64_copts = ["/arch:AVX"],
2839 x86_srcs = FMA3_UKERNELS,
2840 deps = [
2841 ":tables",
2842 "@FP16",
2843 "@pthreadpool",
2844 ],
2845)
2846
2847xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002848 name = "avx2_ukernels",
2849 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002850 gcc_copts = xnnpack_gcc_std_copts(),
2851 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002852 "-mfma",
2853 "-mavx2",
2854 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002855 msvc_copts = xnnpack_msvc_std_copts(),
2856 msvc_x86_32_copts = ["/arch:AVX2"],
2857 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002858 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002859 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002860 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002861 "@FP16",
2862 "@pthreadpool",
2863 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002864)
2865
2866xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002867 name = "avx2_ukernels_test_mode",
2868 hdrs = INTERNAL_HDRS,
2869 copts = [
2870 "-UNDEBUG",
2871 "-DXNN_TEST_MODE=1",
2872 ],
2873 gcc_copts = xnnpack_gcc_std_copts(),
2874 gcc_x86_copts = [
2875 "-mfma",
2876 "-mavx2",
2877 ],
2878 msvc_copts = xnnpack_msvc_std_copts(),
2879 msvc_x86_32_copts = ["/arch:AVX2"],
2880 msvc_x86_64_copts = ["/arch:AVX2"],
2881 x86_srcs = AVX2_UKERNELS,
2882 deps = [
2883 ":tables",
2884 "@FP16",
2885 "@pthreadpool",
2886 ],
2887)
2888
2889xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 name = "avx512f_ukernels",
2891 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002892 gcc_copts = xnnpack_gcc_std_copts(),
2893 gcc_x86_copts = ["-mavx512f"],
2894 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2895 msvc_copts = xnnpack_msvc_std_copts(),
2896 msvc_x86_32_copts = ["/arch:AVX512"],
2897 msvc_x86_64_copts = ["/arch:AVX512"],
2898 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002899 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002900 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002901 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002902 "@FP16",
2903 "@pthreadpool",
2904 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002905)
2906
2907xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002908 name = "avx512f_ukernels_test_mode",
2909 hdrs = INTERNAL_HDRS,
2910 copts = [
2911 "-UNDEBUG",
2912 "-DXNN_TEST_MODE=1",
2913 ],
2914 gcc_copts = xnnpack_gcc_std_copts(),
2915 gcc_x86_copts = ["-mavx512f"],
2916 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2917 msvc_copts = xnnpack_msvc_std_copts(),
2918 msvc_x86_32_copts = ["/arch:AVX512"],
2919 msvc_x86_64_copts = ["/arch:AVX512"],
2920 msys_copts = ["-fno-asynchronous-unwind-tables"],
2921 x86_srcs = AVX512F_UKERNELS,
2922 deps = [
2923 ":tables",
2924 "@FP16",
2925 "@pthreadpool",
2926 ],
2927)
2928
2929xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 name = "asm_ukernels",
2931 hdrs = ["src/xnnpack/assembly.h"],
2932 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002933 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 aarch64_srcs = AARCH64_ASM_UKERNELS,
2935)
2936
Marat Dukhan3b59de22020-06-03 20:15:19 -07002937xnnpack_cc_library(
2938 name = "logging_utils",
2939 srcs = LOGGING_SRCS,
2940 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2941 copts = LOGGING_COPTS + [
2942 "-Isrc",
2943 "-Iinclude",
2944 ] + select({
2945 ":debug_build": [],
2946 "//conditions:default": xnnpack_min_size_copts(),
2947 }),
2948 gcc_copts = xnnpack_gcc_std_copts(),
2949 msvc_copts = xnnpack_msvc_std_copts(),
2950 visibility = xnnpack_visibility(),
2951 deps = [
2952 "@FP16",
2953 "@clog",
2954 "@pthreadpool",
2955 ],
2956)
2957
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958xnnpack_aggregate_library(
2959 name = "ukernels",
2960 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 ":neon_ukernels",
2962 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002963 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 ":asm_ukernels",
2965 ],
2966 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002967 ":neon_ukernels",
2968 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002969 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970 ":neonfp16arith_ukernels",
2971 ":asm_ukernels",
2972 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002973 generic_deps = [
2974 ":scalar_ukernels",
2975 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002976 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002977 ":psimd_fastmath_ukernels",
2978 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002980 wasm_deps = [
2981 ":wasm_ukernels",
2982 ],
2983 wasmsimd_deps = [
2984 ":wasm_ukernels",
2985 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002986 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002988 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002989 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002990 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002991 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002992 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 ":avx512f_ukernels",
2994 ],
2995)
2996
Marat Dukhan33fcf782020-05-24 14:27:15 -07002997xnnpack_aggregate_library(
2998 name = "ukernels_test_mode",
2999 aarch32_deps = [
3000 ":neon_ukernels_test_mode",
3001 ":neonfma_ukernels_test_mode",
3002 ":neonv8_ukernels_test_mode",
3003 ":asm_ukernels",
3004 ],
3005 aarch64_deps = [
3006 ":neon_ukernels_test_mode",
3007 ":neonfma_ukernels_test_mode",
3008 ":neonv8_ukernels_test_mode",
3009 ":neonfp16arith_ukernels_test_mode",
3010 ":asm_ukernels",
3011 ],
3012 generic_deps = [
3013 ":scalar_ukernels_test_mode",
3014 ],
3015 psimd_deps = [
3016 ":psimd_fastmath_ukernels_test_mode",
3017 ":psimd_accmath_ukernels_test_mode",
3018 ],
3019 wasm_deps = [
3020 ":wasm_ukernels_test_mode",
3021 ],
3022 wasmsimd_deps = [
3023 ":wasm_ukernels_test_mode",
3024 ],
3025 x86_deps = [
3026 ":sse2_ukernels_test_mode",
3027 ":ssse3_ukernels_test_mode",
3028 ":sse41_ukernels_test_mode",
3029 ":avx_ukernels_test_mode",
3030 ":fma3_ukernels_test_mode",
3031 ":avx2_ukernels_test_mode",
3032 ":avx512f_ukernels_test_mode",
3033 ],
3034)
3035
Marat Dukhan86df3842020-06-17 14:03:31 -07003036# This target must be compatible with Android NDK r17 and compile with Clang 6.
3037# It is similar to :ukernels target, but excludes microkernels which require
3038# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003039xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003040 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003041 aarch32_deps = [
3042 ":neon_ukernels",
3043 ":neonfma_ukernels",
3044 ":neonv8_ukernels",
3045 ":asm_ukernels",
3046 ],
3047 aarch64_deps = [
3048 ":neon_ukernels",
3049 ":neonfma_ukernels",
3050 ":neonv8_ukernels",
3051 ":asm_ukernels",
3052 ],
3053 generic_deps = [
3054 ":scalar_ukernels",
3055 ],
3056 psimd_deps = [
3057 ":psimd_fastmath_ukernels",
3058 ":psimd_accmath_ukernels",
3059 ],
3060 wasm_deps = [
3061 ":wasm_ukernels",
3062 ],
3063 wasmsimd_deps = [
3064 ":wasm_ukernels",
3065 ],
3066 x86_deps = [
3067 ":sse2_ukernels",
3068 ":ssse3_ukernels",
3069 ":sse41_ukernels",
3070 ":avx_ukernels",
3071 ":fma3_ukernels",
3072 ":avx2_ukernels",
3073 ":avx512f_ukernels",
3074 ],
3075)
3076
Marat Dukhan08c4a432019-10-03 09:29:21 -07003077xnnpack_cc_library(
3078 name = "im2col",
3079 srcs = ["src/im2col.c"],
3080 hdrs = [
3081 "src/xnnpack/common.h",
3082 "src/xnnpack/im2col.h",
3083 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003084 gcc_copts = xnnpack_gcc_std_copts(),
3085 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086)
3087
3088xnnpack_cc_library(
3089 name = "indirection",
3090 srcs = ["src/indirection.c"],
3091 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003092 gcc_copts = xnnpack_gcc_std_copts(),
3093 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003094 deps = [
3095 "@FP16",
3096 "@FXdiv",
3097 "@pthreadpool",
3098 ],
3099)
3100
3101xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003102 name = "indirection_test_mode",
3103 srcs = ["src/indirection.c"],
3104 hdrs = INTERNAL_HDRS,
3105 copts = [
3106 "-UNDEBUG",
3107 "-DXNN_TEST_MODE=1",
3108 ],
3109 gcc_copts = xnnpack_gcc_std_copts(),
3110 msvc_copts = xnnpack_msvc_std_copts(),
3111 deps = [
3112 "@FP16",
3113 "@FXdiv",
3114 "@pthreadpool",
3115 ],
3116)
3117
3118xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003119 name = "packing",
3120 srcs = ["src/packing.c"],
3121 hdrs = INTERNAL_HDRS,
3122 gcc_copts = xnnpack_gcc_std_copts(),
3123 msvc_copts = xnnpack_msvc_std_copts(),
3124 deps = [
3125 "@FP16",
3126 "@FXdiv",
3127 "@pthreadpool",
3128 ],
3129)
3130
3131xnnpack_cc_library(
3132 name = "packing_test_mode",
3133 srcs = ["src/packing.c"],
3134 hdrs = INTERNAL_HDRS,
3135 copts = [
3136 "-UNDEBUG",
3137 "-DXNN_TEST_MODE=1",
3138 ],
3139 gcc_copts = xnnpack_gcc_std_copts(),
3140 msvc_copts = xnnpack_msvc_std_copts(),
3141 deps = [
3142 "@FP16",
3143 "@FXdiv",
3144 "@pthreadpool",
3145 ],
3146)
3147
3148xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003149 name = "operator_run",
3150 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003151 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003152 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003153 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3154 "//conditions:default": [],
3155 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003156 gcc_copts = xnnpack_gcc_std_copts(),
3157 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003158 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003159 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003160 "@FP16",
3161 "@FXdiv",
3162 "@clog",
3163 "@pthreadpool",
3164 ],
3165)
3166
Chao Mei6ddfc602020-05-13 22:29:36 -07003167xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003168 name = "operator_run_test_mode",
3169 srcs = ["src/operator-run.c"],
3170 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3171 copts = LOGGING_COPTS + [
3172 "-UNDEBUG",
3173 "-DXNN_TEST_MODE=1",
3174 ] + select({
3175 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3176 "//conditions:default": [],
3177 }),
3178 gcc_copts = xnnpack_gcc_std_copts(),
3179 msvc_copts = xnnpack_msvc_std_copts(),
3180 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003181 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003182 "@FP16",
3183 "@FXdiv",
3184 "@clog",
3185 "@pthreadpool",
3186 ],
3187)
3188
3189xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003190 name = "memory_planner",
3191 srcs = ["src/memory-planner.c"],
3192 hdrs = INTERNAL_HDRS,
3193 defines = select({
3194 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3195 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3196 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3197 }),
3198 gcc_copts = xnnpack_gcc_std_copts(),
3199 msvc_copts = xnnpack_msvc_std_copts(),
3200 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003201 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003202 "@pthreadpool",
3203 ],
3204)
3205
Marat Dukhan33fcf782020-05-24 14:27:15 -07003206xnnpack_cc_library(
3207 name = "memory_planner_test_mode",
3208 srcs = ["src/memory-planner.c"],
3209 hdrs = INTERNAL_HDRS,
3210 copts = [
3211 "-UNDEBUG",
3212 "-DXNN_TEST_MODE=1",
3213 ],
3214 defines = select({
3215 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3216 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3217 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3218 }),
3219 gcc_copts = xnnpack_gcc_std_copts(),
3220 msvc_copts = xnnpack_msvc_std_copts(),
3221 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003222 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003223 "@pthreadpool",
3224 ],
3225)
3226
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227cc_library(
3228 name = "enable_assembly",
3229 defines = select({
3230 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3231 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003232 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003233 }),
3234)
3235
Marat Dukhan9de90e02020-06-18 16:04:12 -07003236cc_library(
3237 name = "enable_sparse",
3238 defines = select({
3239 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3240 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003241 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003242 }),
3243)
3244
Marat Dukhancf056b22019-10-07 10:26:29 -07003245xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003246 name = "operators",
3247 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003248 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003249 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003250 ],
3251 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003252 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 "-Isrc",
3254 "-Iinclude",
3255 ] + select({
3256 ":debug_build": [],
3257 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003258 }) + select({
3259 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3260 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003261 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003262 gcc_copts = xnnpack_gcc_std_copts(),
3263 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003264 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003266 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003267 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268 "@FP16",
3269 "@FXdiv",
3270 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003272 ],
3273)
3274
Marat Dukhan10a38082020-04-17 03:58:35 -07003275xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003276 name = "operators_test_mode",
3277 srcs = OPERATOR_SRCS + [
3278 "src/memory.c",
3279 "src/operator-delete.c",
3280 ],
3281 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3282 copts = LOGGING_COPTS + [
3283 "-Isrc",
3284 "-Iinclude",
3285 "-UNDEBUG",
3286 "-DXNN_TEST_MODE=1",
3287 ] + select({
3288 ":debug_build": [],
3289 "//conditions:default": xnnpack_min_size_copts(),
3290 }) + select({
3291 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3292 "//conditions:default": [],
3293 }),
3294 gcc_copts = xnnpack_gcc_std_copts(),
3295 msvc_copts = xnnpack_msvc_std_copts(),
3296 deps = [
3297 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003298 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003299 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003300 "@FP16",
3301 "@FXdiv",
3302 "@clog",
3303 "@pthreadpool",
3304 ],
3305)
3306
3307xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003308 name = "XNNPACK",
3309 srcs = [
3310 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003311 "src/runtime.c",
3312 "src/subgraph.c",
3313 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003314 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003315 hdrs = ["include/xnnpack.h"],
3316 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003317 "-Isrc",
3318 "-Iinclude",
3319 ] + select({
3320 ":debug_build": [],
3321 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003322 }) + select({
3323 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3324 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003325 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003326 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003327 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003328 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003329 visibility = xnnpack_visibility(),
3330 deps = [
3331 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003332 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003333 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003334 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003335 ":operator_run",
3336 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003337 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003338 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003339 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003340 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003341 ] + select({
3342 ":emscripten": [],
3343 "//conditions:default": ["@cpuinfo"],
3344 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003345)
3346
Marat Dukhan10a38082020-04-17 03:58:35 -07003347xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003348 name = "XNNPACK_test_mode",
3349 srcs = [
3350 "src/init.c",
3351 "src/runtime.c",
3352 "src/subgraph.c",
3353 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003354 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003355 hdrs = ["include/xnnpack.h"],
3356 copts = LOGGING_COPTS + [
3357 "-Isrc",
3358 "-Iinclude",
3359 "-UNDEBUG",
3360 "-DXNN_TEST_MODE=1",
3361 ] + select({
3362 ":debug_build": [],
3363 "//conditions:default": xnnpack_min_size_copts(),
3364 }) + select({
3365 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3366 "//conditions:default": [],
3367 }),
3368 gcc_copts = xnnpack_gcc_std_copts(),
3369 includes = ["include"],
3370 msvc_copts = xnnpack_msvc_std_copts(),
3371 visibility = xnnpack_visibility(),
3372 deps = [
3373 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003374 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003375 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003376 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003377 ":operator_run_test_mode",
3378 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003379 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003380 "@clog",
3381 "@FP16",
3382 "@pthreadpool",
3383 ] + select({
3384 ":emscripten": [],
3385 "//conditions:default": ["@cpuinfo"],
3386 }),
3387)
3388
3389xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003390 name = "xnnpack_f32",
3391 srcs = [
3392 "src/init.c",
3393 "src/runtime.c",
3394 "src/subgraph.c",
3395 "src/tensor.c",
3396 ] + SUBGRAPH_SRCS,
3397 hdrs = ["include/xnnpack.h"],
3398 copts = LOGGING_COPTS + [
3399 "-Isrc",
3400 "-Iinclude",
3401 ] + select({
3402 ":debug_build": [],
3403 "//conditions:default": xnnpack_min_size_copts(),
3404 }) + select({
3405 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3406 "//conditions:default": [],
3407 }),
3408 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003409 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003410 "XNN_NO_U8_OPERATORS",
3411 "XNN_NO_X8_OPERATORS",
3412 "XNN_NO_F16_OPERATORS",
3413 "XNN_NO_X16_OPERATORS",
3414 ],
3415 gcc_copts = xnnpack_gcc_std_copts(),
3416 includes = ["include"],
3417 msvc_copts = xnnpack_msvc_std_copts(),
3418 visibility = xnnpack_visibility(),
3419 deps = [
3420 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003421 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003422 ":logging_utils",
3423 ":memory_planner",
3424 ":operator_run",
3425 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003426 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003427 "@clog",
3428 "@FP16",
3429 "@pthreadpool",
3430 ] + select({
3431 ":emscripten": [],
3432 "//conditions:default": ["@cpuinfo"],
3433 }),
3434)
3435
3436xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003437 name = "xnnpack_operators_nhwc_f32",
3438 srcs = [
3439 "src/init.c",
3440 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003441 hdrs = ["include/xnnpack.h"],
3442 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003443 "-Isrc",
3444 "-Iinclude",
3445 ] + select({
3446 ":debug_build": [],
3447 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003448 }) + select({
3449 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3450 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003451 }),
3452 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003453 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003454 "XNN_NO_U8_OPERATORS",
3455 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003456 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003457 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003458 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003459 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003460 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003461 visibility = xnnpack_visibility(),
3462 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003463 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003464 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003465 ":operator_run",
3466 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003467 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003468 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003469 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003470 ] + select({
3471 ":emscripten": [],
3472 "//conditions:default": ["@cpuinfo"],
3473 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003474)
3475
Marat Dukhancf056b22019-10-07 10:26:29 -07003476xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003477 name = "bench_utils",
3478 srcs = ["bench/utils.cc"],
3479 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003480 deps = [
3481 "@com_google_benchmark//:benchmark",
3482 "@cpuinfo",
3483 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484)
3485
Frank Barchard7e955972019-10-11 10:34:25 -07003486######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003487
3488xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003489 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490 srcs = [
3491 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003492 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493 "src/xnnpack/AlignedAllocator.h",
3494 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003495 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003496 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003497)
3498
3499xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003500 name = "f16_igemm_bench",
3501 srcs = [
3502 "bench/f16-igemm.cc",
3503 "bench/conv.h",
3504 "bench/google/conv.h",
3505 "src/xnnpack/AlignedAllocator.h",
3506 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003507 deps = MICROKERNEL_BENCHMARK_DEPS + [
3508 ":indirection",
3509 ":packing",
3510 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003511)
3512
3513xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003514 name = "f16_gemm_bench",
3515 srcs = [
3516 "bench/f16-gemm.cc",
3517 "bench/gemm.h",
3518 "src/xnnpack/AlignedAllocator.h",
3519 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003520 deps = MICROKERNEL_BENCHMARK_DEPS + [
3521 ":packing",
3522 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003523)
3524
3525xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003526 name = "f16_spmm_bench",
3527 srcs = [
3528 "bench/f16-spmm.cc",
3529 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003530 "src/xnnpack/AlignedAllocator.h",
3531 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003532 deps = MICROKERNEL_BENCHMARK_DEPS,
3533)
3534
3535xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003536 name = "f32_igemm_bench",
3537 srcs = [
3538 "bench/f32-igemm.cc",
3539 "bench/conv.h",
3540 "src/xnnpack/AlignedAllocator.h",
3541 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003542 deps = MICROKERNEL_BENCHMARK_DEPS + [
3543 ":indirection",
3544 ":packing",
3545 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003546)
3547
3548xnnpack_benchmark(
3549 name = "f32_conv_hwc_bench",
3550 srcs = [
3551 "bench/f32-conv-hwc.cc",
3552 "bench/dconv.h",
3553 "src/xnnpack/AlignedAllocator.h",
3554 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003555 deps = MICROKERNEL_BENCHMARK_DEPS + [
3556 ":packing",
3557 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003558)
3559
3560xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003561 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003562 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003563 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003564 "bench/dconv.h",
3565 "src/xnnpack/AlignedAllocator.h",
3566 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003567 deps = MICROKERNEL_BENCHMARK_DEPS + [
3568 ":packing",
3569 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003570)
3571
3572xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003573 name = "f16_dwconv_bench",
3574 srcs = [
3575 "bench/f16-dwconv.cc",
3576 "bench/dwconv.h",
3577 "bench/google/dwconv.h",
3578 "src/xnnpack/AlignedAllocator.h",
3579 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003580 deps = MICROKERNEL_BENCHMARK_DEPS + [
3581 ":indirection",
3582 ":packing",
3583 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003584)
3585
3586xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003587 name = "f32_dwconv_bench",
3588 srcs = [
3589 "bench/f32-dwconv.cc",
3590 "bench/dwconv.h",
3591 "src/xnnpack/AlignedAllocator.h",
3592 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003593 deps = MICROKERNEL_BENCHMARK_DEPS + [
3594 ":indirection",
3595 ":packing",
3596 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003597)
3598
3599xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003600 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003602 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603 "bench/dwconv.h",
3604 "src/xnnpack/AlignedAllocator.h",
3605 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003606 deps = MICROKERNEL_BENCHMARK_DEPS + [
3607 ":indirection",
3608 ":packing",
3609 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610)
3611
3612xnnpack_benchmark(
3613 name = "f32_gemm_bench",
3614 srcs = [
3615 "bench/f32-gemm.cc",
3616 "bench/gemm.h",
3617 "src/xnnpack/AlignedAllocator.h",
3618 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003619 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003620 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003621)
3622
3623xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003624 name = "f32_hswish_bench",
3625 srcs = [
3626 "bench/f32-hswish.cc",
3627 "src/xnnpack/AlignedAllocator.h",
3628 ] + MICROKERNEL_BENCHMARK_HDRS,
3629 deps = MICROKERNEL_BENCHMARK_DEPS,
3630)
3631
3632xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003633 name = "f32_raddexpminusmax_bench",
3634 srcs = [
3635 "bench/f32-raddexpminusmax.cc",
3636 "src/xnnpack/AlignedAllocator.h",
3637 ] + MICROKERNEL_BENCHMARK_HDRS,
3638 deps = MICROKERNEL_BENCHMARK_DEPS,
3639)
3640
3641xnnpack_benchmark(
3642 name = "f32_raddextexp_bench",
3643 srcs = [
3644 "bench/f32-raddextexp.cc",
3645 "src/xnnpack/AlignedAllocator.h",
3646 ] + MICROKERNEL_BENCHMARK_HDRS,
3647 deps = MICROKERNEL_BENCHMARK_DEPS,
3648)
3649
3650xnnpack_benchmark(
3651 name = "f32_raddstoreexpminusmax_bench",
3652 srcs = [
3653 "bench/f32-raddstoreexpminusmax.cc",
3654 "src/xnnpack/AlignedAllocator.h",
3655 ] + MICROKERNEL_BENCHMARK_HDRS,
3656 deps = MICROKERNEL_BENCHMARK_DEPS,
3657)
3658
3659xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660 name = "f32_rmax_bench",
3661 srcs = [
3662 "bench/f32-rmax.cc",
3663 "src/xnnpack/AlignedAllocator.h",
3664 ] + MICROKERNEL_BENCHMARK_HDRS,
3665 deps = MICROKERNEL_BENCHMARK_DEPS,
3666)
3667
3668xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003669 name = "f32_sigmoid_bench",
3670 srcs = [
3671 "bench/f32-sigmoid.cc",
3672 "src/xnnpack/AlignedAllocator.h",
3673 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003674 deps = MICROKERNEL_BENCHMARK_DEPS,
3675)
3676
3677xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003678 name = "f32_spmm_bench",
3679 srcs = [
3680 "bench/f32-spmm.cc",
3681 "bench/gemm.h",
3682 "src/xnnpack/AlignedAllocator.h",
3683 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003684 deps = MICROKERNEL_BENCHMARK_DEPS,
3685)
3686
3687xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003688 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003689 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003690 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003691 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003692 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003693 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003694)
3695
3696xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003697 name = "f32_vscaleexpminusmax_bench",
3698 srcs = [
3699 "bench/f32-vscaleexpminusmax.cc",
3700 "src/xnnpack/AlignedAllocator.h",
3701 ] + MICROKERNEL_BENCHMARK_HDRS,
3702 deps = MICROKERNEL_BENCHMARK_DEPS,
3703)
3704
3705xnnpack_benchmark(
3706 name = "f32_vscaleextexp_bench",
3707 srcs = [
3708 "bench/f32-vscaleextexp.cc",
3709 "src/xnnpack/AlignedAllocator.h",
3710 ] + MICROKERNEL_BENCHMARK_HDRS,
3711 deps = MICROKERNEL_BENCHMARK_DEPS,
3712)
3713
3714xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003715 name = "f32_vsqrt_bench",
3716 srcs = [
3717 "bench/f32-vsqrt.cc",
3718 "src/xnnpack/AlignedAllocator.h",
3719 ] + MICROKERNEL_BENCHMARK_HDRS,
3720 deps = MICROKERNEL_BENCHMARK_DEPS,
3721)
3722
3723xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003724 name = "f32_im2col_gemm_bench",
3725 srcs = [
3726 "bench/f32-im2col-gemm.cc",
3727 "bench/conv.h",
3728 "src/xnnpack/AlignedAllocator.h",
3729 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003730 deps = MICROKERNEL_BENCHMARK_DEPS + [
3731 ":im2col",
3732 ":packing",
3733 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003734)
3735
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003736xnnpack_benchmark(
3737 name = "requantization_bench",
3738 srcs = [
3739 "bench/requantization.cc",
3740 "src/xnnpack/requantization-stubs.h",
3741 "src/xnnpack/AlignedAllocator.h",
3742 ] + MICROKERNEL_BENCHMARK_HDRS,
3743 deps = MICROKERNEL_BENCHMARK_DEPS,
3744)
3745
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003746xnnpack_benchmark(
3747 name = "rounding_bench",
3748 srcs = [
3749 "bench/rounding.cc",
3750 "src/xnnpack/math-stubs.h",
3751 "src/xnnpack/AlignedAllocator.h",
3752 ] + MICROKERNEL_BENCHMARK_HDRS,
3753 deps = MICROKERNEL_BENCHMARK_DEPS,
3754)
3755
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756########################### Benchmarks for operators ###########################
3757
3758xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759 name = "average_pooling_bench",
3760 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003761 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003762 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003763 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003764)
3765
3766xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003767 name = "bankers_rounding_bench",
3768 srcs = ["bench/bankers-rounding.cc"],
3769 copts = xnnpack_optional_tflite_copts(),
3770 tags = ["nowin32"],
3771 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3772)
3773
3774xnnpack_benchmark(
3775 name = "ceiling_bench",
3776 srcs = ["bench/ceiling.cc"],
3777 copts = xnnpack_optional_tflite_copts(),
3778 tags = ["nowin32"],
3779 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3780)
3781
3782xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003783 name = "channel_shuffle_bench",
3784 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003785 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003786)
3787
3788xnnpack_benchmark(
3789 name = "convolution_bench",
3790 srcs = ["bench/convolution.cc"],
3791 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003792 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003793 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794)
3795
3796xnnpack_benchmark(
3797 name = "deconvolution_bench",
3798 srcs = ["bench/deconvolution.cc"],
3799 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003800 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003801 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802)
3803
3804xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003805 name = "floor_bench",
3806 srcs = ["bench/floor.cc"],
3807 copts = xnnpack_optional_tflite_copts(),
3808 tags = ["nowin32"],
3809 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3810)
3811
3812xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003813 name = "global_average_pooling_bench",
3814 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003815 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816)
3817
3818xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003819 name = "hardswish_bench",
3820 srcs = ["bench/hardswish.cc"],
3821 copts = xnnpack_optional_tflite_copts(),
3822 tags = ["nowin32"],
3823 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3824)
3825
3826xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003827 name = "max_pooling_bench",
3828 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003829 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003830)
3831
3832xnnpack_benchmark(
3833 name = "sigmoid_bench",
3834 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003835 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003836 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003837 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003838)
3839
3840xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003841 name = "prelu_bench",
3842 srcs = ["bench/prelu.cc"],
3843 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003844 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003845 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003846)
3847
3848xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003849 name = "softmax_bench",
3850 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003851 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003852 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003853 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003854)
3855
Marat Dukhan87727142020-06-24 15:24:10 -07003856xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003857 name = "square_root_bench",
3858 srcs = ["bench/square-root.cc"],
3859 copts = xnnpack_optional_tflite_copts(),
3860 tags = ["nowin32"],
3861 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3862)
3863
3864xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003865 name = "truncation_bench",
3866 srcs = ["bench/truncation.cc"],
3867 deps = OPERATOR_BENCHMARK_DEPS,
3868)
3869
Marat Dukhanc068bb62019-10-04 13:24:39 -07003870############################# End-to-end benchmarks ############################
3871
3872cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003873 name = "fp32_mobilenet_v1",
3874 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003875 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003876 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003877 linkstatic = True,
3878 deps = [
3879 ":XNNPACK",
3880 "@pthreadpool",
3881 ],
3882)
3883
3884cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003885 name = "fp16_mobilenet_v1",
3886 srcs = ["models/fp16-mobilenet-v1.cc"],
3887 hdrs = ["models/models.h"],
3888 copts = xnnpack_std_cxxopts(),
3889 linkstatic = True,
3890 deps = [
3891 ":XNNPACK",
3892 "@FP16",
3893 "@pthreadpool",
3894 ],
3895)
3896
3897cc_library(
3898 name = "fp32_mobilenet_v2",
3899 srcs = ["models/fp32-mobilenet-v2.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
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003909cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003910 name = "fp16_mobilenet_v2",
3911 srcs = ["models/fp16-mobilenet-v2.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_v3_large",
3924 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003925 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003926 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003927 linkstatic = True,
3928 deps = [
3929 ":XNNPACK",
3930 "@pthreadpool",
3931 ],
3932)
3933
3934cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003935 name = "fp32_mobilenet_v3_small",
3936 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003937 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003938 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003939 linkstatic = True,
3940 deps = [
3941 ":XNNPACK",
3942 "@pthreadpool",
3943 ],
3944)
3945
Marat Dukhanc068bb62019-10-04 13:24:39 -07003946xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003947 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003948 srcs = [
3949 "bench/f32-dwconv-e2e.cc",
3950 "bench/end2end.h",
3951 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003952 deps = MICROKERNEL_BENCHMARK_DEPS + [
3953 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003954 ":fp32_mobilenet_v1",
3955 ":fp32_mobilenet_v2",
3956 ":fp32_mobilenet_v3_large",
3957 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003958 ],
3959)
3960
3961xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003962 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003963 srcs = [
3964 "bench/f32-gemm-e2e.cc",
3965 "bench/end2end.h",
3966 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003967 deps = MICROKERNEL_BENCHMARK_DEPS + [
3968 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003969 ":fp32_mobilenet_v1",
3970 ":fp32_mobilenet_v2",
3971 ":fp32_mobilenet_v3_large",
3972 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003973 ],
3974)
3975
3976xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003977 name = "end2end_bench",
3978 srcs = ["bench/end2end.cc"],
3979 deps = [
3980 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003981 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003982 ":fp16_mobilenet_v1",
3983 ":fp16_mobilenet_v2",
3984 ":fp32_mobilenet_v1",
3985 ":fp32_mobilenet_v2",
3986 ":fp32_mobilenet_v3_large",
3987 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003988 "@pthreadpool",
3989 ],
3990)
3991
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003992#################### Accuracy evaluation for math functions ####################
3993
3994xnnpack_benchmark(
3995 name = "f32_exp_eval",
3996 srcs = [
3997 "eval/f32-exp.cc",
3998 "src/xnnpack/AlignedAllocator.h",
3999 ] + ACCURACY_EVAL_HDRS,
4000 deps = ACCURACY_EVAL_DEPS,
4001)
4002
Marat Dukhan515c9772019-10-17 18:07:57 -07004003xnnpack_benchmark(
4004 name = "f32_expminus_eval",
4005 srcs = [
4006 "eval/f32-expminus.cc",
4007 "src/xnnpack/AlignedAllocator.h",
4008 ] + ACCURACY_EVAL_HDRS,
4009 deps = ACCURACY_EVAL_DEPS,
4010)
4011
Marat Dukhan98ba4412019-10-23 02:14:28 -07004012xnnpack_benchmark(
4013 name = "f32_extexp_eval",
4014 srcs = [
4015 "eval/f32-extexp.cc",
4016 "src/xnnpack/AlignedAllocator.h",
4017 ] + ACCURACY_EVAL_HDRS,
4018 deps = ACCURACY_EVAL_DEPS,
4019)
4020
Marat Dukhan8853b822020-05-07 12:19:01 -07004021xnnpack_unit_test(
4022 name = "f32_roundne_eval",
4023 srcs = [
4024 "eval/f32-roundne.cc",
4025 "src/xnnpack/AlignedAllocator.h",
4026 "src/xnnpack/math-stubs.h",
4027 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004028 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004029 deps = MICROKERNEL_TEST_DEPS,
4030)
4031
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004032xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004033 name = "f32_roundd_eval",
4034 srcs = [
4035 "eval/f32-roundd.cc",
4036 "src/xnnpack/AlignedAllocator.h",
4037 "src/xnnpack/math-stubs.h",
4038 ] + MICROKERNEL_TEST_HDRS,
4039 automatic = False,
4040 deps = MICROKERNEL_TEST_DEPS,
4041)
4042
4043xnnpack_unit_test(
4044 name = "f32_roundu_eval",
4045 srcs = [
4046 "eval/f32-roundu.cc",
4047 "src/xnnpack/AlignedAllocator.h",
4048 "src/xnnpack/math-stubs.h",
4049 ] + MICROKERNEL_TEST_HDRS,
4050 automatic = False,
4051 deps = MICROKERNEL_TEST_DEPS,
4052)
4053
4054xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004055 name = "f32_roundz_eval",
4056 srcs = [
4057 "eval/f32-roundz.cc",
4058 "src/xnnpack/AlignedAllocator.h",
4059 "src/xnnpack/math-stubs.h",
4060 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004061 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004062 deps = MICROKERNEL_TEST_DEPS,
4063)
4064
Marat Dukhan346a9e52019-11-15 09:06:30 -08004065xnnpack_benchmark(
4066 name = "f32_sigmoid_eval",
4067 srcs = [
4068 "eval/f32-sigmoid.cc",
4069 "src/xnnpack/AlignedAllocator.h",
4070 ] + ACCURACY_EVAL_HDRS,
4071 deps = ACCURACY_EVAL_DEPS,
4072)
4073
Marat Dukhan84000762020-06-29 18:38:43 -07004074xnnpack_benchmark(
4075 name = "f32_sqrt_eval",
4076 srcs = [
4077 "eval/f32-sqrt.cc",
4078 "src/xnnpack/AlignedAllocator.h",
4079 ] + ACCURACY_EVAL_HDRS,
4080 deps = ACCURACY_EVAL_DEPS + [
4081 ":bench_utils",
4082 ],
4083)
4084
Marat Dukhan08c4a432019-10-03 09:29:21 -07004085######################### Unit tests for micro-kernels #########################
4086
4087xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004088 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004089 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004090 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004091 "test/gemm-microkernel-tester.h",
4092 "src/xnnpack/AlignedAllocator.h",
4093 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004094 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004095)
4096
4097xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004098 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004099 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004100 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004101 "test/spmm-microkernel-tester.h",
4102 "src/xnnpack/AlignedAllocator.h",
4103 ] + MICROKERNEL_TEST_HDRS,
4104 deps = MICROKERNEL_TEST_DEPS,
4105)
4106
4107xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004108 name = "f32_argmaxpool_test",
4109 srcs = [
4110 "test/f32-argmaxpool.cc",
4111 "test/argmaxpool-microkernel-tester.h",
4112 "src/xnnpack/AlignedAllocator.h",
4113 ] + MICROKERNEL_TEST_HDRS,
4114 deps = MICROKERNEL_TEST_DEPS,
4115)
4116
4117xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004118 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004119 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004120 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004121 "test/avgpool-microkernel-tester.h",
4122 "src/xnnpack/AlignedAllocator.h",
4123 ] + MICROKERNEL_TEST_HDRS,
4124 deps = MICROKERNEL_TEST_DEPS,
4125)
4126
4127xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004128 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004129 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004130 "test/f32-ibilinear.cc",
4131 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004132 "src/xnnpack/AlignedAllocator.h",
4133 ] + MICROKERNEL_TEST_HDRS,
4134 deps = MICROKERNEL_TEST_DEPS,
4135)
4136
4137xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004138 name = "f16_clamp_test",
4139 srcs = [
4140 "test/f16-clamp.cc",
4141 "test/clamp-microkernel-tester.h",
4142 ] + MICROKERNEL_TEST_HDRS,
4143 deps = MICROKERNEL_TEST_DEPS,
4144)
4145
4146xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004147 name = "f32_clamp_test",
4148 srcs = [
4149 "test/f32-clamp.cc",
4150 "test/clamp-microkernel-tester.h",
4151 ] + MICROKERNEL_TEST_HDRS,
4152 deps = MICROKERNEL_TEST_DEPS,
4153)
4154
4155xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004156 name = "f32_igemm_test",
4157 srcs = [
4158 "test/f32-igemm.cc",
4159 "test/gemm-microkernel-tester.h",
4160 "src/xnnpack/AlignedAllocator.h",
4161 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004162 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004163)
4164
4165xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004166 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004167 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004168 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004169 "test/gemm-microkernel-tester.h",
4170 "src/xnnpack/AlignedAllocator.h",
4171 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004172 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004173)
4174
4175xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004176 name = "f32_igemm_minmax_test",
4177 srcs = [
4178 "test/f32-igemm-minmax.cc",
4179 "test/gemm-microkernel-tester.h",
4180 "src/xnnpack/AlignedAllocator.h",
4181 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004182 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004183)
4184
4185xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004186 name = "f16_igemm_minmax_test",
4187 srcs = [
4188 "test/f16-igemm-minmax.cc",
4189 "test/gemm-microkernel-tester.h",
4190 "src/xnnpack/AlignedAllocator.h",
4191 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004192 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004193)
4194
4195xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004196 name = "f32_conv_hwc_test",
4197 srcs = [
4198 "test/f32-conv-hwc.cc",
4199 "test/conv-hwc-microkernel-tester.h",
4200 "src/xnnpack/AlignedAllocator.h",
4201 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004202 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203)
4204
4205xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004206 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004207 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004208 "test/f32-conv-hwc2chw.cc",
4209 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004210 "src/xnnpack/AlignedAllocator.h",
4211 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004212 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213)
4214
4215xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004216 name = "f32_dwconv_test",
4217 srcs = [
4218 "test/f32-dwconv.cc",
4219 "test/dwconv-microkernel-tester.h",
4220 "src/xnnpack/AlignedAllocator.h",
4221 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004222 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004223)
4224
4225xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004226 name = "f16_dwconv_minmax_test",
4227 srcs = [
4228 "test/f16-dwconv-minmax.cc",
4229 "test/dwconv-microkernel-tester.h",
4230 "src/xnnpack/AlignedAllocator.h",
4231 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004232 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004233)
4234
4235xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004236 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004238 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004239 "test/dwconv-microkernel-tester.h",
4240 "src/xnnpack/AlignedAllocator.h",
4241 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004242 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004243)
4244
4245xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004246 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004247 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004248 "test/f32-dwconv-chw.cc",
4249 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004250 "src/xnnpack/AlignedAllocator.h",
4251 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004252 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253)
4254
4255xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004256 name = "f16_gavgpool_minmax_test",
4257 srcs = [
4258 "test/f16-gavgpool-minmax.cc",
4259 "test/gavgpool-microkernel-tester.h",
4260 "src/xnnpack/AlignedAllocator.h",
4261 ] + MICROKERNEL_TEST_HDRS,
4262 deps = MICROKERNEL_TEST_DEPS,
4263)
4264
4265xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004266 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004267 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004268 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269 "test/gavgpool-microkernel-tester.h",
4270 "src/xnnpack/AlignedAllocator.h",
4271 ] + MICROKERNEL_TEST_HDRS,
4272 deps = MICROKERNEL_TEST_DEPS,
4273)
4274
4275xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004276 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004277 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004278 "test/f32-gavgpool-cw.cc",
4279 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004280 "src/xnnpack/AlignedAllocator.h",
4281 ] + MICROKERNEL_TEST_HDRS,
4282 deps = MICROKERNEL_TEST_DEPS,
4283)
4284
4285xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004286 name = "f32_gemm_test",
4287 srcs = [
4288 "test/f32-gemm.cc",
4289 "test/gemm-microkernel-tester.h",
4290 "src/xnnpack/AlignedAllocator.h",
4291 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004292 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004293)
4294
4295xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004296 name = "f32_gemm_relu_test",
4297 srcs = [
4298 "test/f32-gemm-relu.cc",
4299 "test/gemm-microkernel-tester.h",
4300 "src/xnnpack/AlignedAllocator.h",
4301 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004302 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004303)
4304
4305xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004306 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004308 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004309 "test/gemm-microkernel-tester.h",
4310 "src/xnnpack/AlignedAllocator.h",
4311 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004312 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004313)
4314
4315xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004316 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004317 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004318 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319 "test/gemm-microkernel-tester.h",
4320 "src/xnnpack/AlignedAllocator.h",
4321 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004322 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004323)
4324
4325xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004326 name = "f16_hswish_test",
4327 srcs = [
4328 "test/f16-hswish.cc",
4329 "test/hswish-microkernel-tester.h",
4330 ] + MICROKERNEL_TEST_HDRS,
4331 deps = MICROKERNEL_TEST_DEPS,
4332)
4333
4334xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335 name = "f32_hswish_test",
4336 srcs = [
4337 "test/f32-hswish.cc",
4338 "test/hswish-microkernel-tester.h",
4339 ] + MICROKERNEL_TEST_HDRS,
4340 deps = MICROKERNEL_TEST_DEPS,
4341)
4342
4343xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004344 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004346 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004347 "test/maxpool-microkernel-tester.h",
4348 ] + MICROKERNEL_TEST_HDRS,
4349 deps = MICROKERNEL_TEST_DEPS,
4350)
4351
4352xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004353 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004354 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004355 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004356 "test/avgpool-microkernel-tester.h",
4357 "src/xnnpack/AlignedAllocator.h",
4358 ] + MICROKERNEL_TEST_HDRS,
4359 deps = MICROKERNEL_TEST_DEPS,
4360)
4361
4362xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004363 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004365 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004366 "test/gemm-microkernel-tester.h",
4367 "src/xnnpack/AlignedAllocator.h",
4368 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004369 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004370)
4371
4372xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004373 name = "f16_prelu_test",
4374 srcs = [
4375 "test/f16-prelu.cc",
4376 "test/prelu-microkernel-tester.h",
4377 "src/xnnpack/AlignedAllocator.h",
4378 ] + MICROKERNEL_TEST_HDRS,
4379 deps = MICROKERNEL_TEST_DEPS,
4380)
4381
4382xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 name = "f32_prelu_test",
4384 srcs = [
4385 "test/f32-prelu.cc",
4386 "test/prelu-microkernel-tester.h",
4387 "src/xnnpack/AlignedAllocator.h",
4388 ] + MICROKERNEL_TEST_HDRS,
4389 deps = MICROKERNEL_TEST_DEPS,
4390)
4391
4392xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004393 name = "f32_raddexpminusmax_test",
4394 srcs = [
4395 "test/f32-raddexpminusmax.cc",
4396 "test/raddexpminusmax-microkernel-tester.h",
4397 ] + MICROKERNEL_TEST_HDRS,
4398 deps = MICROKERNEL_TEST_DEPS,
4399)
4400
4401xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004402 name = "f32_raddextexp_test",
4403 srcs = [
4404 "test/f32-raddextexp.cc",
4405 "test/raddextexp-microkernel-tester.h",
4406 ] + MICROKERNEL_TEST_HDRS,
4407 deps = MICROKERNEL_TEST_DEPS,
4408)
4409
4410xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004411 name = "f32_raddstoreexpminusmax_test",
4412 srcs = [
4413 "test/f32-raddstoreexpminusmax.cc",
4414 "test/raddstoreexpminusmax-microkernel-tester.h",
4415 ] + MICROKERNEL_TEST_HDRS,
4416 deps = MICROKERNEL_TEST_DEPS,
4417)
4418
4419xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004420 name = "f32_relu_test",
4421 srcs = [
4422 "test/f32-relu.cc",
4423 "test/vunary-microkernel-tester.h",
4424 ] + MICROKERNEL_TEST_HDRS,
4425 deps = MICROKERNEL_TEST_DEPS,
4426)
4427
4428xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004429 name = "f32_rmax_test",
4430 srcs = [
4431 "test/f32-rmax.cc",
4432 "test/rmax-microkernel-tester.h",
4433 ] + MICROKERNEL_TEST_HDRS,
4434 deps = MICROKERNEL_TEST_DEPS,
4435)
4436
4437xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004438 name = "f32_sigmoid_test",
4439 srcs = [
4440 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004441 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004442 ] + MICROKERNEL_TEST_HDRS,
4443 deps = MICROKERNEL_TEST_DEPS,
4444)
4445
4446xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004447 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004448 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004449 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004450 "test/spmm-microkernel-tester.h",
4451 "src/xnnpack/AlignedAllocator.h",
4452 ] + MICROKERNEL_TEST_HDRS,
4453 deps = MICROKERNEL_TEST_DEPS,
4454)
4455
4456xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004457 name = "f16_vadd_minmax_test",
4458 srcs = [
4459 "test/f16-vadd-minmax.cc",
4460 "test/vbinary-microkernel-tester.h",
4461 ] + MICROKERNEL_TEST_HDRS,
4462 deps = MICROKERNEL_TEST_DEPS,
4463)
4464
4465xnnpack_unit_test(
4466 name = "f16_vaddc_minmax_test",
4467 srcs = [
4468 "test/f16-vaddc-minmax.cc",
4469 "test/vbinaryc-microkernel-tester.h",
4470 ] + MICROKERNEL_TEST_HDRS,
4471 deps = MICROKERNEL_TEST_DEPS,
4472)
4473
4474xnnpack_unit_test(
4475 name = "f16_vdiv_minmax_test",
4476 srcs = [
4477 "test/f16-vdiv-minmax.cc",
4478 "test/vbinary-microkernel-tester.h",
4479 ] + MICROKERNEL_TEST_HDRS,
4480 deps = MICROKERNEL_TEST_DEPS,
4481)
4482
4483xnnpack_unit_test(
4484 name = "f16_vdivc_minmax_test",
4485 srcs = [
4486 "test/f16-vdivc-minmax.cc",
4487 "test/vbinaryc-microkernel-tester.h",
4488 ] + MICROKERNEL_TEST_HDRS,
4489 deps = MICROKERNEL_TEST_DEPS,
4490)
4491
4492xnnpack_unit_test(
4493 name = "f16_vrdivc_minmax_test",
4494 srcs = [
4495 "test/f16-vrdivc-minmax.cc",
4496 "test/vbinaryc-microkernel-tester.h",
4497 ] + MICROKERNEL_TEST_HDRS,
4498 deps = MICROKERNEL_TEST_DEPS,
4499)
4500
4501xnnpack_unit_test(
4502 name = "f16_vmax_test",
4503 srcs = [
4504 "test/f16-vmax.cc",
4505 "test/vbinary-microkernel-tester.h",
4506 ] + MICROKERNEL_TEST_HDRS,
4507 deps = MICROKERNEL_TEST_DEPS,
4508)
4509
4510xnnpack_unit_test(
4511 name = "f16_vmaxc_test",
4512 srcs = [
4513 "test/f16-vmaxc.cc",
4514 "test/vbinaryc-microkernel-tester.h",
4515 ] + MICROKERNEL_TEST_HDRS,
4516 deps = MICROKERNEL_TEST_DEPS,
4517)
4518
4519xnnpack_unit_test(
4520 name = "f16_vmin_test",
4521 srcs = [
4522 "test/f16-vmin.cc",
4523 "test/vbinary-microkernel-tester.h",
4524 ] + MICROKERNEL_TEST_HDRS,
4525 deps = MICROKERNEL_TEST_DEPS,
4526)
4527
4528xnnpack_unit_test(
4529 name = "f16_vminc_test",
4530 srcs = [
4531 "test/f16-vminc.cc",
4532 "test/vbinaryc-microkernel-tester.h",
4533 ] + MICROKERNEL_TEST_HDRS,
4534 deps = MICROKERNEL_TEST_DEPS,
4535)
4536
4537xnnpack_unit_test(
4538 name = "f16_vmul_minmax_test",
4539 srcs = [
4540 "test/f16-vmul-minmax.cc",
4541 "test/vbinary-microkernel-tester.h",
4542 ] + MICROKERNEL_TEST_HDRS,
4543 deps = MICROKERNEL_TEST_DEPS,
4544)
4545
4546xnnpack_unit_test(
4547 name = "f16_vmulc_minmax_test",
4548 srcs = [
4549 "test/f16-vmulc-minmax.cc",
4550 "test/vbinaryc-microkernel-tester.h",
4551 ] + MICROKERNEL_TEST_HDRS,
4552 deps = MICROKERNEL_TEST_DEPS,
4553)
4554
4555xnnpack_unit_test(
4556 name = "f16_vsub_minmax_test",
4557 srcs = [
4558 "test/f16-vsub-minmax.cc",
4559 "test/vbinary-microkernel-tester.h",
4560 ] + MICROKERNEL_TEST_HDRS,
4561 deps = MICROKERNEL_TEST_DEPS,
4562)
4563
4564xnnpack_unit_test(
4565 name = "f16_vsubc_minmax_test",
4566 srcs = [
4567 "test/f16-vsubc-minmax.cc",
4568 "test/vbinaryc-microkernel-tester.h",
4569 ] + MICROKERNEL_TEST_HDRS,
4570 deps = MICROKERNEL_TEST_DEPS,
4571)
4572
4573xnnpack_unit_test(
4574 name = "f16_vrsubc_minmax_test",
4575 srcs = [
4576 "test/f16-vrsubc-minmax.cc",
4577 "test/vbinaryc-microkernel-tester.h",
4578 ] + MICROKERNEL_TEST_HDRS,
4579 deps = MICROKERNEL_TEST_DEPS,
4580)
4581
4582xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004583 name = "f32_vabs_test",
4584 srcs = [
4585 "test/f32-vabs.cc",
4586 "test/vunary-microkernel-tester.h",
4587 ] + MICROKERNEL_TEST_HDRS,
4588 deps = MICROKERNEL_TEST_DEPS,
4589)
4590
4591xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004592 name = "f32_vadd_test",
4593 srcs = [
4594 "test/f32-vadd.cc",
4595 "test/vbinary-microkernel-tester.h",
4596 ] + MICROKERNEL_TEST_HDRS,
4597 deps = MICROKERNEL_TEST_DEPS,
4598)
4599
4600xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004601 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004602 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004603 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004604 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004605 ] + MICROKERNEL_TEST_HDRS,
4606 deps = MICROKERNEL_TEST_DEPS,
4607)
4608
4609xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004610 name = "f32_vaddc_test",
4611 srcs = [
4612 "test/f32-vaddc.cc",
4613 "test/vbinaryc-microkernel-tester.h",
4614 ] + MICROKERNEL_TEST_HDRS,
4615 deps = MICROKERNEL_TEST_DEPS,
4616)
4617
4618xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004619 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004620 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004621 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004622 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004623 ] + MICROKERNEL_TEST_HDRS,
4624 deps = MICROKERNEL_TEST_DEPS,
4625)
4626
4627xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004628 name = "f32_vdiv_test",
4629 srcs = [
4630 "test/f32-vdiv.cc",
4631 "test/vbinary-microkernel-tester.h",
4632 ] + MICROKERNEL_TEST_HDRS,
4633 deps = MICROKERNEL_TEST_DEPS,
4634)
4635
4636xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004637 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004638 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004639 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004640 "test/vbinary-microkernel-tester.h",
4641 ] + MICROKERNEL_TEST_HDRS,
4642 deps = MICROKERNEL_TEST_DEPS,
4643)
4644
4645xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004646 name = "f32_vdivc_test",
4647 srcs = [
4648 "test/f32-vdivc.cc",
4649 "test/vbinaryc-microkernel-tester.h",
4650 ] + MICROKERNEL_TEST_HDRS,
4651 deps = MICROKERNEL_TEST_DEPS,
4652)
4653
4654xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004655 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004656 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004657 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004658 "test/vbinaryc-microkernel-tester.h",
4659 ] + MICROKERNEL_TEST_HDRS,
4660 deps = MICROKERNEL_TEST_DEPS,
4661)
4662
4663xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004664 name = "f32_vrdivc_test",
4665 srcs = [
4666 "test/f32-vrdivc.cc",
4667 "test/vbinaryc-microkernel-tester.h",
4668 ] + MICROKERNEL_TEST_HDRS,
4669 deps = MICROKERNEL_TEST_DEPS,
4670)
4671
4672xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004673 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004674 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004675 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004676 "test/vbinaryc-microkernel-tester.h",
4677 ] + MICROKERNEL_TEST_HDRS,
4678 deps = MICROKERNEL_TEST_DEPS,
4679)
4680
4681xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004682 name = "f32_vmax_test",
4683 srcs = [
4684 "test/f32-vmax.cc",
4685 "test/vbinary-microkernel-tester.h",
4686 ] + MICROKERNEL_TEST_HDRS,
4687 deps = MICROKERNEL_TEST_DEPS,
4688)
4689
4690xnnpack_unit_test(
4691 name = "f32_vmaxc_test",
4692 srcs = [
4693 "test/f32-vmaxc.cc",
4694 "test/vbinaryc-microkernel-tester.h",
4695 ] + MICROKERNEL_TEST_HDRS,
4696 deps = MICROKERNEL_TEST_DEPS,
4697)
4698
4699xnnpack_unit_test(
4700 name = "f32_vmin_test",
4701 srcs = [
4702 "test/f32-vmin.cc",
4703 "test/vbinary-microkernel-tester.h",
4704 ] + MICROKERNEL_TEST_HDRS,
4705 deps = MICROKERNEL_TEST_DEPS,
4706)
4707
4708xnnpack_unit_test(
4709 name = "f32_vminc_test",
4710 srcs = [
4711 "test/f32-vminc.cc",
4712 "test/vbinaryc-microkernel-tester.h",
4713 ] + MICROKERNEL_TEST_HDRS,
4714 deps = MICROKERNEL_TEST_DEPS,
4715)
4716
4717xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004718 name = "f32_vmul_test",
4719 srcs = [
4720 "test/f32-vmul.cc",
4721 "test/vbinary-microkernel-tester.h",
4722 ] + MICROKERNEL_TEST_HDRS,
4723 deps = MICROKERNEL_TEST_DEPS,
4724)
4725
4726xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004727 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004728 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004729 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004730 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004731 ] + MICROKERNEL_TEST_HDRS,
4732 deps = MICROKERNEL_TEST_DEPS,
4733)
4734
4735xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004736 name = "f32_vmulc_test",
4737 srcs = [
4738 "test/f32-vmulc.cc",
4739 "test/vbinaryc-microkernel-tester.h",
4740 ] + MICROKERNEL_TEST_HDRS,
4741 deps = MICROKERNEL_TEST_DEPS,
4742)
4743
4744xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004745 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004746 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004747 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004748 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004749 ] + MICROKERNEL_TEST_HDRS,
4750 deps = MICROKERNEL_TEST_DEPS,
4751)
4752
4753xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004754 name = "f16_vmulcaddc_minmax_test",
4755 srcs = [
4756 "test/f16-vmulcaddc-minmax.cc",
4757 "test/vmulcaddc-microkernel-tester.h",
4758 "src/xnnpack/AlignedAllocator.h",
4759 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004760 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004761)
4762
4763xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004764 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004765 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004766 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004767 "test/vmulcaddc-microkernel-tester.h",
4768 "src/xnnpack/AlignedAllocator.h",
4769 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004770 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004771)
4772
4773xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004774 name = "f32_vlrelu_test",
4775 srcs = [
4776 "test/f32-vlrelu.cc",
4777 "test/vunary-microkernel-tester.h",
4778 ] + MICROKERNEL_TEST_HDRS,
4779 deps = MICROKERNEL_TEST_DEPS,
4780)
4781
4782xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004783 name = "f32_vneg_test",
4784 srcs = [
4785 "test/f32-vneg.cc",
4786 "test/vunary-microkernel-tester.h",
4787 ] + MICROKERNEL_TEST_HDRS,
4788 deps = MICROKERNEL_TEST_DEPS,
4789)
4790
4791xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004792 name = "f32_vrndne_test",
4793 srcs = [
4794 "test/f32-vrndne.cc",
4795 "test/vunary-microkernel-tester.h",
4796 ] + MICROKERNEL_TEST_HDRS,
4797 deps = MICROKERNEL_TEST_DEPS,
4798)
4799
4800xnnpack_unit_test(
4801 name = "f32_vrndz_test",
4802 srcs = [
4803 "test/f32-vrndz.cc",
4804 "test/vunary-microkernel-tester.h",
4805 ] + MICROKERNEL_TEST_HDRS,
4806 deps = MICROKERNEL_TEST_DEPS,
4807)
4808
4809xnnpack_unit_test(
4810 name = "f32_vrndu_test",
4811 srcs = [
4812 "test/f32-vrndu.cc",
4813 "test/vunary-microkernel-tester.h",
4814 ] + MICROKERNEL_TEST_HDRS,
4815 deps = MICROKERNEL_TEST_DEPS,
4816)
4817
4818xnnpack_unit_test(
4819 name = "f32_vrndd_test",
4820 srcs = [
4821 "test/f32-vrndd.cc",
4822 "test/vunary-microkernel-tester.h",
4823 ] + MICROKERNEL_TEST_HDRS,
4824 deps = MICROKERNEL_TEST_DEPS,
4825)
4826
4827xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004828 name = "f32_vscale_test",
4829 srcs = [
4830 "test/f32-vscale.cc",
4831 "test/vscale-microkernel-tester.h",
4832 ] + MICROKERNEL_TEST_HDRS,
4833 deps = MICROKERNEL_TEST_DEPS,
4834)
4835
4836xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004837 name = "f32_vscaleexpminusmax_test",
4838 srcs = [
4839 "test/f32-vscaleexpminusmax.cc",
4840 "test/vscaleexpminusmax-microkernel-tester.h",
4841 ] + MICROKERNEL_TEST_HDRS,
4842 deps = MICROKERNEL_TEST_DEPS,
4843)
4844
4845xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004846 name = "f32_vscaleextexp_test",
4847 srcs = [
4848 "test/f32-vscaleextexp.cc",
4849 "test/vscaleextexp-microkernel-tester.h",
4850 ] + MICROKERNEL_TEST_HDRS,
4851 deps = MICROKERNEL_TEST_DEPS,
4852)
4853
4854xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004855 name = "f32_vsqr_test",
4856 srcs = [
4857 "test/f32-vsqr.cc",
4858 "test/vunary-microkernel-tester.h",
4859 ] + MICROKERNEL_TEST_HDRS,
4860 deps = MICROKERNEL_TEST_DEPS,
4861)
4862
4863xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004864 name = "f32_vsqrdiff_test",
4865 srcs = [
4866 "test/f32-vsqrdiff.cc",
4867 "test/vbinary-microkernel-tester.h",
4868 ] + MICROKERNEL_TEST_HDRS,
4869 deps = MICROKERNEL_TEST_DEPS,
4870)
4871
4872xnnpack_unit_test(
4873 name = "f32_vsqrdiffc_test",
4874 srcs = [
4875 "test/f32-vsqrdiffc.cc",
4876 "test/vbinaryc-microkernel-tester.h",
4877 ] + MICROKERNEL_TEST_HDRS,
4878 deps = MICROKERNEL_TEST_DEPS,
4879)
4880
4881xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004882 name = "f32_vsqrt_test",
4883 srcs = [
4884 "test/f32-vsqrt.cc",
4885 "test/vunary-microkernel-tester.h",
4886 ] + MICROKERNEL_TEST_HDRS,
4887 deps = MICROKERNEL_TEST_DEPS,
4888)
4889
4890xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004891 name = "f32_vsub_test",
4892 srcs = [
4893 "test/f32-vsub.cc",
4894 "test/vbinary-microkernel-tester.h",
4895 ] + MICROKERNEL_TEST_HDRS,
4896 deps = MICROKERNEL_TEST_DEPS,
4897)
4898
4899xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004900 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004901 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004902 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004903 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004904 ] + MICROKERNEL_TEST_HDRS,
4905 deps = MICROKERNEL_TEST_DEPS,
4906)
4907
4908xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004909 name = "f32_vsubc_test",
4910 srcs = [
4911 "test/f32-vsubc.cc",
4912 "test/vbinaryc-microkernel-tester.h",
4913 ] + MICROKERNEL_TEST_HDRS,
4914 deps = MICROKERNEL_TEST_DEPS,
4915)
4916
4917xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004918 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004919 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004920 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004921 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004922 ] + MICROKERNEL_TEST_HDRS,
4923 deps = MICROKERNEL_TEST_DEPS,
4924)
4925
4926xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004927 name = "f32_vrsubc_test",
4928 srcs = [
4929 "test/f32-vrsubc.cc",
4930 "test/vbinaryc-microkernel-tester.h",
4931 ] + MICROKERNEL_TEST_HDRS,
4932 deps = MICROKERNEL_TEST_DEPS,
4933)
4934
4935xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004936 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004937 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004938 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004939 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004940 ] + MICROKERNEL_TEST_HDRS,
4941 deps = MICROKERNEL_TEST_DEPS,
4942)
4943
4944xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004945 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004946 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004947 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004948 "test/avgpool-microkernel-tester.h",
4949 "src/xnnpack/AlignedAllocator.h",
4950 ] + MICROKERNEL_TEST_HDRS,
4951 deps = MICROKERNEL_TEST_DEPS,
4952)
4953
4954xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004955 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004956 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004957 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004958 "test/gemm-microkernel-tester.h",
4959 "src/xnnpack/AlignedAllocator.h",
4960 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004961 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004962)
4963
4964xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004965 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004966 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004967 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004968 "test/dwconv-microkernel-tester.h",
4969 "src/xnnpack/AlignedAllocator.h",
4970 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004971 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004972)
4973
4974xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004975 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004976 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004977 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004978 "test/gavgpool-microkernel-tester.h",
4979 "src/xnnpack/AlignedAllocator.h",
4980 ] + MICROKERNEL_TEST_HDRS,
4981 deps = MICROKERNEL_TEST_DEPS,
4982)
4983
4984xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004985 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004986 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004987 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004988 "test/gemm-microkernel-tester.h",
4989 "src/xnnpack/AlignedAllocator.h",
4990 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004991 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004992)
4993
4994xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004995 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004996 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004997 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004998 "test/vadd-microkernel-tester.h",
4999 ] + MICROKERNEL_TEST_HDRS,
5000 deps = MICROKERNEL_TEST_DEPS,
5001)
5002
5003xnnpack_unit_test(
5004 name = "u8_clamp_test",
5005 srcs = [
5006 "test/u8-clamp.cc",
5007 "test/clamp-microkernel-tester.h",
5008 ] + MICROKERNEL_TEST_HDRS,
5009 deps = MICROKERNEL_TEST_DEPS,
5010)
5011
5012xnnpack_unit_test(
5013 name = "u8_lut32norm_test",
5014 srcs = [
5015 "test/u8-lut32norm.cc",
5016 "test/lut-norm-microkernel-tester.h",
5017 ] + MICROKERNEL_TEST_HDRS,
5018 deps = MICROKERNEL_TEST_DEPS,
5019)
5020
5021xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005022 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005023 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005024 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005025 "test/maxpool-microkernel-tester.h",
5026 ] + MICROKERNEL_TEST_HDRS,
5027 deps = MICROKERNEL_TEST_DEPS,
5028)
5029
5030xnnpack_unit_test(
5031 name = "u8_rmax_test",
5032 srcs = [
5033 "test/u8-rmax.cc",
5034 "test/rmax-microkernel-tester.h",
5035 ] + MICROKERNEL_TEST_HDRS,
5036 deps = MICROKERNEL_TEST_DEPS,
5037)
5038
5039xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005040 name = "x32_fill_test",
5041 srcs = [
5042 "test/x32-fill.cc",
5043 "test/fill-microkernel-tester.h",
5044 ] + MICROKERNEL_TEST_HDRS,
5045 deps = MICROKERNEL_TEST_DEPS,
5046)
5047
5048xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049 name = "x32_packx_test",
5050 srcs = [
5051 "test/x32-packx.cc",
5052 "test/pack-microkernel-tester.h",
5053 "src/xnnpack/AlignedAllocator.h",
5054 ] + MICROKERNEL_TEST_HDRS,
5055 deps = MICROKERNEL_TEST_DEPS,
5056)
5057
5058xnnpack_unit_test(
5059 name = "x32_pad_test",
5060 srcs = [
5061 "test/x32-pad.cc",
5062 "test/pad-microkernel-tester.h",
5063 ] + MICROKERNEL_TEST_HDRS,
5064 deps = MICROKERNEL_TEST_DEPS,
5065)
5066
5067xnnpack_unit_test(
5068 name = "x32_unpool_test",
5069 srcs = [
5070 "test/x32-unpool.cc",
5071 "test/unpool-microkernel-tester.h",
5072 ] + MICROKERNEL_TEST_HDRS,
5073 deps = MICROKERNEL_TEST_DEPS,
5074)
5075
5076xnnpack_unit_test(
5077 name = "x32_zip_test",
5078 srcs = [
5079 "test/x32-zip.cc",
5080 "test/zip-microkernel-tester.h",
5081 ] + MICROKERNEL_TEST_HDRS,
5082 deps = MICROKERNEL_TEST_DEPS,
5083)
5084
5085xnnpack_unit_test(
5086 name = "x8_lut_test",
5087 srcs = [
5088 "test/x8-lut.cc",
5089 "test/lut-microkernel-tester.h",
5090 ] + MICROKERNEL_TEST_HDRS,
5091 deps = MICROKERNEL_TEST_DEPS,
5092)
5093
5094xnnpack_unit_test(
5095 name = "x8_zip_test",
5096 srcs = [
5097 "test/x8-zip.cc",
5098 "test/zip-microkernel-tester.h",
5099 ] + MICROKERNEL_TEST_HDRS,
5100 deps = MICROKERNEL_TEST_DEPS,
5101)
5102
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005103xnnpack_unit_test(
5104 name = "requantization_test",
5105 srcs = [
5106 "src/xnnpack/requantization-stubs.h",
5107 "test/requantization.cc",
5108 "test/requantization-tester.h",
5109 ] + MICROKERNEL_TEST_HDRS,
5110 deps = MICROKERNEL_TEST_DEPS,
5111)
5112
Marat Dukhan20c3b922020-03-10 03:45:06 -07005113########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005114
5115xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005116 name = "operator_size_test",
5117 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005118 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005119)
5120
Marat Dukhan20c3b922020-03-10 03:45:06 -07005121xnnpack_binary(
5122 name = "subgraph_size_test",
5123 srcs = ["test/subgraph-size.c"],
5124 deps = [":XNNPACK"],
5125)
5126
5127########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005128
5129xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005130 name = "abs_nc_test",
5131 srcs = [
5132 "test/abs-nc.cc",
5133 "test/abs-operator-tester.h",
5134 ],
5135 deps = OPERATOR_TEST_DEPS,
5136)
5137
5138xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005139 name = "add_nd_test",
5140 srcs = [
5141 "test/add-nd.cc",
5142 "test/binary-elementwise-operator-tester.h",
5143 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005144 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005145)
5146
5147xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005148 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005149 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005150 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005151 "test/argmax-pooling-operator-tester.h",
5152 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005153 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005154)
5155
5156xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005157 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005158 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005159 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005160 "test/average-pooling-operator-tester.h",
5161 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005162 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005163)
5164
5165xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005166 name = "bankers_rounding_nc_test",
5167 srcs = [
5168 "test/bankers-rounding-nc.cc",
5169 "test/bankers-rounding-operator-tester.h",
5170 ],
5171 deps = OPERATOR_TEST_DEPS,
5172)
5173
5174xnnpack_unit_test(
5175 name = "ceiling_nc_test",
5176 srcs = [
5177 "test/ceiling-nc.cc",
5178 "test/ceiling-operator-tester.h",
5179 ],
5180 deps = OPERATOR_TEST_DEPS,
5181)
5182
5183xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005184 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005185 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005186 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 "test/channel-shuffle-operator-tester.h",
5188 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005189 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190)
5191
5192xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005193 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005195 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005196 "test/clamp-operator-tester.h",
5197 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005198 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005199)
5200
5201xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005202 name = "constant_pad_nd_test",
5203 srcs = [
5204 "test/constant-pad-nd.cc",
5205 "test/constant-pad-operator-tester.h",
5206 ],
5207 deps = OPERATOR_TEST_DEPS,
5208)
5209
5210xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005211 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005212 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005213 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005214 "test/convolution-operator-tester.h",
5215 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005216 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005217)
5218
5219xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005220 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005221 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005222 "test/convolution-nchw.cc",
5223 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005225 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005226)
5227
5228xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005229 name = "copy_nc_test",
5230 srcs = [
5231 "test/copy-nc.cc",
5232 "test/copy-operator-tester.h",
5233 ],
5234 deps = OPERATOR_TEST_DEPS,
5235)
5236
5237xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005238 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005239 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005240 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005241 "test/deconvolution-operator-tester.h",
5242 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005243 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244)
5245
5246xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005247 name = "divide_nd_test",
5248 srcs = [
5249 "test/binary-elementwise-operator-tester.h",
5250 "test/divide-nd.cc",
5251 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005252 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005253)
5254
5255xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005256 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005257 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005258 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005259 "test/fully-connected-operator-tester.h",
5260 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005261 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005262)
5263
5264xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005265 name = "floor_nc_test",
5266 srcs = [
5267 "test/floor-nc.cc",
5268 "test/floor-operator-tester.h",
5269 ],
5270 deps = OPERATOR_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005274 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005275 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005276 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005277 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005278 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005279 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005280)
5281
5282xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005283 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005285 "test/global-average-pooling-ncw.cc",
5286 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005288 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005289)
5290
5291xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005292 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005294 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 "test/hardswish-operator-tester.h",
5296 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005297 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005298)
5299
5300xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005301 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005302 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005303 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005304 "test/leaky-relu-operator-tester.h",
5305 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005306 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307)
5308
5309xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005310 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005311 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005312 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005313 "test/max-pooling-operator-tester.h",
5314 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005315 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316)
5317
5318xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005319 name = "maximum_nd_test",
5320 srcs = [
5321 "test/binary-elementwise-operator-tester.h",
5322 "test/maximum-nd.cc",
5323 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005324 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005325)
5326
5327xnnpack_unit_test(
5328 name = "minimum_nd_test",
5329 srcs = [
5330 "test/binary-elementwise-operator-tester.h",
5331 "test/minimum-nd.cc",
5332 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005333 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005334)
5335
5336xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005337 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005338 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005339 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005340 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005341 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005342 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005343)
5344
5345xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005346 name = "negate_nc_test",
5347 srcs = [
5348 "test/negate-nc.cc",
5349 "test/negate-operator-tester.h",
5350 ],
5351 deps = OPERATOR_TEST_DEPS,
5352)
5353
5354xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005355 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005356 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005357 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005358 "test/prelu-operator-tester.h",
5359 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005360 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005361)
5362
5363xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005364 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005365 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005366 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005367 "test/resize-bilinear-operator-tester.h",
5368 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005369 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005370)
5371
5372xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005373 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005374 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005375 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005376 "test/sigmoid-operator-tester.h",
5377 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005378 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005379)
5380
5381xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005382 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005383 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005384 "test/softmax-nc.cc",
5385 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005386 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005387 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005388)
5389
5390xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005391 name = "square_nc_test",
5392 srcs = [
5393 "test/square-nc.cc",
5394 "test/square-operator-tester.h",
5395 ],
5396 deps = OPERATOR_TEST_DEPS,
5397)
5398
5399xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005400 name = "square_root_nc_test",
5401 srcs = [
5402 "test/square-root-nc.cc",
5403 "test/square-root-operator-tester.h",
5404 ],
5405 deps = OPERATOR_TEST_DEPS,
5406)
5407
5408xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005409 name = "squared_difference_nd_test",
5410 srcs = [
5411 "test/binary-elementwise-operator-tester.h",
5412 "test/squared-difference-nd.cc",
5413 ],
5414 deps = OPERATOR_TEST_DEPS,
5415)
5416
5417xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005418 name = "subtract_nd_test",
5419 srcs = [
5420 "test/binary-elementwise-operator-tester.h",
5421 "test/subtract-nd.cc",
5422 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005423 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005424)
5425
5426xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005427 name = "truncation_nc_test",
5428 srcs = [
5429 "test/truncation-nc.cc",
5430 "test/truncation-operator-tester.h",
5431 ],
5432 deps = OPERATOR_TEST_DEPS,
5433)
5434
5435xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005436 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005437 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005438 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439 "test/unpooling-operator-tester.h",
5440 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005441 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005442)
5443
Chao Mei6ddfc602020-05-13 22:29:36 -07005444############################### Misc unit tests ###############################
5445
5446xnnpack_unit_test(
5447 name = "memory_planner_test",
5448 srcs = [
5449 "test/memory-planner-test.cc",
5450 ],
5451 deps = [
5452 ":XNNPACK",
5453 ":memory_planner",
5454 ],
5455)
5456
Marat Dukhan08c4a432019-10-03 09:29:21 -07005457############################# Build configurations #############################
5458
Marat Dukhanb8642352019-10-30 15:43:02 -07005459# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005460config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005461 name = "xnn_enable_assembly_explicit_true",
5462 define_values = {"xnn_enable_assembly": "true"},
5463)
5464
5465# Disables usage of assembly kernels.
5466config_setting(
5467 name = "xnn_enable_assembly_explicit_false",
5468 define_values = {"xnn_enable_assembly": "false"},
5469)
5470
Marat Dukhan9de90e02020-06-18 16:04:12 -07005471# Enables usage of sparse inference.
5472config_setting(
5473 name = "xnn_enable_sparse_explicit_true",
5474 define_values = {"xnn_enable_sparse": "true"},
5475)
5476
5477# Disables usage of sparse inference.
5478config_setting(
5479 name = "xnn_enable_sparse_explicit_false",
5480 define_values = {"xnn_enable_sparse": "false"},
5481)
5482
Marat Dukhan05702cf2020-03-26 15:41:33 -07005483# Disables usage of HMP-aware optimizations.
5484config_setting(
5485 name = "xnn_enable_hmp_explicit_false",
5486 define_values = {"xnn_enable_hmp": "false"},
5487)
5488
Chao Mei6ddfc602020-05-13 22:29:36 -07005489# Enable usage of optimized memory allocation
5490config_setting(
5491 name = "xnn_enable_memopt_explicit_true",
5492 define_values = {"xnnpack_enable_memopt": "true"},
5493)
5494
5495# Disable usage of optimized memory allocation
5496config_setting(
5497 name = "xnn_enable_memopt_explicit_false",
5498 define_values = {"xnnpack_enable_memopt": "false"},
5499)
5500
Marat Dukhanb8642352019-10-30 15:43:02 -07005501# Builds with -c dbg
5502config_setting(
5503 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005504 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005505 "compilation_mode": "dbg",
5506 },
5507)
5508
5509# Builds with -c opt
5510config_setting(
5511 name = "optimized_build",
5512 values = {
5513 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005514 },
5515)
5516
5517config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005518 name = "linux_k8",
5519 values = {"cpu": "k8"},
5520)
5521
5522config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005523 name = "linux_arm",
5524 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005525)
5526
5527config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005528 name = "linux_armeabi",
5529 values = {"cpu": "armeabi"},
5530)
5531
5532config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005533 name = "linux_armhf",
5534 values = {"cpu": "armhf"},
5535)
5536
5537config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005538 name = "linux_armv7a",
5539 values = {"cpu": "armv7a"},
5540)
5541
5542config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005543 name = "linux_aarch64",
5544 values = {"cpu": "aarch64"},
5545)
5546
5547config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005548 name = "android",
5549 values = {"crosstool_top": "//external:android/crosstool"},
5550)
5551
5552config_setting(
5553 name = "android_armv7",
5554 values = {
5555 "crosstool_top": "//external:android/crosstool",
5556 "cpu": "armeabi-v7a",
5557 },
5558)
5559
5560config_setting(
5561 name = "android_arm64",
5562 values = {
5563 "crosstool_top": "//external:android/crosstool",
5564 "cpu": "arm64-v8a",
5565 },
5566)
5567
5568config_setting(
5569 name = "android_x86",
5570 values = {
5571 "crosstool_top": "//external:android/crosstool",
5572 "cpu": "x86",
5573 },
5574)
5575
5576config_setting(
5577 name = "android_x86_64",
5578 values = {
5579 "crosstool_top": "//external:android/crosstool",
5580 "cpu": "x86_64",
5581 },
5582)
5583
5584config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005585 name = "windows_x86_64",
5586 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005587)
5588
5589config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005590 name = "windows_x86_64_clang",
5591 values = {
5592 "compiler": "clang-cl",
5593 "cpu": "x64_windows",
5594 },
5595)
5596
5597config_setting(
5598 name = "windows_x86_64_mingw",
5599 values = {
5600 "compiler": "mingw-gcc",
5601 "cpu": "x64_windows",
5602 },
5603)
5604
5605config_setting(
5606 name = "windows_x86_64_msys",
5607 values = {
5608 "compiler": "msys-gcc",
5609 "cpu": "x64_windows",
5610 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005611)
5612
5613config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005614 name = "macos_x86_64",
5615 values = {
5616 "apple_platform_type": "macos",
5617 "cpu": "darwin",
5618 },
5619)
5620
5621config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005622 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005623 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005624)
5625
5626config_setting(
5627 name = "emscripten_wasm",
5628 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005629 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005630 "cpu": "wasm",
5631 },
5632)
5633
5634config_setting(
5635 name = "emscripten_wasmsimd",
5636 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005637 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005638 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005639 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005640 },
5641)
5642
5643config_setting(
5644 name = "emscripten_asmjs",
5645 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005646 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005647 "cpu": "asmjs",
5648 },
5649)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005650
5651config_setting(
5652 name = "ios_armv7",
5653 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005654 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005655 "cpu": "ios_armv7",
5656 },
5657)
5658
5659config_setting(
5660 name = "ios_arm64",
5661 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005662 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005663 "cpu": "ios_arm64",
5664 },
5665)
5666
5667config_setting(
5668 name = "ios_arm64e",
5669 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005670 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005671 "cpu": "ios_arm64e",
5672 },
5673)
5674
5675config_setting(
5676 name = "ios_x86",
5677 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005678 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005679 "cpu": "ios_i386",
5680 },
5681)
5682
5683config_setting(
5684 name = "ios_x86_64",
5685 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005686 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005687 "cpu": "ios_x86_64",
5688 },
5689)
5690
5691config_setting(
5692 name = "watchos_armv7k",
5693 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005694 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005695 "cpu": "watchos_armv7k",
5696 },
5697)
5698
5699config_setting(
5700 name = "watchos_arm64_32",
5701 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005702 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005703 "cpu": "watchos_arm64_32",
5704 },
5705)
5706
5707config_setting(
5708 name = "watchos_x86",
5709 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005710 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005711 "cpu": "watchos_i386",
5712 },
5713)
5714
5715config_setting(
5716 name = "watchos_x86_64",
5717 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005718 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005719 "cpu": "watchos_x86_64",
5720 },
5721)
5722
5723config_setting(
5724 name = "tvos_arm64",
5725 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005726 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005727 "cpu": "tvos_arm64",
5728 },
5729)
5730
5731config_setting(
5732 name = "tvos_x86_64",
5733 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005734 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005735 "cpu": "tvos_x86_64",
5736 },
5737)