blob: 3187399126fe92cddace2e04fe727e8f006fe659 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700102 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700103 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700104 "src/subgraph/squared-difference.c",
105 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700106 "src/subgraph/unpooling-2d.c",
107]
108
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800109TABLE_SRCS = [
110 "src/tables/exp2-k-over-64.c",
111 "src/tables/exp2-k-over-2048.c",
112]
113
Marat Dukhan08c4a432019-10-03 09:29:21 -0700114SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800115 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800116 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800117 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700118 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
119 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700120 "src/f32-clamp/gen/scalar-x1.c",
121 "src/f32-clamp/gen/scalar-x2.c",
122 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800123 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700125 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
126 "src/f32-dwconv-chw/3x3p1-scalar.c",
127 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
128 "src/f32-dwconv-chw/5x5p2-scalar.c",
129 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700130 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
131 "src/f32-dwconv/gen/up1x4-scalar.c",
132 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
133 "src/f32-dwconv/gen/up1x9-scalar.c",
134 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
135 "src/f32-dwconv/gen/up1x25-scalar.c",
136 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
137 "src/f32-dwconv/gen/up2x4-scalar.c",
138 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
139 "src/f32-dwconv/gen/up2x9-scalar.c",
140 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
141 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700142 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
144 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700148 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
150 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700152 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700154 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700155 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
156 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700157 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
158 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700160 "src/f32-gemm/gen/1x4-scalar.c",
161 "src/f32-gemm/gen/2x4-scalar.c",
162 "src/f32-gemm/gen/4x2-scalar.c",
163 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700164 "src/f32-gemm/gen/1x4-relu-scalar.c",
165 "src/f32-gemm/gen/2x4-relu-scalar.c",
166 "src/f32-gemm/gen/4x2-relu-scalar.c",
167 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700168 "src/f32-gemm/gen/1x4-minmax-scalar.c",
169 "src/f32-gemm/gen/2x4-minmax-scalar.c",
170 "src/f32-gemm/gen/4x2-minmax-scalar.c",
171 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800172 "src/f32-hswish/gen/scalar-x1.c",
173 "src/f32-hswish/gen/scalar-x2.c",
174 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700175 "src/f32-ibilinear/gen/scalar-c1.c",
176 "src/f32-ibilinear/gen/scalar-c2.c",
177 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700178 "src/f32-igemm/gen/1x4-scalar.c",
179 "src/f32-igemm/gen/2x4-scalar.c",
180 "src/f32-igemm/gen/4x2-scalar.c",
181 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700182 "src/f32-igemm/gen/1x4-relu-scalar.c",
183 "src/f32-igemm/gen/2x4-relu-scalar.c",
184 "src/f32-igemm/gen/4x2-relu-scalar.c",
185 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700186 "src/f32-igemm/gen/1x4-minmax-scalar.c",
187 "src/f32-igemm/gen/2x4-minmax-scalar.c",
188 "src/f32-igemm/gen/4x2-minmax-scalar.c",
189 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700190 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
191 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700193 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
194 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
195 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800197 "src/f32-prelu/gen/scalar-2x1.c",
198 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800199 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
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 Dukhan662faa02019-12-09 22:48:16 -0800442 "src/f32-hswish/gen/wasm-x1.c",
443 "src/f32-hswish/gen/wasm-x2.c",
444 "src/f32-hswish/gen/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 Dukhan3fa52c82020-07-08 12:54:33 -0700516 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
517 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
518 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
519 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700520 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
521 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
522 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
523 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
524 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
525 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700526 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
527 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
528 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
529 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
530 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
531 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
532 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
533 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
534 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
535 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
536 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
537 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
538 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
539 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
540 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
541 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
542 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
543 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
544 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
545 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
546 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
547 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
548 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
549 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700550 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
551 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
552 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700553 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700554 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
555 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
556 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
557 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700558 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
559 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
560 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
561 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
562 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
563 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
564 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
565 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
566 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
567 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
568 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
569 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
570 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
571 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
572 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
573 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
574 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
575 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
576 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
577 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
578 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
579 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
580 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
581 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
582 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
583 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
584 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
585 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
586 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
587 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700588 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
589 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700590 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
591 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
592 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
593 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
594 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
595 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
596 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
597 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
598 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
599 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
600 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
601 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
602 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
603 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
604 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
605 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
606 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
607 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
608 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
609 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
610 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
611 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
612 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
613 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
614 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
615 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
616 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
617 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
618 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
619 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700620 "src/f32-ibilinear/gen/wasmsimd-c4.c",
621 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700622 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
623 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
624 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700625 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700626 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
627 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
628 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
629 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700630 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
631 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
632 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
633 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
634 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
635 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
636 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
637 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
638 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
639 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
640 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
641 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
642 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
643 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
644 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
645 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
646 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
647 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
648 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
649 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
650 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
651 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
652 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
653 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
654 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
655 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
656 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
657 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
658 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
659 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700660 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
661 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan9df9dc62020-07-10 20:08:49 -0700662 "src/f32-hswish/gen/wasmsimd-x4.c",
663 "src/f32-hswish/gen/wasmsimd-x8.c",
664 "src/f32-hswish/gen/wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700665 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
666 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700667 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
668 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
669 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
670 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700671 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
672 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
673 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
674 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
675 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
676 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
677 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
678 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
679 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
680 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
681 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
682 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700683 "src/f32-relu/gen/wasmsimd-x4.c",
684 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700685 "src/f32-rmax/wasmsimd-arm.c",
686 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700687 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
688 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
689 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
690 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
691 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
692 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700693 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
694 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
695 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
696 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700697 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
698 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700699 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
700 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
701 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
702 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700703 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
704 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700705 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
706 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
707 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
708 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700709 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
710 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700711 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
712 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
713 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
714 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700715 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
716 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700717 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
718 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
719 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
720 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700721 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
722 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
723 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
724 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700725 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
726 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
727 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
728 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700729 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
730 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
731 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
732 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700733 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
734 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
735 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
736 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700737 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
738 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700739 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
740 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
741 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
742 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700743 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
744 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700745 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
746 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
747 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
748 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700749 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
750 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700751 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
752 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
753 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
754 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700755 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
756 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700757 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
760 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700761 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
762 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
763 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
764 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700765 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
766 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700767 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
768 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
769 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
770 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700771 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
772 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700773 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
774 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
775 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
776 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700777 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
778 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
779 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
780 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700781 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
782 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
783 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
784 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
785 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
786 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
787 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
788 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
789 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
790 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
791 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
792 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
793 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
794 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700795 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
796 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700797 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
798 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
799 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
800 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
801 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
802 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700803 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700804 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700805 "src/x32-pad/wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700806 "src/math/roundne-wasmsimd-addsub.c",
807 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700808 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700809 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700810 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700811 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700812 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700813 "src/requantization/fp32-wasmsimd.c",
814]
815
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800816PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800817 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800818 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800819 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700820 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
821 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700822 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700823 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700824 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700825 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700826 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700827 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
828 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
829 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
830 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
831 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
832 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
833 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
834 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
835 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
836 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
837 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
838 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700839 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700840 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
841 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700842 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
843 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
844 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
845 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
846 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
847 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
848 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
849 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
850 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
851 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
852 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
853 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
854 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
855 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
856 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
857 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
858 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
859 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
860 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700861 "src/f32-ibilinear/gen/psimd-c4.c",
862 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700863 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
864 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
865 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
866 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
867 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
868 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
869 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
870 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
871 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
872 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700873 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700874 "src/f32-spmm/gen/4x1-minmax-psimd.c",
875 "src/f32-spmm/gen/8x1-minmax-psimd.c",
876 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700877 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
878 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700879 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800880 "src/f32-prelu/gen/psimd-2x4.c",
881 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800882 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700883 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
884 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
885 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
886 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
887 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
888 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
889 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
890 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800891 "src/f32-vbinary/gen/vmax-psimd-x4.c",
892 "src/f32-vbinary/gen/vmax-psimd-x8.c",
893 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
894 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
895 "src/f32-vbinary/gen/vmin-psimd-x4.c",
896 "src/f32-vbinary/gen/vmin-psimd-x8.c",
897 "src/f32-vbinary/gen/vminc-psimd-x4.c",
898 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700899 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
900 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
901 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
902 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
903 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
904 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
905 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
906 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700907 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
908 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
909 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
910 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700911 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
912 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
913 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
914 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700915 "src/f32-vlrelu/gen/vlrelu-psimd-x4.c",
916 "src/f32-vlrelu/gen/vlrelu-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700917 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
918 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700919 "src/f32-vunary/gen/vabs-psimd-x4.c",
920 "src/f32-vunary/gen/vabs-psimd-x8.c",
921 "src/f32-vunary/gen/vneg-psimd-x4.c",
922 "src/f32-vunary/gen/vneg-psimd-x8.c",
923 "src/f32-vunary/gen/vsqr-psimd-x4.c",
924 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700925 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700926 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700927 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700928 "src/x32-unpool/psimd.c",
929 "src/x32-zip/x2-psimd.c",
930 "src/x32-zip/x3-psimd.c",
931 "src/x32-zip/x4-psimd.c",
932 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700933 "src/requantization/precise-psimd.c",
934 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700935]
936
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800937PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800938 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
939 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
940 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
941 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
942 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
943 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
944 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
945 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
946 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
947 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
948 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
949 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800950 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
951 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
952 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
953 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
954 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
955 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700956 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
957 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
958 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
959 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
960 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
961 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
962 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
963 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700964 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700965 "src/math/roundd-psimd-addsub.c",
966 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700967 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800968 "src/math/sigmoid-psimd-p5-div.c",
969]
970
Marat Dukhan08c4a432019-10-03 09:29:21 -0700971# ISA-specific micro-kernels
972NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700973 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
974 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700975 "src/f32-clamp/gen/neon-x4.c",
976 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700977 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
978 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700979 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
980 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700981 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
982 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700983 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
984 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700985 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
986 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
987 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
988 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700989 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
990 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
991 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
992 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700993 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
994 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
995 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
996 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700997 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700998 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
999 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001000 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1001 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1002 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1003 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1004 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1005 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1006 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1007 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1008 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1009 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1010 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1011 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1012 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1013 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1014 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1015 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1016 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1017 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1018 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1019 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1020 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1021 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1022 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1023 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1024 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1025 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1026 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1027 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1028 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1029 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1030 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001031 "src/f32-hswish/gen/neon-x4.c",
1032 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan55dde5b2020-07-10 22:48:54 -07001033 "src/f32-hswish/gen/neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001034 "src/f32-ibilinear/gen/neon-c4.c",
1035 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001036 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1037 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1038 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1039 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1040 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1041 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1042 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1043 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1044 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1045 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1046 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1047 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1048 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1049 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1050 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1051 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001052 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1053 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1054 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001055 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1056 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001057 "src/f32-prelu/gen/neon-2x4.c",
1058 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001059 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1060 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1061 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1062 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1063 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1064 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1065 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1066 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1067 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1068 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1069 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1070 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1071 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1072 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1073 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1074 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1075 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1076 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1077 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1078 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1079 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1080 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1081 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1082 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001083 "src/f32-relu/gen/neon-x4.c",
1084 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001085 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001086 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001087 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1088 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1089 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1090 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1091 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1092 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1093 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1094 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1095 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1096 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1097 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1098 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1099 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1100 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1101 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1102 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1103 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1104 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001105 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1106 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1107 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1108 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001109 "src/f32-vbinary/gen/vmax-neon-x4.c",
1110 "src/f32-vbinary/gen/vmax-neon-x8.c",
1111 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1112 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1113 "src/f32-vbinary/gen/vmin-neon-x4.c",
1114 "src/f32-vbinary/gen/vmin-neon-x8.c",
1115 "src/f32-vbinary/gen/vminc-neon-x4.c",
1116 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001117 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1118 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1119 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1120 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1121 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1122 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001123 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1124 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1125 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1126 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001127 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1128 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1129 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1130 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001131 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1132 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001133 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1134 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001135 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1136 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1137 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1138 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1139 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1140 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1141 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1142 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001143 "src/f32-vunary/gen/vabs-neon-x4.c",
1144 "src/f32-vunary/gen/vabs-neon-x8.c",
1145 "src/f32-vunary/gen/vneg-neon-x4.c",
1146 "src/f32-vunary/gen/vneg-neon-x8.c",
1147 "src/f32-vunary/gen/vsqr-neon-x4.c",
1148 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001149 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1150 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1151 "src/qu8-dwconv/up8x9-minmax-neon.c",
1152 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1153 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1154 "src/qu8-gemm/4x8-minmax-neon.c",
1155 "src/qu8-gemm/8x8-minmax-neon.c",
1156 "src/qu8-igemm/4x8-minmax-neon.c",
1157 "src/qu8-igemm/8x8-minmax-neon.c",
1158 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001159 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001160 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001161 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001162 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001163 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001164 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001165 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001166 "src/x32-zip/x2-neon.c",
1167 "src/x32-zip/x3-neon.c",
1168 "src/x32-zip/x4-neon.c",
1169 "src/x32-zip/xm-neon.c",
1170 "src/x8-zip/x2-neon.c",
1171 "src/x8-zip/x3-neon.c",
1172 "src/x8-zip/x4-neon.c",
1173 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001174 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001175 "src/math/roundd-neon-addsub.c",
1176 "src/math/roundd-neon-cvt.c",
1177 "src/math/roundu-neon-addsub.c",
1178 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001179 "src/math/roundz-neon-addsub.c",
1180 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001181 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001182 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1183 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1184 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1185 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1186 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1187 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001188 "src/math/sqrt-neon-nr1rsqrts.c",
1189 "src/math/sqrt-neon-nr2rsqrts.c",
1190 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001191 "src/requantization/precise-neon.c",
1192 "src/requantization/fp32-neon.c",
1193 "src/requantization/q31-neon.c",
1194 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001195]
1196
1197NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001198 "src/f32-ibilinear/gen/neonfma-c4.c",
1199 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001200 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1201 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1202 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1203 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1204 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1205 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1206 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1207 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1208 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001209 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1210 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1211 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1212 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001213 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1214 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1215 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1216 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001217 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1218 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1219 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1220 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001221 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1222 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1223 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1224 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1225 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1226 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1227 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1228 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1229 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1230 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1231 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1232 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1233 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1234 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1235 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1236 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1237 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1238 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001239 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1240 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001241 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1242 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1243 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1244 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1245 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1246 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1247 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1248 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1249 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1250 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1251 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1252 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1253 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1254 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1255 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1256 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1257 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1258 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1259 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1260 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1261 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1262 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1263 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1264 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001265 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1269 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1270 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1271 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1272 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1273 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1275 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1276 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1277 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1278 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1279 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1280 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1281 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1282 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1283 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1284 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1285 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1286 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1287 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1288 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1289 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1290 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1291 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1292 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1293 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1294 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1295 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1296 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1297 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1298 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1299 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1300 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1301 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1302 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1303 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1304 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1305 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1306 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1307 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1308 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1309 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1310 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1311 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1312 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1313 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1314 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1315 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1316 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1318 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001319 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1320 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001321 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1322 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1323 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1324 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1325 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1326 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1327 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1328 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1329 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1330 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1331 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1332 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1333 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1334 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1335 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1336 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1337 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1338 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1339 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1340 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001341 "src/math/exp-neonfma-lut64-p2.c",
1342 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001343 "src/math/expminus-neonfma-lut2048-p1.c",
1344 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001345 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001346 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1347 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1348 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1349 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1350 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1351 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1352 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1353 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1354 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1355 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1356 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1357 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1358 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1359 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1360 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1361 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1362 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1363 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001364 "src/math/sqrt-neonfma-nr1fma.c",
1365 "src/math/sqrt-neonfma-nr2fma.c",
1366 "src/math/sqrt-neonfma-nr3fma.c",
1367 "src/math/sqrt-neonfma-nr2fma1adj.c",
1368 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001369]
1370
1371AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001372 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1373 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1374 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1375 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1376 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1377 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001378 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1379 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1380 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1381 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1382 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1383 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1384 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1385 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1386 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1387 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1388 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1389 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1390 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1391 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1392 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1393 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1394 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1395 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1396 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1397 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001398 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1399 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001400 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1401 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001402 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1403 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001404 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1405 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001406 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1407 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1408 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1409 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1410 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001411 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1412 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1413 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1414 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1415 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1416 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1417 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1418 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1419 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1420 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1421 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1422 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1423 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1424 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1425 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1426 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1427 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1428 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001429 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1430 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1431 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1432 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1433 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1434 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1435 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1436 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1437 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1438 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1439 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1440 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1441 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1442 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1443 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1444 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1445 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1446 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001447 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1448 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001449 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1450 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1451 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1452 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1453 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1454 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001455]
1456
Marat Dukhan8853b822020-05-07 12:19:01 -07001457NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001458 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1459 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1460 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1461 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1462 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1463 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1464 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1465 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001466 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001467 "src/math/roundd-neonv8.c",
1468 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001469 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001470]
1471
Marat Dukhan08c4a432019-10-03 09:29:21 -07001472AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001473 "src/f16-clamp/gen/neonfp16arith-x16.c",
1474 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001475 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1476 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1477 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1478 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1479 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1480 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1481 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1482 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1483 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1484 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1485 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1486 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001487 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1488 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001489 "src/f16-hswish/gen/neonfp16arith-x16.c",
1490 "src/f16-hswish/gen/neonfp16arith-x8.c",
1491 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1492 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001493 "src/f16-relu/gen/neonfp16arith-x16.c",
1494 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001495 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1496 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1497 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1498 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1499 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1500 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1501 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1502 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1503 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1504 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1505 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1506 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1507 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1508 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1509 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1510 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1511 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1512 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1513 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1514 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1515 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1516 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1517 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1518 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1519 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1520 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1521 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1522 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001523 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1524 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001525 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1526 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1527 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1528 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1529 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1530 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1531 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1532 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001533 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1534 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1535 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1536 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001537 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1538 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1539 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1540 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1541 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1542 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1543 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1544 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1545 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1546 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1547 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1548 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001549 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1550 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1551 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1552 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1553 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1554 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1555 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1556 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001557]
1558
1559SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001560 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1561 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001562 "src/f32-clamp/gen/sse-x4.c",
1563 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001564 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1565 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001566 "src/f32-dwconv-chw/3x3p1-sse.c",
1567 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001568 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1569 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1570 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1571 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1572 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1573 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1574 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1575 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1576 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1577 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1578 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1579 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001580 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001581 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1582 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001583 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1584 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1585 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1586 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1587 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1588 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1589 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1590 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1591 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1592 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1593 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1594 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1595 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001596 "src/f32-hswish/gen/sse-x4.c",
1597 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001598 "src/f32-ibilinear/gen/sse-c4.c",
1599 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001600 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1601 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1602 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1603 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1604 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1605 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1606 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001607 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1608 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1609 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001610 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001611 "src/f32-prelu/gen/sse-2x4.c",
1612 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001613 "src/f32-relu/gen/sse-x4.c",
1614 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001615 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001616 "src/f32-spmm/gen/4x1-minmax-sse.c",
1617 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001618 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001619 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1620 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1621 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1622 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1623 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1624 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1625 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1626 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001627 "src/f32-vbinary/gen/vmax-sse-x4.c",
1628 "src/f32-vbinary/gen/vmax-sse-x8.c",
1629 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1630 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1631 "src/f32-vbinary/gen/vmin-sse-x4.c",
1632 "src/f32-vbinary/gen/vmin-sse-x8.c",
1633 "src/f32-vbinary/gen/vminc-sse-x4.c",
1634 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001635 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1636 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1637 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1638 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1639 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1640 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1641 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1642 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001643 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1644 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1645 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1646 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001647 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1648 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1649 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1650 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001651 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1652 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001653 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1654 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001655 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1656 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001657 "src/f32-vunary/gen/vabs-sse-x4.c",
1658 "src/f32-vunary/gen/vabs-sse-x8.c",
1659 "src/f32-vunary/gen/vneg-sse-x4.c",
1660 "src/f32-vunary/gen/vneg-sse-x8.c",
1661 "src/f32-vunary/gen/vsqr-sse-x4.c",
1662 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001663 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001665 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001666 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001667 "src/math/roundd-sse-addsub.c",
1668 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001669 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001670 "src/math/sqrt-sse-hh1mac.c",
1671 "src/math/sqrt-sse-nr1mac.c",
1672 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001673]
1674
1675SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001676 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1677 "src/f32-argmaxpool/4x-sse2-c4.c",
1678 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001679 "src/f32-prelu/gen/sse2-2x4.c",
1680 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001681 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1682 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1683 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1684 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1685 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1686 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1687 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1688 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1689 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1690 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1691 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1692 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001693 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001694 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001695 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001696 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001697 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1698 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001699 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1700 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001701 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1702 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1703 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1704 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1705 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1706 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1707 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1708 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001709 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1710 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1711 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1712 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1713 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1714 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1715 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1716 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1717 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001718 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001719 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001720 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001721 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001722 "src/x32-zip/x2-sse2.c",
1723 "src/x32-zip/x3-sse2.c",
1724 "src/x32-zip/x4-sse2.c",
1725 "src/x32-zip/xm-sse2.c",
1726 "src/x8-zip/x2-sse2.c",
1727 "src/x8-zip/x3-sse2.c",
1728 "src/x8-zip/x4-sse2.c",
1729 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001730 "src/math/exp-sse2-p5.c",
1731 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001732 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001733 "src/math/roundd-sse2-cvt.c",
1734 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001735 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001736 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001737 "src/requantization/precise-sse2.c",
1738 "src/requantization/fp32-sse2.c",
1739 "src/requantization/q31-sse2.c",
1740 "src/requantization/gemmlowp-sse2.c",
1741]
1742
1743SSSE3_UKERNELS = [
1744 "src/requantization/precise-ssse3.c",
1745 "src/requantization/q31-ssse3.c",
1746 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747]
1748
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001749SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001750 "src/f32-prelu/gen/sse41-2x4.c",
1751 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001752 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001753 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001754 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001755 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001756 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1757 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001758 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1759 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001760 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1761 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1762 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1763 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1764 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1765 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1766 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1767 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001768 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001769 "src/math/roundd-sse41.c",
1770 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001771 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001772 "src/requantization/precise-sse4.c",
1773 "src/requantization/q31-sse4.c",
1774 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001775]
1776
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001778 "src/f32-clamp/gen/avx-x8.c",
1779 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001780 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1781 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1782 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1783 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1784 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1785 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1786 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1787 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1788 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1789 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1790 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1791 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1792 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1793 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1794 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1795 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1796 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1797 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1798 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1799 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1800 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1801 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1802 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1803 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1804 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1805 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1806 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1807 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1808 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1809 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001810 "src/f32-hswish/gen/avx-x8.c",
1811 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001812 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1813 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1814 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1815 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1816 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1817 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1818 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1819 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1820 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001821 "src/f32-prelu/gen/avx-2x8.c",
1822 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001823 "src/f32-relu/gen/avx-x16.c",
1824 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001825 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001826 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1827 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1828 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1829 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1830 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1831 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1832 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1833 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001834 "src/f32-vbinary/gen/vmax-avx-x8.c",
1835 "src/f32-vbinary/gen/vmax-avx-x16.c",
1836 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1837 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1838 "src/f32-vbinary/gen/vmin-avx-x8.c",
1839 "src/f32-vbinary/gen/vmin-avx-x16.c",
1840 "src/f32-vbinary/gen/vminc-avx-x8.c",
1841 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001842 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1843 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1844 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1845 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1846 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1847 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1848 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1849 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001850 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1851 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1852 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1853 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001854 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1855 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1856 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1857 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001858 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1859 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001860 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001861 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1862 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1863 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1864 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1865 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1866 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1867 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1868 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001869 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1870 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001871 "src/f32-vunary/gen/vabs-avx-x8.c",
1872 "src/f32-vunary/gen/vabs-avx-x16.c",
1873 "src/f32-vunary/gen/vneg-avx-x8.c",
1874 "src/f32-vunary/gen/vneg-avx-x16.c",
1875 "src/f32-vunary/gen/vsqr-avx-x8.c",
1876 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001877]
1878
Marat Dukhanfda12b82019-11-21 12:27:59 -08001879FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001880 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1881 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1882 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1883 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1884 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1885 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1886 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1887 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1888 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1889 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1890 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1891 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1892 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1893 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1894 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1895 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1896 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1897 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1898 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1899 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1900 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1901 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1902 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1903 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1904 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1905 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1906 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1907 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1908 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1909 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1910 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1911 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1912 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1913 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1914 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1915 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1916 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1917 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1918 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1919 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001920 "src/f32-hswish/gen/fma3-x8.c",
1921 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001922 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1923 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1924 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1925 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1926 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1927 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1928 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1929 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1930 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1931 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1932 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1933 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1934 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1935 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001936 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1937 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1938 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1939 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1940 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1941 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1942 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1943 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001944 "src/math/sqrt-fma3-nr1fma.c",
1945 "src/math/sqrt-fma3-nr2fma.c",
1946 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001947]
1948
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001949AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001950 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1951 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1952 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1953 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1954 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1955 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1956 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1957 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1958 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1959 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1960 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1961 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1962 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1963 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1964 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1965 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1966 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1967 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1968 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1969 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1970 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1971 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1972 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1973 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1974 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1975 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1976 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1977 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1978 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1979 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1980 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1981 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1982 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1983 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1984 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1985 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001986 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1987 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1988 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1989 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1990 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1991 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1992 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1993 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1994 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1995 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1996 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1997 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1998 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1999 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2000 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2001 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2002 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2003 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2004 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2005 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2006 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2007 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2008 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2009 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2010 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2011 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2012 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2013 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2014 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2015 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002016 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2017 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2018 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2019 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2020 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2021 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2022 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2023 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2024 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2025 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2026 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2027 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2028 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2029 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2030 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2031 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2032 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2033 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2034 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2035 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2036 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2037 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2038 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2039 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002040 "src/math/exp-avx2-p5.c",
2041 "src/math/exp-avx2-perm-p3.c",
2042 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002043 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002044 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002045 "src/math/sigmoid-avx2-rr2-p5-div.c",
2046 "src/math/sigmoid-avx2-rr1-p5-div.c",
2047 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2048 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2049 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2050 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002051]
2052
Marat Dukhan08c4a432019-10-03 09:29:21 -07002053AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002054 "src/f32-clamp/gen/avx512f-x16.c",
2055 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002056 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2057 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2058 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2059 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2060 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2061 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2062 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2063 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2064 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2065 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2066 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2067 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2068 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2069 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2070 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2071 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2072 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2073 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2074 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2075 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2076 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2077 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2078 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2079 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08002080 "src/f32-hswish/gen/avx512f-x16.c",
2081 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002082 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2083 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2084 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2085 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2086 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2087 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002088 "src/f32-prelu/gen/avx512f-2x16.c",
2089 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002090 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2091 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2092 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2093 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2094 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2095 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2096 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2097 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2098 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2099 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2100 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2101 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2102 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2103 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2104 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2105 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2106 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2107 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2108 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2109 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2110 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2111 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2112 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2113 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2114 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2115 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2116 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2117 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2118 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2119 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2120 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2121 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2122 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2123 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2124 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2125 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002126 "src/f32-relu/gen/avx512f-x16.c",
2127 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002128 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002129 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2130 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2131 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2132 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2133 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2134 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2135 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2136 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002137 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2138 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2139 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2140 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2141 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2142 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2143 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2144 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002145 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2146 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2147 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2148 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2149 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2150 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2151 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2152 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002153 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2154 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2155 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2156 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002157 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2158 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2159 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2160 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002161 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2162 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002163 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002164 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2165 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2166 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2167 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2168 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2169 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2170 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2171 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2172 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2173 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2174 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2175 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2176 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2177 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2178 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2179 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2180 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2181 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2182 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2183 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2184 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2185 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2186 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2187 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002188 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2189 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2190 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2191 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2192 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2193 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2194 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2195 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002196 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2197 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2198 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2199 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2200 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2201 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2202 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2203 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002204 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2205 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2206 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2207 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2208 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2209 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002210 "src/math/exp-avx512f-p5-scalef.c",
2211 "src/math/exp-avx512f-p5.c",
2212 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002213 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002214 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002215 "src/math/sqrt-avx512f-nr1fma.c",
2216 "src/math/sqrt-avx512f-nr2fma.c",
2217 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002218]
2219
2220AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002221 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002222 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002223 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2224 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2225 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2226 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002227 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2228 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2229 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2230 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002231 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2232 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2233 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2234 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002235]
2236
2237AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002238 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002239 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2240 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002241 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002242 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2243 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002244 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2245 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2246 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002247 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002248 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2249 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2250 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002251 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002252 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2253 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002254 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002255 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2256 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2257 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2258 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2259 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2260 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2261 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2262 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2263 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2264 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2265 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2266 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2267 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2268 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2269 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2270 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2271 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2272 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002273 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2274 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002275 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002276 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2277 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2278 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2279 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2280 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2281 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2282 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2283 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2284 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2285 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2286 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2287 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2288 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2289 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2290 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2291 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2292 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2293 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002294 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2295 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2296 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2297 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2298 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2299 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2300 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2301 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2302 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2303 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2304 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2305 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2306 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2307 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2308 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2309 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2310 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2311 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002312]
2313
Marat Dukhan1b354632020-03-23 12:50:22 -07002314INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002315 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002316 "src/xnnpack/argmaxpool.h",
2317 "src/xnnpack/avgpool.h",
2318 "src/xnnpack/clamp.h",
2319 "src/xnnpack/common.h",
2320 "src/xnnpack/conv.h",
2321 "src/xnnpack/dwconv.h",
2322 "src/xnnpack/gavgpool.h",
2323 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002324 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002326 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002327 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002328 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002329 "src/xnnpack/lut.h",
2330 "src/xnnpack/math.h",
2331 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002332 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002333 "src/xnnpack/packx.h",
2334 "src/xnnpack/pad.h",
2335 "src/xnnpack/params.h",
2336 "src/xnnpack/pavgpool.h",
2337 "src/xnnpack/ppmm.h",
2338 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002339 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002340 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002341 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002342 "src/xnnpack/rmax.h",
2343 "src/xnnpack/scalar-utils.h",
2344 "src/xnnpack/spmm.h",
2345 "src/xnnpack/unpool.h",
2346 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002347 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002348 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002349 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002350 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002351 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002352 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002353 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002354]
2355
2356INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357 "include/xnnpack.h",
2358 "src/xnnpack/allocator.h",
2359 "src/xnnpack/compute.h",
2360 "src/xnnpack/im2col.h",
2361 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002362 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002363 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002364 "src/xnnpack/operator.h",
2365 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002366 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002367 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002368 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002369 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002370]
2371
Marat Dukhan1b354632020-03-23 12:50:22 -07002372ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002373 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002374]
2375
Marat Dukhan1b354632020-03-23 12:50:22 -07002376MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002377 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002378 "include/xnnpack.h",
2379]
2380
Marat Dukhan1b354632020-03-23 12:50:22 -07002381MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002382 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002383 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002384 "src/xnnpack/requantization.h",
2385 "include/xnnpack.h",
2386]
2387
2388OPERATOR_TEST_PARAMS_HDRS = [
2389 "src/xnnpack/params.h",
2390 "src/xnnpack/common.h",
2391]
2392
2393WEIGHTS_PACK_HDRS = [
2394 "src/xnnpack/pack.h",
2395 "src/xnnpack/operator.h",
2396 "src/xnnpack/compute.h",
2397]
2398
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002399LOGGING_COPTS = select({
2400 # No logging in optimized mode
2401 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2402 # Full logging in debug mode
2403 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2404 # Error-only logging in default (fastbuild) mode
2405 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2406})
2407
Marat Dukhan3b59de22020-06-03 20:15:19 -07002408LOGGING_SRCS = select({
2409 # No logging in optimized mode
2410 ":optimized_build": [],
2411 "//conditions:default": [
2412 "src/operator-strings.c",
2413 "src/subgraph-strings.c",
2414 ],
2415})
2416
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002417LOGGING_HDRS = [
2418 "src/xnnpack/log.h",
2419]
2420
Marat Dukhan08c4a432019-10-03 09:29:21 -07002421xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002422 name = "tables",
2423 srcs = TABLE_SRCS,
2424 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002425 gcc_copts = xnnpack_gcc_std_copts(),
2426 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002427)
2428
2429xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002430 name = "scalar_ukernels",
2431 srcs = SCALAR_UKERNELS,
2432 hdrs = INTERNAL_HDRS,
2433 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002434 gcc_copts = xnnpack_gcc_std_copts(),
2435 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002436 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002437 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002438 "@FP16",
2439 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002440 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002441 ],
2442)
2443
2444xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002445 name = "scalar_ukernels_test_mode",
2446 srcs = SCALAR_UKERNELS,
2447 hdrs = INTERNAL_HDRS,
2448 aarch32_copts = ["-marm"],
2449 copts = [
2450 "-UNDEBUG",
2451 "-DXNN_TEST_MODE=1",
2452 ],
2453 gcc_copts = xnnpack_gcc_std_copts(),
2454 msvc_copts = xnnpack_msvc_std_copts(),
2455 deps = [
2456 ":tables",
2457 "@FP16",
2458 "@FXdiv",
2459 "@pthreadpool",
2460 ],
2461)
2462
2463xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002464 name = "wasm_ukernels",
2465 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002466 gcc_copts = xnnpack_gcc_std_copts(),
2467 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002468 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002469 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002470 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002471 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002472 "@FP16",
2473 "@FXdiv",
2474 "@pthreadpool",
2475 ],
2476)
2477
2478xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002479 name = "wasm_ukernels_test_mode",
2480 hdrs = INTERNAL_HDRS,
2481 copts = [
2482 "-UNDEBUG",
2483 "-DXNN_TEST_MODE=1",
2484 ],
2485 gcc_copts = xnnpack_gcc_std_copts(),
2486 msvc_copts = xnnpack_msvc_std_copts(),
2487 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002488 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002489 deps = [
2490 ":tables",
2491 "@FP16",
2492 "@FXdiv",
2493 "@pthreadpool",
2494 ],
2495)
2496
2497xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002498 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002499 hdrs = INTERNAL_HDRS,
2500 aarch32_copts = [
2501 "-marm",
2502 "-mfpu=neon",
2503 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002504 gcc_copts = xnnpack_gcc_std_copts(),
2505 gcc_x86_copts = ["-msse2"],
2506 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002507 optimized_copts = [
2508 "-O3",
2509 "-ffast-math",
2510 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002511 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002512 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002513 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002514 "@FP16",
2515 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002516 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002517 ],
2518)
2519
2520xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002521 name = "psimd_fastmath_ukernels_test_mode",
2522 hdrs = INTERNAL_HDRS,
2523 aarch32_copts = [
2524 "-marm",
2525 "-mfpu=neon",
2526 ],
2527 copts = [
2528 "-UNDEBUG",
2529 "-DXNN_TEST_MODE=1",
2530 ],
2531 gcc_copts = xnnpack_gcc_std_copts(),
2532 gcc_x86_copts = ["-msse2"],
2533 msvc_copts = xnnpack_msvc_std_copts(),
2534 optimized_copts = [
2535 "-O3",
2536 "-ffast-math",
2537 ],
2538 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2539 deps = [
2540 ":tables",
2541 "@FP16",
2542 "@psimd",
2543 "@pthreadpool",
2544 ],
2545)
2546
2547xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002548 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002549 hdrs = INTERNAL_HDRS,
2550 aarch32_copts = [
2551 "-marm",
2552 "-mfpu=neon",
2553 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002554 gcc_copts = xnnpack_gcc_std_copts(),
2555 gcc_x86_copts = ["-msse2"],
2556 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002557 optimized_copts = [
2558 "-O3",
2559 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002560 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002561 deps = [
2562 ":tables",
2563 "@FP16",
2564 "@psimd",
2565 "@pthreadpool",
2566 ],
2567)
2568
2569xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002570 name = "psimd_accmath_ukernels_test_mode",
2571 hdrs = INTERNAL_HDRS,
2572 aarch32_copts = [
2573 "-marm",
2574 "-mfpu=neon",
2575 ],
2576 copts = [
2577 "-UNDEBUG",
2578 "-DXNN_TEST_MODE=1",
2579 ],
2580 gcc_copts = xnnpack_gcc_std_copts(),
2581 gcc_x86_copts = ["-msse2"],
2582 msvc_copts = xnnpack_msvc_std_copts(),
2583 optimized_copts = [
2584 "-O3",
2585 ],
2586 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2587 deps = [
2588 ":tables",
2589 "@FP16",
2590 "@psimd",
2591 "@pthreadpool",
2592 ],
2593)
2594
2595xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002596 name = "neon_ukernels",
2597 hdrs = INTERNAL_HDRS,
2598 aarch32_copts = [
2599 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002600 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002601 "-mfpu=neon",
2602 ],
2603 aarch32_srcs = NEON_UKERNELS,
2604 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002605 gcc_copts = xnnpack_gcc_std_copts(),
2606 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002607 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002608 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002609 "@FP16",
2610 "@pthreadpool",
2611 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002612)
2613
2614xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002615 name = "neon_ukernels_test_mode",
2616 hdrs = INTERNAL_HDRS,
2617 aarch32_copts = [
2618 "-marm",
2619 "-march=armv7-a",
2620 "-mfpu=neon",
2621 ],
2622 aarch32_srcs = NEON_UKERNELS,
2623 aarch64_srcs = NEON_UKERNELS,
2624 copts = [
2625 "-UNDEBUG",
2626 "-DXNN_TEST_MODE=1",
2627 ],
2628 gcc_copts = xnnpack_gcc_std_copts(),
2629 msvc_copts = xnnpack_msvc_std_copts(),
2630 deps = [
2631 ":tables",
2632 "@FP16",
2633 "@pthreadpool",
2634 ],
2635)
2636
2637xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002638 name = "neonfma_ukernels",
2639 hdrs = INTERNAL_HDRS,
2640 aarch32_copts = [
2641 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002642 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002643 "-mfpu=neon-vfpv4",
2644 ],
2645 aarch32_srcs = NEONFMA_UKERNELS,
2646 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002647 apple_aarch32_copts = [
2648 "-mcpu=swift",
2649 "-mtune=generic",
2650 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002651 gcc_copts = xnnpack_gcc_std_copts(),
2652 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002653 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002654 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002655 "@FP16",
2656 "@pthreadpool",
2657 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002658)
2659
2660xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002661 name = "neonfma_ukernels_test_mode",
2662 hdrs = INTERNAL_HDRS,
2663 aarch32_copts = [
2664 "-marm",
2665 "-march=armv7-a",
2666 "-mfpu=neon-vfpv4",
2667 ],
2668 aarch32_srcs = NEONFMA_UKERNELS,
2669 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002670 apple_aarch32_copts = [
2671 "-mcpu=swift",
2672 "-mtune=generic",
2673 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002674 copts = [
2675 "-UNDEBUG",
2676 "-DXNN_TEST_MODE=1",
2677 ],
2678 gcc_copts = xnnpack_gcc_std_copts(),
2679 msvc_copts = xnnpack_msvc_std_copts(),
2680 deps = [
2681 ":tables",
2682 "@FP16",
2683 "@pthreadpool",
2684 ],
2685)
2686
2687xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002688 name = "neonv8_ukernels",
2689 hdrs = INTERNAL_HDRS,
2690 aarch32_copts = [
2691 "-marm",
2692 "-march=armv8-a",
2693 "-mfpu=neon-fp-armv8",
2694 ],
2695 aarch32_srcs = NEONV8_UKERNELS,
2696 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002697 apple_aarch32_copts = [
2698 "-mcpu=cyclone",
2699 "-mtune=generic",
2700 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002701 gcc_copts = xnnpack_gcc_std_copts(),
2702 msvc_copts = xnnpack_msvc_std_copts(),
2703 deps = [
2704 ":tables",
2705 "@FP16",
2706 "@pthreadpool",
2707 ],
2708)
2709
2710xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002711 name = "neonv8_ukernels_test_mode",
2712 hdrs = INTERNAL_HDRS,
2713 aarch32_copts = [
2714 "-marm",
2715 "-march=armv8-a",
2716 "-mfpu=neon-fp-armv8",
2717 ],
2718 aarch32_srcs = NEONV8_UKERNELS,
2719 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002720 apple_aarch32_copts = [
2721 "-mcpu=cyclone",
2722 "-mtune=generic",
2723 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002724 copts = [
2725 "-UNDEBUG",
2726 "-DXNN_TEST_MODE=1",
2727 ],
2728 gcc_copts = xnnpack_gcc_std_copts(),
2729 msvc_copts = xnnpack_msvc_std_copts(),
2730 deps = [
2731 ":tables",
2732 "@FP16",
2733 "@pthreadpool",
2734 ],
2735)
2736
2737xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002738 name = "neonfp16arith_ukernels",
2739 hdrs = INTERNAL_HDRS,
2740 aarch64_copts = ["-march=armv8.2-a+fp16"],
2741 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002742 gcc_copts = xnnpack_gcc_std_copts(),
2743 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002744 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002745 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002746 "@FP16",
2747 "@pthreadpool",
2748 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002749)
2750
2751xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002752 name = "neonfp16arith_ukernels_test_mode",
2753 hdrs = INTERNAL_HDRS,
2754 aarch64_copts = ["-march=armv8.2-a+fp16"],
2755 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2756 copts = [
2757 "-UNDEBUG",
2758 "-DXNN_TEST_MODE=1",
2759 ],
2760 gcc_copts = xnnpack_gcc_std_copts(),
2761 msvc_copts = xnnpack_msvc_std_copts(),
2762 deps = [
2763 ":tables",
2764 "@FP16",
2765 "@pthreadpool",
2766 ],
2767)
2768
2769xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002770 name = "sse2_ukernels",
2771 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002772 gcc_copts = xnnpack_gcc_std_copts(),
2773 gcc_x86_copts = ["-msse2"],
2774 msvc_copts = xnnpack_msvc_std_copts(),
2775 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002776 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002777 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002778 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002779 "@FP16",
2780 "@pthreadpool",
2781 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002782)
2783
2784xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002785 name = "sse2_ukernels_test_mode",
2786 hdrs = INTERNAL_HDRS,
2787 copts = [
2788 "-UNDEBUG",
2789 "-DXNN_TEST_MODE=1",
2790 ],
2791 gcc_copts = xnnpack_gcc_std_copts(),
2792 gcc_x86_copts = ["-msse2"],
2793 msvc_copts = xnnpack_msvc_std_copts(),
2794 msvc_x86_32_copts = ["/arch:SSE2"],
2795 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2796 deps = [
2797 ":tables",
2798 "@FP16",
2799 "@pthreadpool",
2800 ],
2801)
2802
2803xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002804 name = "ssse3_ukernels",
2805 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002806 gcc_copts = xnnpack_gcc_std_copts(),
2807 gcc_x86_copts = ["-mssse3"],
2808 msvc_copts = xnnpack_msvc_std_copts(),
2809 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002810 x86_srcs = SSSE3_UKERNELS,
2811 deps = [
2812 ":tables",
2813 "@FP16",
2814 "@pthreadpool",
2815 ],
2816)
2817
2818xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002819 name = "ssse3_ukernels_test_mode",
2820 hdrs = INTERNAL_HDRS,
2821 copts = [
2822 "-UNDEBUG",
2823 "-DXNN_TEST_MODE=1",
2824 ],
2825 gcc_copts = xnnpack_gcc_std_copts(),
2826 gcc_x86_copts = ["-mssse3"],
2827 msvc_copts = xnnpack_msvc_std_copts(),
2828 msvc_x86_32_copts = ["/arch:SSE2"],
2829 x86_srcs = SSSE3_UKERNELS,
2830 deps = [
2831 ":tables",
2832 "@FP16",
2833 "@pthreadpool",
2834 ],
2835)
2836
2837xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002838 name = "sse41_ukernels",
2839 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002840 gcc_copts = xnnpack_gcc_std_copts(),
2841 gcc_x86_copts = ["-msse4.1"],
2842 msvc_copts = xnnpack_msvc_std_copts(),
2843 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002844 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002845 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002846 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002847 "@FP16",
2848 "@pthreadpool",
2849 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002850)
2851
2852xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002853 name = "sse41_ukernels_test_mode",
2854 hdrs = INTERNAL_HDRS,
2855 copts = [
2856 "-UNDEBUG",
2857 "-DXNN_TEST_MODE=1",
2858 ],
2859 gcc_copts = xnnpack_gcc_std_copts(),
2860 gcc_x86_copts = ["-msse4.1"],
2861 msvc_copts = xnnpack_msvc_std_copts(),
2862 msvc_x86_32_copts = ["/arch:SSE2"],
2863 x86_srcs = SSE41_UKERNELS,
2864 deps = [
2865 ":tables",
2866 "@FP16",
2867 "@pthreadpool",
2868 ],
2869)
2870
2871xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872 name = "avx_ukernels",
2873 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002874 gcc_copts = xnnpack_gcc_std_copts(),
2875 gcc_x86_copts = ["-mavx"],
2876 msvc_copts = xnnpack_msvc_std_copts(),
2877 msvc_x86_32_copts = ["/arch:AVX"],
2878 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002879 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002880 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002881 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002882 "@FP16",
2883 "@pthreadpool",
2884 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885)
2886
2887xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002888 name = "avx_ukernels_test_mode",
2889 hdrs = INTERNAL_HDRS,
2890 copts = [
2891 "-UNDEBUG",
2892 "-DXNN_TEST_MODE=1",
2893 ],
2894 gcc_copts = xnnpack_gcc_std_copts(),
2895 gcc_x86_copts = ["-mavx"],
2896 msvc_copts = xnnpack_msvc_std_copts(),
2897 msvc_x86_32_copts = ["/arch:AVX"],
2898 msvc_x86_64_copts = ["/arch:AVX"],
2899 x86_srcs = AVX_UKERNELS,
2900 deps = [
2901 ":tables",
2902 "@FP16",
2903 "@pthreadpool",
2904 ],
2905)
2906
2907xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002908 name = "fma3_ukernels",
2909 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002910 gcc_copts = xnnpack_gcc_std_copts(),
2911 gcc_x86_copts = ["-mfma"],
2912 msvc_copts = xnnpack_msvc_std_copts(),
2913 msvc_x86_32_copts = ["/arch:AVX"],
2914 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002915 x86_srcs = FMA3_UKERNELS,
2916 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002917 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002918 "@FP16",
2919 "@pthreadpool",
2920 ],
2921)
2922
2923xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002924 name = "fma3_ukernels_test_mode",
2925 hdrs = INTERNAL_HDRS,
2926 copts = [
2927 "-UNDEBUG",
2928 "-DXNN_TEST_MODE=1",
2929 ],
2930 gcc_copts = xnnpack_gcc_std_copts(),
2931 gcc_x86_copts = ["-mfma"],
2932 msvc_copts = xnnpack_msvc_std_copts(),
2933 msvc_x86_32_copts = ["/arch:AVX"],
2934 msvc_x86_64_copts = ["/arch:AVX"],
2935 x86_srcs = FMA3_UKERNELS,
2936 deps = [
2937 ":tables",
2938 "@FP16",
2939 "@pthreadpool",
2940 ],
2941)
2942
2943xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002944 name = "avx2_ukernels",
2945 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002946 gcc_copts = xnnpack_gcc_std_copts(),
2947 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002948 "-mfma",
2949 "-mavx2",
2950 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002951 msvc_copts = xnnpack_msvc_std_copts(),
2952 msvc_x86_32_copts = ["/arch:AVX2"],
2953 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002954 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002955 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002956 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002957 "@FP16",
2958 "@pthreadpool",
2959 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002960)
2961
2962xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002963 name = "avx2_ukernels_test_mode",
2964 hdrs = INTERNAL_HDRS,
2965 copts = [
2966 "-UNDEBUG",
2967 "-DXNN_TEST_MODE=1",
2968 ],
2969 gcc_copts = xnnpack_gcc_std_copts(),
2970 gcc_x86_copts = [
2971 "-mfma",
2972 "-mavx2",
2973 ],
2974 msvc_copts = xnnpack_msvc_std_copts(),
2975 msvc_x86_32_copts = ["/arch:AVX2"],
2976 msvc_x86_64_copts = ["/arch:AVX2"],
2977 x86_srcs = AVX2_UKERNELS,
2978 deps = [
2979 ":tables",
2980 "@FP16",
2981 "@pthreadpool",
2982 ],
2983)
2984
2985xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002986 name = "avx512f_ukernels",
2987 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002988 gcc_copts = xnnpack_gcc_std_copts(),
2989 gcc_x86_copts = ["-mavx512f"],
2990 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2991 msvc_copts = xnnpack_msvc_std_copts(),
2992 msvc_x86_32_copts = ["/arch:AVX512"],
2993 msvc_x86_64_copts = ["/arch:AVX512"],
2994 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002996 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002997 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002998 "@FP16",
2999 "@pthreadpool",
3000 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003001)
3002
3003xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003004 name = "avx512f_ukernels_test_mode",
3005 hdrs = INTERNAL_HDRS,
3006 copts = [
3007 "-UNDEBUG",
3008 "-DXNN_TEST_MODE=1",
3009 ],
3010 gcc_copts = xnnpack_gcc_std_copts(),
3011 gcc_x86_copts = ["-mavx512f"],
3012 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3013 msvc_copts = xnnpack_msvc_std_copts(),
3014 msvc_x86_32_copts = ["/arch:AVX512"],
3015 msvc_x86_64_copts = ["/arch:AVX512"],
3016 msys_copts = ["-fno-asynchronous-unwind-tables"],
3017 x86_srcs = AVX512F_UKERNELS,
3018 deps = [
3019 ":tables",
3020 "@FP16",
3021 "@pthreadpool",
3022 ],
3023)
3024
3025xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003026 name = "asm_ukernels",
3027 hdrs = ["src/xnnpack/assembly.h"],
3028 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003029 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030 aarch64_srcs = AARCH64_ASM_UKERNELS,
3031)
3032
Marat Dukhan3b59de22020-06-03 20:15:19 -07003033xnnpack_cc_library(
3034 name = "logging_utils",
3035 srcs = LOGGING_SRCS,
3036 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3037 copts = LOGGING_COPTS + [
3038 "-Isrc",
3039 "-Iinclude",
3040 ] + select({
3041 ":debug_build": [],
3042 "//conditions:default": xnnpack_min_size_copts(),
3043 }),
3044 gcc_copts = xnnpack_gcc_std_copts(),
3045 msvc_copts = xnnpack_msvc_std_copts(),
3046 visibility = xnnpack_visibility(),
3047 deps = [
3048 "@FP16",
3049 "@clog",
3050 "@pthreadpool",
3051 ],
3052)
3053
Marat Dukhan08c4a432019-10-03 09:29:21 -07003054xnnpack_aggregate_library(
3055 name = "ukernels",
3056 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003057 ":neon_ukernels",
3058 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003059 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003060 ":asm_ukernels",
3061 ],
3062 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003063 ":neon_ukernels",
3064 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003065 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003066 ":neonfp16arith_ukernels",
3067 ":asm_ukernels",
3068 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003069 generic_deps = [
3070 ":scalar_ukernels",
3071 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003072 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003073 ":psimd_fastmath_ukernels",
3074 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003075 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003076 wasm_deps = [
3077 ":wasm_ukernels",
3078 ],
3079 wasmsimd_deps = [
3080 ":wasm_ukernels",
3081 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003082 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003083 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003084 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003085 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003087 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003088 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003089 ":avx512f_ukernels",
3090 ],
3091)
3092
Marat Dukhan33fcf782020-05-24 14:27:15 -07003093xnnpack_aggregate_library(
3094 name = "ukernels_test_mode",
3095 aarch32_deps = [
3096 ":neon_ukernels_test_mode",
3097 ":neonfma_ukernels_test_mode",
3098 ":neonv8_ukernels_test_mode",
3099 ":asm_ukernels",
3100 ],
3101 aarch64_deps = [
3102 ":neon_ukernels_test_mode",
3103 ":neonfma_ukernels_test_mode",
3104 ":neonv8_ukernels_test_mode",
3105 ":neonfp16arith_ukernels_test_mode",
3106 ":asm_ukernels",
3107 ],
3108 generic_deps = [
3109 ":scalar_ukernels_test_mode",
3110 ],
3111 psimd_deps = [
3112 ":psimd_fastmath_ukernels_test_mode",
3113 ":psimd_accmath_ukernels_test_mode",
3114 ],
3115 wasm_deps = [
3116 ":wasm_ukernels_test_mode",
3117 ],
3118 wasmsimd_deps = [
3119 ":wasm_ukernels_test_mode",
3120 ],
3121 x86_deps = [
3122 ":sse2_ukernels_test_mode",
3123 ":ssse3_ukernels_test_mode",
3124 ":sse41_ukernels_test_mode",
3125 ":avx_ukernels_test_mode",
3126 ":fma3_ukernels_test_mode",
3127 ":avx2_ukernels_test_mode",
3128 ":avx512f_ukernels_test_mode",
3129 ],
3130)
3131
Marat Dukhan86df3842020-06-17 14:03:31 -07003132# This target must be compatible with Android NDK r17 and compile with Clang 6.
3133# It is similar to :ukernels target, but excludes microkernels which require
3134# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003135xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003136 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003137 aarch32_deps = [
3138 ":neon_ukernels",
3139 ":neonfma_ukernels",
3140 ":neonv8_ukernels",
3141 ":asm_ukernels",
3142 ],
3143 aarch64_deps = [
3144 ":neon_ukernels",
3145 ":neonfma_ukernels",
3146 ":neonv8_ukernels",
3147 ":asm_ukernels",
3148 ],
3149 generic_deps = [
3150 ":scalar_ukernels",
3151 ],
3152 psimd_deps = [
3153 ":psimd_fastmath_ukernels",
3154 ":psimd_accmath_ukernels",
3155 ],
3156 wasm_deps = [
3157 ":wasm_ukernels",
3158 ],
3159 wasmsimd_deps = [
3160 ":wasm_ukernels",
3161 ],
3162 x86_deps = [
3163 ":sse2_ukernels",
3164 ":ssse3_ukernels",
3165 ":sse41_ukernels",
3166 ":avx_ukernels",
3167 ":fma3_ukernels",
3168 ":avx2_ukernels",
3169 ":avx512f_ukernels",
3170 ],
3171)
3172
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173xnnpack_cc_library(
3174 name = "im2col",
3175 srcs = ["src/im2col.c"],
3176 hdrs = [
3177 "src/xnnpack/common.h",
3178 "src/xnnpack/im2col.h",
3179 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003180 gcc_copts = xnnpack_gcc_std_copts(),
3181 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003182)
3183
3184xnnpack_cc_library(
3185 name = "indirection",
3186 srcs = ["src/indirection.c"],
3187 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003188 gcc_copts = xnnpack_gcc_std_copts(),
3189 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190 deps = [
3191 "@FP16",
3192 "@FXdiv",
3193 "@pthreadpool",
3194 ],
3195)
3196
3197xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003198 name = "indirection_test_mode",
3199 srcs = ["src/indirection.c"],
3200 hdrs = INTERNAL_HDRS,
3201 copts = [
3202 "-UNDEBUG",
3203 "-DXNN_TEST_MODE=1",
3204 ],
3205 gcc_copts = xnnpack_gcc_std_copts(),
3206 msvc_copts = xnnpack_msvc_std_copts(),
3207 deps = [
3208 "@FP16",
3209 "@FXdiv",
3210 "@pthreadpool",
3211 ],
3212)
3213
3214xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003215 name = "packing",
3216 srcs = ["src/packing.c"],
3217 hdrs = INTERNAL_HDRS,
3218 gcc_copts = xnnpack_gcc_std_copts(),
3219 msvc_copts = xnnpack_msvc_std_copts(),
3220 deps = [
3221 "@FP16",
3222 "@FXdiv",
3223 "@pthreadpool",
3224 ],
3225)
3226
3227xnnpack_cc_library(
3228 name = "packing_test_mode",
3229 srcs = ["src/packing.c"],
3230 hdrs = INTERNAL_HDRS,
3231 copts = [
3232 "-UNDEBUG",
3233 "-DXNN_TEST_MODE=1",
3234 ],
3235 gcc_copts = xnnpack_gcc_std_copts(),
3236 msvc_copts = xnnpack_msvc_std_copts(),
3237 deps = [
3238 "@FP16",
3239 "@FXdiv",
3240 "@pthreadpool",
3241 ],
3242)
3243
3244xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245 name = "operator_run",
3246 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003247 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003248 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003249 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3250 "//conditions:default": [],
3251 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003252 gcc_copts = xnnpack_gcc_std_copts(),
3253 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003254 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003255 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003256 "@FP16",
3257 "@FXdiv",
3258 "@clog",
3259 "@pthreadpool",
3260 ],
3261)
3262
Chao Mei6ddfc602020-05-13 22:29:36 -07003263xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003264 name = "operator_run_test_mode",
3265 srcs = ["src/operator-run.c"],
3266 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3267 copts = LOGGING_COPTS + [
3268 "-UNDEBUG",
3269 "-DXNN_TEST_MODE=1",
3270 ] + select({
3271 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3272 "//conditions:default": [],
3273 }),
3274 gcc_copts = xnnpack_gcc_std_copts(),
3275 msvc_copts = xnnpack_msvc_std_copts(),
3276 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003277 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003278 "@FP16",
3279 "@FXdiv",
3280 "@clog",
3281 "@pthreadpool",
3282 ],
3283)
3284
3285xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003286 name = "memory_planner",
3287 srcs = ["src/memory-planner.c"],
3288 hdrs = INTERNAL_HDRS,
3289 defines = select({
3290 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3291 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3292 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3293 }),
3294 gcc_copts = xnnpack_gcc_std_copts(),
3295 msvc_copts = xnnpack_msvc_std_copts(),
3296 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003297 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003298 "@pthreadpool",
3299 ],
3300)
3301
Marat Dukhan33fcf782020-05-24 14:27:15 -07003302xnnpack_cc_library(
3303 name = "memory_planner_test_mode",
3304 srcs = ["src/memory-planner.c"],
3305 hdrs = INTERNAL_HDRS,
3306 copts = [
3307 "-UNDEBUG",
3308 "-DXNN_TEST_MODE=1",
3309 ],
3310 defines = select({
3311 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3312 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3313 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3314 }),
3315 gcc_copts = xnnpack_gcc_std_copts(),
3316 msvc_copts = xnnpack_msvc_std_copts(),
3317 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003318 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003319 "@pthreadpool",
3320 ],
3321)
3322
Marat Dukhan08c4a432019-10-03 09:29:21 -07003323cc_library(
3324 name = "enable_assembly",
3325 defines = select({
3326 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3327 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003328 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003329 }),
3330)
3331
Marat Dukhan9de90e02020-06-18 16:04:12 -07003332cc_library(
3333 name = "enable_sparse",
3334 defines = select({
3335 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3336 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003337 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003338 }),
3339)
3340
Marat Dukhancf056b22019-10-07 10:26:29 -07003341xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003342 name = "operators",
3343 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003344 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003345 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003346 ],
3347 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003348 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003349 "-Isrc",
3350 "-Iinclude",
3351 ] + select({
3352 ":debug_build": [],
3353 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003354 }) + select({
3355 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3356 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003357 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003358 gcc_copts = xnnpack_gcc_std_copts(),
3359 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003360 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003361 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003362 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003363 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364 "@FP16",
3365 "@FXdiv",
3366 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003367 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003368 ],
3369)
3370
Marat Dukhan10a38082020-04-17 03:58:35 -07003371xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003372 name = "operators_test_mode",
3373 srcs = OPERATOR_SRCS + [
3374 "src/memory.c",
3375 "src/operator-delete.c",
3376 ],
3377 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3378 copts = LOGGING_COPTS + [
3379 "-Isrc",
3380 "-Iinclude",
3381 "-UNDEBUG",
3382 "-DXNN_TEST_MODE=1",
3383 ] + select({
3384 ":debug_build": [],
3385 "//conditions:default": xnnpack_min_size_copts(),
3386 }) + select({
3387 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3388 "//conditions:default": [],
3389 }),
3390 gcc_copts = xnnpack_gcc_std_copts(),
3391 msvc_copts = xnnpack_msvc_std_copts(),
3392 deps = [
3393 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003394 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003395 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003396 "@FP16",
3397 "@FXdiv",
3398 "@clog",
3399 "@pthreadpool",
3400 ],
3401)
3402
3403xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003404 name = "XNNPACK",
3405 srcs = [
3406 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003407 "src/runtime.c",
3408 "src/subgraph.c",
3409 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003410 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003411 hdrs = ["include/xnnpack.h"],
3412 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003413 "-Isrc",
3414 "-Iinclude",
3415 ] + select({
3416 ":debug_build": [],
3417 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003418 }) + select({
3419 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3420 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003421 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003422 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003423 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003424 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003425 visibility = xnnpack_visibility(),
3426 deps = [
3427 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003428 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003429 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003430 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003431 ":operator_run",
3432 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003433 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003434 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003435 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003436 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003437 ] + select({
3438 ":emscripten": [],
3439 "//conditions:default": ["@cpuinfo"],
3440 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003441)
3442
Marat Dukhan10a38082020-04-17 03:58:35 -07003443xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003444 name = "XNNPACK_test_mode",
3445 srcs = [
3446 "src/init.c",
3447 "src/runtime.c",
3448 "src/subgraph.c",
3449 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003450 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003451 hdrs = ["include/xnnpack.h"],
3452 copts = LOGGING_COPTS + [
3453 "-Isrc",
3454 "-Iinclude",
3455 "-UNDEBUG",
3456 "-DXNN_TEST_MODE=1",
3457 ] + select({
3458 ":debug_build": [],
3459 "//conditions:default": xnnpack_min_size_copts(),
3460 }) + select({
3461 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3462 "//conditions:default": [],
3463 }),
3464 gcc_copts = xnnpack_gcc_std_copts(),
3465 includes = ["include"],
3466 msvc_copts = xnnpack_msvc_std_copts(),
3467 visibility = xnnpack_visibility(),
3468 deps = [
3469 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003470 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003471 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003472 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003473 ":operator_run_test_mode",
3474 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003475 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003476 "@clog",
3477 "@FP16",
3478 "@pthreadpool",
3479 ] + select({
3480 ":emscripten": [],
3481 "//conditions:default": ["@cpuinfo"],
3482 }),
3483)
3484
3485xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003486 name = "xnnpack_f32",
3487 srcs = [
3488 "src/init.c",
3489 "src/runtime.c",
3490 "src/subgraph.c",
3491 "src/tensor.c",
3492 ] + SUBGRAPH_SRCS,
3493 hdrs = ["include/xnnpack.h"],
3494 copts = LOGGING_COPTS + [
3495 "-Isrc",
3496 "-Iinclude",
3497 ] + select({
3498 ":debug_build": [],
3499 "//conditions:default": xnnpack_min_size_copts(),
3500 }) + select({
3501 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3502 "//conditions:default": [],
3503 }),
3504 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003505 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003506 "XNN_NO_U8_OPERATORS",
3507 "XNN_NO_X8_OPERATORS",
3508 "XNN_NO_F16_OPERATORS",
3509 "XNN_NO_X16_OPERATORS",
3510 ],
3511 gcc_copts = xnnpack_gcc_std_copts(),
3512 includes = ["include"],
3513 msvc_copts = xnnpack_msvc_std_copts(),
3514 visibility = xnnpack_visibility(),
3515 deps = [
3516 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003517 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003518 ":logging_utils",
3519 ":memory_planner",
3520 ":operator_run",
3521 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003522 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003523 "@clog",
3524 "@FP16",
3525 "@pthreadpool",
3526 ] + select({
3527 ":emscripten": [],
3528 "//conditions:default": ["@cpuinfo"],
3529 }),
3530)
3531
3532xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003533 name = "xnnpack_operators_nhwc_f32",
3534 srcs = [
3535 "src/init.c",
3536 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003537 hdrs = ["include/xnnpack.h"],
3538 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003539 "-Isrc",
3540 "-Iinclude",
3541 ] + select({
3542 ":debug_build": [],
3543 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003544 }) + select({
3545 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3546 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003547 }),
3548 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003549 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003550 "XNN_NO_U8_OPERATORS",
3551 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003552 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003553 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003554 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003556 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557 visibility = xnnpack_visibility(),
3558 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003559 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003560 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561 ":operator_run",
3562 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003563 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003564 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003565 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003566 ] + select({
3567 ":emscripten": [],
3568 "//conditions:default": ["@cpuinfo"],
3569 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003570)
3571
Marat Dukhancf056b22019-10-07 10:26:29 -07003572xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003573 name = "bench_utils",
3574 srcs = ["bench/utils.cc"],
3575 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003576 deps = [
3577 "@com_google_benchmark//:benchmark",
3578 "@cpuinfo",
3579 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003580)
3581
Frank Barchard7e955972019-10-11 10:34:25 -07003582######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583
3584xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003585 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 srcs = [
3587 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003588 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589 "src/xnnpack/AlignedAllocator.h",
3590 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003591 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003592 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593)
3594
3595xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003596 name = "f16_igemm_bench",
3597 srcs = [
3598 "bench/f16-igemm.cc",
3599 "bench/conv.h",
3600 "bench/google/conv.h",
3601 "src/xnnpack/AlignedAllocator.h",
3602 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003603 deps = MICROKERNEL_BENCHMARK_DEPS + [
3604 ":indirection",
3605 ":packing",
3606 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003607)
3608
3609xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610 name = "f16_gemm_bench",
3611 srcs = [
3612 "bench/f16-gemm.cc",
3613 "bench/gemm.h",
3614 "src/xnnpack/AlignedAllocator.h",
3615 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003616 deps = MICROKERNEL_BENCHMARK_DEPS + [
3617 ":packing",
3618 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003619)
3620
3621xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003622 name = "f16_spmm_bench",
3623 srcs = [
3624 "bench/f16-spmm.cc",
3625 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003626 "src/xnnpack/AlignedAllocator.h",
3627 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003628 deps = MICROKERNEL_BENCHMARK_DEPS,
3629)
3630
3631xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632 name = "f32_igemm_bench",
3633 srcs = [
3634 "bench/f32-igemm.cc",
3635 "bench/conv.h",
3636 "src/xnnpack/AlignedAllocator.h",
3637 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003638 deps = MICROKERNEL_BENCHMARK_DEPS + [
3639 ":indirection",
3640 ":packing",
3641 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642)
3643
3644xnnpack_benchmark(
3645 name = "f32_conv_hwc_bench",
3646 srcs = [
3647 "bench/f32-conv-hwc.cc",
3648 "bench/dconv.h",
3649 "src/xnnpack/AlignedAllocator.h",
3650 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003651 deps = MICROKERNEL_BENCHMARK_DEPS + [
3652 ":packing",
3653 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003654)
3655
3656xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003657 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003658 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003659 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003660 "bench/dconv.h",
3661 "src/xnnpack/AlignedAllocator.h",
3662 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003663 deps = MICROKERNEL_BENCHMARK_DEPS + [
3664 ":packing",
3665 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003666)
3667
3668xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003669 name = "f16_dwconv_bench",
3670 srcs = [
3671 "bench/f16-dwconv.cc",
3672 "bench/dwconv.h",
3673 "bench/google/dwconv.h",
3674 "src/xnnpack/AlignedAllocator.h",
3675 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003676 deps = MICROKERNEL_BENCHMARK_DEPS + [
3677 ":indirection",
3678 ":packing",
3679 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003680)
3681
3682xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003683 name = "f32_dwconv_bench",
3684 srcs = [
3685 "bench/f32-dwconv.cc",
3686 "bench/dwconv.h",
3687 "src/xnnpack/AlignedAllocator.h",
3688 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003689 deps = MICROKERNEL_BENCHMARK_DEPS + [
3690 ":indirection",
3691 ":packing",
3692 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003693)
3694
3695xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003696 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003697 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003698 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003699 "bench/dwconv.h",
3700 "src/xnnpack/AlignedAllocator.h",
3701 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003702 deps = MICROKERNEL_BENCHMARK_DEPS + [
3703 ":indirection",
3704 ":packing",
3705 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003706)
3707
3708xnnpack_benchmark(
3709 name = "f32_gemm_bench",
3710 srcs = [
3711 "bench/f32-gemm.cc",
3712 "bench/gemm.h",
3713 "src/xnnpack/AlignedAllocator.h",
3714 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003715 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003716 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003717)
3718
3719xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003720 name = "f32_hswish_bench",
3721 srcs = [
3722 "bench/f32-hswish.cc",
3723 "src/xnnpack/AlignedAllocator.h",
3724 ] + MICROKERNEL_BENCHMARK_HDRS,
3725 deps = MICROKERNEL_BENCHMARK_DEPS,
3726)
3727
3728xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003729 name = "f32_raddexpminusmax_bench",
3730 srcs = [
3731 "bench/f32-raddexpminusmax.cc",
3732 "src/xnnpack/AlignedAllocator.h",
3733 ] + MICROKERNEL_BENCHMARK_HDRS,
3734 deps = MICROKERNEL_BENCHMARK_DEPS,
3735)
3736
3737xnnpack_benchmark(
3738 name = "f32_raddextexp_bench",
3739 srcs = [
3740 "bench/f32-raddextexp.cc",
3741 "src/xnnpack/AlignedAllocator.h",
3742 ] + MICROKERNEL_BENCHMARK_HDRS,
3743 deps = MICROKERNEL_BENCHMARK_DEPS,
3744)
3745
3746xnnpack_benchmark(
3747 name = "f32_raddstoreexpminusmax_bench",
3748 srcs = [
3749 "bench/f32-raddstoreexpminusmax.cc",
3750 "src/xnnpack/AlignedAllocator.h",
3751 ] + MICROKERNEL_BENCHMARK_HDRS,
3752 deps = MICROKERNEL_BENCHMARK_DEPS,
3753)
3754
3755xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756 name = "f32_rmax_bench",
3757 srcs = [
3758 "bench/f32-rmax.cc",
3759 "src/xnnpack/AlignedAllocator.h",
3760 ] + MICROKERNEL_BENCHMARK_HDRS,
3761 deps = MICROKERNEL_BENCHMARK_DEPS,
3762)
3763
3764xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003765 name = "f32_sigmoid_bench",
3766 srcs = [
3767 "bench/f32-sigmoid.cc",
3768 "src/xnnpack/AlignedAllocator.h",
3769 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003770 deps = MICROKERNEL_BENCHMARK_DEPS,
3771)
3772
3773xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003774 name = "f32_spmm_bench",
3775 srcs = [
3776 "bench/f32-spmm.cc",
3777 "bench/gemm.h",
3778 "src/xnnpack/AlignedAllocator.h",
3779 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003780 deps = MICROKERNEL_BENCHMARK_DEPS,
3781)
3782
3783xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003784 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003785 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003786 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003787 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003788 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003789 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003790)
3791
3792xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003793 name = "f32_vscaleexpminusmax_bench",
3794 srcs = [
3795 "bench/f32-vscaleexpminusmax.cc",
3796 "src/xnnpack/AlignedAllocator.h",
3797 ] + MICROKERNEL_BENCHMARK_HDRS,
3798 deps = MICROKERNEL_BENCHMARK_DEPS,
3799)
3800
3801xnnpack_benchmark(
3802 name = "f32_vscaleextexp_bench",
3803 srcs = [
3804 "bench/f32-vscaleextexp.cc",
3805 "src/xnnpack/AlignedAllocator.h",
3806 ] + MICROKERNEL_BENCHMARK_HDRS,
3807 deps = MICROKERNEL_BENCHMARK_DEPS,
3808)
3809
3810xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003811 name = "f32_vsqrt_bench",
3812 srcs = [
3813 "bench/f32-vsqrt.cc",
3814 "src/xnnpack/AlignedAllocator.h",
3815 ] + MICROKERNEL_BENCHMARK_HDRS,
3816 deps = MICROKERNEL_BENCHMARK_DEPS,
3817)
3818
3819xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003820 name = "f32_im2col_gemm_bench",
3821 srcs = [
3822 "bench/f32-im2col-gemm.cc",
3823 "bench/conv.h",
3824 "src/xnnpack/AlignedAllocator.h",
3825 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003826 deps = MICROKERNEL_BENCHMARK_DEPS + [
3827 ":im2col",
3828 ":packing",
3829 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003830)
3831
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003832xnnpack_benchmark(
3833 name = "requantization_bench",
3834 srcs = [
3835 "bench/requantization.cc",
3836 "src/xnnpack/requantization-stubs.h",
3837 "src/xnnpack/AlignedAllocator.h",
3838 ] + MICROKERNEL_BENCHMARK_HDRS,
3839 deps = MICROKERNEL_BENCHMARK_DEPS,
3840)
3841
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003842xnnpack_benchmark(
3843 name = "rounding_bench",
3844 srcs = [
3845 "bench/rounding.cc",
3846 "src/xnnpack/math-stubs.h",
3847 "src/xnnpack/AlignedAllocator.h",
3848 ] + MICROKERNEL_BENCHMARK_HDRS,
3849 deps = MICROKERNEL_BENCHMARK_DEPS,
3850)
3851
Marat Dukhan08c4a432019-10-03 09:29:21 -07003852########################### Benchmarks for operators ###########################
3853
3854xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003855 name = "average_pooling_bench",
3856 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003857 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003858 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003859 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003860)
3861
3862xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003863 name = "bankers_rounding_bench",
3864 srcs = ["bench/bankers-rounding.cc"],
3865 copts = xnnpack_optional_tflite_copts(),
3866 tags = ["nowin32"],
3867 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3868)
3869
3870xnnpack_benchmark(
3871 name = "ceiling_bench",
3872 srcs = ["bench/ceiling.cc"],
3873 copts = xnnpack_optional_tflite_copts(),
3874 tags = ["nowin32"],
3875 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3876)
3877
3878xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003879 name = "channel_shuffle_bench",
3880 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003881 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003882)
3883
3884xnnpack_benchmark(
3885 name = "convolution_bench",
3886 srcs = ["bench/convolution.cc"],
3887 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003888 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003889 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890)
3891
3892xnnpack_benchmark(
3893 name = "deconvolution_bench",
3894 srcs = ["bench/deconvolution.cc"],
3895 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003896 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003897 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003898)
3899
3900xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003901 name = "floor_bench",
3902 srcs = ["bench/floor.cc"],
3903 copts = xnnpack_optional_tflite_copts(),
3904 tags = ["nowin32"],
3905 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3906)
3907
3908xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909 name = "global_average_pooling_bench",
3910 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003911 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003912)
3913
3914xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003915 name = "hardswish_bench",
3916 srcs = ["bench/hardswish.cc"],
3917 copts = xnnpack_optional_tflite_copts(),
3918 tags = ["nowin32"],
3919 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3920)
3921
3922xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003923 name = "max_pooling_bench",
3924 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003925 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003926)
3927
3928xnnpack_benchmark(
3929 name = "sigmoid_bench",
3930 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003931 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003932 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003933 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003934)
3935
3936xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003937 name = "prelu_bench",
3938 srcs = ["bench/prelu.cc"],
3939 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003940 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003941 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003942)
3943
3944xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003945 name = "softmax_bench",
3946 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003947 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003948 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003949 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003950)
3951
Marat Dukhan87727142020-06-24 15:24:10 -07003952xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003953 name = "square_root_bench",
3954 srcs = ["bench/square-root.cc"],
3955 copts = xnnpack_optional_tflite_copts(),
3956 tags = ["nowin32"],
3957 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3958)
3959
3960xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003961 name = "truncation_bench",
3962 srcs = ["bench/truncation.cc"],
3963 deps = OPERATOR_BENCHMARK_DEPS,
3964)
3965
Marat Dukhanc068bb62019-10-04 13:24:39 -07003966############################# End-to-end benchmarks ############################
3967
3968cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003969 name = "fp32_mobilenet_v1",
3970 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003971 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003972 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003973 linkstatic = True,
3974 deps = [
3975 ":XNNPACK",
3976 "@pthreadpool",
3977 ],
3978)
3979
3980cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003981 name = "fp16_mobilenet_v1",
3982 srcs = ["models/fp16-mobilenet-v1.cc"],
3983 hdrs = ["models/models.h"],
3984 copts = xnnpack_std_cxxopts(),
3985 linkstatic = True,
3986 deps = [
3987 ":XNNPACK",
3988 "@FP16",
3989 "@pthreadpool",
3990 ],
3991)
3992
3993cc_library(
3994 name = "fp32_mobilenet_v2",
3995 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003996 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003997 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003998 linkstatic = True,
3999 deps = [
4000 ":XNNPACK",
4001 "@pthreadpool",
4002 ],
4003)
4004
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004005cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004006 name = "fp16_mobilenet_v2",
4007 srcs = ["models/fp16-mobilenet-v2.cc"],
4008 hdrs = ["models/models.h"],
4009 copts = xnnpack_std_cxxopts(),
4010 linkstatic = True,
4011 deps = [
4012 ":XNNPACK",
4013 "@FP16",
4014 "@pthreadpool",
4015 ],
4016)
4017
4018cc_library(
4019 name = "fp32_mobilenet_v3_large",
4020 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004021 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004022 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004023 linkstatic = True,
4024 deps = [
4025 ":XNNPACK",
4026 "@pthreadpool",
4027 ],
4028)
4029
4030cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004031 name = "fp32_mobilenet_v3_small",
4032 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004033 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004034 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004035 linkstatic = True,
4036 deps = [
4037 ":XNNPACK",
4038 "@pthreadpool",
4039 ],
4040)
4041
Marat Dukhanc068bb62019-10-04 13:24:39 -07004042xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004043 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004044 srcs = [
4045 "bench/f32-dwconv-e2e.cc",
4046 "bench/end2end.h",
4047 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004048 deps = MICROKERNEL_BENCHMARK_DEPS + [
4049 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004050 ":fp32_mobilenet_v1",
4051 ":fp32_mobilenet_v2",
4052 ":fp32_mobilenet_v3_large",
4053 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004054 ],
4055)
4056
4057xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004058 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004059 srcs = [
4060 "bench/f32-gemm-e2e.cc",
4061 "bench/end2end.h",
4062 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004063 deps = MICROKERNEL_BENCHMARK_DEPS + [
4064 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004065 ":fp32_mobilenet_v1",
4066 ":fp32_mobilenet_v2",
4067 ":fp32_mobilenet_v3_large",
4068 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004069 ],
4070)
4071
4072xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004073 name = "end2end_bench",
4074 srcs = ["bench/end2end.cc"],
4075 deps = [
4076 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004077 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004078 ":fp16_mobilenet_v1",
4079 ":fp16_mobilenet_v2",
4080 ":fp32_mobilenet_v1",
4081 ":fp32_mobilenet_v2",
4082 ":fp32_mobilenet_v3_large",
4083 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004084 "@pthreadpool",
4085 ],
4086)
4087
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004088#################### Accuracy evaluation for math functions ####################
4089
4090xnnpack_benchmark(
4091 name = "f32_exp_eval",
4092 srcs = [
4093 "eval/f32-exp.cc",
4094 "src/xnnpack/AlignedAllocator.h",
4095 ] + ACCURACY_EVAL_HDRS,
4096 deps = ACCURACY_EVAL_DEPS,
4097)
4098
Marat Dukhan515c9772019-10-17 18:07:57 -07004099xnnpack_benchmark(
4100 name = "f32_expminus_eval",
4101 srcs = [
4102 "eval/f32-expminus.cc",
4103 "src/xnnpack/AlignedAllocator.h",
4104 ] + ACCURACY_EVAL_HDRS,
4105 deps = ACCURACY_EVAL_DEPS,
4106)
4107
Marat Dukhan98ba4412019-10-23 02:14:28 -07004108xnnpack_benchmark(
4109 name = "f32_extexp_eval",
4110 srcs = [
4111 "eval/f32-extexp.cc",
4112 "src/xnnpack/AlignedAllocator.h",
4113 ] + ACCURACY_EVAL_HDRS,
4114 deps = ACCURACY_EVAL_DEPS,
4115)
4116
Marat Dukhan8853b822020-05-07 12:19:01 -07004117xnnpack_unit_test(
4118 name = "f32_roundne_eval",
4119 srcs = [
4120 "eval/f32-roundne.cc",
4121 "src/xnnpack/AlignedAllocator.h",
4122 "src/xnnpack/math-stubs.h",
4123 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004124 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004125 deps = MICROKERNEL_TEST_DEPS,
4126)
4127
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004128xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004129 name = "f32_roundd_eval",
4130 srcs = [
4131 "eval/f32-roundd.cc",
4132 "src/xnnpack/AlignedAllocator.h",
4133 "src/xnnpack/math-stubs.h",
4134 ] + MICROKERNEL_TEST_HDRS,
4135 automatic = False,
4136 deps = MICROKERNEL_TEST_DEPS,
4137)
4138
4139xnnpack_unit_test(
4140 name = "f32_roundu_eval",
4141 srcs = [
4142 "eval/f32-roundu.cc",
4143 "src/xnnpack/AlignedAllocator.h",
4144 "src/xnnpack/math-stubs.h",
4145 ] + MICROKERNEL_TEST_HDRS,
4146 automatic = False,
4147 deps = MICROKERNEL_TEST_DEPS,
4148)
4149
4150xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004151 name = "f32_roundz_eval",
4152 srcs = [
4153 "eval/f32-roundz.cc",
4154 "src/xnnpack/AlignedAllocator.h",
4155 "src/xnnpack/math-stubs.h",
4156 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004157 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004158 deps = MICROKERNEL_TEST_DEPS,
4159)
4160
Marat Dukhan346a9e52019-11-15 09:06:30 -08004161xnnpack_benchmark(
4162 name = "f32_sigmoid_eval",
4163 srcs = [
4164 "eval/f32-sigmoid.cc",
4165 "src/xnnpack/AlignedAllocator.h",
4166 ] + ACCURACY_EVAL_HDRS,
4167 deps = ACCURACY_EVAL_DEPS,
4168)
4169
Marat Dukhan84000762020-06-29 18:38:43 -07004170xnnpack_benchmark(
4171 name = "f32_sqrt_eval",
4172 srcs = [
4173 "eval/f32-sqrt.cc",
4174 "src/xnnpack/AlignedAllocator.h",
4175 ] + ACCURACY_EVAL_HDRS,
4176 deps = ACCURACY_EVAL_DEPS + [
4177 ":bench_utils",
4178 ],
4179)
4180
Marat Dukhan08c4a432019-10-03 09:29:21 -07004181######################### Unit tests for micro-kernels #########################
4182
4183xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004184 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004185 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004186 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004187 "test/gemm-microkernel-tester.h",
4188 "src/xnnpack/AlignedAllocator.h",
4189 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004190 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004191)
4192
4193xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004194 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004195 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004196 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004197 "test/spmm-microkernel-tester.h",
4198 "src/xnnpack/AlignedAllocator.h",
4199 ] + MICROKERNEL_TEST_HDRS,
4200 deps = MICROKERNEL_TEST_DEPS,
4201)
4202
4203xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004204 name = "f32_argmaxpool_test",
4205 srcs = [
4206 "test/f32-argmaxpool.cc",
4207 "test/argmaxpool-microkernel-tester.h",
4208 "src/xnnpack/AlignedAllocator.h",
4209 ] + MICROKERNEL_TEST_HDRS,
4210 deps = MICROKERNEL_TEST_DEPS,
4211)
4212
4213xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004214 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004215 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004216 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004217 "test/avgpool-microkernel-tester.h",
4218 "src/xnnpack/AlignedAllocator.h",
4219 ] + MICROKERNEL_TEST_HDRS,
4220 deps = MICROKERNEL_TEST_DEPS,
4221)
4222
4223xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004224 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004225 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004226 "test/f32-ibilinear.cc",
4227 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004228 "src/xnnpack/AlignedAllocator.h",
4229 ] + MICROKERNEL_TEST_HDRS,
4230 deps = MICROKERNEL_TEST_DEPS,
4231)
4232
4233xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004234 name = "f16_clamp_test",
4235 srcs = [
4236 "test/f16-clamp.cc",
4237 "test/clamp-microkernel-tester.h",
4238 ] + MICROKERNEL_TEST_HDRS,
4239 deps = MICROKERNEL_TEST_DEPS,
4240)
4241
4242xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004243 name = "f32_clamp_test",
4244 srcs = [
4245 "test/f32-clamp.cc",
4246 "test/clamp-microkernel-tester.h",
4247 ] + MICROKERNEL_TEST_HDRS,
4248 deps = MICROKERNEL_TEST_DEPS,
4249)
4250
4251xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004252 name = "f32_igemm_test",
4253 srcs = [
4254 "test/f32-igemm.cc",
4255 "test/gemm-microkernel-tester.h",
4256 "src/xnnpack/AlignedAllocator.h",
4257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004258 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004259)
4260
4261xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004262 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004264 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004265 "test/gemm-microkernel-tester.h",
4266 "src/xnnpack/AlignedAllocator.h",
4267 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004268 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269)
4270
4271xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004272 name = "f32_igemm_minmax_test",
4273 srcs = [
4274 "test/f32-igemm-minmax.cc",
4275 "test/gemm-microkernel-tester.h",
4276 "src/xnnpack/AlignedAllocator.h",
4277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004278 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004279)
4280
4281xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004282 name = "f16_igemm_minmax_test",
4283 srcs = [
4284 "test/f16-igemm-minmax.cc",
4285 "test/gemm-microkernel-tester.h",
4286 "src/xnnpack/AlignedAllocator.h",
4287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004288 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004289)
4290
4291xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004292 name = "f32_conv_hwc_test",
4293 srcs = [
4294 "test/f32-conv-hwc.cc",
4295 "test/conv-hwc-microkernel-tester.h",
4296 "src/xnnpack/AlignedAllocator.h",
4297 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004298 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299)
4300
4301xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004302 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004303 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004304 "test/f32-conv-hwc2chw.cc",
4305 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004306 "src/xnnpack/AlignedAllocator.h",
4307 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004308 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004309)
4310
4311xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004312 name = "f32_dwconv_test",
4313 srcs = [
4314 "test/f32-dwconv.cc",
4315 "test/dwconv-microkernel-tester.h",
4316 "src/xnnpack/AlignedAllocator.h",
4317 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004318 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004319)
4320
4321xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004322 name = "f16_dwconv_minmax_test",
4323 srcs = [
4324 "test/f16-dwconv-minmax.cc",
4325 "test/dwconv-microkernel-tester.h",
4326 "src/xnnpack/AlignedAllocator.h",
4327 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004328 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004329)
4330
4331xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004332 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004333 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004334 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335 "test/dwconv-microkernel-tester.h",
4336 "src/xnnpack/AlignedAllocator.h",
4337 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004338 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004339)
4340
4341xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004342 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004343 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004344 "test/f32-dwconv-chw.cc",
4345 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004346 "src/xnnpack/AlignedAllocator.h",
4347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004348 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349)
4350
4351xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004352 name = "f16_gavgpool_minmax_test",
4353 srcs = [
4354 "test/f16-gavgpool-minmax.cc",
4355 "test/gavgpool-microkernel-tester.h",
4356 "src/xnnpack/AlignedAllocator.h",
4357 ] + MICROKERNEL_TEST_HDRS,
4358 deps = MICROKERNEL_TEST_DEPS,
4359)
4360
4361xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004362 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004363 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004364 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 "test/gavgpool-microkernel-tester.h",
4366 "src/xnnpack/AlignedAllocator.h",
4367 ] + MICROKERNEL_TEST_HDRS,
4368 deps = MICROKERNEL_TEST_DEPS,
4369)
4370
4371xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004372 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004373 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004374 "test/f32-gavgpool-cw.cc",
4375 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004376 "src/xnnpack/AlignedAllocator.h",
4377 ] + MICROKERNEL_TEST_HDRS,
4378 deps = MICROKERNEL_TEST_DEPS,
4379)
4380
4381xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004382 name = "f32_gemm_test",
4383 srcs = [
4384 "test/f32-gemm.cc",
4385 "test/gemm-microkernel-tester.h",
4386 "src/xnnpack/AlignedAllocator.h",
4387 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004388 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004389)
4390
4391xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004392 name = "f32_gemm_relu_test",
4393 srcs = [
4394 "test/f32-gemm-relu.cc",
4395 "test/gemm-microkernel-tester.h",
4396 "src/xnnpack/AlignedAllocator.h",
4397 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004398 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004399)
4400
4401xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004402 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004403 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004404 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004405 "test/gemm-microkernel-tester.h",
4406 "src/xnnpack/AlignedAllocator.h",
4407 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004408 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004409)
4410
4411xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004412 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004414 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004415 "test/gemm-microkernel-tester.h",
4416 "src/xnnpack/AlignedAllocator.h",
4417 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004418 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004419)
4420
4421xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004422 name = "f16_hswish_test",
4423 srcs = [
4424 "test/f16-hswish.cc",
4425 "test/hswish-microkernel-tester.h",
4426 ] + MICROKERNEL_TEST_HDRS,
4427 deps = MICROKERNEL_TEST_DEPS,
4428)
4429
4430xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004431 name = "f32_hswish_test",
4432 srcs = [
4433 "test/f32-hswish.cc",
4434 "test/hswish-microkernel-tester.h",
4435 ] + MICROKERNEL_TEST_HDRS,
4436 deps = MICROKERNEL_TEST_DEPS,
4437)
4438
4439xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004440 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004441 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004442 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004443 "test/maxpool-microkernel-tester.h",
4444 ] + MICROKERNEL_TEST_HDRS,
4445 deps = MICROKERNEL_TEST_DEPS,
4446)
4447
4448xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004449 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004450 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004451 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452 "test/avgpool-microkernel-tester.h",
4453 "src/xnnpack/AlignedAllocator.h",
4454 ] + MICROKERNEL_TEST_HDRS,
4455 deps = MICROKERNEL_TEST_DEPS,
4456)
4457
4458xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004459 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004461 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004462 "test/gemm-microkernel-tester.h",
4463 "src/xnnpack/AlignedAllocator.h",
4464 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004465 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466)
4467
4468xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004469 name = "f16_prelu_test",
4470 srcs = [
4471 "test/f16-prelu.cc",
4472 "test/prelu-microkernel-tester.h",
4473 "src/xnnpack/AlignedAllocator.h",
4474 ] + MICROKERNEL_TEST_HDRS,
4475 deps = MICROKERNEL_TEST_DEPS,
4476)
4477
4478xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479 name = "f32_prelu_test",
4480 srcs = [
4481 "test/f32-prelu.cc",
4482 "test/prelu-microkernel-tester.h",
4483 "src/xnnpack/AlignedAllocator.h",
4484 ] + MICROKERNEL_TEST_HDRS,
4485 deps = MICROKERNEL_TEST_DEPS,
4486)
4487
4488xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004489 name = "f32_raddexpminusmax_test",
4490 srcs = [
4491 "test/f32-raddexpminusmax.cc",
4492 "test/raddexpminusmax-microkernel-tester.h",
4493 ] + MICROKERNEL_TEST_HDRS,
4494 deps = MICROKERNEL_TEST_DEPS,
4495)
4496
4497xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004498 name = "f32_raddextexp_test",
4499 srcs = [
4500 "test/f32-raddextexp.cc",
4501 "test/raddextexp-microkernel-tester.h",
4502 ] + MICROKERNEL_TEST_HDRS,
4503 deps = MICROKERNEL_TEST_DEPS,
4504)
4505
4506xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004507 name = "f32_raddstoreexpminusmax_test",
4508 srcs = [
4509 "test/f32-raddstoreexpminusmax.cc",
4510 "test/raddstoreexpminusmax-microkernel-tester.h",
4511 ] + MICROKERNEL_TEST_HDRS,
4512 deps = MICROKERNEL_TEST_DEPS,
4513)
4514
4515xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004516 name = "f32_relu_test",
4517 srcs = [
4518 "test/f32-relu.cc",
4519 "test/vunary-microkernel-tester.h",
4520 ] + MICROKERNEL_TEST_HDRS,
4521 deps = MICROKERNEL_TEST_DEPS,
4522)
4523
4524xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525 name = "f32_rmax_test",
4526 srcs = [
4527 "test/f32-rmax.cc",
4528 "test/rmax-microkernel-tester.h",
4529 ] + MICROKERNEL_TEST_HDRS,
4530 deps = MICROKERNEL_TEST_DEPS,
4531)
4532
4533xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004534 name = "f32_sigmoid_test",
4535 srcs = [
4536 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004537 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004538 ] + MICROKERNEL_TEST_HDRS,
4539 deps = MICROKERNEL_TEST_DEPS,
4540)
4541
4542xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004543 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004544 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004545 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004546 "test/spmm-microkernel-tester.h",
4547 "src/xnnpack/AlignedAllocator.h",
4548 ] + MICROKERNEL_TEST_HDRS,
4549 deps = MICROKERNEL_TEST_DEPS,
4550)
4551
4552xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004553 name = "f16_vadd_minmax_test",
4554 srcs = [
4555 "test/f16-vadd-minmax.cc",
4556 "test/vbinary-microkernel-tester.h",
4557 ] + MICROKERNEL_TEST_HDRS,
4558 deps = MICROKERNEL_TEST_DEPS,
4559)
4560
4561xnnpack_unit_test(
4562 name = "f16_vaddc_minmax_test",
4563 srcs = [
4564 "test/f16-vaddc-minmax.cc",
4565 "test/vbinaryc-microkernel-tester.h",
4566 ] + MICROKERNEL_TEST_HDRS,
4567 deps = MICROKERNEL_TEST_DEPS,
4568)
4569
4570xnnpack_unit_test(
4571 name = "f16_vdiv_minmax_test",
4572 srcs = [
4573 "test/f16-vdiv-minmax.cc",
4574 "test/vbinary-microkernel-tester.h",
4575 ] + MICROKERNEL_TEST_HDRS,
4576 deps = MICROKERNEL_TEST_DEPS,
4577)
4578
4579xnnpack_unit_test(
4580 name = "f16_vdivc_minmax_test",
4581 srcs = [
4582 "test/f16-vdivc-minmax.cc",
4583 "test/vbinaryc-microkernel-tester.h",
4584 ] + MICROKERNEL_TEST_HDRS,
4585 deps = MICROKERNEL_TEST_DEPS,
4586)
4587
4588xnnpack_unit_test(
4589 name = "f16_vrdivc_minmax_test",
4590 srcs = [
4591 "test/f16-vrdivc-minmax.cc",
4592 "test/vbinaryc-microkernel-tester.h",
4593 ] + MICROKERNEL_TEST_HDRS,
4594 deps = MICROKERNEL_TEST_DEPS,
4595)
4596
4597xnnpack_unit_test(
4598 name = "f16_vmax_test",
4599 srcs = [
4600 "test/f16-vmax.cc",
4601 "test/vbinary-microkernel-tester.h",
4602 ] + MICROKERNEL_TEST_HDRS,
4603 deps = MICROKERNEL_TEST_DEPS,
4604)
4605
4606xnnpack_unit_test(
4607 name = "f16_vmaxc_test",
4608 srcs = [
4609 "test/f16-vmaxc.cc",
4610 "test/vbinaryc-microkernel-tester.h",
4611 ] + MICROKERNEL_TEST_HDRS,
4612 deps = MICROKERNEL_TEST_DEPS,
4613)
4614
4615xnnpack_unit_test(
4616 name = "f16_vmin_test",
4617 srcs = [
4618 "test/f16-vmin.cc",
4619 "test/vbinary-microkernel-tester.h",
4620 ] + MICROKERNEL_TEST_HDRS,
4621 deps = MICROKERNEL_TEST_DEPS,
4622)
4623
4624xnnpack_unit_test(
4625 name = "f16_vminc_test",
4626 srcs = [
4627 "test/f16-vminc.cc",
4628 "test/vbinaryc-microkernel-tester.h",
4629 ] + MICROKERNEL_TEST_HDRS,
4630 deps = MICROKERNEL_TEST_DEPS,
4631)
4632
4633xnnpack_unit_test(
4634 name = "f16_vmul_minmax_test",
4635 srcs = [
4636 "test/f16-vmul-minmax.cc",
4637 "test/vbinary-microkernel-tester.h",
4638 ] + MICROKERNEL_TEST_HDRS,
4639 deps = MICROKERNEL_TEST_DEPS,
4640)
4641
4642xnnpack_unit_test(
4643 name = "f16_vmulc_minmax_test",
4644 srcs = [
4645 "test/f16-vmulc-minmax.cc",
4646 "test/vbinaryc-microkernel-tester.h",
4647 ] + MICROKERNEL_TEST_HDRS,
4648 deps = MICROKERNEL_TEST_DEPS,
4649)
4650
4651xnnpack_unit_test(
4652 name = "f16_vsub_minmax_test",
4653 srcs = [
4654 "test/f16-vsub-minmax.cc",
4655 "test/vbinary-microkernel-tester.h",
4656 ] + MICROKERNEL_TEST_HDRS,
4657 deps = MICROKERNEL_TEST_DEPS,
4658)
4659
4660xnnpack_unit_test(
4661 name = "f16_vsubc_minmax_test",
4662 srcs = [
4663 "test/f16-vsubc-minmax.cc",
4664 "test/vbinaryc-microkernel-tester.h",
4665 ] + MICROKERNEL_TEST_HDRS,
4666 deps = MICROKERNEL_TEST_DEPS,
4667)
4668
4669xnnpack_unit_test(
4670 name = "f16_vrsubc_minmax_test",
4671 srcs = [
4672 "test/f16-vrsubc-minmax.cc",
4673 "test/vbinaryc-microkernel-tester.h",
4674 ] + MICROKERNEL_TEST_HDRS,
4675 deps = MICROKERNEL_TEST_DEPS,
4676)
4677
4678xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004679 name = "f32_vabs_test",
4680 srcs = [
4681 "test/f32-vabs.cc",
4682 "test/vunary-microkernel-tester.h",
4683 ] + MICROKERNEL_TEST_HDRS,
4684 deps = MICROKERNEL_TEST_DEPS,
4685)
4686
4687xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004688 name = "f32_vadd_test",
4689 srcs = [
4690 "test/f32-vadd.cc",
4691 "test/vbinary-microkernel-tester.h",
4692 ] + MICROKERNEL_TEST_HDRS,
4693 deps = MICROKERNEL_TEST_DEPS,
4694)
4695
4696xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004697 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004698 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004699 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004700 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004701 ] + MICROKERNEL_TEST_HDRS,
4702 deps = MICROKERNEL_TEST_DEPS,
4703)
4704
4705xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004706 name = "f32_vaddc_test",
4707 srcs = [
4708 "test/f32-vaddc.cc",
4709 "test/vbinaryc-microkernel-tester.h",
4710 ] + MICROKERNEL_TEST_HDRS,
4711 deps = MICROKERNEL_TEST_DEPS,
4712)
4713
4714xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004715 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004716 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004717 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004718 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719 ] + MICROKERNEL_TEST_HDRS,
4720 deps = MICROKERNEL_TEST_DEPS,
4721)
4722
4723xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004724 name = "f32_vdiv_test",
4725 srcs = [
4726 "test/f32-vdiv.cc",
4727 "test/vbinary-microkernel-tester.h",
4728 ] + MICROKERNEL_TEST_HDRS,
4729 deps = MICROKERNEL_TEST_DEPS,
4730)
4731
4732xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004733 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004734 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004735 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004736 "test/vbinary-microkernel-tester.h",
4737 ] + MICROKERNEL_TEST_HDRS,
4738 deps = MICROKERNEL_TEST_DEPS,
4739)
4740
4741xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004742 name = "f32_vdivc_test",
4743 srcs = [
4744 "test/f32-vdivc.cc",
4745 "test/vbinaryc-microkernel-tester.h",
4746 ] + MICROKERNEL_TEST_HDRS,
4747 deps = MICROKERNEL_TEST_DEPS,
4748)
4749
4750xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004751 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004752 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004753 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004754 "test/vbinaryc-microkernel-tester.h",
4755 ] + MICROKERNEL_TEST_HDRS,
4756 deps = MICROKERNEL_TEST_DEPS,
4757)
4758
4759xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004760 name = "f32_vrdivc_test",
4761 srcs = [
4762 "test/f32-vrdivc.cc",
4763 "test/vbinaryc-microkernel-tester.h",
4764 ] + MICROKERNEL_TEST_HDRS,
4765 deps = MICROKERNEL_TEST_DEPS,
4766)
4767
4768xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004769 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004770 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004771 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004772 "test/vbinaryc-microkernel-tester.h",
4773 ] + MICROKERNEL_TEST_HDRS,
4774 deps = MICROKERNEL_TEST_DEPS,
4775)
4776
4777xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004778 name = "f32_vmax_test",
4779 srcs = [
4780 "test/f32-vmax.cc",
4781 "test/vbinary-microkernel-tester.h",
4782 ] + MICROKERNEL_TEST_HDRS,
4783 deps = MICROKERNEL_TEST_DEPS,
4784)
4785
4786xnnpack_unit_test(
4787 name = "f32_vmaxc_test",
4788 srcs = [
4789 "test/f32-vmaxc.cc",
4790 "test/vbinaryc-microkernel-tester.h",
4791 ] + MICROKERNEL_TEST_HDRS,
4792 deps = MICROKERNEL_TEST_DEPS,
4793)
4794
4795xnnpack_unit_test(
4796 name = "f32_vmin_test",
4797 srcs = [
4798 "test/f32-vmin.cc",
4799 "test/vbinary-microkernel-tester.h",
4800 ] + MICROKERNEL_TEST_HDRS,
4801 deps = MICROKERNEL_TEST_DEPS,
4802)
4803
4804xnnpack_unit_test(
4805 name = "f32_vminc_test",
4806 srcs = [
4807 "test/f32-vminc.cc",
4808 "test/vbinaryc-microkernel-tester.h",
4809 ] + MICROKERNEL_TEST_HDRS,
4810 deps = MICROKERNEL_TEST_DEPS,
4811)
4812
4813xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004814 name = "f32_vmul_test",
4815 srcs = [
4816 "test/f32-vmul.cc",
4817 "test/vbinary-microkernel-tester.h",
4818 ] + MICROKERNEL_TEST_HDRS,
4819 deps = MICROKERNEL_TEST_DEPS,
4820)
4821
4822xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004823 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004824 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004825 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004826 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004827 ] + MICROKERNEL_TEST_HDRS,
4828 deps = MICROKERNEL_TEST_DEPS,
4829)
4830
4831xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004832 name = "f32_vmulc_test",
4833 srcs = [
4834 "test/f32-vmulc.cc",
4835 "test/vbinaryc-microkernel-tester.h",
4836 ] + MICROKERNEL_TEST_HDRS,
4837 deps = MICROKERNEL_TEST_DEPS,
4838)
4839
4840xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004841 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004842 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004843 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004844 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004845 ] + MICROKERNEL_TEST_HDRS,
4846 deps = MICROKERNEL_TEST_DEPS,
4847)
4848
4849xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004850 name = "f16_vmulcaddc_minmax_test",
4851 srcs = [
4852 "test/f16-vmulcaddc-minmax.cc",
4853 "test/vmulcaddc-microkernel-tester.h",
4854 "src/xnnpack/AlignedAllocator.h",
4855 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004856 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004857)
4858
4859xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004860 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004862 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004863 "test/vmulcaddc-microkernel-tester.h",
4864 "src/xnnpack/AlignedAllocator.h",
4865 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004866 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004867)
4868
4869xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004870 name = "f32_vlrelu_test",
4871 srcs = [
4872 "test/f32-vlrelu.cc",
4873 "test/vunary-microkernel-tester.h",
4874 ] + MICROKERNEL_TEST_HDRS,
4875 deps = MICROKERNEL_TEST_DEPS,
4876)
4877
4878xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004879 name = "f32_vneg_test",
4880 srcs = [
4881 "test/f32-vneg.cc",
4882 "test/vunary-microkernel-tester.h",
4883 ] + MICROKERNEL_TEST_HDRS,
4884 deps = MICROKERNEL_TEST_DEPS,
4885)
4886
4887xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004888 name = "f32_vrndne_test",
4889 srcs = [
4890 "test/f32-vrndne.cc",
4891 "test/vunary-microkernel-tester.h",
4892 ] + MICROKERNEL_TEST_HDRS,
4893 deps = MICROKERNEL_TEST_DEPS,
4894)
4895
4896xnnpack_unit_test(
4897 name = "f32_vrndz_test",
4898 srcs = [
4899 "test/f32-vrndz.cc",
4900 "test/vunary-microkernel-tester.h",
4901 ] + MICROKERNEL_TEST_HDRS,
4902 deps = MICROKERNEL_TEST_DEPS,
4903)
4904
4905xnnpack_unit_test(
4906 name = "f32_vrndu_test",
4907 srcs = [
4908 "test/f32-vrndu.cc",
4909 "test/vunary-microkernel-tester.h",
4910 ] + MICROKERNEL_TEST_HDRS,
4911 deps = MICROKERNEL_TEST_DEPS,
4912)
4913
4914xnnpack_unit_test(
4915 name = "f32_vrndd_test",
4916 srcs = [
4917 "test/f32-vrndd.cc",
4918 "test/vunary-microkernel-tester.h",
4919 ] + MICROKERNEL_TEST_HDRS,
4920 deps = MICROKERNEL_TEST_DEPS,
4921)
4922
4923xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004924 name = "f32_vscale_test",
4925 srcs = [
4926 "test/f32-vscale.cc",
4927 "test/vscale-microkernel-tester.h",
4928 ] + MICROKERNEL_TEST_HDRS,
4929 deps = MICROKERNEL_TEST_DEPS,
4930)
4931
4932xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004933 name = "f32_vscaleexpminusmax_test",
4934 srcs = [
4935 "test/f32-vscaleexpminusmax.cc",
4936 "test/vscaleexpminusmax-microkernel-tester.h",
4937 ] + MICROKERNEL_TEST_HDRS,
4938 deps = MICROKERNEL_TEST_DEPS,
4939)
4940
4941xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004942 name = "f32_vscaleextexp_test",
4943 srcs = [
4944 "test/f32-vscaleextexp.cc",
4945 "test/vscaleextexp-microkernel-tester.h",
4946 ] + MICROKERNEL_TEST_HDRS,
4947 deps = MICROKERNEL_TEST_DEPS,
4948)
4949
4950xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004951 name = "f32_vsqr_test",
4952 srcs = [
4953 "test/f32-vsqr.cc",
4954 "test/vunary-microkernel-tester.h",
4955 ] + MICROKERNEL_TEST_HDRS,
4956 deps = MICROKERNEL_TEST_DEPS,
4957)
4958
4959xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004960 name = "f32_vsqrdiff_test",
4961 srcs = [
4962 "test/f32-vsqrdiff.cc",
4963 "test/vbinary-microkernel-tester.h",
4964 ] + MICROKERNEL_TEST_HDRS,
4965 deps = MICROKERNEL_TEST_DEPS,
4966)
4967
4968xnnpack_unit_test(
4969 name = "f32_vsqrdiffc_test",
4970 srcs = [
4971 "test/f32-vsqrdiffc.cc",
4972 "test/vbinaryc-microkernel-tester.h",
4973 ] + MICROKERNEL_TEST_HDRS,
4974 deps = MICROKERNEL_TEST_DEPS,
4975)
4976
4977xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004978 name = "f32_vsqrt_test",
4979 srcs = [
4980 "test/f32-vsqrt.cc",
4981 "test/vunary-microkernel-tester.h",
4982 ] + MICROKERNEL_TEST_HDRS,
4983 deps = MICROKERNEL_TEST_DEPS,
4984)
4985
4986xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004987 name = "f32_vsub_test",
4988 srcs = [
4989 "test/f32-vsub.cc",
4990 "test/vbinary-microkernel-tester.h",
4991 ] + MICROKERNEL_TEST_HDRS,
4992 deps = MICROKERNEL_TEST_DEPS,
4993)
4994
4995xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004996 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004997 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004998 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004999 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005000 ] + MICROKERNEL_TEST_HDRS,
5001 deps = MICROKERNEL_TEST_DEPS,
5002)
5003
5004xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005005 name = "f32_vsubc_test",
5006 srcs = [
5007 "test/f32-vsubc.cc",
5008 "test/vbinaryc-microkernel-tester.h",
5009 ] + MICROKERNEL_TEST_HDRS,
5010 deps = MICROKERNEL_TEST_DEPS,
5011)
5012
5013xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005014 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005015 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005016 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005017 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005018 ] + MICROKERNEL_TEST_HDRS,
5019 deps = MICROKERNEL_TEST_DEPS,
5020)
5021
5022xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005023 name = "f32_vrsubc_test",
5024 srcs = [
5025 "test/f32-vrsubc.cc",
5026 "test/vbinaryc-microkernel-tester.h",
5027 ] + MICROKERNEL_TEST_HDRS,
5028 deps = MICROKERNEL_TEST_DEPS,
5029)
5030
5031xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005032 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005033 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005034 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005035 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005036 ] + MICROKERNEL_TEST_HDRS,
5037 deps = MICROKERNEL_TEST_DEPS,
5038)
5039
5040xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005041 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005042 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005043 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005044 "test/avgpool-microkernel-tester.h",
5045 "src/xnnpack/AlignedAllocator.h",
5046 ] + MICROKERNEL_TEST_HDRS,
5047 deps = MICROKERNEL_TEST_DEPS,
5048)
5049
5050xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005051 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005052 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005053 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054 "test/gemm-microkernel-tester.h",
5055 "src/xnnpack/AlignedAllocator.h",
5056 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005057 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005058)
5059
5060xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005061 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005062 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005063 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005064 "test/dwconv-microkernel-tester.h",
5065 "src/xnnpack/AlignedAllocator.h",
5066 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005067 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005068)
5069
5070xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005071 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005072 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005073 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005074 "test/gavgpool-microkernel-tester.h",
5075 "src/xnnpack/AlignedAllocator.h",
5076 ] + MICROKERNEL_TEST_HDRS,
5077 deps = MICROKERNEL_TEST_DEPS,
5078)
5079
5080xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005081 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005082 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005083 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084 "test/gemm-microkernel-tester.h",
5085 "src/xnnpack/AlignedAllocator.h",
5086 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005087 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005088)
5089
5090xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005091 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005093 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005094 "test/vadd-microkernel-tester.h",
5095 ] + MICROKERNEL_TEST_HDRS,
5096 deps = MICROKERNEL_TEST_DEPS,
5097)
5098
5099xnnpack_unit_test(
5100 name = "u8_clamp_test",
5101 srcs = [
5102 "test/u8-clamp.cc",
5103 "test/clamp-microkernel-tester.h",
5104 ] + MICROKERNEL_TEST_HDRS,
5105 deps = MICROKERNEL_TEST_DEPS,
5106)
5107
5108xnnpack_unit_test(
5109 name = "u8_lut32norm_test",
5110 srcs = [
5111 "test/u8-lut32norm.cc",
5112 "test/lut-norm-microkernel-tester.h",
5113 ] + MICROKERNEL_TEST_HDRS,
5114 deps = MICROKERNEL_TEST_DEPS,
5115)
5116
5117xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005118 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005119 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005120 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005121 "test/maxpool-microkernel-tester.h",
5122 ] + MICROKERNEL_TEST_HDRS,
5123 deps = MICROKERNEL_TEST_DEPS,
5124)
5125
5126xnnpack_unit_test(
5127 name = "u8_rmax_test",
5128 srcs = [
5129 "test/u8-rmax.cc",
5130 "test/rmax-microkernel-tester.h",
5131 ] + MICROKERNEL_TEST_HDRS,
5132 deps = MICROKERNEL_TEST_DEPS,
5133)
5134
5135xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005136 name = "x32_fill_test",
5137 srcs = [
5138 "test/x32-fill.cc",
5139 "test/fill-microkernel-tester.h",
5140 ] + MICROKERNEL_TEST_HDRS,
5141 deps = MICROKERNEL_TEST_DEPS,
5142)
5143
5144xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005145 name = "x32_packx_test",
5146 srcs = [
5147 "test/x32-packx.cc",
5148 "test/pack-microkernel-tester.h",
5149 "src/xnnpack/AlignedAllocator.h",
5150 ] + MICROKERNEL_TEST_HDRS,
5151 deps = MICROKERNEL_TEST_DEPS,
5152)
5153
5154xnnpack_unit_test(
5155 name = "x32_pad_test",
5156 srcs = [
5157 "test/x32-pad.cc",
5158 "test/pad-microkernel-tester.h",
5159 ] + MICROKERNEL_TEST_HDRS,
5160 deps = MICROKERNEL_TEST_DEPS,
5161)
5162
5163xnnpack_unit_test(
5164 name = "x32_unpool_test",
5165 srcs = [
5166 "test/x32-unpool.cc",
5167 "test/unpool-microkernel-tester.h",
5168 ] + MICROKERNEL_TEST_HDRS,
5169 deps = MICROKERNEL_TEST_DEPS,
5170)
5171
5172xnnpack_unit_test(
5173 name = "x32_zip_test",
5174 srcs = [
5175 "test/x32-zip.cc",
5176 "test/zip-microkernel-tester.h",
5177 ] + MICROKERNEL_TEST_HDRS,
5178 deps = MICROKERNEL_TEST_DEPS,
5179)
5180
5181xnnpack_unit_test(
5182 name = "x8_lut_test",
5183 srcs = [
5184 "test/x8-lut.cc",
5185 "test/lut-microkernel-tester.h",
5186 ] + MICROKERNEL_TEST_HDRS,
5187 deps = MICROKERNEL_TEST_DEPS,
5188)
5189
5190xnnpack_unit_test(
5191 name = "x8_zip_test",
5192 srcs = [
5193 "test/x8-zip.cc",
5194 "test/zip-microkernel-tester.h",
5195 ] + MICROKERNEL_TEST_HDRS,
5196 deps = MICROKERNEL_TEST_DEPS,
5197)
5198
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005199xnnpack_unit_test(
5200 name = "requantization_test",
5201 srcs = [
5202 "src/xnnpack/requantization-stubs.h",
5203 "test/requantization.cc",
5204 "test/requantization-tester.h",
5205 ] + MICROKERNEL_TEST_HDRS,
5206 deps = MICROKERNEL_TEST_DEPS,
5207)
5208
Marat Dukhan20c3b922020-03-10 03:45:06 -07005209########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210
5211xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005212 name = "operator_size_test",
5213 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005214 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215)
5216
Marat Dukhan20c3b922020-03-10 03:45:06 -07005217xnnpack_binary(
5218 name = "subgraph_size_test",
5219 srcs = ["test/subgraph-size.c"],
5220 deps = [":XNNPACK"],
5221)
5222
5223########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224
5225xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005226 name = "abs_nc_test",
5227 srcs = [
5228 "test/abs-nc.cc",
5229 "test/abs-operator-tester.h",
5230 ],
5231 deps = OPERATOR_TEST_DEPS,
5232)
5233
5234xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005235 name = "add_nd_test",
5236 srcs = [
5237 "test/add-nd.cc",
5238 "test/binary-elementwise-operator-tester.h",
5239 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005240 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005241)
5242
5243xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005244 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005245 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005246 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005247 "test/argmax-pooling-operator-tester.h",
5248 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005249 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005250)
5251
5252xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005253 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005254 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005255 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005256 "test/average-pooling-operator-tester.h",
5257 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005258 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005259)
5260
5261xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005262 name = "bankers_rounding_nc_test",
5263 srcs = [
5264 "test/bankers-rounding-nc.cc",
5265 "test/bankers-rounding-operator-tester.h",
5266 ],
5267 deps = OPERATOR_TEST_DEPS,
5268)
5269
5270xnnpack_unit_test(
5271 name = "ceiling_nc_test",
5272 srcs = [
5273 "test/ceiling-nc.cc",
5274 "test/ceiling-operator-tester.h",
5275 ],
5276 deps = OPERATOR_TEST_DEPS,
5277)
5278
5279xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005280 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005281 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005282 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005283 "test/channel-shuffle-operator-tester.h",
5284 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005285 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005286)
5287
5288xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005289 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005290 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005291 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005292 "test/clamp-operator-tester.h",
5293 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005294 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295)
5296
5297xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005298 name = "constant_pad_nd_test",
5299 srcs = [
5300 "test/constant-pad-nd.cc",
5301 "test/constant-pad-operator-tester.h",
5302 ],
5303 deps = OPERATOR_TEST_DEPS,
5304)
5305
5306xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005307 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005308 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005309 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310 "test/convolution-operator-tester.h",
5311 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005312 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005313)
5314
5315xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005316 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005318 "test/convolution-nchw.cc",
5319 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005320 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005321 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322)
5323
5324xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005325 name = "copy_nc_test",
5326 srcs = [
5327 "test/copy-nc.cc",
5328 "test/copy-operator-tester.h",
5329 ],
5330 deps = OPERATOR_TEST_DEPS,
5331)
5332
5333xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005334 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005335 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005336 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005337 "test/deconvolution-operator-tester.h",
5338 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005339 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005340)
5341
5342xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005343 name = "divide_nd_test",
5344 srcs = [
5345 "test/binary-elementwise-operator-tester.h",
5346 "test/divide-nd.cc",
5347 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005348 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005349)
5350
5351xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005352 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005353 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005354 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005355 "test/fully-connected-operator-tester.h",
5356 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005357 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005358)
5359
5360xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005361 name = "floor_nc_test",
5362 srcs = [
5363 "test/floor-nc.cc",
5364 "test/floor-operator-tester.h",
5365 ],
5366 deps = OPERATOR_TEST_DEPS,
5367)
5368
5369xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005370 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005371 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005372 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005373 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005374 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005375 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005376)
5377
5378xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005379 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005380 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005381 "test/global-average-pooling-ncw.cc",
5382 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005383 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005384 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005385)
5386
5387xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005388 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005389 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005390 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005391 "test/hardswish-operator-tester.h",
5392 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005393 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005394)
5395
5396xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005397 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005398 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005399 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005400 "test/leaky-relu-operator-tester.h",
5401 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005402 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403)
5404
5405xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005406 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005407 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005408 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005409 "test/max-pooling-operator-tester.h",
5410 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005411 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412)
5413
5414xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005415 name = "maximum_nd_test",
5416 srcs = [
5417 "test/binary-elementwise-operator-tester.h",
5418 "test/maximum-nd.cc",
5419 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005420 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005421)
5422
5423xnnpack_unit_test(
5424 name = "minimum_nd_test",
5425 srcs = [
5426 "test/binary-elementwise-operator-tester.h",
5427 "test/minimum-nd.cc",
5428 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005429 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005430)
5431
5432xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005433 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005434 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005435 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005436 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005437 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005438 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005439)
5440
5441xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005442 name = "negate_nc_test",
5443 srcs = [
5444 "test/negate-nc.cc",
5445 "test/negate-operator-tester.h",
5446 ],
5447 deps = OPERATOR_TEST_DEPS,
5448)
5449
5450xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005451 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005452 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005453 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005454 "test/prelu-operator-tester.h",
5455 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005456 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005457)
5458
5459xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005460 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005461 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005462 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005463 "test/resize-bilinear-operator-tester.h",
5464 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005465 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005466)
5467
5468xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005469 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005470 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005471 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005472 "test/sigmoid-operator-tester.h",
5473 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005474 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475)
5476
5477xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005478 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005479 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005480 "test/softmax-nc.cc",
5481 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005482 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005483 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005484)
5485
5486xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005487 name = "square_nc_test",
5488 srcs = [
5489 "test/square-nc.cc",
5490 "test/square-operator-tester.h",
5491 ],
5492 deps = OPERATOR_TEST_DEPS,
5493)
5494
5495xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005496 name = "square_root_nc_test",
5497 srcs = [
5498 "test/square-root-nc.cc",
5499 "test/square-root-operator-tester.h",
5500 ],
5501 deps = OPERATOR_TEST_DEPS,
5502)
5503
5504xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005505 name = "squared_difference_nd_test",
5506 srcs = [
5507 "test/binary-elementwise-operator-tester.h",
5508 "test/squared-difference-nd.cc",
5509 ],
5510 deps = OPERATOR_TEST_DEPS,
5511)
5512
5513xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005514 name = "subtract_nd_test",
5515 srcs = [
5516 "test/binary-elementwise-operator-tester.h",
5517 "test/subtract-nd.cc",
5518 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005519 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005520)
5521
5522xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005523 name = "truncation_nc_test",
5524 srcs = [
5525 "test/truncation-nc.cc",
5526 "test/truncation-operator-tester.h",
5527 ],
5528 deps = OPERATOR_TEST_DEPS,
5529)
5530
5531xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005532 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005533 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005534 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005535 "test/unpooling-operator-tester.h",
5536 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005537 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005538)
5539
Chao Mei6ddfc602020-05-13 22:29:36 -07005540############################### Misc unit tests ###############################
5541
5542xnnpack_unit_test(
5543 name = "memory_planner_test",
5544 srcs = [
5545 "test/memory-planner-test.cc",
5546 ],
5547 deps = [
5548 ":XNNPACK",
5549 ":memory_planner",
5550 ],
5551)
5552
Marat Dukhan08c4a432019-10-03 09:29:21 -07005553############################# Build configurations #############################
5554
Marat Dukhanb8642352019-10-30 15:43:02 -07005555# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005556config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005557 name = "xnn_enable_assembly_explicit_true",
5558 define_values = {"xnn_enable_assembly": "true"},
5559)
5560
5561# Disables usage of assembly kernels.
5562config_setting(
5563 name = "xnn_enable_assembly_explicit_false",
5564 define_values = {"xnn_enable_assembly": "false"},
5565)
5566
Marat Dukhan9de90e02020-06-18 16:04:12 -07005567# Enables usage of sparse inference.
5568config_setting(
5569 name = "xnn_enable_sparse_explicit_true",
5570 define_values = {"xnn_enable_sparse": "true"},
5571)
5572
5573# Disables usage of sparse inference.
5574config_setting(
5575 name = "xnn_enable_sparse_explicit_false",
5576 define_values = {"xnn_enable_sparse": "false"},
5577)
5578
Marat Dukhan05702cf2020-03-26 15:41:33 -07005579# Disables usage of HMP-aware optimizations.
5580config_setting(
5581 name = "xnn_enable_hmp_explicit_false",
5582 define_values = {"xnn_enable_hmp": "false"},
5583)
5584
Chao Mei6ddfc602020-05-13 22:29:36 -07005585# Enable usage of optimized memory allocation
5586config_setting(
5587 name = "xnn_enable_memopt_explicit_true",
5588 define_values = {"xnnpack_enable_memopt": "true"},
5589)
5590
5591# Disable usage of optimized memory allocation
5592config_setting(
5593 name = "xnn_enable_memopt_explicit_false",
5594 define_values = {"xnnpack_enable_memopt": "false"},
5595)
5596
Marat Dukhanb8642352019-10-30 15:43:02 -07005597# Builds with -c dbg
5598config_setting(
5599 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005600 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005601 "compilation_mode": "dbg",
5602 },
5603)
5604
5605# Builds with -c opt
5606config_setting(
5607 name = "optimized_build",
5608 values = {
5609 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005610 },
5611)
5612
5613config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005614 name = "linux_k8",
5615 values = {"cpu": "k8"},
5616)
5617
5618config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005619 name = "linux_arm",
5620 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005621)
5622
5623config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005624 name = "linux_armeabi",
5625 values = {"cpu": "armeabi"},
5626)
5627
5628config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005629 name = "linux_armhf",
5630 values = {"cpu": "armhf"},
5631)
5632
5633config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005634 name = "linux_armv7a",
5635 values = {"cpu": "armv7a"},
5636)
5637
5638config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005639 name = "linux_aarch64",
5640 values = {"cpu": "aarch64"},
5641)
5642
5643config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005644 name = "android",
5645 values = {"crosstool_top": "//external:android/crosstool"},
5646)
5647
5648config_setting(
5649 name = "android_armv7",
5650 values = {
5651 "crosstool_top": "//external:android/crosstool",
5652 "cpu": "armeabi-v7a",
5653 },
5654)
5655
5656config_setting(
5657 name = "android_arm64",
5658 values = {
5659 "crosstool_top": "//external:android/crosstool",
5660 "cpu": "arm64-v8a",
5661 },
5662)
5663
5664config_setting(
5665 name = "android_x86",
5666 values = {
5667 "crosstool_top": "//external:android/crosstool",
5668 "cpu": "x86",
5669 },
5670)
5671
5672config_setting(
5673 name = "android_x86_64",
5674 values = {
5675 "crosstool_top": "//external:android/crosstool",
5676 "cpu": "x86_64",
5677 },
5678)
5679
5680config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005681 name = "windows_x86_64",
5682 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005683)
5684
5685config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005686 name = "windows_x86_64_clang",
5687 values = {
5688 "compiler": "clang-cl",
5689 "cpu": "x64_windows",
5690 },
5691)
5692
5693config_setting(
5694 name = "windows_x86_64_mingw",
5695 values = {
5696 "compiler": "mingw-gcc",
5697 "cpu": "x64_windows",
5698 },
5699)
5700
5701config_setting(
5702 name = "windows_x86_64_msys",
5703 values = {
5704 "compiler": "msys-gcc",
5705 "cpu": "x64_windows",
5706 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005707)
5708
5709config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005710 name = "macos_x86_64",
5711 values = {
5712 "apple_platform_type": "macos",
5713 "cpu": "darwin",
5714 },
5715)
5716
5717config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005718 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005719 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005720)
5721
5722config_setting(
5723 name = "emscripten_wasm",
5724 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005725 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005726 "cpu": "wasm",
5727 },
5728)
5729
5730config_setting(
5731 name = "emscripten_wasmsimd",
5732 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005733 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005734 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005735 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005736 },
5737)
5738
5739config_setting(
5740 name = "emscripten_asmjs",
5741 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005742 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005743 "cpu": "asmjs",
5744 },
5745)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005746
5747config_setting(
5748 name = "ios_armv7",
5749 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005750 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005751 "cpu": "ios_armv7",
5752 },
5753)
5754
5755config_setting(
5756 name = "ios_arm64",
5757 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005758 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005759 "cpu": "ios_arm64",
5760 },
5761)
5762
5763config_setting(
5764 name = "ios_arm64e",
5765 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005766 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005767 "cpu": "ios_arm64e",
5768 },
5769)
5770
5771config_setting(
5772 name = "ios_x86",
5773 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005774 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005775 "cpu": "ios_i386",
5776 },
5777)
5778
5779config_setting(
5780 name = "ios_x86_64",
5781 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005782 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005783 "cpu": "ios_x86_64",
5784 },
5785)
5786
5787config_setting(
5788 name = "watchos_armv7k",
5789 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005790 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005791 "cpu": "watchos_armv7k",
5792 },
5793)
5794
5795config_setting(
5796 name = "watchos_arm64_32",
5797 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005798 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005799 "cpu": "watchos_arm64_32",
5800 },
5801)
5802
5803config_setting(
5804 name = "watchos_x86",
5805 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005806 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005807 "cpu": "watchos_i386",
5808 },
5809)
5810
5811config_setting(
5812 name = "watchos_x86_64",
5813 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005814 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005815 "cpu": "watchos_x86_64",
5816 },
5817)
5818
5819config_setting(
5820 name = "tvos_arm64",
5821 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005822 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005823 "cpu": "tvos_arm64",
5824 },
5825)
5826
5827config_setting(
5828 name = "tvos_x86_64",
5829 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005830 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005831 "cpu": "tvos_x86_64",
5832 },
5833)