blob: ad472db19a5807a3d5e672751df2aa4e9a0b0e0c [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 Dukhand3f3d872020-06-24 13:08:25 -0700805 "src/math/roundne-wasmsimd-addsub.c",
806 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700807 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700808 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700809 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700810 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700811 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700812 "src/requantization/fp32-wasmsimd.c",
813]
814
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800815PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800816 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800817 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800818 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700819 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
820 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700821 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700822 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700823 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700824 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700825 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700826 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
827 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
828 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
829 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
830 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
831 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
832 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
833 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
834 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
835 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
836 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
837 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700838 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700839 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
840 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700841 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
842 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
843 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
844 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
845 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
846 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
847 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
848 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
849 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
850 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
851 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
852 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
853 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
854 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
855 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
856 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
857 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
858 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
859 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700860 "src/f32-ibilinear/gen/psimd-c4.c",
861 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700862 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
863 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
864 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
865 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
866 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
867 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
868 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
869 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
870 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
871 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700872 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700873 "src/f32-spmm/gen/4x1-minmax-psimd.c",
874 "src/f32-spmm/gen/8x1-minmax-psimd.c",
875 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700876 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
877 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700878 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800879 "src/f32-prelu/gen/psimd-2x4.c",
880 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800881 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700882 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
883 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
884 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
885 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
886 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
887 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
888 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
889 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800890 "src/f32-vbinary/gen/vmax-psimd-x4.c",
891 "src/f32-vbinary/gen/vmax-psimd-x8.c",
892 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
893 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
894 "src/f32-vbinary/gen/vmin-psimd-x4.c",
895 "src/f32-vbinary/gen/vmin-psimd-x8.c",
896 "src/f32-vbinary/gen/vminc-psimd-x4.c",
897 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700898 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
899 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
900 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
901 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
902 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
903 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
904 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
905 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700906 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
907 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
908 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
909 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700910 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
911 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
912 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
913 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700914 "src/f32-vlrelu/gen/vlrelu-psimd-x4.c",
915 "src/f32-vlrelu/gen/vlrelu-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700916 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
917 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700918 "src/f32-vunary/gen/vabs-psimd-x4.c",
919 "src/f32-vunary/gen/vabs-psimd-x8.c",
920 "src/f32-vunary/gen/vneg-psimd-x4.c",
921 "src/f32-vunary/gen/vneg-psimd-x8.c",
922 "src/f32-vunary/gen/vsqr-psimd-x4.c",
923 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700924 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700925 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700926 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700927 "src/x32-unpool/psimd.c",
928 "src/x32-zip/x2-psimd.c",
929 "src/x32-zip/x3-psimd.c",
930 "src/x32-zip/x4-psimd.c",
931 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700932 "src/requantization/precise-psimd.c",
933 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700934]
935
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800936PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800937 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
938 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
939 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
940 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
941 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
942 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
943 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
944 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
945 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
946 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
947 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
948 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800949 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
950 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
951 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
952 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
953 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
954 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700955 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
956 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
957 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
958 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
959 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
960 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
961 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
962 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700963 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700964 "src/math/roundd-psimd-addsub.c",
965 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700966 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800967 "src/math/sigmoid-psimd-p5-div.c",
968]
969
Marat Dukhan08c4a432019-10-03 09:29:21 -0700970# ISA-specific micro-kernels
971NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700972 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
973 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700974 "src/f32-clamp/gen/neon-x4.c",
975 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700976 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
977 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700978 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
979 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700980 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
981 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700982 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
983 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700984 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
985 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
986 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
987 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700988 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
989 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
990 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
991 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700992 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
993 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
994 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
995 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700996 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700997 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
998 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700999 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1000 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1001 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1002 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1003 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1004 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1005 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1006 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1007 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1008 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1009 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1010 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1011 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1012 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1013 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1014 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1015 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1016 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1017 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1018 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1019 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1020 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1021 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1022 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1023 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1024 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1025 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1026 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1027 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1028 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1029 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001030 "src/f32-hswish/gen/neon-x4.c",
1031 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan55dde5b2020-07-10 22:48:54 -07001032 "src/f32-hswish/gen/neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001033 "src/f32-ibilinear/gen/neon-c4.c",
1034 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001035 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1036 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1037 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1038 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1039 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1040 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1041 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1042 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1043 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1044 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1045 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1046 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1047 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1048 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1049 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1050 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001051 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1052 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1053 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001054 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1055 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001056 "src/f32-prelu/gen/neon-2x4.c",
1057 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001058 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1059 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1060 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1061 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1062 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1063 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1064 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1065 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1066 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1067 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1068 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1069 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1070 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1071 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1072 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1073 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1074 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1075 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1076 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1077 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1078 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1079 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1080 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1081 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001082 "src/f32-relu/gen/neon-x4.c",
1083 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001084 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001085 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001086 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1087 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1088 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1089 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1090 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1091 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1092 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1093 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1094 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1095 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1096 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1097 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1098 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1099 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1100 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1101 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1102 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1103 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001104 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1105 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1106 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1107 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001108 "src/f32-vbinary/gen/vmax-neon-x4.c",
1109 "src/f32-vbinary/gen/vmax-neon-x8.c",
1110 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1111 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1112 "src/f32-vbinary/gen/vmin-neon-x4.c",
1113 "src/f32-vbinary/gen/vmin-neon-x8.c",
1114 "src/f32-vbinary/gen/vminc-neon-x4.c",
1115 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001116 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1117 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1118 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1119 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1120 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1121 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001122 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1123 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1124 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1125 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001126 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1127 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1128 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1129 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001130 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1131 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001132 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1133 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001134 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1135 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1136 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1137 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1138 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1139 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1140 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1141 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001142 "src/f32-vunary/gen/vabs-neon-x4.c",
1143 "src/f32-vunary/gen/vabs-neon-x8.c",
1144 "src/f32-vunary/gen/vneg-neon-x4.c",
1145 "src/f32-vunary/gen/vneg-neon-x8.c",
1146 "src/f32-vunary/gen/vsqr-neon-x4.c",
1147 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001148 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1149 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1150 "src/qu8-dwconv/up8x9-minmax-neon.c",
1151 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1152 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1153 "src/qu8-gemm/4x8-minmax-neon.c",
1154 "src/qu8-gemm/8x8-minmax-neon.c",
1155 "src/qu8-igemm/4x8-minmax-neon.c",
1156 "src/qu8-igemm/8x8-minmax-neon.c",
1157 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001158 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001159 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001160 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001161 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001162 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001163 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001164 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001165 "src/x32-zip/x2-neon.c",
1166 "src/x32-zip/x3-neon.c",
1167 "src/x32-zip/x4-neon.c",
1168 "src/x32-zip/xm-neon.c",
1169 "src/x8-zip/x2-neon.c",
1170 "src/x8-zip/x3-neon.c",
1171 "src/x8-zip/x4-neon.c",
1172 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001173 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001174 "src/math/roundd-neon-addsub.c",
1175 "src/math/roundd-neon-cvt.c",
1176 "src/math/roundu-neon-addsub.c",
1177 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001178 "src/math/roundz-neon-addsub.c",
1179 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001180 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001181 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1182 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1183 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1184 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1185 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1186 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001187 "src/math/sqrt-neon-nr1rsqrts.c",
1188 "src/math/sqrt-neon-nr2rsqrts.c",
1189 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001190 "src/requantization/precise-neon.c",
1191 "src/requantization/fp32-neon.c",
1192 "src/requantization/q31-neon.c",
1193 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001194]
1195
1196NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001197 "src/f32-ibilinear/gen/neonfma-c4.c",
1198 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001199 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1200 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1201 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1202 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1203 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1204 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1205 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1206 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1207 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001208 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1209 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1210 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1211 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001212 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1213 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1214 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1215 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001216 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1217 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1218 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1219 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001220 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1221 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1222 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1223 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1224 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1225 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1226 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1227 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1228 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1229 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1230 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1231 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1232 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1233 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1234 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1235 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1236 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1237 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001238 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1239 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001240 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1241 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1242 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1243 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1244 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1245 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1246 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1247 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1248 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1249 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1250 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1251 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1252 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1253 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1254 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1255 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1256 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1257 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1258 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1259 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1260 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1261 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1262 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1263 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001264 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1265 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1269 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1270 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1271 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1272 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1273 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1275 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1276 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1277 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1278 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1279 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1280 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1281 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1282 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1283 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1284 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1285 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1286 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1287 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1288 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1289 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1290 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1291 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1292 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1293 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1294 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1295 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1296 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1297 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1298 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1299 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1300 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1301 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1302 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1303 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1304 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1305 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1306 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1307 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1308 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1309 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1310 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1311 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1312 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1313 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1314 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1315 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1316 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001318 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1319 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001320 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1321 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1322 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1323 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1324 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1325 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1326 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1327 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1328 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1329 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1330 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1331 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1332 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1333 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1334 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1335 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1336 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1337 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1338 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1339 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001340 "src/math/exp-neonfma-lut64-p2.c",
1341 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001342 "src/math/expminus-neonfma-lut2048-p1.c",
1343 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001344 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001345 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1346 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1347 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1348 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1349 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1350 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1351 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1352 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1353 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1354 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1355 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1356 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1357 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1358 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1359 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1360 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1361 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1362 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001363 "src/math/sqrt-neonfma-nr1fma.c",
1364 "src/math/sqrt-neonfma-nr2fma.c",
1365 "src/math/sqrt-neonfma-nr3fma.c",
1366 "src/math/sqrt-neonfma-nr2fma1adj.c",
1367 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001368]
1369
1370AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001371 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1372 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1373 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1374 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1375 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1376 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001377 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1378 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1379 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1380 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1381 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1382 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1383 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1384 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1385 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1386 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1387 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1388 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1389 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1390 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1391 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1392 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1393 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1394 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1395 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1396 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001397 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1398 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001399 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1400 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001401 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1402 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001403 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1404 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001405 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1406 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1407 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1408 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1409 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001410 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1411 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1412 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1413 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1414 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1415 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1416 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1417 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1418 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1419 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1420 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1421 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1422 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1423 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1424 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1425 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1426 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1427 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001428 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1429 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1430 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1431 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1432 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1433 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1434 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1435 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1436 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1437 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1438 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1439 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1440 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1441 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1442 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1443 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1444 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1445 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001446 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1447 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001448 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1449 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1450 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1451 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1452 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1453 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001454]
1455
Marat Dukhan8853b822020-05-07 12:19:01 -07001456NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001457 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1458 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1459 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1460 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1461 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1462 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1463 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1464 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001465 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001466 "src/math/roundd-neonv8.c",
1467 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001468 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001469]
1470
Marat Dukhan08c4a432019-10-03 09:29:21 -07001471AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001472 "src/f16-clamp/gen/neonfp16arith-x16.c",
1473 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001474 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1475 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1476 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1477 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1478 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1479 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1480 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1481 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1482 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1483 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1484 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1485 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001486 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1487 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001488 "src/f16-hswish/gen/neonfp16arith-x16.c",
1489 "src/f16-hswish/gen/neonfp16arith-x8.c",
1490 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1491 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001492 "src/f16-relu/gen/neonfp16arith-x16.c",
1493 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001494 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1495 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1496 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1497 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1498 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1499 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1500 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1501 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1502 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1503 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1504 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1505 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1506 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1507 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1508 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1509 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1510 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1511 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1512 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1513 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1514 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1515 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1516 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1517 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1518 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1519 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1520 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1521 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001522 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1523 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001524 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1525 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1526 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1527 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1528 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1529 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1530 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1531 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001532 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1533 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1534 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1535 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001536 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1537 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1538 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1539 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1540 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1541 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1542 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1543 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1544 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1545 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1546 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1547 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001548 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1549 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1550 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1551 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1552 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1553 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1554 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1555 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001556]
1557
1558SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001559 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1560 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001561 "src/f32-clamp/gen/sse-x4.c",
1562 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001563 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1564 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001565 "src/f32-dwconv-chw/3x3p1-sse.c",
1566 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001567 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1568 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1569 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1570 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1571 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1572 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1573 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1574 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1575 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1576 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1577 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1578 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001579 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001580 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1581 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001582 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1583 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1584 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1585 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1586 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1587 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1588 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1589 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1590 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1591 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1592 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1593 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1594 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001595 "src/f32-hswish/gen/sse-x4.c",
1596 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001597 "src/f32-ibilinear/gen/sse-c4.c",
1598 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001599 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1600 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1601 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1602 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1603 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1604 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1605 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001606 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1607 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1608 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001609 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001610 "src/f32-prelu/gen/sse-2x4.c",
1611 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001612 "src/f32-relu/gen/sse-x4.c",
1613 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001614 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001615 "src/f32-spmm/gen/4x1-minmax-sse.c",
1616 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001617 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001618 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1619 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1620 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1621 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1622 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1623 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1624 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1625 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001626 "src/f32-vbinary/gen/vmax-sse-x4.c",
1627 "src/f32-vbinary/gen/vmax-sse-x8.c",
1628 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1629 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1630 "src/f32-vbinary/gen/vmin-sse-x4.c",
1631 "src/f32-vbinary/gen/vmin-sse-x8.c",
1632 "src/f32-vbinary/gen/vminc-sse-x4.c",
1633 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001634 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1635 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1636 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1637 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1638 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1639 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1640 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1641 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001642 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1643 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1644 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1645 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001646 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1647 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1648 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1649 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001650 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1651 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001652 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1653 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001654 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1655 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001656 "src/f32-vunary/gen/vabs-sse-x4.c",
1657 "src/f32-vunary/gen/vabs-sse-x8.c",
1658 "src/f32-vunary/gen/vneg-sse-x4.c",
1659 "src/f32-vunary/gen/vneg-sse-x8.c",
1660 "src/f32-vunary/gen/vsqr-sse-x4.c",
1661 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001662 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001663 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001664 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001665 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001666 "src/math/roundd-sse-addsub.c",
1667 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001668 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001669 "src/math/sqrt-sse-hh1mac.c",
1670 "src/math/sqrt-sse-nr1mac.c",
1671 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001672]
1673
1674SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001675 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1676 "src/f32-argmaxpool/4x-sse2-c4.c",
1677 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001678 "src/f32-prelu/gen/sse2-2x4.c",
1679 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001680 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1681 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1682 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1683 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1684 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1685 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1686 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1687 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1688 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1689 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1690 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1691 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001692 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001693 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001694 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001695 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001696 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1697 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001698 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1699 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001700 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1701 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1702 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1703 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1704 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1705 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1706 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1707 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001708 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1709 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1710 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1711 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1712 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1713 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1714 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1715 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1716 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001717 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001718 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001719 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001720 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721 "src/x32-zip/x2-sse2.c",
1722 "src/x32-zip/x3-sse2.c",
1723 "src/x32-zip/x4-sse2.c",
1724 "src/x32-zip/xm-sse2.c",
1725 "src/x8-zip/x2-sse2.c",
1726 "src/x8-zip/x3-sse2.c",
1727 "src/x8-zip/x4-sse2.c",
1728 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001729 "src/math/exp-sse2-p5.c",
1730 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001731 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001732 "src/math/roundd-sse2-cvt.c",
1733 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001734 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001735 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001736 "src/requantization/precise-sse2.c",
1737 "src/requantization/fp32-sse2.c",
1738 "src/requantization/q31-sse2.c",
1739 "src/requantization/gemmlowp-sse2.c",
1740]
1741
1742SSSE3_UKERNELS = [
1743 "src/requantization/precise-ssse3.c",
1744 "src/requantization/q31-ssse3.c",
1745 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001746]
1747
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001748SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001749 "src/f32-prelu/gen/sse41-2x4.c",
1750 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001751 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001752 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001753 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001754 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001755 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1756 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001757 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1758 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001759 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1760 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1761 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1762 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1763 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1764 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1765 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1766 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001767 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001768 "src/math/roundd-sse41.c",
1769 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001770 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001771 "src/requantization/precise-sse4.c",
1772 "src/requantization/q31-sse4.c",
1773 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001774]
1775
Marat Dukhan08c4a432019-10-03 09:29:21 -07001776AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001777 "src/f32-clamp/gen/avx-x8.c",
1778 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001779 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1780 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1781 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1782 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1783 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1784 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1785 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1786 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1787 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1788 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1789 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1790 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1791 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1792 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1793 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1794 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1795 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1796 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1797 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1798 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1799 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1800 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1801 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1802 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1803 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1804 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1805 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1806 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1807 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1808 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001809 "src/f32-hswish/gen/avx-x8.c",
1810 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001811 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1812 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1813 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1814 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1815 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1816 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1817 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1818 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1819 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001820 "src/f32-prelu/gen/avx-2x8.c",
1821 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001822 "src/f32-relu/gen/avx-x16.c",
1823 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001824 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001825 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1826 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1827 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1828 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1829 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1830 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1831 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1832 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001833 "src/f32-vbinary/gen/vmax-avx-x8.c",
1834 "src/f32-vbinary/gen/vmax-avx-x16.c",
1835 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1836 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1837 "src/f32-vbinary/gen/vmin-avx-x8.c",
1838 "src/f32-vbinary/gen/vmin-avx-x16.c",
1839 "src/f32-vbinary/gen/vminc-avx-x8.c",
1840 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001841 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1842 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1843 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1844 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1845 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1846 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1847 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1848 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001849 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1850 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1851 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1852 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001853 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1854 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1855 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1856 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001857 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1858 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001859 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001860 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1861 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1862 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1863 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1864 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1865 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1866 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1867 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001868 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1869 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001870 "src/f32-vunary/gen/vabs-avx-x8.c",
1871 "src/f32-vunary/gen/vabs-avx-x16.c",
1872 "src/f32-vunary/gen/vneg-avx-x8.c",
1873 "src/f32-vunary/gen/vneg-avx-x16.c",
1874 "src/f32-vunary/gen/vsqr-avx-x8.c",
1875 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001876]
1877
Marat Dukhanfda12b82019-11-21 12:27:59 -08001878FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001879 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1880 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1881 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1882 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1883 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1884 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1885 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1886 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1887 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1888 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1889 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1890 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1891 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1892 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1893 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1894 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1895 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1896 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1897 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1898 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1899 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1900 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1901 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1902 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1903 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1904 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1905 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1906 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1907 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1908 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1909 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1910 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1911 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1912 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1913 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1914 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1915 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1916 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1917 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1918 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001919 "src/f32-hswish/gen/fma3-x8.c",
1920 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001921 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1922 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1923 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1924 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1925 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1926 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1927 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1928 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1929 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1930 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1931 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1932 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1933 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1934 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001935 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1936 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1937 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1938 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1939 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1940 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1941 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1942 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001943 "src/math/sqrt-fma3-nr1fma.c",
1944 "src/math/sqrt-fma3-nr2fma.c",
1945 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001946]
1947
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001948AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001949 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1950 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1951 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1952 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1953 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1954 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1955 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1956 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1957 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1958 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1959 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1960 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1961 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1962 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1963 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1964 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1965 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1966 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1967 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1968 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1969 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1970 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1971 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1972 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1973 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1974 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1975 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1976 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1977 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1978 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1979 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1980 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1981 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1982 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1983 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1984 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001985 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1986 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1987 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1988 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1989 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1990 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1991 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1992 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1993 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1994 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1995 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1996 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1997 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1998 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1999 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2000 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2001 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2002 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2003 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2004 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2005 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2006 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2007 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2008 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2009 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2010 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2011 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2012 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2013 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2014 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002015 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2016 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2017 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2018 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2019 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2020 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2021 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2022 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2023 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2024 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2025 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2026 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2027 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2028 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2029 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2030 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2031 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2032 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2033 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2034 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2035 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2036 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2037 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2038 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002039 "src/math/exp-avx2-p5.c",
2040 "src/math/exp-avx2-perm-p3.c",
2041 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002042 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002043 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002044 "src/math/sigmoid-avx2-rr2-p5-div.c",
2045 "src/math/sigmoid-avx2-rr1-p5-div.c",
2046 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2047 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2048 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2049 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002050]
2051
Marat Dukhan08c4a432019-10-03 09:29:21 -07002052AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002053 "src/f32-clamp/gen/avx512f-x16.c",
2054 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002055 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2056 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2057 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2058 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2059 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2060 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2061 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2062 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2063 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2064 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2065 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2066 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2067 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2068 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2069 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2070 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2071 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2072 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2073 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2074 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2075 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2076 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2077 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2078 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08002079 "src/f32-hswish/gen/avx512f-x16.c",
2080 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002081 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2082 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2083 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2084 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2085 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2086 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002087 "src/f32-prelu/gen/avx512f-2x16.c",
2088 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002089 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2090 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2091 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2092 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2093 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2094 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2095 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2096 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2097 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2098 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2099 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2100 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2101 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2102 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2103 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2104 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2105 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2106 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2107 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2108 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2109 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2110 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2111 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2112 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2113 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2114 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2115 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2116 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2117 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2118 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2119 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2120 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2121 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2122 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2123 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2124 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002125 "src/f32-relu/gen/avx512f-x16.c",
2126 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002127 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002128 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2129 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2130 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2131 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2132 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2133 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2134 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2135 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002136 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2137 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2138 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2139 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2140 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2141 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2142 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2143 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002144 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2145 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2146 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2147 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2148 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2149 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2150 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2151 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002152 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2153 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2154 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2155 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002156 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2157 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2158 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2159 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002160 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2161 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002162 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002163 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2164 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2165 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2166 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2167 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2168 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2169 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2170 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2171 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2172 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2173 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2174 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2175 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2176 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2177 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2178 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2179 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2180 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2181 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2182 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2183 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2184 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2185 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2186 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002187 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2188 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2189 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2190 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2191 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2192 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2193 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2194 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002195 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2196 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2197 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2198 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2199 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2200 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2201 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2202 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002203 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2204 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2205 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2206 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2207 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2208 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002209 "src/math/exp-avx512f-p5-scalef.c",
2210 "src/math/exp-avx512f-p5.c",
2211 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002212 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002213 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002214 "src/math/sqrt-avx512f-nr1fma.c",
2215 "src/math/sqrt-avx512f-nr2fma.c",
2216 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002217]
2218
2219AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002220 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002221 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002222 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2223 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2224 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2225 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002226 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2227 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2228 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2229 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002230 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2231 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2232 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2233 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002234]
2235
2236AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002237 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002238 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2239 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002240 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002241 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2242 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002243 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2244 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2245 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002246 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002247 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2248 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2249 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002250 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002251 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2252 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002253 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002254 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2255 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2256 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2257 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2258 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2259 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2260 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2261 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2262 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2263 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2264 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2265 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2266 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2267 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2268 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2269 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2270 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2271 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002272 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2273 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002274 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002275 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2276 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2277 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2278 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2279 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2280 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2281 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2282 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2283 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2284 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2285 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2286 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2287 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2288 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2289 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2290 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2291 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2292 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002293 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2294 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2295 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2296 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2297 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2298 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2299 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2300 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2301 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2302 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2303 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2304 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2305 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2306 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2307 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2308 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2309 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2310 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002311]
2312
Marat Dukhan1b354632020-03-23 12:50:22 -07002313INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002314 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002315 "src/xnnpack/argmaxpool.h",
2316 "src/xnnpack/avgpool.h",
2317 "src/xnnpack/clamp.h",
2318 "src/xnnpack/common.h",
2319 "src/xnnpack/conv.h",
2320 "src/xnnpack/dwconv.h",
2321 "src/xnnpack/gavgpool.h",
2322 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002323 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002324 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002325 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002326 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002327 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002328 "src/xnnpack/lut.h",
2329 "src/xnnpack/math.h",
2330 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002331 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002332 "src/xnnpack/packx.h",
2333 "src/xnnpack/pad.h",
2334 "src/xnnpack/params.h",
2335 "src/xnnpack/pavgpool.h",
2336 "src/xnnpack/ppmm.h",
2337 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002338 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002339 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002340 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002341 "src/xnnpack/rmax.h",
2342 "src/xnnpack/scalar-utils.h",
2343 "src/xnnpack/spmm.h",
2344 "src/xnnpack/unpool.h",
2345 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002346 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002347 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002348 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002349 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002350 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002351 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002353]
2354
2355INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002356 "include/xnnpack.h",
2357 "src/xnnpack/allocator.h",
2358 "src/xnnpack/compute.h",
2359 "src/xnnpack/im2col.h",
2360 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002361 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002362 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002363 "src/xnnpack/operator.h",
2364 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002365 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002366 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002367 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002368 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002369]
2370
Marat Dukhan1b354632020-03-23 12:50:22 -07002371ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002372 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002373]
2374
Marat Dukhan1b354632020-03-23 12:50:22 -07002375MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002376 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002377 "include/xnnpack.h",
2378]
2379
Marat Dukhan1b354632020-03-23 12:50:22 -07002380MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002381 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002382 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002383 "src/xnnpack/requantization.h",
2384 "include/xnnpack.h",
2385]
2386
2387OPERATOR_TEST_PARAMS_HDRS = [
2388 "src/xnnpack/params.h",
2389 "src/xnnpack/common.h",
2390]
2391
2392WEIGHTS_PACK_HDRS = [
2393 "src/xnnpack/pack.h",
2394 "src/xnnpack/operator.h",
2395 "src/xnnpack/compute.h",
2396]
2397
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002398LOGGING_COPTS = select({
2399 # No logging in optimized mode
2400 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2401 # Full logging in debug mode
2402 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2403 # Error-only logging in default (fastbuild) mode
2404 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2405})
2406
Marat Dukhan3b59de22020-06-03 20:15:19 -07002407LOGGING_SRCS = select({
2408 # No logging in optimized mode
2409 ":optimized_build": [],
2410 "//conditions:default": [
2411 "src/operator-strings.c",
2412 "src/subgraph-strings.c",
2413 ],
2414})
2415
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002416LOGGING_HDRS = [
2417 "src/xnnpack/log.h",
2418]
2419
Marat Dukhan08c4a432019-10-03 09:29:21 -07002420xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002421 name = "tables",
2422 srcs = TABLE_SRCS,
2423 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002424 gcc_copts = xnnpack_gcc_std_copts(),
2425 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002426)
2427
2428xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002429 name = "scalar_ukernels",
2430 srcs = SCALAR_UKERNELS,
2431 hdrs = INTERNAL_HDRS,
2432 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002433 gcc_copts = xnnpack_gcc_std_copts(),
2434 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002435 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002436 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002437 "@FP16",
2438 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002439 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002440 ],
2441)
2442
2443xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002444 name = "scalar_ukernels_test_mode",
2445 srcs = SCALAR_UKERNELS,
2446 hdrs = INTERNAL_HDRS,
2447 aarch32_copts = ["-marm"],
2448 copts = [
2449 "-UNDEBUG",
2450 "-DXNN_TEST_MODE=1",
2451 ],
2452 gcc_copts = xnnpack_gcc_std_copts(),
2453 msvc_copts = xnnpack_msvc_std_copts(),
2454 deps = [
2455 ":tables",
2456 "@FP16",
2457 "@FXdiv",
2458 "@pthreadpool",
2459 ],
2460)
2461
2462xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002463 name = "wasm_ukernels",
2464 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002465 gcc_copts = xnnpack_gcc_std_copts(),
2466 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002467 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002468 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002469 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002470 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002471 "@FP16",
2472 "@FXdiv",
2473 "@pthreadpool",
2474 ],
2475)
2476
2477xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002478 name = "wasm_ukernels_test_mode",
2479 hdrs = INTERNAL_HDRS,
2480 copts = [
2481 "-UNDEBUG",
2482 "-DXNN_TEST_MODE=1",
2483 ],
2484 gcc_copts = xnnpack_gcc_std_copts(),
2485 msvc_copts = xnnpack_msvc_std_copts(),
2486 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002487 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002488 deps = [
2489 ":tables",
2490 "@FP16",
2491 "@FXdiv",
2492 "@pthreadpool",
2493 ],
2494)
2495
2496xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002497 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002498 hdrs = INTERNAL_HDRS,
2499 aarch32_copts = [
2500 "-marm",
2501 "-mfpu=neon",
2502 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002503 gcc_copts = xnnpack_gcc_std_copts(),
2504 gcc_x86_copts = ["-msse2"],
2505 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002506 optimized_copts = [
2507 "-O3",
2508 "-ffast-math",
2509 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002510 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002511 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002512 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002513 "@FP16",
2514 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002515 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002516 ],
2517)
2518
2519xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002520 name = "psimd_fastmath_ukernels_test_mode",
2521 hdrs = INTERNAL_HDRS,
2522 aarch32_copts = [
2523 "-marm",
2524 "-mfpu=neon",
2525 ],
2526 copts = [
2527 "-UNDEBUG",
2528 "-DXNN_TEST_MODE=1",
2529 ],
2530 gcc_copts = xnnpack_gcc_std_copts(),
2531 gcc_x86_copts = ["-msse2"],
2532 msvc_copts = xnnpack_msvc_std_copts(),
2533 optimized_copts = [
2534 "-O3",
2535 "-ffast-math",
2536 ],
2537 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2538 deps = [
2539 ":tables",
2540 "@FP16",
2541 "@psimd",
2542 "@pthreadpool",
2543 ],
2544)
2545
2546xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002547 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002548 hdrs = INTERNAL_HDRS,
2549 aarch32_copts = [
2550 "-marm",
2551 "-mfpu=neon",
2552 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002553 gcc_copts = xnnpack_gcc_std_copts(),
2554 gcc_x86_copts = ["-msse2"],
2555 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002556 optimized_copts = [
2557 "-O3",
2558 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002559 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002560 deps = [
2561 ":tables",
2562 "@FP16",
2563 "@psimd",
2564 "@pthreadpool",
2565 ],
2566)
2567
2568xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002569 name = "psimd_accmath_ukernels_test_mode",
2570 hdrs = INTERNAL_HDRS,
2571 aarch32_copts = [
2572 "-marm",
2573 "-mfpu=neon",
2574 ],
2575 copts = [
2576 "-UNDEBUG",
2577 "-DXNN_TEST_MODE=1",
2578 ],
2579 gcc_copts = xnnpack_gcc_std_copts(),
2580 gcc_x86_copts = ["-msse2"],
2581 msvc_copts = xnnpack_msvc_std_copts(),
2582 optimized_copts = [
2583 "-O3",
2584 ],
2585 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2586 deps = [
2587 ":tables",
2588 "@FP16",
2589 "@psimd",
2590 "@pthreadpool",
2591 ],
2592)
2593
2594xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002595 name = "neon_ukernels",
2596 hdrs = INTERNAL_HDRS,
2597 aarch32_copts = [
2598 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002599 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002600 "-mfpu=neon",
2601 ],
2602 aarch32_srcs = NEON_UKERNELS,
2603 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002604 gcc_copts = xnnpack_gcc_std_copts(),
2605 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002606 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002607 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002608 "@FP16",
2609 "@pthreadpool",
2610 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002611)
2612
2613xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002614 name = "neon_ukernels_test_mode",
2615 hdrs = INTERNAL_HDRS,
2616 aarch32_copts = [
2617 "-marm",
2618 "-march=armv7-a",
2619 "-mfpu=neon",
2620 ],
2621 aarch32_srcs = NEON_UKERNELS,
2622 aarch64_srcs = NEON_UKERNELS,
2623 copts = [
2624 "-UNDEBUG",
2625 "-DXNN_TEST_MODE=1",
2626 ],
2627 gcc_copts = xnnpack_gcc_std_copts(),
2628 msvc_copts = xnnpack_msvc_std_copts(),
2629 deps = [
2630 ":tables",
2631 "@FP16",
2632 "@pthreadpool",
2633 ],
2634)
2635
2636xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002637 name = "neonfma_ukernels",
2638 hdrs = INTERNAL_HDRS,
2639 aarch32_copts = [
2640 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002641 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002642 "-mfpu=neon-vfpv4",
2643 ],
2644 aarch32_srcs = NEONFMA_UKERNELS,
2645 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002646 apple_aarch32_copts = [
2647 "-mcpu=swift",
2648 "-mtune=generic",
2649 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002650 gcc_copts = xnnpack_gcc_std_copts(),
2651 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002652 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002653 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002654 "@FP16",
2655 "@pthreadpool",
2656 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002657)
2658
2659xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002660 name = "neonfma_ukernels_test_mode",
2661 hdrs = INTERNAL_HDRS,
2662 aarch32_copts = [
2663 "-marm",
2664 "-march=armv7-a",
2665 "-mfpu=neon-vfpv4",
2666 ],
2667 aarch32_srcs = NEONFMA_UKERNELS,
2668 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002669 apple_aarch32_copts = [
2670 "-mcpu=swift",
2671 "-mtune=generic",
2672 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002673 copts = [
2674 "-UNDEBUG",
2675 "-DXNN_TEST_MODE=1",
2676 ],
2677 gcc_copts = xnnpack_gcc_std_copts(),
2678 msvc_copts = xnnpack_msvc_std_copts(),
2679 deps = [
2680 ":tables",
2681 "@FP16",
2682 "@pthreadpool",
2683 ],
2684)
2685
2686xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002687 name = "neonv8_ukernels",
2688 hdrs = INTERNAL_HDRS,
2689 aarch32_copts = [
2690 "-marm",
2691 "-march=armv8-a",
2692 "-mfpu=neon-fp-armv8",
2693 ],
2694 aarch32_srcs = NEONV8_UKERNELS,
2695 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002696 apple_aarch32_copts = [
2697 "-mcpu=cyclone",
2698 "-mtune=generic",
2699 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002700 gcc_copts = xnnpack_gcc_std_copts(),
2701 msvc_copts = xnnpack_msvc_std_copts(),
2702 deps = [
2703 ":tables",
2704 "@FP16",
2705 "@pthreadpool",
2706 ],
2707)
2708
2709xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002710 name = "neonv8_ukernels_test_mode",
2711 hdrs = INTERNAL_HDRS,
2712 aarch32_copts = [
2713 "-marm",
2714 "-march=armv8-a",
2715 "-mfpu=neon-fp-armv8",
2716 ],
2717 aarch32_srcs = NEONV8_UKERNELS,
2718 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002719 apple_aarch32_copts = [
2720 "-mcpu=cyclone",
2721 "-mtune=generic",
2722 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002723 copts = [
2724 "-UNDEBUG",
2725 "-DXNN_TEST_MODE=1",
2726 ],
2727 gcc_copts = xnnpack_gcc_std_copts(),
2728 msvc_copts = xnnpack_msvc_std_copts(),
2729 deps = [
2730 ":tables",
2731 "@FP16",
2732 "@pthreadpool",
2733 ],
2734)
2735
2736xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737 name = "neonfp16arith_ukernels",
2738 hdrs = INTERNAL_HDRS,
2739 aarch64_copts = ["-march=armv8.2-a+fp16"],
2740 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002741 gcc_copts = xnnpack_gcc_std_copts(),
2742 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002743 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002744 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002745 "@FP16",
2746 "@pthreadpool",
2747 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002748)
2749
2750xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002751 name = "neonfp16arith_ukernels_test_mode",
2752 hdrs = INTERNAL_HDRS,
2753 aarch64_copts = ["-march=armv8.2-a+fp16"],
2754 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2755 copts = [
2756 "-UNDEBUG",
2757 "-DXNN_TEST_MODE=1",
2758 ],
2759 gcc_copts = xnnpack_gcc_std_copts(),
2760 msvc_copts = xnnpack_msvc_std_copts(),
2761 deps = [
2762 ":tables",
2763 "@FP16",
2764 "@pthreadpool",
2765 ],
2766)
2767
2768xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002769 name = "sse2_ukernels",
2770 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002771 gcc_copts = xnnpack_gcc_std_copts(),
2772 gcc_x86_copts = ["-msse2"],
2773 msvc_copts = xnnpack_msvc_std_copts(),
2774 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002775 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002776 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002777 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002778 "@FP16",
2779 "@pthreadpool",
2780 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002781)
2782
2783xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002784 name = "sse2_ukernels_test_mode",
2785 hdrs = INTERNAL_HDRS,
2786 copts = [
2787 "-UNDEBUG",
2788 "-DXNN_TEST_MODE=1",
2789 ],
2790 gcc_copts = xnnpack_gcc_std_copts(),
2791 gcc_x86_copts = ["-msse2"],
2792 msvc_copts = xnnpack_msvc_std_copts(),
2793 msvc_x86_32_copts = ["/arch:SSE2"],
2794 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2795 deps = [
2796 ":tables",
2797 "@FP16",
2798 "@pthreadpool",
2799 ],
2800)
2801
2802xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002803 name = "ssse3_ukernels",
2804 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002805 gcc_copts = xnnpack_gcc_std_copts(),
2806 gcc_x86_copts = ["-mssse3"],
2807 msvc_copts = xnnpack_msvc_std_copts(),
2808 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002809 x86_srcs = SSSE3_UKERNELS,
2810 deps = [
2811 ":tables",
2812 "@FP16",
2813 "@pthreadpool",
2814 ],
2815)
2816
2817xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002818 name = "ssse3_ukernels_test_mode",
2819 hdrs = INTERNAL_HDRS,
2820 copts = [
2821 "-UNDEBUG",
2822 "-DXNN_TEST_MODE=1",
2823 ],
2824 gcc_copts = xnnpack_gcc_std_copts(),
2825 gcc_x86_copts = ["-mssse3"],
2826 msvc_copts = xnnpack_msvc_std_copts(),
2827 msvc_x86_32_copts = ["/arch:SSE2"],
2828 x86_srcs = SSSE3_UKERNELS,
2829 deps = [
2830 ":tables",
2831 "@FP16",
2832 "@pthreadpool",
2833 ],
2834)
2835
2836xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002837 name = "sse41_ukernels",
2838 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002839 gcc_copts = xnnpack_gcc_std_copts(),
2840 gcc_x86_copts = ["-msse4.1"],
2841 msvc_copts = xnnpack_msvc_std_copts(),
2842 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002843 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002844 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002845 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002846 "@FP16",
2847 "@pthreadpool",
2848 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002849)
2850
2851xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002852 name = "sse41_ukernels_test_mode",
2853 hdrs = INTERNAL_HDRS,
2854 copts = [
2855 "-UNDEBUG",
2856 "-DXNN_TEST_MODE=1",
2857 ],
2858 gcc_copts = xnnpack_gcc_std_copts(),
2859 gcc_x86_copts = ["-msse4.1"],
2860 msvc_copts = xnnpack_msvc_std_copts(),
2861 msvc_x86_32_copts = ["/arch:SSE2"],
2862 x86_srcs = SSE41_UKERNELS,
2863 deps = [
2864 ":tables",
2865 "@FP16",
2866 "@pthreadpool",
2867 ],
2868)
2869
2870xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002871 name = "avx_ukernels",
2872 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002873 gcc_copts = xnnpack_gcc_std_copts(),
2874 gcc_x86_copts = ["-mavx"],
2875 msvc_copts = xnnpack_msvc_std_copts(),
2876 msvc_x86_32_copts = ["/arch:AVX"],
2877 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002878 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002879 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002880 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002881 "@FP16",
2882 "@pthreadpool",
2883 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002884)
2885
2886xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002887 name = "avx_ukernels_test_mode",
2888 hdrs = INTERNAL_HDRS,
2889 copts = [
2890 "-UNDEBUG",
2891 "-DXNN_TEST_MODE=1",
2892 ],
2893 gcc_copts = xnnpack_gcc_std_copts(),
2894 gcc_x86_copts = ["-mavx"],
2895 msvc_copts = xnnpack_msvc_std_copts(),
2896 msvc_x86_32_copts = ["/arch:AVX"],
2897 msvc_x86_64_copts = ["/arch:AVX"],
2898 x86_srcs = AVX_UKERNELS,
2899 deps = [
2900 ":tables",
2901 "@FP16",
2902 "@pthreadpool",
2903 ],
2904)
2905
2906xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002907 name = "fma3_ukernels",
2908 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002909 gcc_copts = xnnpack_gcc_std_copts(),
2910 gcc_x86_copts = ["-mfma"],
2911 msvc_copts = xnnpack_msvc_std_copts(),
2912 msvc_x86_32_copts = ["/arch:AVX"],
2913 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002914 x86_srcs = FMA3_UKERNELS,
2915 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002916 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002917 "@FP16",
2918 "@pthreadpool",
2919 ],
2920)
2921
2922xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002923 name = "fma3_ukernels_test_mode",
2924 hdrs = INTERNAL_HDRS,
2925 copts = [
2926 "-UNDEBUG",
2927 "-DXNN_TEST_MODE=1",
2928 ],
2929 gcc_copts = xnnpack_gcc_std_copts(),
2930 gcc_x86_copts = ["-mfma"],
2931 msvc_copts = xnnpack_msvc_std_copts(),
2932 msvc_x86_32_copts = ["/arch:AVX"],
2933 msvc_x86_64_copts = ["/arch:AVX"],
2934 x86_srcs = FMA3_UKERNELS,
2935 deps = [
2936 ":tables",
2937 "@FP16",
2938 "@pthreadpool",
2939 ],
2940)
2941
2942xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002943 name = "avx2_ukernels",
2944 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002945 gcc_copts = xnnpack_gcc_std_copts(),
2946 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002947 "-mfma",
2948 "-mavx2",
2949 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002950 msvc_copts = xnnpack_msvc_std_copts(),
2951 msvc_x86_32_copts = ["/arch:AVX2"],
2952 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002953 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002954 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002955 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002956 "@FP16",
2957 "@pthreadpool",
2958 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002959)
2960
2961xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002962 name = "avx2_ukernels_test_mode",
2963 hdrs = INTERNAL_HDRS,
2964 copts = [
2965 "-UNDEBUG",
2966 "-DXNN_TEST_MODE=1",
2967 ],
2968 gcc_copts = xnnpack_gcc_std_copts(),
2969 gcc_x86_copts = [
2970 "-mfma",
2971 "-mavx2",
2972 ],
2973 msvc_copts = xnnpack_msvc_std_copts(),
2974 msvc_x86_32_copts = ["/arch:AVX2"],
2975 msvc_x86_64_copts = ["/arch:AVX2"],
2976 x86_srcs = AVX2_UKERNELS,
2977 deps = [
2978 ":tables",
2979 "@FP16",
2980 "@pthreadpool",
2981 ],
2982)
2983
2984xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 name = "avx512f_ukernels",
2986 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002987 gcc_copts = xnnpack_gcc_std_copts(),
2988 gcc_x86_copts = ["-mavx512f"],
2989 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2990 msvc_copts = xnnpack_msvc_std_copts(),
2991 msvc_x86_32_copts = ["/arch:AVX512"],
2992 msvc_x86_64_copts = ["/arch:AVX512"],
2993 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002994 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002995 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002996 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002997 "@FP16",
2998 "@pthreadpool",
2999 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000)
3001
3002xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003003 name = "avx512f_ukernels_test_mode",
3004 hdrs = INTERNAL_HDRS,
3005 copts = [
3006 "-UNDEBUG",
3007 "-DXNN_TEST_MODE=1",
3008 ],
3009 gcc_copts = xnnpack_gcc_std_copts(),
3010 gcc_x86_copts = ["-mavx512f"],
3011 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3012 msvc_copts = xnnpack_msvc_std_copts(),
3013 msvc_x86_32_copts = ["/arch:AVX512"],
3014 msvc_x86_64_copts = ["/arch:AVX512"],
3015 msys_copts = ["-fno-asynchronous-unwind-tables"],
3016 x86_srcs = AVX512F_UKERNELS,
3017 deps = [
3018 ":tables",
3019 "@FP16",
3020 "@pthreadpool",
3021 ],
3022)
3023
3024xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 name = "asm_ukernels",
3026 hdrs = ["src/xnnpack/assembly.h"],
3027 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003028 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003029 aarch64_srcs = AARCH64_ASM_UKERNELS,
3030)
3031
Marat Dukhan3b59de22020-06-03 20:15:19 -07003032xnnpack_cc_library(
3033 name = "logging_utils",
3034 srcs = LOGGING_SRCS,
3035 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3036 copts = LOGGING_COPTS + [
3037 "-Isrc",
3038 "-Iinclude",
3039 ] + select({
3040 ":debug_build": [],
3041 "//conditions:default": xnnpack_min_size_copts(),
3042 }),
3043 gcc_copts = xnnpack_gcc_std_copts(),
3044 msvc_copts = xnnpack_msvc_std_copts(),
3045 visibility = xnnpack_visibility(),
3046 deps = [
3047 "@FP16",
3048 "@clog",
3049 "@pthreadpool",
3050 ],
3051)
3052
Marat Dukhan08c4a432019-10-03 09:29:21 -07003053xnnpack_aggregate_library(
3054 name = "ukernels",
3055 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003056 ":neon_ukernels",
3057 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003058 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003059 ":asm_ukernels",
3060 ],
3061 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003062 ":neon_ukernels",
3063 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003064 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003065 ":neonfp16arith_ukernels",
3066 ":asm_ukernels",
3067 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003068 generic_deps = [
3069 ":scalar_ukernels",
3070 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003071 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003072 ":psimd_fastmath_ukernels",
3073 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003074 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003075 wasm_deps = [
3076 ":wasm_ukernels",
3077 ],
3078 wasmsimd_deps = [
3079 ":wasm_ukernels",
3080 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003081 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003082 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003083 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003084 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003085 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003086 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003087 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003088 ":avx512f_ukernels",
3089 ],
3090)
3091
Marat Dukhan33fcf782020-05-24 14:27:15 -07003092xnnpack_aggregate_library(
3093 name = "ukernels_test_mode",
3094 aarch32_deps = [
3095 ":neon_ukernels_test_mode",
3096 ":neonfma_ukernels_test_mode",
3097 ":neonv8_ukernels_test_mode",
3098 ":asm_ukernels",
3099 ],
3100 aarch64_deps = [
3101 ":neon_ukernels_test_mode",
3102 ":neonfma_ukernels_test_mode",
3103 ":neonv8_ukernels_test_mode",
3104 ":neonfp16arith_ukernels_test_mode",
3105 ":asm_ukernels",
3106 ],
3107 generic_deps = [
3108 ":scalar_ukernels_test_mode",
3109 ],
3110 psimd_deps = [
3111 ":psimd_fastmath_ukernels_test_mode",
3112 ":psimd_accmath_ukernels_test_mode",
3113 ],
3114 wasm_deps = [
3115 ":wasm_ukernels_test_mode",
3116 ],
3117 wasmsimd_deps = [
3118 ":wasm_ukernels_test_mode",
3119 ],
3120 x86_deps = [
3121 ":sse2_ukernels_test_mode",
3122 ":ssse3_ukernels_test_mode",
3123 ":sse41_ukernels_test_mode",
3124 ":avx_ukernels_test_mode",
3125 ":fma3_ukernels_test_mode",
3126 ":avx2_ukernels_test_mode",
3127 ":avx512f_ukernels_test_mode",
3128 ],
3129)
3130
Marat Dukhan86df3842020-06-17 14:03:31 -07003131# This target must be compatible with Android NDK r17 and compile with Clang 6.
3132# It is similar to :ukernels target, but excludes microkernels which require
3133# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003134xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003135 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003136 aarch32_deps = [
3137 ":neon_ukernels",
3138 ":neonfma_ukernels",
3139 ":neonv8_ukernels",
3140 ":asm_ukernels",
3141 ],
3142 aarch64_deps = [
3143 ":neon_ukernels",
3144 ":neonfma_ukernels",
3145 ":neonv8_ukernels",
3146 ":asm_ukernels",
3147 ],
3148 generic_deps = [
3149 ":scalar_ukernels",
3150 ],
3151 psimd_deps = [
3152 ":psimd_fastmath_ukernels",
3153 ":psimd_accmath_ukernels",
3154 ],
3155 wasm_deps = [
3156 ":wasm_ukernels",
3157 ],
3158 wasmsimd_deps = [
3159 ":wasm_ukernels",
3160 ],
3161 x86_deps = [
3162 ":sse2_ukernels",
3163 ":ssse3_ukernels",
3164 ":sse41_ukernels",
3165 ":avx_ukernels",
3166 ":fma3_ukernels",
3167 ":avx2_ukernels",
3168 ":avx512f_ukernels",
3169 ],
3170)
3171
Marat Dukhan08c4a432019-10-03 09:29:21 -07003172xnnpack_cc_library(
3173 name = "im2col",
3174 srcs = ["src/im2col.c"],
3175 hdrs = [
3176 "src/xnnpack/common.h",
3177 "src/xnnpack/im2col.h",
3178 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003179 gcc_copts = xnnpack_gcc_std_copts(),
3180 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181)
3182
3183xnnpack_cc_library(
3184 name = "indirection",
3185 srcs = ["src/indirection.c"],
3186 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003187 gcc_copts = xnnpack_gcc_std_copts(),
3188 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189 deps = [
3190 "@FP16",
3191 "@FXdiv",
3192 "@pthreadpool",
3193 ],
3194)
3195
3196xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003197 name = "indirection_test_mode",
3198 srcs = ["src/indirection.c"],
3199 hdrs = INTERNAL_HDRS,
3200 copts = [
3201 "-UNDEBUG",
3202 "-DXNN_TEST_MODE=1",
3203 ],
3204 gcc_copts = xnnpack_gcc_std_copts(),
3205 msvc_copts = xnnpack_msvc_std_copts(),
3206 deps = [
3207 "@FP16",
3208 "@FXdiv",
3209 "@pthreadpool",
3210 ],
3211)
3212
3213xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003214 name = "packing",
3215 srcs = ["src/packing.c"],
3216 hdrs = INTERNAL_HDRS,
3217 gcc_copts = xnnpack_gcc_std_copts(),
3218 msvc_copts = xnnpack_msvc_std_copts(),
3219 deps = [
3220 "@FP16",
3221 "@FXdiv",
3222 "@pthreadpool",
3223 ],
3224)
3225
3226xnnpack_cc_library(
3227 name = "packing_test_mode",
3228 srcs = ["src/packing.c"],
3229 hdrs = INTERNAL_HDRS,
3230 copts = [
3231 "-UNDEBUG",
3232 "-DXNN_TEST_MODE=1",
3233 ],
3234 gcc_copts = xnnpack_gcc_std_copts(),
3235 msvc_copts = xnnpack_msvc_std_copts(),
3236 deps = [
3237 "@FP16",
3238 "@FXdiv",
3239 "@pthreadpool",
3240 ],
3241)
3242
3243xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003244 name = "operator_run",
3245 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003246 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003247 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003248 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3249 "//conditions:default": [],
3250 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003251 gcc_copts = xnnpack_gcc_std_copts(),
3252 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003254 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003255 "@FP16",
3256 "@FXdiv",
3257 "@clog",
3258 "@pthreadpool",
3259 ],
3260)
3261
Chao Mei6ddfc602020-05-13 22:29:36 -07003262xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003263 name = "operator_run_test_mode",
3264 srcs = ["src/operator-run.c"],
3265 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3266 copts = LOGGING_COPTS + [
3267 "-UNDEBUG",
3268 "-DXNN_TEST_MODE=1",
3269 ] + select({
3270 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3271 "//conditions:default": [],
3272 }),
3273 gcc_copts = xnnpack_gcc_std_copts(),
3274 msvc_copts = xnnpack_msvc_std_copts(),
3275 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003276 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003277 "@FP16",
3278 "@FXdiv",
3279 "@clog",
3280 "@pthreadpool",
3281 ],
3282)
3283
3284xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003285 name = "memory_planner",
3286 srcs = ["src/memory-planner.c"],
3287 hdrs = INTERNAL_HDRS,
3288 defines = select({
3289 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3290 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3291 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3292 }),
3293 gcc_copts = xnnpack_gcc_std_copts(),
3294 msvc_copts = xnnpack_msvc_std_copts(),
3295 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003296 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003297 "@pthreadpool",
3298 ],
3299)
3300
Marat Dukhan33fcf782020-05-24 14:27:15 -07003301xnnpack_cc_library(
3302 name = "memory_planner_test_mode",
3303 srcs = ["src/memory-planner.c"],
3304 hdrs = INTERNAL_HDRS,
3305 copts = [
3306 "-UNDEBUG",
3307 "-DXNN_TEST_MODE=1",
3308 ],
3309 defines = select({
3310 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3311 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3312 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3313 }),
3314 gcc_copts = xnnpack_gcc_std_copts(),
3315 msvc_copts = xnnpack_msvc_std_copts(),
3316 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003317 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003318 "@pthreadpool",
3319 ],
3320)
3321
Marat Dukhan08c4a432019-10-03 09:29:21 -07003322cc_library(
3323 name = "enable_assembly",
3324 defines = select({
3325 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3326 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003327 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003328 }),
3329)
3330
Marat Dukhan9de90e02020-06-18 16:04:12 -07003331cc_library(
3332 name = "enable_sparse",
3333 defines = select({
3334 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3335 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003336 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003337 }),
3338)
3339
Marat Dukhancf056b22019-10-07 10:26:29 -07003340xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003341 name = "operators",
3342 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003343 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003344 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003345 ],
3346 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003347 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003348 "-Isrc",
3349 "-Iinclude",
3350 ] + select({
3351 ":debug_build": [],
3352 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003353 }) + select({
3354 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3355 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003356 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003357 gcc_copts = xnnpack_gcc_std_copts(),
3358 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003359 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003360 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003361 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003362 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003363 "@FP16",
3364 "@FXdiv",
3365 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003366 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003367 ],
3368)
3369
Marat Dukhan10a38082020-04-17 03:58:35 -07003370xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003371 name = "operators_test_mode",
3372 srcs = OPERATOR_SRCS + [
3373 "src/memory.c",
3374 "src/operator-delete.c",
3375 ],
3376 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3377 copts = LOGGING_COPTS + [
3378 "-Isrc",
3379 "-Iinclude",
3380 "-UNDEBUG",
3381 "-DXNN_TEST_MODE=1",
3382 ] + select({
3383 ":debug_build": [],
3384 "//conditions:default": xnnpack_min_size_copts(),
3385 }) + select({
3386 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3387 "//conditions:default": [],
3388 }),
3389 gcc_copts = xnnpack_gcc_std_copts(),
3390 msvc_copts = xnnpack_msvc_std_copts(),
3391 deps = [
3392 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003393 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003394 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003395 "@FP16",
3396 "@FXdiv",
3397 "@clog",
3398 "@pthreadpool",
3399 ],
3400)
3401
3402xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003403 name = "XNNPACK",
3404 srcs = [
3405 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003406 "src/runtime.c",
3407 "src/subgraph.c",
3408 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003409 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003410 hdrs = ["include/xnnpack.h"],
3411 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003412 "-Isrc",
3413 "-Iinclude",
3414 ] + select({
3415 ":debug_build": [],
3416 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003417 }) + select({
3418 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3419 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003420 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003421 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003422 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003423 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003424 visibility = xnnpack_visibility(),
3425 deps = [
3426 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003427 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003428 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003429 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003430 ":operator_run",
3431 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003432 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003433 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003434 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003435 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003436 ] + select({
3437 ":emscripten": [],
3438 "//conditions:default": ["@cpuinfo"],
3439 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003440)
3441
Marat Dukhan10a38082020-04-17 03:58:35 -07003442xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003443 name = "XNNPACK_test_mode",
3444 srcs = [
3445 "src/init.c",
3446 "src/runtime.c",
3447 "src/subgraph.c",
3448 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003449 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003450 hdrs = ["include/xnnpack.h"],
3451 copts = LOGGING_COPTS + [
3452 "-Isrc",
3453 "-Iinclude",
3454 "-UNDEBUG",
3455 "-DXNN_TEST_MODE=1",
3456 ] + select({
3457 ":debug_build": [],
3458 "//conditions:default": xnnpack_min_size_copts(),
3459 }) + select({
3460 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3461 "//conditions:default": [],
3462 }),
3463 gcc_copts = xnnpack_gcc_std_copts(),
3464 includes = ["include"],
3465 msvc_copts = xnnpack_msvc_std_copts(),
3466 visibility = xnnpack_visibility(),
3467 deps = [
3468 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003469 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003470 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003471 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003472 ":operator_run_test_mode",
3473 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003474 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003475 "@clog",
3476 "@FP16",
3477 "@pthreadpool",
3478 ] + select({
3479 ":emscripten": [],
3480 "//conditions:default": ["@cpuinfo"],
3481 }),
3482)
3483
3484xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003485 name = "xnnpack_f32",
3486 srcs = [
3487 "src/init.c",
3488 "src/runtime.c",
3489 "src/subgraph.c",
3490 "src/tensor.c",
3491 ] + SUBGRAPH_SRCS,
3492 hdrs = ["include/xnnpack.h"],
3493 copts = LOGGING_COPTS + [
3494 "-Isrc",
3495 "-Iinclude",
3496 ] + select({
3497 ":debug_build": [],
3498 "//conditions:default": xnnpack_min_size_copts(),
3499 }) + select({
3500 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3501 "//conditions:default": [],
3502 }),
3503 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003504 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003505 "XNN_NO_U8_OPERATORS",
3506 "XNN_NO_X8_OPERATORS",
3507 "XNN_NO_F16_OPERATORS",
3508 "XNN_NO_X16_OPERATORS",
3509 ],
3510 gcc_copts = xnnpack_gcc_std_copts(),
3511 includes = ["include"],
3512 msvc_copts = xnnpack_msvc_std_copts(),
3513 visibility = xnnpack_visibility(),
3514 deps = [
3515 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003516 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003517 ":logging_utils",
3518 ":memory_planner",
3519 ":operator_run",
3520 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003521 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003522 "@clog",
3523 "@FP16",
3524 "@pthreadpool",
3525 ] + select({
3526 ":emscripten": [],
3527 "//conditions:default": ["@cpuinfo"],
3528 }),
3529)
3530
3531xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003532 name = "xnnpack_operators_nhwc_f32",
3533 srcs = [
3534 "src/init.c",
3535 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003536 hdrs = ["include/xnnpack.h"],
3537 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003538 "-Isrc",
3539 "-Iinclude",
3540 ] + select({
3541 ":debug_build": [],
3542 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003543 }) + select({
3544 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3545 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003546 }),
3547 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003548 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003549 "XNN_NO_U8_OPERATORS",
3550 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003551 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003552 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003553 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003555 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003556 visibility = xnnpack_visibility(),
3557 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003558 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003559 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003560 ":operator_run",
3561 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003562 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003563 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003565 ] + select({
3566 ":emscripten": [],
3567 "//conditions:default": ["@cpuinfo"],
3568 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003569)
3570
Marat Dukhancf056b22019-10-07 10:26:29 -07003571xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003572 name = "bench_utils",
3573 srcs = ["bench/utils.cc"],
3574 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003575 deps = [
3576 "@com_google_benchmark//:benchmark",
3577 "@cpuinfo",
3578 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579)
3580
Frank Barchard7e955972019-10-11 10:34:25 -07003581######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582
3583xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003584 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585 srcs = [
3586 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003587 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 "src/xnnpack/AlignedAllocator.h",
3589 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003590 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003591 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003592)
3593
3594xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003595 name = "f16_igemm_bench",
3596 srcs = [
3597 "bench/f16-igemm.cc",
3598 "bench/conv.h",
3599 "bench/google/conv.h",
3600 "src/xnnpack/AlignedAllocator.h",
3601 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003602 deps = MICROKERNEL_BENCHMARK_DEPS + [
3603 ":indirection",
3604 ":packing",
3605 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003606)
3607
3608xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003609 name = "f16_gemm_bench",
3610 srcs = [
3611 "bench/f16-gemm.cc",
3612 "bench/gemm.h",
3613 "src/xnnpack/AlignedAllocator.h",
3614 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003615 deps = MICROKERNEL_BENCHMARK_DEPS + [
3616 ":packing",
3617 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618)
3619
3620xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003621 name = "f16_spmm_bench",
3622 srcs = [
3623 "bench/f16-spmm.cc",
3624 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003625 "src/xnnpack/AlignedAllocator.h",
3626 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003627 deps = MICROKERNEL_BENCHMARK_DEPS,
3628)
3629
3630xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003631 name = "f32_igemm_bench",
3632 srcs = [
3633 "bench/f32-igemm.cc",
3634 "bench/conv.h",
3635 "src/xnnpack/AlignedAllocator.h",
3636 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003637 deps = MICROKERNEL_BENCHMARK_DEPS + [
3638 ":indirection",
3639 ":packing",
3640 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003641)
3642
3643xnnpack_benchmark(
3644 name = "f32_conv_hwc_bench",
3645 srcs = [
3646 "bench/f32-conv-hwc.cc",
3647 "bench/dconv.h",
3648 "src/xnnpack/AlignedAllocator.h",
3649 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003650 deps = MICROKERNEL_BENCHMARK_DEPS + [
3651 ":packing",
3652 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653)
3654
3655xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003656 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003657 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003658 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003659 "bench/dconv.h",
3660 "src/xnnpack/AlignedAllocator.h",
3661 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003662 deps = MICROKERNEL_BENCHMARK_DEPS + [
3663 ":packing",
3664 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003665)
3666
3667xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003668 name = "f16_dwconv_bench",
3669 srcs = [
3670 "bench/f16-dwconv.cc",
3671 "bench/dwconv.h",
3672 "bench/google/dwconv.h",
3673 "src/xnnpack/AlignedAllocator.h",
3674 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003675 deps = MICROKERNEL_BENCHMARK_DEPS + [
3676 ":indirection",
3677 ":packing",
3678 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003679)
3680
3681xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003682 name = "f32_dwconv_bench",
3683 srcs = [
3684 "bench/f32-dwconv.cc",
3685 "bench/dwconv.h",
3686 "src/xnnpack/AlignedAllocator.h",
3687 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003688 deps = MICROKERNEL_BENCHMARK_DEPS + [
3689 ":indirection",
3690 ":packing",
3691 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003692)
3693
3694xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003695 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003697 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003698 "bench/dwconv.h",
3699 "src/xnnpack/AlignedAllocator.h",
3700 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003701 deps = MICROKERNEL_BENCHMARK_DEPS + [
3702 ":indirection",
3703 ":packing",
3704 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003705)
3706
3707xnnpack_benchmark(
3708 name = "f32_gemm_bench",
3709 srcs = [
3710 "bench/f32-gemm.cc",
3711 "bench/gemm.h",
3712 "src/xnnpack/AlignedAllocator.h",
3713 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003714 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003715 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003716)
3717
3718xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003719 name = "f32_hswish_bench",
3720 srcs = [
3721 "bench/f32-hswish.cc",
3722 "src/xnnpack/AlignedAllocator.h",
3723 ] + MICROKERNEL_BENCHMARK_HDRS,
3724 deps = MICROKERNEL_BENCHMARK_DEPS,
3725)
3726
3727xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003728 name = "f32_raddexpminusmax_bench",
3729 srcs = [
3730 "bench/f32-raddexpminusmax.cc",
3731 "src/xnnpack/AlignedAllocator.h",
3732 ] + MICROKERNEL_BENCHMARK_HDRS,
3733 deps = MICROKERNEL_BENCHMARK_DEPS,
3734)
3735
3736xnnpack_benchmark(
3737 name = "f32_raddextexp_bench",
3738 srcs = [
3739 "bench/f32-raddextexp.cc",
3740 "src/xnnpack/AlignedAllocator.h",
3741 ] + MICROKERNEL_BENCHMARK_HDRS,
3742 deps = MICROKERNEL_BENCHMARK_DEPS,
3743)
3744
3745xnnpack_benchmark(
3746 name = "f32_raddstoreexpminusmax_bench",
3747 srcs = [
3748 "bench/f32-raddstoreexpminusmax.cc",
3749 "src/xnnpack/AlignedAllocator.h",
3750 ] + MICROKERNEL_BENCHMARK_HDRS,
3751 deps = MICROKERNEL_BENCHMARK_DEPS,
3752)
3753
3754xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003755 name = "f32_rmax_bench",
3756 srcs = [
3757 "bench/f32-rmax.cc",
3758 "src/xnnpack/AlignedAllocator.h",
3759 ] + MICROKERNEL_BENCHMARK_HDRS,
3760 deps = MICROKERNEL_BENCHMARK_DEPS,
3761)
3762
3763xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003764 name = "f32_sigmoid_bench",
3765 srcs = [
3766 "bench/f32-sigmoid.cc",
3767 "src/xnnpack/AlignedAllocator.h",
3768 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003769 deps = MICROKERNEL_BENCHMARK_DEPS,
3770)
3771
3772xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003773 name = "f32_spmm_bench",
3774 srcs = [
3775 "bench/f32-spmm.cc",
3776 "bench/gemm.h",
3777 "src/xnnpack/AlignedAllocator.h",
3778 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003779 deps = MICROKERNEL_BENCHMARK_DEPS,
3780)
3781
3782xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003783 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003784 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003785 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003786 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003787 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003788 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003789)
3790
3791xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003792 name = "f32_vscaleexpminusmax_bench",
3793 srcs = [
3794 "bench/f32-vscaleexpminusmax.cc",
3795 "src/xnnpack/AlignedAllocator.h",
3796 ] + MICROKERNEL_BENCHMARK_HDRS,
3797 deps = MICROKERNEL_BENCHMARK_DEPS,
3798)
3799
3800xnnpack_benchmark(
3801 name = "f32_vscaleextexp_bench",
3802 srcs = [
3803 "bench/f32-vscaleextexp.cc",
3804 "src/xnnpack/AlignedAllocator.h",
3805 ] + MICROKERNEL_BENCHMARK_HDRS,
3806 deps = MICROKERNEL_BENCHMARK_DEPS,
3807)
3808
3809xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003810 name = "f32_vsqrt_bench",
3811 srcs = [
3812 "bench/f32-vsqrt.cc",
3813 "src/xnnpack/AlignedAllocator.h",
3814 ] + MICROKERNEL_BENCHMARK_HDRS,
3815 deps = MICROKERNEL_BENCHMARK_DEPS,
3816)
3817
3818xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819 name = "f32_im2col_gemm_bench",
3820 srcs = [
3821 "bench/f32-im2col-gemm.cc",
3822 "bench/conv.h",
3823 "src/xnnpack/AlignedAllocator.h",
3824 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003825 deps = MICROKERNEL_BENCHMARK_DEPS + [
3826 ":im2col",
3827 ":packing",
3828 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003829)
3830
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003831xnnpack_benchmark(
3832 name = "requantization_bench",
3833 srcs = [
3834 "bench/requantization.cc",
3835 "src/xnnpack/requantization-stubs.h",
3836 "src/xnnpack/AlignedAllocator.h",
3837 ] + MICROKERNEL_BENCHMARK_HDRS,
3838 deps = MICROKERNEL_BENCHMARK_DEPS,
3839)
3840
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003841xnnpack_benchmark(
3842 name = "rounding_bench",
3843 srcs = [
3844 "bench/rounding.cc",
3845 "src/xnnpack/math-stubs.h",
3846 "src/xnnpack/AlignedAllocator.h",
3847 ] + MICROKERNEL_BENCHMARK_HDRS,
3848 deps = MICROKERNEL_BENCHMARK_DEPS,
3849)
3850
Marat Dukhan08c4a432019-10-03 09:29:21 -07003851########################### Benchmarks for operators ###########################
3852
3853xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003854 name = "average_pooling_bench",
3855 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003856 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003857 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003858 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003859)
3860
3861xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003862 name = "bankers_rounding_bench",
3863 srcs = ["bench/bankers-rounding.cc"],
3864 copts = xnnpack_optional_tflite_copts(),
3865 tags = ["nowin32"],
3866 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3867)
3868
3869xnnpack_benchmark(
3870 name = "ceiling_bench",
3871 srcs = ["bench/ceiling.cc"],
3872 copts = xnnpack_optional_tflite_copts(),
3873 tags = ["nowin32"],
3874 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3875)
3876
3877xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003878 name = "channel_shuffle_bench",
3879 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003880 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003881)
3882
3883xnnpack_benchmark(
3884 name = "convolution_bench",
3885 srcs = ["bench/convolution.cc"],
3886 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003887 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003888 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003889)
3890
3891xnnpack_benchmark(
3892 name = "deconvolution_bench",
3893 srcs = ["bench/deconvolution.cc"],
3894 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003895 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003896 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003897)
3898
3899xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003900 name = "floor_bench",
3901 srcs = ["bench/floor.cc"],
3902 copts = xnnpack_optional_tflite_copts(),
3903 tags = ["nowin32"],
3904 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3905)
3906
3907xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003908 name = "global_average_pooling_bench",
3909 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003910 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003911)
3912
3913xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003914 name = "hardswish_bench",
3915 srcs = ["bench/hardswish.cc"],
3916 copts = xnnpack_optional_tflite_copts(),
3917 tags = ["nowin32"],
3918 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3919)
3920
3921xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003922 name = "max_pooling_bench",
3923 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003924 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003925)
3926
3927xnnpack_benchmark(
3928 name = "sigmoid_bench",
3929 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003930 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003931 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003932 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003933)
3934
3935xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003936 name = "prelu_bench",
3937 srcs = ["bench/prelu.cc"],
3938 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003939 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003940 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003941)
3942
3943xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003944 name = "softmax_bench",
3945 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003946 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003947 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003948 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003949)
3950
Marat Dukhan87727142020-06-24 15:24:10 -07003951xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003952 name = "square_root_bench",
3953 srcs = ["bench/square-root.cc"],
3954 copts = xnnpack_optional_tflite_copts(),
3955 tags = ["nowin32"],
3956 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3957)
3958
3959xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003960 name = "truncation_bench",
3961 srcs = ["bench/truncation.cc"],
3962 deps = OPERATOR_BENCHMARK_DEPS,
3963)
3964
Marat Dukhanc068bb62019-10-04 13:24:39 -07003965############################# End-to-end benchmarks ############################
3966
3967cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003968 name = "fp32_mobilenet_v1",
3969 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003970 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003971 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003972 linkstatic = True,
3973 deps = [
3974 ":XNNPACK",
3975 "@pthreadpool",
3976 ],
3977)
3978
3979cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003980 name = "fp16_mobilenet_v1",
3981 srcs = ["models/fp16-mobilenet-v1.cc"],
3982 hdrs = ["models/models.h"],
3983 copts = xnnpack_std_cxxopts(),
3984 linkstatic = True,
3985 deps = [
3986 ":XNNPACK",
3987 "@FP16",
3988 "@pthreadpool",
3989 ],
3990)
3991
3992cc_library(
3993 name = "fp32_mobilenet_v2",
3994 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003995 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003996 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003997 linkstatic = True,
3998 deps = [
3999 ":XNNPACK",
4000 "@pthreadpool",
4001 ],
4002)
4003
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004004cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004005 name = "fp16_mobilenet_v2",
4006 srcs = ["models/fp16-mobilenet-v2.cc"],
4007 hdrs = ["models/models.h"],
4008 copts = xnnpack_std_cxxopts(),
4009 linkstatic = True,
4010 deps = [
4011 ":XNNPACK",
4012 "@FP16",
4013 "@pthreadpool",
4014 ],
4015)
4016
4017cc_library(
4018 name = "fp32_mobilenet_v3_large",
4019 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004020 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004021 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004022 linkstatic = True,
4023 deps = [
4024 ":XNNPACK",
4025 "@pthreadpool",
4026 ],
4027)
4028
4029cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004030 name = "fp32_mobilenet_v3_small",
4031 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004032 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004033 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004034 linkstatic = True,
4035 deps = [
4036 ":XNNPACK",
4037 "@pthreadpool",
4038 ],
4039)
4040
Marat Dukhanc068bb62019-10-04 13:24:39 -07004041xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004042 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004043 srcs = [
4044 "bench/f32-dwconv-e2e.cc",
4045 "bench/end2end.h",
4046 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004047 deps = MICROKERNEL_BENCHMARK_DEPS + [
4048 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004049 ":fp32_mobilenet_v1",
4050 ":fp32_mobilenet_v2",
4051 ":fp32_mobilenet_v3_large",
4052 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004053 ],
4054)
4055
4056xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004057 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004058 srcs = [
4059 "bench/f32-gemm-e2e.cc",
4060 "bench/end2end.h",
4061 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004062 deps = MICROKERNEL_BENCHMARK_DEPS + [
4063 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004064 ":fp32_mobilenet_v1",
4065 ":fp32_mobilenet_v2",
4066 ":fp32_mobilenet_v3_large",
4067 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004068 ],
4069)
4070
4071xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004072 name = "end2end_bench",
4073 srcs = ["bench/end2end.cc"],
4074 deps = [
4075 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004076 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004077 ":fp16_mobilenet_v1",
4078 ":fp16_mobilenet_v2",
4079 ":fp32_mobilenet_v1",
4080 ":fp32_mobilenet_v2",
4081 ":fp32_mobilenet_v3_large",
4082 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004083 "@pthreadpool",
4084 ],
4085)
4086
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004087#################### Accuracy evaluation for math functions ####################
4088
4089xnnpack_benchmark(
4090 name = "f32_exp_eval",
4091 srcs = [
4092 "eval/f32-exp.cc",
4093 "src/xnnpack/AlignedAllocator.h",
4094 ] + ACCURACY_EVAL_HDRS,
4095 deps = ACCURACY_EVAL_DEPS,
4096)
4097
Marat Dukhan515c9772019-10-17 18:07:57 -07004098xnnpack_benchmark(
4099 name = "f32_expminus_eval",
4100 srcs = [
4101 "eval/f32-expminus.cc",
4102 "src/xnnpack/AlignedAllocator.h",
4103 ] + ACCURACY_EVAL_HDRS,
4104 deps = ACCURACY_EVAL_DEPS,
4105)
4106
Marat Dukhan98ba4412019-10-23 02:14:28 -07004107xnnpack_benchmark(
4108 name = "f32_extexp_eval",
4109 srcs = [
4110 "eval/f32-extexp.cc",
4111 "src/xnnpack/AlignedAllocator.h",
4112 ] + ACCURACY_EVAL_HDRS,
4113 deps = ACCURACY_EVAL_DEPS,
4114)
4115
Marat Dukhan8853b822020-05-07 12:19:01 -07004116xnnpack_unit_test(
4117 name = "f32_roundne_eval",
4118 srcs = [
4119 "eval/f32-roundne.cc",
4120 "src/xnnpack/AlignedAllocator.h",
4121 "src/xnnpack/math-stubs.h",
4122 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004123 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004124 deps = MICROKERNEL_TEST_DEPS,
4125)
4126
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004127xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004128 name = "f32_roundd_eval",
4129 srcs = [
4130 "eval/f32-roundd.cc",
4131 "src/xnnpack/AlignedAllocator.h",
4132 "src/xnnpack/math-stubs.h",
4133 ] + MICROKERNEL_TEST_HDRS,
4134 automatic = False,
4135 deps = MICROKERNEL_TEST_DEPS,
4136)
4137
4138xnnpack_unit_test(
4139 name = "f32_roundu_eval",
4140 srcs = [
4141 "eval/f32-roundu.cc",
4142 "src/xnnpack/AlignedAllocator.h",
4143 "src/xnnpack/math-stubs.h",
4144 ] + MICROKERNEL_TEST_HDRS,
4145 automatic = False,
4146 deps = MICROKERNEL_TEST_DEPS,
4147)
4148
4149xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004150 name = "f32_roundz_eval",
4151 srcs = [
4152 "eval/f32-roundz.cc",
4153 "src/xnnpack/AlignedAllocator.h",
4154 "src/xnnpack/math-stubs.h",
4155 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004156 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004157 deps = MICROKERNEL_TEST_DEPS,
4158)
4159
Marat Dukhan346a9e52019-11-15 09:06:30 -08004160xnnpack_benchmark(
4161 name = "f32_sigmoid_eval",
4162 srcs = [
4163 "eval/f32-sigmoid.cc",
4164 "src/xnnpack/AlignedAllocator.h",
4165 ] + ACCURACY_EVAL_HDRS,
4166 deps = ACCURACY_EVAL_DEPS,
4167)
4168
Marat Dukhan84000762020-06-29 18:38:43 -07004169xnnpack_benchmark(
4170 name = "f32_sqrt_eval",
4171 srcs = [
4172 "eval/f32-sqrt.cc",
4173 "src/xnnpack/AlignedAllocator.h",
4174 ] + ACCURACY_EVAL_HDRS,
4175 deps = ACCURACY_EVAL_DEPS + [
4176 ":bench_utils",
4177 ],
4178)
4179
Marat Dukhan08c4a432019-10-03 09:29:21 -07004180######################### Unit tests for micro-kernels #########################
4181
4182xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004183 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004184 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004185 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004186 "test/gemm-microkernel-tester.h",
4187 "src/xnnpack/AlignedAllocator.h",
4188 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004189 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004190)
4191
4192xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004193 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004194 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004195 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004196 "test/spmm-microkernel-tester.h",
4197 "src/xnnpack/AlignedAllocator.h",
4198 ] + MICROKERNEL_TEST_HDRS,
4199 deps = MICROKERNEL_TEST_DEPS,
4200)
4201
4202xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203 name = "f32_argmaxpool_test",
4204 srcs = [
4205 "test/f32-argmaxpool.cc",
4206 "test/argmaxpool-microkernel-tester.h",
4207 "src/xnnpack/AlignedAllocator.h",
4208 ] + MICROKERNEL_TEST_HDRS,
4209 deps = MICROKERNEL_TEST_DEPS,
4210)
4211
4212xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004213 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004214 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004215 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004216 "test/avgpool-microkernel-tester.h",
4217 "src/xnnpack/AlignedAllocator.h",
4218 ] + MICROKERNEL_TEST_HDRS,
4219 deps = MICROKERNEL_TEST_DEPS,
4220)
4221
4222xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004223 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004224 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004225 "test/f32-ibilinear.cc",
4226 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004227 "src/xnnpack/AlignedAllocator.h",
4228 ] + MICROKERNEL_TEST_HDRS,
4229 deps = MICROKERNEL_TEST_DEPS,
4230)
4231
4232xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004233 name = "f16_clamp_test",
4234 srcs = [
4235 "test/f16-clamp.cc",
4236 "test/clamp-microkernel-tester.h",
4237 ] + MICROKERNEL_TEST_HDRS,
4238 deps = MICROKERNEL_TEST_DEPS,
4239)
4240
4241xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004242 name = "f32_clamp_test",
4243 srcs = [
4244 "test/f32-clamp.cc",
4245 "test/clamp-microkernel-tester.h",
4246 ] + MICROKERNEL_TEST_HDRS,
4247 deps = MICROKERNEL_TEST_DEPS,
4248)
4249
4250xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004251 name = "f32_igemm_test",
4252 srcs = [
4253 "test/f32-igemm.cc",
4254 "test/gemm-microkernel-tester.h",
4255 "src/xnnpack/AlignedAllocator.h",
4256 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004257 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004258)
4259
4260xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004261 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004262 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004263 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004264 "test/gemm-microkernel-tester.h",
4265 "src/xnnpack/AlignedAllocator.h",
4266 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004267 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004268)
4269
4270xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004271 name = "f32_igemm_minmax_test",
4272 srcs = [
4273 "test/f32-igemm-minmax.cc",
4274 "test/gemm-microkernel-tester.h",
4275 "src/xnnpack/AlignedAllocator.h",
4276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004277 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004278)
4279
4280xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004281 name = "f16_igemm_minmax_test",
4282 srcs = [
4283 "test/f16-igemm-minmax.cc",
4284 "test/gemm-microkernel-tester.h",
4285 "src/xnnpack/AlignedAllocator.h",
4286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004287 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004288)
4289
4290xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004291 name = "f32_conv_hwc_test",
4292 srcs = [
4293 "test/f32-conv-hwc.cc",
4294 "test/conv-hwc-microkernel-tester.h",
4295 "src/xnnpack/AlignedAllocator.h",
4296 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004297 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004298)
4299
4300xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004301 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004302 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004303 "test/f32-conv-hwc2chw.cc",
4304 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305 "src/xnnpack/AlignedAllocator.h",
4306 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004307 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308)
4309
4310xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004311 name = "f32_dwconv_test",
4312 srcs = [
4313 "test/f32-dwconv.cc",
4314 "test/dwconv-microkernel-tester.h",
4315 "src/xnnpack/AlignedAllocator.h",
4316 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004317 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004318)
4319
4320xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004321 name = "f16_dwconv_minmax_test",
4322 srcs = [
4323 "test/f16-dwconv-minmax.cc",
4324 "test/dwconv-microkernel-tester.h",
4325 "src/xnnpack/AlignedAllocator.h",
4326 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004327 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004328)
4329
4330xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004331 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004332 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004333 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004334 "test/dwconv-microkernel-tester.h",
4335 "src/xnnpack/AlignedAllocator.h",
4336 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004337 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004338)
4339
4340xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004341 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004342 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004343 "test/f32-dwconv-chw.cc",
4344 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 "src/xnnpack/AlignedAllocator.h",
4346 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004347 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348)
4349
4350xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004351 name = "f16_gavgpool_minmax_test",
4352 srcs = [
4353 "test/f16-gavgpool-minmax.cc",
4354 "test/gavgpool-microkernel-tester.h",
4355 "src/xnnpack/AlignedAllocator.h",
4356 ] + MICROKERNEL_TEST_HDRS,
4357 deps = MICROKERNEL_TEST_DEPS,
4358)
4359
4360xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004361 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004362 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004363 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364 "test/gavgpool-microkernel-tester.h",
4365 "src/xnnpack/AlignedAllocator.h",
4366 ] + MICROKERNEL_TEST_HDRS,
4367 deps = MICROKERNEL_TEST_DEPS,
4368)
4369
4370xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004371 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004372 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004373 "test/f32-gavgpool-cw.cc",
4374 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004375 "src/xnnpack/AlignedAllocator.h",
4376 ] + MICROKERNEL_TEST_HDRS,
4377 deps = MICROKERNEL_TEST_DEPS,
4378)
4379
4380xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004381 name = "f32_gemm_test",
4382 srcs = [
4383 "test/f32-gemm.cc",
4384 "test/gemm-microkernel-tester.h",
4385 "src/xnnpack/AlignedAllocator.h",
4386 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004387 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004388)
4389
4390xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004391 name = "f32_gemm_relu_test",
4392 srcs = [
4393 "test/f32-gemm-relu.cc",
4394 "test/gemm-microkernel-tester.h",
4395 "src/xnnpack/AlignedAllocator.h",
4396 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004397 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004398)
4399
4400xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004401 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004402 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004403 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004404 "test/gemm-microkernel-tester.h",
4405 "src/xnnpack/AlignedAllocator.h",
4406 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004407 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004408)
4409
4410xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004411 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004412 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004413 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004414 "test/gemm-microkernel-tester.h",
4415 "src/xnnpack/AlignedAllocator.h",
4416 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004417 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004418)
4419
4420xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004421 name = "f16_hswish_test",
4422 srcs = [
4423 "test/f16-hswish.cc",
4424 "test/hswish-microkernel-tester.h",
4425 ] + MICROKERNEL_TEST_HDRS,
4426 deps = MICROKERNEL_TEST_DEPS,
4427)
4428
4429xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004430 name = "f32_hswish_test",
4431 srcs = [
4432 "test/f32-hswish.cc",
4433 "test/hswish-microkernel-tester.h",
4434 ] + MICROKERNEL_TEST_HDRS,
4435 deps = MICROKERNEL_TEST_DEPS,
4436)
4437
4438xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004439 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004440 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004441 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004442 "test/maxpool-microkernel-tester.h",
4443 ] + MICROKERNEL_TEST_HDRS,
4444 deps = MICROKERNEL_TEST_DEPS,
4445)
4446
4447xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004448 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004449 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004450 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 "test/avgpool-microkernel-tester.h",
4452 "src/xnnpack/AlignedAllocator.h",
4453 ] + MICROKERNEL_TEST_HDRS,
4454 deps = MICROKERNEL_TEST_DEPS,
4455)
4456
4457xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004458 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004460 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461 "test/gemm-microkernel-tester.h",
4462 "src/xnnpack/AlignedAllocator.h",
4463 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004464 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465)
4466
4467xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004468 name = "f16_prelu_test",
4469 srcs = [
4470 "test/f16-prelu.cc",
4471 "test/prelu-microkernel-tester.h",
4472 "src/xnnpack/AlignedAllocator.h",
4473 ] + MICROKERNEL_TEST_HDRS,
4474 deps = MICROKERNEL_TEST_DEPS,
4475)
4476
4477xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004478 name = "f32_prelu_test",
4479 srcs = [
4480 "test/f32-prelu.cc",
4481 "test/prelu-microkernel-tester.h",
4482 "src/xnnpack/AlignedAllocator.h",
4483 ] + MICROKERNEL_TEST_HDRS,
4484 deps = MICROKERNEL_TEST_DEPS,
4485)
4486
4487xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004488 name = "f32_raddexpminusmax_test",
4489 srcs = [
4490 "test/f32-raddexpminusmax.cc",
4491 "test/raddexpminusmax-microkernel-tester.h",
4492 ] + MICROKERNEL_TEST_HDRS,
4493 deps = MICROKERNEL_TEST_DEPS,
4494)
4495
4496xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004497 name = "f32_raddextexp_test",
4498 srcs = [
4499 "test/f32-raddextexp.cc",
4500 "test/raddextexp-microkernel-tester.h",
4501 ] + MICROKERNEL_TEST_HDRS,
4502 deps = MICROKERNEL_TEST_DEPS,
4503)
4504
4505xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004506 name = "f32_raddstoreexpminusmax_test",
4507 srcs = [
4508 "test/f32-raddstoreexpminusmax.cc",
4509 "test/raddstoreexpminusmax-microkernel-tester.h",
4510 ] + MICROKERNEL_TEST_HDRS,
4511 deps = MICROKERNEL_TEST_DEPS,
4512)
4513
4514xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004515 name = "f32_relu_test",
4516 srcs = [
4517 "test/f32-relu.cc",
4518 "test/vunary-microkernel-tester.h",
4519 ] + MICROKERNEL_TEST_HDRS,
4520 deps = MICROKERNEL_TEST_DEPS,
4521)
4522
4523xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004524 name = "f32_rmax_test",
4525 srcs = [
4526 "test/f32-rmax.cc",
4527 "test/rmax-microkernel-tester.h",
4528 ] + MICROKERNEL_TEST_HDRS,
4529 deps = MICROKERNEL_TEST_DEPS,
4530)
4531
4532xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004533 name = "f32_sigmoid_test",
4534 srcs = [
4535 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004536 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004537 ] + MICROKERNEL_TEST_HDRS,
4538 deps = MICROKERNEL_TEST_DEPS,
4539)
4540
4541xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004542 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004543 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004544 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004545 "test/spmm-microkernel-tester.h",
4546 "src/xnnpack/AlignedAllocator.h",
4547 ] + MICROKERNEL_TEST_HDRS,
4548 deps = MICROKERNEL_TEST_DEPS,
4549)
4550
4551xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004552 name = "f16_vadd_minmax_test",
4553 srcs = [
4554 "test/f16-vadd-minmax.cc",
4555 "test/vbinary-microkernel-tester.h",
4556 ] + MICROKERNEL_TEST_HDRS,
4557 deps = MICROKERNEL_TEST_DEPS,
4558)
4559
4560xnnpack_unit_test(
4561 name = "f16_vaddc_minmax_test",
4562 srcs = [
4563 "test/f16-vaddc-minmax.cc",
4564 "test/vbinaryc-microkernel-tester.h",
4565 ] + MICROKERNEL_TEST_HDRS,
4566 deps = MICROKERNEL_TEST_DEPS,
4567)
4568
4569xnnpack_unit_test(
4570 name = "f16_vdiv_minmax_test",
4571 srcs = [
4572 "test/f16-vdiv-minmax.cc",
4573 "test/vbinary-microkernel-tester.h",
4574 ] + MICROKERNEL_TEST_HDRS,
4575 deps = MICROKERNEL_TEST_DEPS,
4576)
4577
4578xnnpack_unit_test(
4579 name = "f16_vdivc_minmax_test",
4580 srcs = [
4581 "test/f16-vdivc-minmax.cc",
4582 "test/vbinaryc-microkernel-tester.h",
4583 ] + MICROKERNEL_TEST_HDRS,
4584 deps = MICROKERNEL_TEST_DEPS,
4585)
4586
4587xnnpack_unit_test(
4588 name = "f16_vrdivc_minmax_test",
4589 srcs = [
4590 "test/f16-vrdivc-minmax.cc",
4591 "test/vbinaryc-microkernel-tester.h",
4592 ] + MICROKERNEL_TEST_HDRS,
4593 deps = MICROKERNEL_TEST_DEPS,
4594)
4595
4596xnnpack_unit_test(
4597 name = "f16_vmax_test",
4598 srcs = [
4599 "test/f16-vmax.cc",
4600 "test/vbinary-microkernel-tester.h",
4601 ] + MICROKERNEL_TEST_HDRS,
4602 deps = MICROKERNEL_TEST_DEPS,
4603)
4604
4605xnnpack_unit_test(
4606 name = "f16_vmaxc_test",
4607 srcs = [
4608 "test/f16-vmaxc.cc",
4609 "test/vbinaryc-microkernel-tester.h",
4610 ] + MICROKERNEL_TEST_HDRS,
4611 deps = MICROKERNEL_TEST_DEPS,
4612)
4613
4614xnnpack_unit_test(
4615 name = "f16_vmin_test",
4616 srcs = [
4617 "test/f16-vmin.cc",
4618 "test/vbinary-microkernel-tester.h",
4619 ] + MICROKERNEL_TEST_HDRS,
4620 deps = MICROKERNEL_TEST_DEPS,
4621)
4622
4623xnnpack_unit_test(
4624 name = "f16_vminc_test",
4625 srcs = [
4626 "test/f16-vminc.cc",
4627 "test/vbinaryc-microkernel-tester.h",
4628 ] + MICROKERNEL_TEST_HDRS,
4629 deps = MICROKERNEL_TEST_DEPS,
4630)
4631
4632xnnpack_unit_test(
4633 name = "f16_vmul_minmax_test",
4634 srcs = [
4635 "test/f16-vmul-minmax.cc",
4636 "test/vbinary-microkernel-tester.h",
4637 ] + MICROKERNEL_TEST_HDRS,
4638 deps = MICROKERNEL_TEST_DEPS,
4639)
4640
4641xnnpack_unit_test(
4642 name = "f16_vmulc_minmax_test",
4643 srcs = [
4644 "test/f16-vmulc-minmax.cc",
4645 "test/vbinaryc-microkernel-tester.h",
4646 ] + MICROKERNEL_TEST_HDRS,
4647 deps = MICROKERNEL_TEST_DEPS,
4648)
4649
4650xnnpack_unit_test(
4651 name = "f16_vsub_minmax_test",
4652 srcs = [
4653 "test/f16-vsub-minmax.cc",
4654 "test/vbinary-microkernel-tester.h",
4655 ] + MICROKERNEL_TEST_HDRS,
4656 deps = MICROKERNEL_TEST_DEPS,
4657)
4658
4659xnnpack_unit_test(
4660 name = "f16_vsubc_minmax_test",
4661 srcs = [
4662 "test/f16-vsubc-minmax.cc",
4663 "test/vbinaryc-microkernel-tester.h",
4664 ] + MICROKERNEL_TEST_HDRS,
4665 deps = MICROKERNEL_TEST_DEPS,
4666)
4667
4668xnnpack_unit_test(
4669 name = "f16_vrsubc_minmax_test",
4670 srcs = [
4671 "test/f16-vrsubc-minmax.cc",
4672 "test/vbinaryc-microkernel-tester.h",
4673 ] + MICROKERNEL_TEST_HDRS,
4674 deps = MICROKERNEL_TEST_DEPS,
4675)
4676
4677xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004678 name = "f32_vabs_test",
4679 srcs = [
4680 "test/f32-vabs.cc",
4681 "test/vunary-microkernel-tester.h",
4682 ] + MICROKERNEL_TEST_HDRS,
4683 deps = MICROKERNEL_TEST_DEPS,
4684)
4685
4686xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004687 name = "f32_vadd_test",
4688 srcs = [
4689 "test/f32-vadd.cc",
4690 "test/vbinary-microkernel-tester.h",
4691 ] + MICROKERNEL_TEST_HDRS,
4692 deps = MICROKERNEL_TEST_DEPS,
4693)
4694
4695xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004696 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004697 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004698 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004699 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004700 ] + MICROKERNEL_TEST_HDRS,
4701 deps = MICROKERNEL_TEST_DEPS,
4702)
4703
4704xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004705 name = "f32_vaddc_test",
4706 srcs = [
4707 "test/f32-vaddc.cc",
4708 "test/vbinaryc-microkernel-tester.h",
4709 ] + MICROKERNEL_TEST_HDRS,
4710 deps = MICROKERNEL_TEST_DEPS,
4711)
4712
4713xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004714 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004715 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004716 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004717 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004718 ] + MICROKERNEL_TEST_HDRS,
4719 deps = MICROKERNEL_TEST_DEPS,
4720)
4721
4722xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004723 name = "f32_vdiv_test",
4724 srcs = [
4725 "test/f32-vdiv.cc",
4726 "test/vbinary-microkernel-tester.h",
4727 ] + MICROKERNEL_TEST_HDRS,
4728 deps = MICROKERNEL_TEST_DEPS,
4729)
4730
4731xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004732 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004733 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004734 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004735 "test/vbinary-microkernel-tester.h",
4736 ] + MICROKERNEL_TEST_HDRS,
4737 deps = MICROKERNEL_TEST_DEPS,
4738)
4739
4740xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004741 name = "f32_vdivc_test",
4742 srcs = [
4743 "test/f32-vdivc.cc",
4744 "test/vbinaryc-microkernel-tester.h",
4745 ] + MICROKERNEL_TEST_HDRS,
4746 deps = MICROKERNEL_TEST_DEPS,
4747)
4748
4749xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004750 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004751 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004752 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004753 "test/vbinaryc-microkernel-tester.h",
4754 ] + MICROKERNEL_TEST_HDRS,
4755 deps = MICROKERNEL_TEST_DEPS,
4756)
4757
4758xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004759 name = "f32_vrdivc_test",
4760 srcs = [
4761 "test/f32-vrdivc.cc",
4762 "test/vbinaryc-microkernel-tester.h",
4763 ] + MICROKERNEL_TEST_HDRS,
4764 deps = MICROKERNEL_TEST_DEPS,
4765)
4766
4767xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004768 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004769 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004770 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004771 "test/vbinaryc-microkernel-tester.h",
4772 ] + MICROKERNEL_TEST_HDRS,
4773 deps = MICROKERNEL_TEST_DEPS,
4774)
4775
4776xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004777 name = "f32_vmax_test",
4778 srcs = [
4779 "test/f32-vmax.cc",
4780 "test/vbinary-microkernel-tester.h",
4781 ] + MICROKERNEL_TEST_HDRS,
4782 deps = MICROKERNEL_TEST_DEPS,
4783)
4784
4785xnnpack_unit_test(
4786 name = "f32_vmaxc_test",
4787 srcs = [
4788 "test/f32-vmaxc.cc",
4789 "test/vbinaryc-microkernel-tester.h",
4790 ] + MICROKERNEL_TEST_HDRS,
4791 deps = MICROKERNEL_TEST_DEPS,
4792)
4793
4794xnnpack_unit_test(
4795 name = "f32_vmin_test",
4796 srcs = [
4797 "test/f32-vmin.cc",
4798 "test/vbinary-microkernel-tester.h",
4799 ] + MICROKERNEL_TEST_HDRS,
4800 deps = MICROKERNEL_TEST_DEPS,
4801)
4802
4803xnnpack_unit_test(
4804 name = "f32_vminc_test",
4805 srcs = [
4806 "test/f32-vminc.cc",
4807 "test/vbinaryc-microkernel-tester.h",
4808 ] + MICROKERNEL_TEST_HDRS,
4809 deps = MICROKERNEL_TEST_DEPS,
4810)
4811
4812xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004813 name = "f32_vmul_test",
4814 srcs = [
4815 "test/f32-vmul.cc",
4816 "test/vbinary-microkernel-tester.h",
4817 ] + MICROKERNEL_TEST_HDRS,
4818 deps = MICROKERNEL_TEST_DEPS,
4819)
4820
4821xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004822 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004823 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004824 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004825 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004826 ] + MICROKERNEL_TEST_HDRS,
4827 deps = MICROKERNEL_TEST_DEPS,
4828)
4829
4830xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004831 name = "f32_vmulc_test",
4832 srcs = [
4833 "test/f32-vmulc.cc",
4834 "test/vbinaryc-microkernel-tester.h",
4835 ] + MICROKERNEL_TEST_HDRS,
4836 deps = MICROKERNEL_TEST_DEPS,
4837)
4838
4839xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004840 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004841 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004842 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004843 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004844 ] + MICROKERNEL_TEST_HDRS,
4845 deps = MICROKERNEL_TEST_DEPS,
4846)
4847
4848xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004849 name = "f16_vmulcaddc_minmax_test",
4850 srcs = [
4851 "test/f16-vmulcaddc-minmax.cc",
4852 "test/vmulcaddc-microkernel-tester.h",
4853 "src/xnnpack/AlignedAllocator.h",
4854 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004855 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004856)
4857
4858xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004859 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004860 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004861 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004862 "test/vmulcaddc-microkernel-tester.h",
4863 "src/xnnpack/AlignedAllocator.h",
4864 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004865 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004866)
4867
4868xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004869 name = "f32_vlrelu_test",
4870 srcs = [
4871 "test/f32-vlrelu.cc",
4872 "test/vunary-microkernel-tester.h",
4873 ] + MICROKERNEL_TEST_HDRS,
4874 deps = MICROKERNEL_TEST_DEPS,
4875)
4876
4877xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004878 name = "f32_vneg_test",
4879 srcs = [
4880 "test/f32-vneg.cc",
4881 "test/vunary-microkernel-tester.h",
4882 ] + MICROKERNEL_TEST_HDRS,
4883 deps = MICROKERNEL_TEST_DEPS,
4884)
4885
4886xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004887 name = "f32_vrndne_test",
4888 srcs = [
4889 "test/f32-vrndne.cc",
4890 "test/vunary-microkernel-tester.h",
4891 ] + MICROKERNEL_TEST_HDRS,
4892 deps = MICROKERNEL_TEST_DEPS,
4893)
4894
4895xnnpack_unit_test(
4896 name = "f32_vrndz_test",
4897 srcs = [
4898 "test/f32-vrndz.cc",
4899 "test/vunary-microkernel-tester.h",
4900 ] + MICROKERNEL_TEST_HDRS,
4901 deps = MICROKERNEL_TEST_DEPS,
4902)
4903
4904xnnpack_unit_test(
4905 name = "f32_vrndu_test",
4906 srcs = [
4907 "test/f32-vrndu.cc",
4908 "test/vunary-microkernel-tester.h",
4909 ] + MICROKERNEL_TEST_HDRS,
4910 deps = MICROKERNEL_TEST_DEPS,
4911)
4912
4913xnnpack_unit_test(
4914 name = "f32_vrndd_test",
4915 srcs = [
4916 "test/f32-vrndd.cc",
4917 "test/vunary-microkernel-tester.h",
4918 ] + MICROKERNEL_TEST_HDRS,
4919 deps = MICROKERNEL_TEST_DEPS,
4920)
4921
4922xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004923 name = "f32_vscale_test",
4924 srcs = [
4925 "test/f32-vscale.cc",
4926 "test/vscale-microkernel-tester.h",
4927 ] + MICROKERNEL_TEST_HDRS,
4928 deps = MICROKERNEL_TEST_DEPS,
4929)
4930
4931xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004932 name = "f32_vscaleexpminusmax_test",
4933 srcs = [
4934 "test/f32-vscaleexpminusmax.cc",
4935 "test/vscaleexpminusmax-microkernel-tester.h",
4936 ] + MICROKERNEL_TEST_HDRS,
4937 deps = MICROKERNEL_TEST_DEPS,
4938)
4939
4940xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004941 name = "f32_vscaleextexp_test",
4942 srcs = [
4943 "test/f32-vscaleextexp.cc",
4944 "test/vscaleextexp-microkernel-tester.h",
4945 ] + MICROKERNEL_TEST_HDRS,
4946 deps = MICROKERNEL_TEST_DEPS,
4947)
4948
4949xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004950 name = "f32_vsqr_test",
4951 srcs = [
4952 "test/f32-vsqr.cc",
4953 "test/vunary-microkernel-tester.h",
4954 ] + MICROKERNEL_TEST_HDRS,
4955 deps = MICROKERNEL_TEST_DEPS,
4956)
4957
4958xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004959 name = "f32_vsqrdiff_test",
4960 srcs = [
4961 "test/f32-vsqrdiff.cc",
4962 "test/vbinary-microkernel-tester.h",
4963 ] + MICROKERNEL_TEST_HDRS,
4964 deps = MICROKERNEL_TEST_DEPS,
4965)
4966
4967xnnpack_unit_test(
4968 name = "f32_vsqrdiffc_test",
4969 srcs = [
4970 "test/f32-vsqrdiffc.cc",
4971 "test/vbinaryc-microkernel-tester.h",
4972 ] + MICROKERNEL_TEST_HDRS,
4973 deps = MICROKERNEL_TEST_DEPS,
4974)
4975
4976xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004977 name = "f32_vsqrt_test",
4978 srcs = [
4979 "test/f32-vsqrt.cc",
4980 "test/vunary-microkernel-tester.h",
4981 ] + MICROKERNEL_TEST_HDRS,
4982 deps = MICROKERNEL_TEST_DEPS,
4983)
4984
4985xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004986 name = "f32_vsub_test",
4987 srcs = [
4988 "test/f32-vsub.cc",
4989 "test/vbinary-microkernel-tester.h",
4990 ] + MICROKERNEL_TEST_HDRS,
4991 deps = MICROKERNEL_TEST_DEPS,
4992)
4993
4994xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004995 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004996 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004997 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004998 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004999 ] + MICROKERNEL_TEST_HDRS,
5000 deps = MICROKERNEL_TEST_DEPS,
5001)
5002
5003xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005004 name = "f32_vsubc_test",
5005 srcs = [
5006 "test/f32-vsubc.cc",
5007 "test/vbinaryc-microkernel-tester.h",
5008 ] + MICROKERNEL_TEST_HDRS,
5009 deps = MICROKERNEL_TEST_DEPS,
5010)
5011
5012xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005013 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005014 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005015 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005016 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005017 ] + MICROKERNEL_TEST_HDRS,
5018 deps = MICROKERNEL_TEST_DEPS,
5019)
5020
5021xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005022 name = "f32_vrsubc_test",
5023 srcs = [
5024 "test/f32-vrsubc.cc",
5025 "test/vbinaryc-microkernel-tester.h",
5026 ] + MICROKERNEL_TEST_HDRS,
5027 deps = MICROKERNEL_TEST_DEPS,
5028)
5029
5030xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005031 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005032 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005033 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005034 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005035 ] + MICROKERNEL_TEST_HDRS,
5036 deps = MICROKERNEL_TEST_DEPS,
5037)
5038
5039xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005040 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005041 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005042 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005043 "test/avgpool-microkernel-tester.h",
5044 "src/xnnpack/AlignedAllocator.h",
5045 ] + MICROKERNEL_TEST_HDRS,
5046 deps = MICROKERNEL_TEST_DEPS,
5047)
5048
5049xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005050 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005051 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005052 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005053 "test/gemm-microkernel-tester.h",
5054 "src/xnnpack/AlignedAllocator.h",
5055 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005056 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057)
5058
5059xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005060 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005061 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005062 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063 "test/dwconv-microkernel-tester.h",
5064 "src/xnnpack/AlignedAllocator.h",
5065 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005066 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005067)
5068
5069xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005070 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005071 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005072 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005073 "test/gavgpool-microkernel-tester.h",
5074 "src/xnnpack/AlignedAllocator.h",
5075 ] + MICROKERNEL_TEST_HDRS,
5076 deps = MICROKERNEL_TEST_DEPS,
5077)
5078
5079xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005080 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005081 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005082 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005083 "test/gemm-microkernel-tester.h",
5084 "src/xnnpack/AlignedAllocator.h",
5085 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005086 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005087)
5088
5089xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005090 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005091 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005092 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005093 "test/vadd-microkernel-tester.h",
5094 ] + MICROKERNEL_TEST_HDRS,
5095 deps = MICROKERNEL_TEST_DEPS,
5096)
5097
5098xnnpack_unit_test(
5099 name = "u8_clamp_test",
5100 srcs = [
5101 "test/u8-clamp.cc",
5102 "test/clamp-microkernel-tester.h",
5103 ] + MICROKERNEL_TEST_HDRS,
5104 deps = MICROKERNEL_TEST_DEPS,
5105)
5106
5107xnnpack_unit_test(
5108 name = "u8_lut32norm_test",
5109 srcs = [
5110 "test/u8-lut32norm.cc",
5111 "test/lut-norm-microkernel-tester.h",
5112 ] + MICROKERNEL_TEST_HDRS,
5113 deps = MICROKERNEL_TEST_DEPS,
5114)
5115
5116xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005117 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005118 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005119 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005120 "test/maxpool-microkernel-tester.h",
5121 ] + MICROKERNEL_TEST_HDRS,
5122 deps = MICROKERNEL_TEST_DEPS,
5123)
5124
5125xnnpack_unit_test(
5126 name = "u8_rmax_test",
5127 srcs = [
5128 "test/u8-rmax.cc",
5129 "test/rmax-microkernel-tester.h",
5130 ] + MICROKERNEL_TEST_HDRS,
5131 deps = MICROKERNEL_TEST_DEPS,
5132)
5133
5134xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005135 name = "x32_fill_test",
5136 srcs = [
5137 "test/x32-fill.cc",
5138 "test/fill-microkernel-tester.h",
5139 ] + MICROKERNEL_TEST_HDRS,
5140 deps = MICROKERNEL_TEST_DEPS,
5141)
5142
5143xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005144 name = "x32_packx_test",
5145 srcs = [
5146 "test/x32-packx.cc",
5147 "test/pack-microkernel-tester.h",
5148 "src/xnnpack/AlignedAllocator.h",
5149 ] + MICROKERNEL_TEST_HDRS,
5150 deps = MICROKERNEL_TEST_DEPS,
5151)
5152
5153xnnpack_unit_test(
5154 name = "x32_pad_test",
5155 srcs = [
5156 "test/x32-pad.cc",
5157 "test/pad-microkernel-tester.h",
5158 ] + MICROKERNEL_TEST_HDRS,
5159 deps = MICROKERNEL_TEST_DEPS,
5160)
5161
5162xnnpack_unit_test(
5163 name = "x32_unpool_test",
5164 srcs = [
5165 "test/x32-unpool.cc",
5166 "test/unpool-microkernel-tester.h",
5167 ] + MICROKERNEL_TEST_HDRS,
5168 deps = MICROKERNEL_TEST_DEPS,
5169)
5170
5171xnnpack_unit_test(
5172 name = "x32_zip_test",
5173 srcs = [
5174 "test/x32-zip.cc",
5175 "test/zip-microkernel-tester.h",
5176 ] + MICROKERNEL_TEST_HDRS,
5177 deps = MICROKERNEL_TEST_DEPS,
5178)
5179
5180xnnpack_unit_test(
5181 name = "x8_lut_test",
5182 srcs = [
5183 "test/x8-lut.cc",
5184 "test/lut-microkernel-tester.h",
5185 ] + MICROKERNEL_TEST_HDRS,
5186 deps = MICROKERNEL_TEST_DEPS,
5187)
5188
5189xnnpack_unit_test(
5190 name = "x8_zip_test",
5191 srcs = [
5192 "test/x8-zip.cc",
5193 "test/zip-microkernel-tester.h",
5194 ] + MICROKERNEL_TEST_HDRS,
5195 deps = MICROKERNEL_TEST_DEPS,
5196)
5197
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005198xnnpack_unit_test(
5199 name = "requantization_test",
5200 srcs = [
5201 "src/xnnpack/requantization-stubs.h",
5202 "test/requantization.cc",
5203 "test/requantization-tester.h",
5204 ] + MICROKERNEL_TEST_HDRS,
5205 deps = MICROKERNEL_TEST_DEPS,
5206)
5207
Marat Dukhan20c3b922020-03-10 03:45:06 -07005208########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005209
5210xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005211 name = "operator_size_test",
5212 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005213 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005214)
5215
Marat Dukhan20c3b922020-03-10 03:45:06 -07005216xnnpack_binary(
5217 name = "subgraph_size_test",
5218 srcs = ["test/subgraph-size.c"],
5219 deps = [":XNNPACK"],
5220)
5221
5222########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223
5224xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005225 name = "abs_nc_test",
5226 srcs = [
5227 "test/abs-nc.cc",
5228 "test/abs-operator-tester.h",
5229 ],
5230 deps = OPERATOR_TEST_DEPS,
5231)
5232
5233xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005234 name = "add_nd_test",
5235 srcs = [
5236 "test/add-nd.cc",
5237 "test/binary-elementwise-operator-tester.h",
5238 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005239 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005240)
5241
5242xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005243 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005245 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246 "test/argmax-pooling-operator-tester.h",
5247 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005248 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005249)
5250
5251xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005252 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005254 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005255 "test/average-pooling-operator-tester.h",
5256 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005257 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258)
5259
5260xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005261 name = "bankers_rounding_nc_test",
5262 srcs = [
5263 "test/bankers-rounding-nc.cc",
5264 "test/bankers-rounding-operator-tester.h",
5265 ],
5266 deps = OPERATOR_TEST_DEPS,
5267)
5268
5269xnnpack_unit_test(
5270 name = "ceiling_nc_test",
5271 srcs = [
5272 "test/ceiling-nc.cc",
5273 "test/ceiling-operator-tester.h",
5274 ],
5275 deps = OPERATOR_TEST_DEPS,
5276)
5277
5278xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005279 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005280 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005281 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 "test/channel-shuffle-operator-tester.h",
5283 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005284 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285)
5286
5287xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005288 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005289 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005290 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291 "test/clamp-operator-tester.h",
5292 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005293 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294)
5295
5296xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005297 name = "constant_pad_nd_test",
5298 srcs = [
5299 "test/constant-pad-nd.cc",
5300 "test/constant-pad-operator-tester.h",
5301 ],
5302 deps = OPERATOR_TEST_DEPS,
5303)
5304
5305xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005306 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005308 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005309 "test/convolution-operator-tester.h",
5310 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005311 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312)
5313
5314xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005315 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005317 "test/convolution-nchw.cc",
5318 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005319 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005320 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321)
5322
5323xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005324 name = "copy_nc_test",
5325 srcs = [
5326 "test/copy-nc.cc",
5327 "test/copy-operator-tester.h",
5328 ],
5329 deps = OPERATOR_TEST_DEPS,
5330)
5331
5332xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005333 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005335 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336 "test/deconvolution-operator-tester.h",
5337 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005338 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339)
5340
5341xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005342 name = "divide_nd_test",
5343 srcs = [
5344 "test/binary-elementwise-operator-tester.h",
5345 "test/divide-nd.cc",
5346 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005347 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005348)
5349
5350xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005351 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005352 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005353 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005354 "test/fully-connected-operator-tester.h",
5355 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005356 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005357)
5358
5359xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005360 name = "floor_nc_test",
5361 srcs = [
5362 "test/floor-nc.cc",
5363 "test/floor-operator-tester.h",
5364 ],
5365 deps = OPERATOR_TEST_DEPS,
5366)
5367
5368xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005369 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005370 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005371 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005372 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005373 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005374 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005375)
5376
5377xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005378 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005379 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005380 "test/global-average-pooling-ncw.cc",
5381 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005382 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005383 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005384)
5385
5386xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005387 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005388 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005389 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005390 "test/hardswish-operator-tester.h",
5391 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005392 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005393)
5394
5395xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005396 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005397 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005398 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399 "test/leaky-relu-operator-tester.h",
5400 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005401 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005402)
5403
5404xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005405 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005407 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005408 "test/max-pooling-operator-tester.h",
5409 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005410 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411)
5412
5413xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005414 name = "maximum_nd_test",
5415 srcs = [
5416 "test/binary-elementwise-operator-tester.h",
5417 "test/maximum-nd.cc",
5418 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005419 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005420)
5421
5422xnnpack_unit_test(
5423 name = "minimum_nd_test",
5424 srcs = [
5425 "test/binary-elementwise-operator-tester.h",
5426 "test/minimum-nd.cc",
5427 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005428 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005429)
5430
5431xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005432 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005433 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005434 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005435 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005436 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005437 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005438)
5439
5440xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005441 name = "negate_nc_test",
5442 srcs = [
5443 "test/negate-nc.cc",
5444 "test/negate-operator-tester.h",
5445 ],
5446 deps = OPERATOR_TEST_DEPS,
5447)
5448
5449xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005450 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005451 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005452 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005453 "test/prelu-operator-tester.h",
5454 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005455 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005456)
5457
5458xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005459 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005460 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005461 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005462 "test/resize-bilinear-operator-tester.h",
5463 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005464 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005465)
5466
5467xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005468 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005469 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005470 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005471 "test/sigmoid-operator-tester.h",
5472 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005473 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005474)
5475
5476xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005477 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005479 "test/softmax-nc.cc",
5480 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005481 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005482 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005483)
5484
5485xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005486 name = "square_nc_test",
5487 srcs = [
5488 "test/square-nc.cc",
5489 "test/square-operator-tester.h",
5490 ],
5491 deps = OPERATOR_TEST_DEPS,
5492)
5493
5494xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005495 name = "square_root_nc_test",
5496 srcs = [
5497 "test/square-root-nc.cc",
5498 "test/square-root-operator-tester.h",
5499 ],
5500 deps = OPERATOR_TEST_DEPS,
5501)
5502
5503xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005504 name = "squared_difference_nd_test",
5505 srcs = [
5506 "test/binary-elementwise-operator-tester.h",
5507 "test/squared-difference-nd.cc",
5508 ],
5509 deps = OPERATOR_TEST_DEPS,
5510)
5511
5512xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005513 name = "subtract_nd_test",
5514 srcs = [
5515 "test/binary-elementwise-operator-tester.h",
5516 "test/subtract-nd.cc",
5517 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005518 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005519)
5520
5521xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005522 name = "truncation_nc_test",
5523 srcs = [
5524 "test/truncation-nc.cc",
5525 "test/truncation-operator-tester.h",
5526 ],
5527 deps = OPERATOR_TEST_DEPS,
5528)
5529
5530xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005531 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005532 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005533 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005534 "test/unpooling-operator-tester.h",
5535 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005536 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005537)
5538
Chao Mei6ddfc602020-05-13 22:29:36 -07005539############################### Misc unit tests ###############################
5540
5541xnnpack_unit_test(
5542 name = "memory_planner_test",
5543 srcs = [
5544 "test/memory-planner-test.cc",
5545 ],
5546 deps = [
5547 ":XNNPACK",
5548 ":memory_planner",
5549 ],
5550)
5551
Marat Dukhan08c4a432019-10-03 09:29:21 -07005552############################# Build configurations #############################
5553
Marat Dukhanb8642352019-10-30 15:43:02 -07005554# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005555config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005556 name = "xnn_enable_assembly_explicit_true",
5557 define_values = {"xnn_enable_assembly": "true"},
5558)
5559
5560# Disables usage of assembly kernels.
5561config_setting(
5562 name = "xnn_enable_assembly_explicit_false",
5563 define_values = {"xnn_enable_assembly": "false"},
5564)
5565
Marat Dukhan9de90e02020-06-18 16:04:12 -07005566# Enables usage of sparse inference.
5567config_setting(
5568 name = "xnn_enable_sparse_explicit_true",
5569 define_values = {"xnn_enable_sparse": "true"},
5570)
5571
5572# Disables usage of sparse inference.
5573config_setting(
5574 name = "xnn_enable_sparse_explicit_false",
5575 define_values = {"xnn_enable_sparse": "false"},
5576)
5577
Marat Dukhan05702cf2020-03-26 15:41:33 -07005578# Disables usage of HMP-aware optimizations.
5579config_setting(
5580 name = "xnn_enable_hmp_explicit_false",
5581 define_values = {"xnn_enable_hmp": "false"},
5582)
5583
Chao Mei6ddfc602020-05-13 22:29:36 -07005584# Enable usage of optimized memory allocation
5585config_setting(
5586 name = "xnn_enable_memopt_explicit_true",
5587 define_values = {"xnnpack_enable_memopt": "true"},
5588)
5589
5590# Disable usage of optimized memory allocation
5591config_setting(
5592 name = "xnn_enable_memopt_explicit_false",
5593 define_values = {"xnnpack_enable_memopt": "false"},
5594)
5595
Marat Dukhanb8642352019-10-30 15:43:02 -07005596# Builds with -c dbg
5597config_setting(
5598 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005599 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005600 "compilation_mode": "dbg",
5601 },
5602)
5603
5604# Builds with -c opt
5605config_setting(
5606 name = "optimized_build",
5607 values = {
5608 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005609 },
5610)
5611
5612config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005613 name = "linux_k8",
5614 values = {"cpu": "k8"},
5615)
5616
5617config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005618 name = "linux_arm",
5619 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005620)
5621
5622config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005623 name = "linux_armeabi",
5624 values = {"cpu": "armeabi"},
5625)
5626
5627config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005628 name = "linux_armhf",
5629 values = {"cpu": "armhf"},
5630)
5631
5632config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005633 name = "linux_armv7a",
5634 values = {"cpu": "armv7a"},
5635)
5636
5637config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005638 name = "linux_aarch64",
5639 values = {"cpu": "aarch64"},
5640)
5641
5642config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005643 name = "android",
5644 values = {"crosstool_top": "//external:android/crosstool"},
5645)
5646
5647config_setting(
5648 name = "android_armv7",
5649 values = {
5650 "crosstool_top": "//external:android/crosstool",
5651 "cpu": "armeabi-v7a",
5652 },
5653)
5654
5655config_setting(
5656 name = "android_arm64",
5657 values = {
5658 "crosstool_top": "//external:android/crosstool",
5659 "cpu": "arm64-v8a",
5660 },
5661)
5662
5663config_setting(
5664 name = "android_x86",
5665 values = {
5666 "crosstool_top": "//external:android/crosstool",
5667 "cpu": "x86",
5668 },
5669)
5670
5671config_setting(
5672 name = "android_x86_64",
5673 values = {
5674 "crosstool_top": "//external:android/crosstool",
5675 "cpu": "x86_64",
5676 },
5677)
5678
5679config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005680 name = "windows_x86_64",
5681 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005682)
5683
5684config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005685 name = "windows_x86_64_clang",
5686 values = {
5687 "compiler": "clang-cl",
5688 "cpu": "x64_windows",
5689 },
5690)
5691
5692config_setting(
5693 name = "windows_x86_64_mingw",
5694 values = {
5695 "compiler": "mingw-gcc",
5696 "cpu": "x64_windows",
5697 },
5698)
5699
5700config_setting(
5701 name = "windows_x86_64_msys",
5702 values = {
5703 "compiler": "msys-gcc",
5704 "cpu": "x64_windows",
5705 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005706)
5707
5708config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005709 name = "macos_x86_64",
5710 values = {
5711 "apple_platform_type": "macos",
5712 "cpu": "darwin",
5713 },
5714)
5715
5716config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005717 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005718 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005719)
5720
5721config_setting(
5722 name = "emscripten_wasm",
5723 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005724 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005725 "cpu": "wasm",
5726 },
5727)
5728
5729config_setting(
5730 name = "emscripten_wasmsimd",
5731 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005732 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005733 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005734 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005735 },
5736)
5737
5738config_setting(
5739 name = "emscripten_asmjs",
5740 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005741 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005742 "cpu": "asmjs",
5743 },
5744)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005745
5746config_setting(
5747 name = "ios_armv7",
5748 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005749 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005750 "cpu": "ios_armv7",
5751 },
5752)
5753
5754config_setting(
5755 name = "ios_arm64",
5756 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005757 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005758 "cpu": "ios_arm64",
5759 },
5760)
5761
5762config_setting(
5763 name = "ios_arm64e",
5764 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005765 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005766 "cpu": "ios_arm64e",
5767 },
5768)
5769
5770config_setting(
5771 name = "ios_x86",
5772 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005773 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005774 "cpu": "ios_i386",
5775 },
5776)
5777
5778config_setting(
5779 name = "ios_x86_64",
5780 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005781 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005782 "cpu": "ios_x86_64",
5783 },
5784)
5785
5786config_setting(
5787 name = "watchos_armv7k",
5788 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005789 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005790 "cpu": "watchos_armv7k",
5791 },
5792)
5793
5794config_setting(
5795 name = "watchos_arm64_32",
5796 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005797 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005798 "cpu": "watchos_arm64_32",
5799 },
5800)
5801
5802config_setting(
5803 name = "watchos_x86",
5804 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005805 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005806 "cpu": "watchos_i386",
5807 },
5808)
5809
5810config_setting(
5811 name = "watchos_x86_64",
5812 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005813 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005814 "cpu": "watchos_x86_64",
5815 },
5816)
5817
5818config_setting(
5819 name = "tvos_arm64",
5820 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005821 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005822 "cpu": "tvos_arm64",
5823 },
5824)
5825
5826config_setting(
5827 name = "tvos_x86_64",
5828 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005829 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005830 "cpu": "tvos_x86_64",
5831 },
5832)