blob: 309748a60f386826aaf2b6794d30c327fdee91b3 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700102 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700103 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700104 "src/subgraph/squared-difference.c",
105 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700106 "src/subgraph/unpooling-2d.c",
107]
108
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800109TABLE_SRCS = [
110 "src/tables/exp2-k-over-64.c",
111 "src/tables/exp2-k-over-2048.c",
112]
113
Marat Dukhan08c4a432019-10-03 09:29:21 -0700114SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800115 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800116 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800117 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700118 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
119 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700120 "src/f32-clamp/gen/scalar-x1.c",
121 "src/f32-clamp/gen/scalar-x2.c",
122 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800123 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700125 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
126 "src/f32-dwconv-chw/3x3p1-scalar.c",
127 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
128 "src/f32-dwconv-chw/5x5p2-scalar.c",
129 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700130 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
131 "src/f32-dwconv/gen/up1x4-scalar.c",
132 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
133 "src/f32-dwconv/gen/up1x9-scalar.c",
134 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
135 "src/f32-dwconv/gen/up1x25-scalar.c",
136 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
137 "src/f32-dwconv/gen/up2x4-scalar.c",
138 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
139 "src/f32-dwconv/gen/up2x9-scalar.c",
140 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
141 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700142 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
144 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700148 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
150 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700152 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700154 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700155 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
156 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700157 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
158 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700160 "src/f32-gemm/gen/1x4-scalar.c",
161 "src/f32-gemm/gen/2x4-scalar.c",
162 "src/f32-gemm/gen/4x2-scalar.c",
163 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700164 "src/f32-gemm/gen/1x4-relu-scalar.c",
165 "src/f32-gemm/gen/2x4-relu-scalar.c",
166 "src/f32-gemm/gen/4x2-relu-scalar.c",
167 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700168 "src/f32-gemm/gen/1x4-minmax-scalar.c",
169 "src/f32-gemm/gen/2x4-minmax-scalar.c",
170 "src/f32-gemm/gen/4x2-minmax-scalar.c",
171 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700172 "src/f32-hswish/gen/hswish-scalar-x1.c",
173 "src/f32-hswish/gen/hswish-scalar-x2.c",
174 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700175 "src/f32-ibilinear/gen/scalar-c1.c",
176 "src/f32-ibilinear/gen/scalar-c2.c",
177 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700178 "src/f32-igemm/gen/1x4-scalar.c",
179 "src/f32-igemm/gen/2x4-scalar.c",
180 "src/f32-igemm/gen/4x2-scalar.c",
181 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700182 "src/f32-igemm/gen/1x4-relu-scalar.c",
183 "src/f32-igemm/gen/2x4-relu-scalar.c",
184 "src/f32-igemm/gen/4x2-relu-scalar.c",
185 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700186 "src/f32-igemm/gen/1x4-minmax-scalar.c",
187 "src/f32-igemm/gen/2x4-minmax-scalar.c",
188 "src/f32-igemm/gen/4x2-minmax-scalar.c",
189 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700190 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
191 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700193 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
194 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
195 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800197 "src/f32-prelu/gen/scalar-2x1.c",
198 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800199 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700211 "src/f32-relu/gen/scalar-x1.c",
212 "src/f32-relu/gen/scalar-x2.c",
213 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700214 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800215 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
218 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
221 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700224 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
225 "src/f32-spmm/gen/1x1-minmax-scalar.c",
226 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar.c",
228 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar.c",
230 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar.c",
232 "src/f32-spmm/gen/8x2-minmax-scalar.c",
233 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700234 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
235 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700237 "src/f32-vbinary/gen/vadd-scalar-x1.c",
238 "src/f32-vbinary/gen/vadd-scalar-x2.c",
239 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700240 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700243 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
244 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700246 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700249 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
250 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700252 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700255 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
256 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800258 "src/f32-vbinary/gen/vmax-scalar-x1.c",
259 "src/f32-vbinary/gen/vmax-scalar-x2.c",
260 "src/f32-vbinary/gen/vmax-scalar-x4.c",
261 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
264 "src/f32-vbinary/gen/vmin-scalar-x1.c",
265 "src/f32-vbinary/gen/vmin-scalar-x2.c",
266 "src/f32-vbinary/gen/vmin-scalar-x4.c",
267 "src/f32-vbinary/gen/vminc-scalar-x1.c",
268 "src/f32-vbinary/gen/vminc-scalar-x2.c",
269 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700270 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
271 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700273 "src/f32-vbinary/gen/vmul-scalar-x1.c",
274 "src/f32-vbinary/gen/vmul-scalar-x2.c",
275 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700276 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700279 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
280 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700282 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700285 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
286 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700288 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
292 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700294 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
295 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
297 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700300 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
301 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vsub-scalar-x1.c",
304 "src/f32-vbinary/gen/vsub-scalar-x2.c",
305 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700306 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700309 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
310 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700312 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
313 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700315 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
316 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700318 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
319 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
321 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
324 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
327 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700330 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
331 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700333 "src/f32-vunary/gen/vabs-scalar-x1.c",
334 "src/f32-vunary/gen/vabs-scalar-x2.c",
335 "src/f32-vunary/gen/vabs-scalar-x4.c",
336 "src/f32-vunary/gen/vneg-scalar-x1.c",
337 "src/f32-vunary/gen/vneg-scalar-x2.c",
338 "src/f32-vunary/gen/vneg-scalar-x4.c",
339 "src/f32-vunary/gen/vsqr-scalar-x1.c",
340 "src/f32-vunary/gen/vsqr-scalar-x2.c",
341 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800342 "src/math/expminus-scalar-lut2048-p1.c",
343 "src/math/expminus-scalar-lut64-p2.c",
344 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700345 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700346 "src/math/roundne-scalar-nearbyint.c",
347 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700348 "src/math/roundd-scalar-addsub.c",
349 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700350 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700351 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700352 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700353 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700354 "src/math/roundz-scalar-addsub.c",
355 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700356 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800357 "src/math/sigmoid-scalar-lut2048-p1-div.c",
358 "src/math/sigmoid-scalar-lut64-p2-div.c",
359 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700360 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
361 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
362 "src/qu8-dwconv/up1x9-minmax-scalar.c",
363 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
364 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
365 "src/qu8-gemm/2x2-minmax-scalar.c",
366 "src/qu8-igemm/2x2-minmax-scalar.c",
367 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700368 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700369 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700370 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700371 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700372 "src/x32-fill/scalar-float.c",
373 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700374 "src/x32-packx/x2-scalar.c",
375 "src/x32-packx/x3-scalar.c",
376 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700377 "src/x32-pad/scalar-int.c",
378 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700379 "src/x32-unpool/scalar.c",
380 "src/x32-zip/x2-scalar.c",
381 "src/x32-zip/x3-scalar.c",
382 "src/x32-zip/x4-scalar.c",
383 "src/x32-zip/xm-scalar.c",
384 "src/x8-lut/scalar.c",
385 "src/x8-zip/x2-scalar.c",
386 "src/x8-zip/x3-scalar.c",
387 "src/x8-zip/x4-scalar.c",
388 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700389 "src/requantization/precise-scalar.c",
390 "src/requantization/fp32-scalar.c",
391 "src/requantization/q31-scalar.c",
392 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700393]
394
Marat Dukhan436ebe62019-12-04 15:10:12 -0800395WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700396 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
397 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700398 "src/f32-clamp/gen/wasm-x1.c",
399 "src/f32-clamp/gen/wasm-x2.c",
400 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700401 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
402 "src/f32-dwconv/gen/up1x4-wasm.c",
403 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
404 "src/f32-dwconv/gen/up1x9-wasm.c",
405 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
406 "src/f32-dwconv/gen/up1x25-wasm.c",
407 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
408 "src/f32-dwconv/gen/up2x4-wasm.c",
409 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
410 "src/f32-dwconv/gen/up2x9-wasm.c",
411 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
412 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700413 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
414 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
415 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
416 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700417 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
418 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700419 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
420 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
421 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
422 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700423 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
424 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700425 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
426 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700427 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
428 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
429 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700430 "src/f32-gemm/gen/1x4-wasm.c",
431 "src/f32-gemm/gen/2x4-wasm.c",
432 "src/f32-gemm/gen/4x2-wasm.c",
433 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700434 "src/f32-gemm/gen/1x4-relu-wasm.c",
435 "src/f32-gemm/gen/2x4-relu-wasm.c",
436 "src/f32-gemm/gen/4x2-relu-wasm.c",
437 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700438 "src/f32-gemm/gen/1x4-minmax-wasm.c",
439 "src/f32-gemm/gen/2x4-minmax-wasm.c",
440 "src/f32-gemm/gen/4x2-minmax-wasm.c",
441 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700442 "src/f32-hswish/gen/hswish-wasm-x1.c",
443 "src/f32-hswish/gen/hswish-wasm-x2.c",
444 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700445 "src/f32-igemm/gen/1x4-wasm.c",
446 "src/f32-igemm/gen/2x4-wasm.c",
447 "src/f32-igemm/gen/4x2-wasm.c",
448 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700449 "src/f32-igemm/gen/1x4-relu-wasm.c",
450 "src/f32-igemm/gen/2x4-relu-wasm.c",
451 "src/f32-igemm/gen/4x2-relu-wasm.c",
452 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700453 "src/f32-igemm/gen/1x4-minmax-wasm.c",
454 "src/f32-igemm/gen/2x4-minmax-wasm.c",
455 "src/f32-igemm/gen/4x2-minmax-wasm.c",
456 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700457 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
458 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
459 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700460 "src/f32-prelu/gen/wasm-2x1.c",
461 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700462 "src/f32-relu/gen/wasm-x1.c",
463 "src/f32-relu/gen/wasm-x2.c",
464 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700465 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
466 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
467 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
468 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
469 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
470 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
471 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
472 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
473 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
474 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800477 "src/f32-vbinary/gen/vmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
481 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
482 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
483 "src/f32-vbinary/gen/vmin-wasm-x1.c",
484 "src/f32-vbinary/gen/vmin-wasm-x2.c",
485 "src/f32-vbinary/gen/vmin-wasm-x4.c",
486 "src/f32-vbinary/gen/vminc-wasm-x1.c",
487 "src/f32-vbinary/gen/vminc-wasm-x2.c",
488 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700489 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
490 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
491 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
492 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
493 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
494 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
495 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
496 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
497 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
498 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700507 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
508 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
509 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700510 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
511 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
512 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800513]
514
Marat Dukhan290055c2020-06-09 12:24:29 -0700515WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700516 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
517 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
518 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700519 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
520 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
521 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
522 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700523 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
524 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
525 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
526 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700527 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
528 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
529 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
530 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
531 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
532 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700533 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
534 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
535 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
536 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
537 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
538 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
539 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
540 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
541 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
542 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
543 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
544 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
545 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
546 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
547 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
548 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
549 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
550 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
551 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
552 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
553 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
554 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
555 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
556 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700557 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
558 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
559 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700560 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700561 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
562 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
563 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
564 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700565 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
566 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
567 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
568 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
569 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
570 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
571 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
572 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
573 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
574 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
581 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
582 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
583 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
584 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
585 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
586 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
587 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
588 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
589 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700595 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
596 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700597 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
598 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
599 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
600 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
601 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
602 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
603 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
604 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
605 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
606 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
613 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
614 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
615 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
616 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
617 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
618 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
619 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
620 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
621 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700627 "src/f32-ibilinear/gen/wasmsimd-c4.c",
628 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700629 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
630 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
631 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700632 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700633 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
634 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
635 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
636 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700637 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
638 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
639 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
640 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
641 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
642 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
643 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
644 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
645 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
646 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
653 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
654 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
655 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
656 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
657 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
658 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
659 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
660 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
661 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700667 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
668 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700669 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
670 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
671 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
672 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700673 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
674 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
675 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700676 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
677 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700678 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700682 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
683 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
684 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
685 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700686 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
687 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
688 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
689 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
690 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
691 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
692 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
693 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
694 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
695 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
696 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
697 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700698 "src/f32-relu/gen/wasmsimd-x4.c",
699 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700700 "src/f32-rmax/wasmsimd-arm.c",
701 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700702 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
703 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
704 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
705 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
706 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
707 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700708 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
709 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
710 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
711 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700712 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
713 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700714 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
715 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
716 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
717 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700718 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
719 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700720 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
721 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
722 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
723 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700724 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
725 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700726 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
727 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
728 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
729 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700730 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
731 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700732 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
733 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
734 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
735 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700736 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
737 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
738 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
739 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700740 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
741 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
742 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
743 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700744 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
745 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
746 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
747 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700748 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
749 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
750 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
751 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700752 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
753 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700754 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
755 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
756 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
757 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700758 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
759 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700760 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
761 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
762 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
763 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700764 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
765 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700766 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
767 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
768 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
769 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700770 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
771 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700772 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
773 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700774 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
775 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700776 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
777 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
778 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
779 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700780 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
781 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700782 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
783 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
784 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
785 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700786 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
787 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700788 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
789 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
790 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
791 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700792 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
793 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
794 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
795 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700796 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
797 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
798 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
799 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
800 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
801 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
802 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
803 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
804 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
805 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
806 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
807 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
808 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
809 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700810 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
811 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700812 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
813 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
814 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
815 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
816 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
817 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700818 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700819 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700820 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700821 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700822 "src/x32-zip/x2-wasmsimd.c",
823 "src/x32-zip/x3-wasmsimd.c",
824 "src/x32-zip/x4-wasmsimd.c",
825 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700826 "src/math/roundne-wasmsimd-addsub.c",
827 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700828 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700829 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700830 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700831 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700832 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700833 "src/requantization/fp32-wasmsimd.c",
834]
835
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800836PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800837 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800838 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800839 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700840 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
841 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700842 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700843 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700844 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700845 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700846 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700847 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
848 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
849 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
850 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
851 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
852 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
853 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
854 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
855 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
856 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
857 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
858 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700859 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700860 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
861 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700862 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
863 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
864 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
865 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
866 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
867 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
868 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
869 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
870 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
871 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
872 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
873 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
874 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
875 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
876 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
877 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
878 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
879 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
880 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700881 "src/f32-ibilinear/gen/psimd-c4.c",
882 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700883 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
884 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
885 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
886 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
887 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
888 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
889 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
890 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
891 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
892 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700893 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700894 "src/f32-spmm/gen/4x1-minmax-psimd.c",
895 "src/f32-spmm/gen/8x1-minmax-psimd.c",
896 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700897 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
898 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700899 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800900 "src/f32-prelu/gen/psimd-2x4.c",
901 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800902 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700903 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
904 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
905 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
906 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
907 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
908 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
909 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
910 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800911 "src/f32-vbinary/gen/vmax-psimd-x4.c",
912 "src/f32-vbinary/gen/vmax-psimd-x8.c",
913 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
914 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
915 "src/f32-vbinary/gen/vmin-psimd-x4.c",
916 "src/f32-vbinary/gen/vmin-psimd-x8.c",
917 "src/f32-vbinary/gen/vminc-psimd-x4.c",
918 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700919 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
920 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
921 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
922 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
923 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
924 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
925 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
926 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700927 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
928 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
929 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
930 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700931 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
932 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
933 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
934 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700935 "src/f32-vlrelu/gen/vlrelu-psimd-x4.c",
936 "src/f32-vlrelu/gen/vlrelu-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700937 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
938 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700939 "src/f32-vunary/gen/vabs-psimd-x4.c",
940 "src/f32-vunary/gen/vabs-psimd-x8.c",
941 "src/f32-vunary/gen/vneg-psimd-x4.c",
942 "src/f32-vunary/gen/vneg-psimd-x8.c",
943 "src/f32-vunary/gen/vsqr-psimd-x4.c",
944 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700945 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700946 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700947 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700948 "src/x32-unpool/psimd.c",
949 "src/x32-zip/x2-psimd.c",
950 "src/x32-zip/x3-psimd.c",
951 "src/x32-zip/x4-psimd.c",
952 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700953 "src/requantization/precise-psimd.c",
954 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700955]
956
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800957PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800958 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
959 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
960 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
961 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
962 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
963 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
964 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
965 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
966 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
967 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
968 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
969 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800970 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
971 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
972 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
973 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
974 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
975 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700976 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
977 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
978 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
979 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
980 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
981 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
982 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
983 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700984 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700985 "src/math/roundd-psimd-addsub.c",
986 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700987 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800988 "src/math/sigmoid-psimd-p5-div.c",
989]
990
Marat Dukhan08c4a432019-10-03 09:29:21 -0700991# ISA-specific micro-kernels
992NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700993 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
994 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700995 "src/f32-clamp/gen/neon-x4.c",
996 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700997 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
998 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700999 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1000 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001001 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1002 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001003 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1004 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001005 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1006 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1007 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1008 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001009 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1010 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1011 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1012 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001013 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1014 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1015 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1016 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001017 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001018 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1019 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001020 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1021 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1022 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1023 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1024 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1025 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1026 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1027 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1028 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1029 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1030 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1031 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1032 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1033 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1034 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1035 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1036 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1037 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1038 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1039 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1040 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1041 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1042 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1043 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1044 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1045 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1046 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1047 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1048 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1049 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1050 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001051 "src/f32-hswish/gen/hswish-neon-x4.c",
1052 "src/f32-hswish/gen/hswish-neon-x8.c",
1053 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001054 "src/f32-ibilinear/gen/neon-c4.c",
1055 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001056 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1057 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1058 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1059 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1060 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1061 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1062 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1063 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1064 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1065 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1066 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1067 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1068 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1069 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1070 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1071 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001072 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1073 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1074 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001075 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1076 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001077 "src/f32-prelu/gen/neon-2x4.c",
1078 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001079 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1080 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1081 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1082 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1083 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1084 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1085 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1086 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1087 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1088 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1089 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1090 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1091 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1092 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1093 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1094 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1095 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1096 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1097 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1098 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1099 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1100 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1101 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1102 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001103 "src/f32-relu/gen/neon-x4.c",
1104 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001105 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001106 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001107 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1108 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1109 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1110 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1111 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1112 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1113 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1114 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1115 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1116 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1117 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1118 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1119 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1120 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1121 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1122 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1123 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1124 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001125 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1126 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1127 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1128 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001129 "src/f32-vbinary/gen/vmax-neon-x4.c",
1130 "src/f32-vbinary/gen/vmax-neon-x8.c",
1131 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1132 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1133 "src/f32-vbinary/gen/vmin-neon-x4.c",
1134 "src/f32-vbinary/gen/vmin-neon-x8.c",
1135 "src/f32-vbinary/gen/vminc-neon-x4.c",
1136 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001137 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1138 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1139 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1140 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1141 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1142 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001143 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1144 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1145 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1146 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001147 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1148 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1149 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1150 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001151 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1152 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001153 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1154 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001155 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1156 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1157 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1158 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1159 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1160 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1161 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1162 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001163 "src/f32-vunary/gen/vabs-neon-x4.c",
1164 "src/f32-vunary/gen/vabs-neon-x8.c",
1165 "src/f32-vunary/gen/vneg-neon-x4.c",
1166 "src/f32-vunary/gen/vneg-neon-x8.c",
1167 "src/f32-vunary/gen/vsqr-neon-x4.c",
1168 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001169 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1170 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1171 "src/qu8-dwconv/up8x9-minmax-neon.c",
1172 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1173 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1174 "src/qu8-gemm/4x8-minmax-neon.c",
1175 "src/qu8-gemm/8x8-minmax-neon.c",
1176 "src/qu8-igemm/4x8-minmax-neon.c",
1177 "src/qu8-igemm/8x8-minmax-neon.c",
1178 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001179 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001180 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001181 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001182 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001183 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001184 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001185 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001186 "src/x32-zip/x2-neon.c",
1187 "src/x32-zip/x3-neon.c",
1188 "src/x32-zip/x4-neon.c",
1189 "src/x32-zip/xm-neon.c",
1190 "src/x8-zip/x2-neon.c",
1191 "src/x8-zip/x3-neon.c",
1192 "src/x8-zip/x4-neon.c",
1193 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001194 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001195 "src/math/roundd-neon-addsub.c",
1196 "src/math/roundd-neon-cvt.c",
1197 "src/math/roundu-neon-addsub.c",
1198 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001199 "src/math/roundz-neon-addsub.c",
1200 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001201 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001202 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1203 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1204 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1205 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1206 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1207 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001208 "src/math/sqrt-neon-nr1rsqrts.c",
1209 "src/math/sqrt-neon-nr2rsqrts.c",
1210 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001211 "src/requantization/precise-neon.c",
1212 "src/requantization/fp32-neon.c",
1213 "src/requantization/q31-neon.c",
1214 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001215]
1216
1217NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001218 "src/f32-ibilinear/gen/neonfma-c4.c",
1219 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001220 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1221 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1222 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1223 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1224 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1225 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1226 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1227 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1228 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001229 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1230 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1231 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1232 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001233 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1234 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1235 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1236 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001237 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1238 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1239 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1240 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001241 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1242 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1243 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1244 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1245 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1246 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1247 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1248 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1249 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1250 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1251 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1252 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1253 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1254 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1255 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1256 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1257 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1258 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001259 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1260 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001261 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1262 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1263 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1264 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1265 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1266 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1267 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1268 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1269 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1270 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1271 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1272 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1273 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1274 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1275 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1276 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1277 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1278 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1279 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1280 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1281 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1282 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1283 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1284 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001285 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1286 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1287 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1288 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1289 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1290 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1291 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1292 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1293 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1294 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1295 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1296 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1297 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1298 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1299 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1300 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1301 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1302 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1303 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1304 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1305 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1306 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1307 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1308 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1309 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1310 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1311 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1312 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1313 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1314 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1315 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1316 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1318 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1319 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1320 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1321 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1322 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1323 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1325 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1326 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1327 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001339 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1340 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001341 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1342 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1343 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1344 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1345 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1346 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1347 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1348 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1349 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1350 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1351 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1352 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1353 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1354 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1355 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1356 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1357 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1358 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1359 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1360 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001361 "src/math/exp-neonfma-lut64-p2.c",
1362 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001363 "src/math/expminus-neonfma-lut2048-p1.c",
1364 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001365 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001366 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1367 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1368 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1369 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1370 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1371 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1372 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1373 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1374 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1375 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1376 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1377 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1378 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1379 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1380 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1381 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1382 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1383 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001384 "src/math/sqrt-neonfma-nr1fma.c",
1385 "src/math/sqrt-neonfma-nr2fma.c",
1386 "src/math/sqrt-neonfma-nr3fma.c",
1387 "src/math/sqrt-neonfma-nr2fma1adj.c",
1388 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001389]
1390
1391AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001392 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1393 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1394 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1395 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1396 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1397 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001398 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1399 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1400 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1401 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1402 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1403 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1404 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1405 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1406 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1407 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1408 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1409 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1410 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1411 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1412 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1413 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1414 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1415 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1416 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1417 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001418 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1419 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001420 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1421 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001422 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1423 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001424 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1425 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001426 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1427 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1428 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1429 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1430 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001431 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1432 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1433 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1434 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1435 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1436 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1437 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1438 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1439 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1440 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1441 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1442 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1443 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1444 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1445 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1446 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1447 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1448 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001449 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1450 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1451 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1452 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1453 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1454 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1455 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1456 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1457 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1458 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1459 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1460 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1461 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1462 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1463 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1464 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1465 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1466 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001467 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1468 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001469 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1470 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1471 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1472 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1473 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1474 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001475]
1476
Marat Dukhan8853b822020-05-07 12:19:01 -07001477NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001478 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1479 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1480 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1481 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1482 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1483 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1484 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1485 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001486 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001487 "src/math/roundd-neonv8.c",
1488 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001489 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001490]
1491
Marat Dukhan08c4a432019-10-03 09:29:21 -07001492AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001493 "src/f16-clamp/gen/neonfp16arith-x16.c",
1494 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001495 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1496 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1497 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1498 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1499 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1500 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1501 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1502 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1503 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1504 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1505 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1506 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001507 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1508 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001509 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1510 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001511 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1512 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001513 "src/f16-relu/gen/neonfp16arith-x16.c",
1514 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001515 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1516 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1517 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1518 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1519 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1520 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1521 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1522 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1523 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1524 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1525 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1526 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1527 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1528 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1529 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1530 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1531 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1532 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1533 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1534 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1535 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1536 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1537 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1538 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1539 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1540 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1541 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1542 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001543 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1544 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001545 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1546 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1547 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1548 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1549 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1550 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1551 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1552 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001553 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1554 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1555 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1556 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001557 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1558 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1559 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1560 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1561 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1562 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1563 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1564 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1565 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1566 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1567 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1568 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001569 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1570 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1571 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1572 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1573 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1574 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1575 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1576 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001577]
1578
1579SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001580 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1581 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001582 "src/f32-clamp/gen/sse-x4.c",
1583 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001584 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1585 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001586 "src/f32-dwconv-chw/3x3p1-sse.c",
1587 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001588 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1589 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1590 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1591 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1592 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1593 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1594 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1595 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1596 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1597 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1598 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1599 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001600 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001601 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1602 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001603 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1604 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1605 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1606 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1607 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1608 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1609 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1610 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1611 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1612 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1613 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1614 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1615 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001616 "src/f32-hswish/gen/hswish-sse-x4.c",
1617 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001618 "src/f32-ibilinear/gen/sse-c4.c",
1619 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001620 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1621 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1622 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1623 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1624 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1625 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1626 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001627 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1628 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1629 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001630 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001631 "src/f32-prelu/gen/sse-2x4.c",
1632 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001633 "src/f32-relu/gen/sse-x4.c",
1634 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001635 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001636 "src/f32-spmm/gen/4x1-minmax-sse.c",
1637 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001638 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001639 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1640 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1641 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1642 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1643 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1644 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1645 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1646 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001647 "src/f32-vbinary/gen/vmax-sse-x4.c",
1648 "src/f32-vbinary/gen/vmax-sse-x8.c",
1649 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1650 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1651 "src/f32-vbinary/gen/vmin-sse-x4.c",
1652 "src/f32-vbinary/gen/vmin-sse-x8.c",
1653 "src/f32-vbinary/gen/vminc-sse-x4.c",
1654 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001655 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1656 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1657 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1658 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1659 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1660 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1661 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1662 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001663 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1664 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1665 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1666 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001667 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1668 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1669 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1670 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001671 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1672 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001673 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1674 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001675 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1676 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001677 "src/f32-vunary/gen/vabs-sse-x4.c",
1678 "src/f32-vunary/gen/vabs-sse-x8.c",
1679 "src/f32-vunary/gen/vneg-sse-x4.c",
1680 "src/f32-vunary/gen/vneg-sse-x8.c",
1681 "src/f32-vunary/gen/vsqr-sse-x4.c",
1682 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001683 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001684 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001685 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001686 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001687 "src/math/roundd-sse-addsub.c",
1688 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001689 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001690 "src/math/sqrt-sse-hh1mac.c",
1691 "src/math/sqrt-sse-nr1mac.c",
1692 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693]
1694
1695SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001696 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1697 "src/f32-argmaxpool/4x-sse2-c4.c",
1698 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001699 "src/f32-prelu/gen/sse2-2x4.c",
1700 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001701 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1702 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1703 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1704 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1705 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1706 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1707 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1708 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1709 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1710 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1711 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1712 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001713 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001714 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001715 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001716 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001717 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1718 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001719 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1720 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001721 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1722 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1723 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1724 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1725 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1726 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1727 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1728 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001729 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1730 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1731 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1732 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1733 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1734 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1735 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1736 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1737 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001738 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001739 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001740 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001741 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001742 "src/x32-zip/x2-sse2.c",
1743 "src/x32-zip/x3-sse2.c",
1744 "src/x32-zip/x4-sse2.c",
1745 "src/x32-zip/xm-sse2.c",
1746 "src/x8-zip/x2-sse2.c",
1747 "src/x8-zip/x3-sse2.c",
1748 "src/x8-zip/x4-sse2.c",
1749 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001750 "src/math/exp-sse2-p5.c",
1751 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001752 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001753 "src/math/roundd-sse2-cvt.c",
1754 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001755 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001756 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001757 "src/requantization/precise-sse2.c",
1758 "src/requantization/fp32-sse2.c",
1759 "src/requantization/q31-sse2.c",
1760 "src/requantization/gemmlowp-sse2.c",
1761]
1762
1763SSSE3_UKERNELS = [
1764 "src/requantization/precise-ssse3.c",
1765 "src/requantization/q31-ssse3.c",
1766 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001767]
1768
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001769SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001770 "src/f32-prelu/gen/sse41-2x4.c",
1771 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001772 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001773 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001774 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001775 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001776 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1777 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001778 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1779 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001780 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1781 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1782 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1783 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1784 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1785 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1786 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1787 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001788 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001789 "src/math/roundd-sse41.c",
1790 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001791 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001792 "src/requantization/precise-sse4.c",
1793 "src/requantization/q31-sse4.c",
1794 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001795]
1796
Marat Dukhan08c4a432019-10-03 09:29:21 -07001797AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001798 "src/f32-clamp/gen/avx-x8.c",
1799 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001800 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1801 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1802 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1803 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1804 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1805 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1806 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1807 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1808 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1809 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1810 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1811 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1812 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1813 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1814 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1815 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1816 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1817 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1818 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1819 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1820 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1821 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1822 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1823 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1824 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1825 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1826 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1827 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1828 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1829 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001830 "src/f32-hswish/gen/hswish-avx-x8.c",
1831 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001832 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1833 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1834 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1835 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1836 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1837 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1838 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1839 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1840 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001841 "src/f32-prelu/gen/avx-2x8.c",
1842 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001843 "src/f32-relu/gen/avx-x16.c",
1844 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001845 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001846 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1847 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1848 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1849 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1850 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1851 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1852 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1853 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001854 "src/f32-vbinary/gen/vmax-avx-x8.c",
1855 "src/f32-vbinary/gen/vmax-avx-x16.c",
1856 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1857 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1858 "src/f32-vbinary/gen/vmin-avx-x8.c",
1859 "src/f32-vbinary/gen/vmin-avx-x16.c",
1860 "src/f32-vbinary/gen/vminc-avx-x8.c",
1861 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001862 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1863 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1864 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1865 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1866 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1867 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1868 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1869 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001870 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1871 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1872 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1873 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001874 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1875 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1876 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1877 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001878 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1879 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001880 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001881 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1882 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1883 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1884 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1885 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1886 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1887 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1888 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001889 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1890 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001891 "src/f32-vunary/gen/vabs-avx-x8.c",
1892 "src/f32-vunary/gen/vabs-avx-x16.c",
1893 "src/f32-vunary/gen/vneg-avx-x8.c",
1894 "src/f32-vunary/gen/vneg-avx-x16.c",
1895 "src/f32-vunary/gen/vsqr-avx-x8.c",
1896 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001897]
1898
Marat Dukhanfda12b82019-11-21 12:27:59 -08001899FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001900 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1901 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1902 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1903 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1904 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1905 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1906 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1907 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1908 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1909 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1910 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1911 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1912 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1913 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1914 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1915 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1916 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1917 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1918 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1919 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1920 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1921 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1922 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1923 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1924 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1925 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1926 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1927 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1928 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1929 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1930 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1931 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1932 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1933 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1934 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1935 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1936 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1937 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1938 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1939 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001940 "src/f32-hswish/gen/hswish-fma3-x8.c",
1941 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001942 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1943 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1944 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1945 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1946 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1947 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1948 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1949 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1950 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1951 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1952 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1953 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1954 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1955 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001956 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1957 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1958 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1959 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1960 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1961 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1962 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1963 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001964 "src/math/sqrt-fma3-nr1fma.c",
1965 "src/math/sqrt-fma3-nr2fma.c",
1966 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001967]
1968
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001969AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001970 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1971 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1972 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1973 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1974 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1975 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1976 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1977 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1978 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1979 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1980 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1981 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1982 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1983 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1984 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1985 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1986 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1987 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1988 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1989 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1990 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1991 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1992 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1993 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1994 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1995 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1996 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1997 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1998 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1999 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2000 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2001 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2002 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2003 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2004 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2005 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002006 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2007 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2008 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2009 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2010 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2011 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2012 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2013 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2014 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2015 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2016 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2017 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2018 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2019 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2020 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2021 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2022 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2023 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2024 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2025 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2026 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2027 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2028 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2029 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2030 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2031 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2032 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2033 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2034 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2035 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002036 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2037 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2038 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2039 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2040 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2041 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2042 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2043 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2044 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2045 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2046 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2047 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2048 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2049 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2050 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2051 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2052 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2053 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2054 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2055 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2056 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2057 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2058 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2059 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002060 "src/math/exp-avx2-p5.c",
2061 "src/math/exp-avx2-perm-p3.c",
2062 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002063 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002064 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002065 "src/math/sigmoid-avx2-rr2-p5-div.c",
2066 "src/math/sigmoid-avx2-rr1-p5-div.c",
2067 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2068 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2069 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2070 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002071]
2072
Marat Dukhan08c4a432019-10-03 09:29:21 -07002073AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002074 "src/f32-clamp/gen/avx512f-x16.c",
2075 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002076 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2077 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2078 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2079 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2080 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2081 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2082 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2083 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2084 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2085 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2086 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2087 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2088 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2089 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2090 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2091 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2092 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2093 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2094 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2095 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2096 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2097 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2098 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2099 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002100 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2101 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002102 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2103 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2104 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2105 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2106 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2107 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002108 "src/f32-prelu/gen/avx512f-2x16.c",
2109 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002110 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2111 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2112 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2113 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2114 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2115 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2116 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2117 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2118 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2119 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2120 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2121 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2122 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2123 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2124 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2125 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2126 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2127 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2128 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2129 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2130 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2131 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2132 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2133 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2134 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2135 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2136 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2137 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2138 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2139 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2140 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2141 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2142 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2143 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2144 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2145 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002146 "src/f32-relu/gen/avx512f-x16.c",
2147 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002148 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002149 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2150 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2151 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2152 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2153 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2154 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2155 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2156 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002157 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2158 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2159 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2160 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2161 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2162 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2163 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2164 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002165 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2166 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2167 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2168 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2169 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2170 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2171 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2172 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002173 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2174 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2175 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2176 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002177 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2178 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2179 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2180 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002181 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2182 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002183 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002184 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2185 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2186 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2187 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2188 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2189 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2190 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2191 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2192 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2193 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2194 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2195 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2196 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2197 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2198 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2199 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2200 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2201 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2202 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2203 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2204 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2205 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2206 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2207 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002208 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2209 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2210 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2211 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2212 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2213 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2214 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2215 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002216 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2217 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2218 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2219 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2220 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2221 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2222 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2223 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002224 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2225 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2226 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2227 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2228 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2229 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002230 "src/math/exp-avx512f-p5-scalef.c",
2231 "src/math/exp-avx512f-p5.c",
2232 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002233 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002234 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002235 "src/math/sqrt-avx512f-nr1fma.c",
2236 "src/math/sqrt-avx512f-nr2fma.c",
2237 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238]
2239
2240AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002241 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002242 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002243 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2244 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002245 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002246 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002247 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002248 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002249 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2250 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002251 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2252 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2253 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2254 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002255]
2256
2257AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002258 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002259 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2260 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002261 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002262 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2263 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002264 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2265 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2266 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002267 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002268 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2269 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2270 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002271 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002272 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2273 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002274 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002275 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2276 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2277 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2278 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2279 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2280 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2281 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2282 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2283 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2284 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2285 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2286 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2287 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2288 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2289 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2290 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2291 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2292 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002293 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2294 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002295 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002296 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2297 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2298 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2299 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2300 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2301 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2302 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2303 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2304 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2305 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2306 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2307 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2308 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2309 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2310 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2311 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2312 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2313 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002314 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2315 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2316 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2317 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2318 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2319 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2320 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2321 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2322 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2323 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2324 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2325 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2326 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2327 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2328 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2329 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2330 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2331 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002332]
2333
Marat Dukhan1b354632020-03-23 12:50:22 -07002334INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002335 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002336 "src/xnnpack/argmaxpool.h",
2337 "src/xnnpack/avgpool.h",
2338 "src/xnnpack/clamp.h",
2339 "src/xnnpack/common.h",
2340 "src/xnnpack/conv.h",
2341 "src/xnnpack/dwconv.h",
2342 "src/xnnpack/gavgpool.h",
2343 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002344 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002345 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002346 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002347 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002348 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002349 "src/xnnpack/lut.h",
2350 "src/xnnpack/math.h",
2351 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002352 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002353 "src/xnnpack/packx.h",
2354 "src/xnnpack/pad.h",
2355 "src/xnnpack/params.h",
2356 "src/xnnpack/pavgpool.h",
2357 "src/xnnpack/ppmm.h",
2358 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002359 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002360 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002361 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002362 "src/xnnpack/rmax.h",
2363 "src/xnnpack/scalar-utils.h",
2364 "src/xnnpack/spmm.h",
2365 "src/xnnpack/unpool.h",
2366 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002367 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002368 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002369 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002370 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002371 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002372 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002373 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002374]
2375
2376INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002377 "include/xnnpack.h",
2378 "src/xnnpack/allocator.h",
2379 "src/xnnpack/compute.h",
2380 "src/xnnpack/im2col.h",
2381 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002382 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002383 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002384 "src/xnnpack/operator.h",
2385 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002386 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002387 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002388 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002389 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002390]
2391
Marat Dukhan1b354632020-03-23 12:50:22 -07002392ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002393 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002394]
2395
Marat Dukhan1b354632020-03-23 12:50:22 -07002396MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002397 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002398 "include/xnnpack.h",
2399]
2400
Marat Dukhan1b354632020-03-23 12:50:22 -07002401MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002402 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002403 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404 "src/xnnpack/requantization.h",
2405 "include/xnnpack.h",
2406]
2407
2408OPERATOR_TEST_PARAMS_HDRS = [
2409 "src/xnnpack/params.h",
2410 "src/xnnpack/common.h",
2411]
2412
2413WEIGHTS_PACK_HDRS = [
2414 "src/xnnpack/pack.h",
2415 "src/xnnpack/operator.h",
2416 "src/xnnpack/compute.h",
2417]
2418
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002419LOGGING_COPTS = select({
2420 # No logging in optimized mode
2421 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2422 # Full logging in debug mode
2423 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2424 # Error-only logging in default (fastbuild) mode
2425 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2426})
2427
Marat Dukhan3b59de22020-06-03 20:15:19 -07002428LOGGING_SRCS = select({
2429 # No logging in optimized mode
2430 ":optimized_build": [],
2431 "//conditions:default": [
2432 "src/operator-strings.c",
2433 "src/subgraph-strings.c",
2434 ],
2435})
2436
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002437LOGGING_HDRS = [
2438 "src/xnnpack/log.h",
2439]
2440
Marat Dukhan08c4a432019-10-03 09:29:21 -07002441xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002442 name = "tables",
2443 srcs = TABLE_SRCS,
2444 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002445 gcc_copts = xnnpack_gcc_std_copts(),
2446 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002447)
2448
2449xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002450 name = "scalar_ukernels",
2451 srcs = SCALAR_UKERNELS,
2452 hdrs = INTERNAL_HDRS,
2453 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002454 gcc_copts = xnnpack_gcc_std_copts(),
2455 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002456 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002457 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002458 "@FP16",
2459 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002460 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002461 ],
2462)
2463
2464xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002465 name = "scalar_ukernels_test_mode",
2466 srcs = SCALAR_UKERNELS,
2467 hdrs = INTERNAL_HDRS,
2468 aarch32_copts = ["-marm"],
2469 copts = [
2470 "-UNDEBUG",
2471 "-DXNN_TEST_MODE=1",
2472 ],
2473 gcc_copts = xnnpack_gcc_std_copts(),
2474 msvc_copts = xnnpack_msvc_std_copts(),
2475 deps = [
2476 ":tables",
2477 "@FP16",
2478 "@FXdiv",
2479 "@pthreadpool",
2480 ],
2481)
2482
2483xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002484 name = "wasm_ukernels",
2485 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002486 gcc_copts = xnnpack_gcc_std_copts(),
2487 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002488 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002489 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002490 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002491 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002492 "@FP16",
2493 "@FXdiv",
2494 "@pthreadpool",
2495 ],
2496)
2497
2498xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002499 name = "wasm_ukernels_test_mode",
2500 hdrs = INTERNAL_HDRS,
2501 copts = [
2502 "-UNDEBUG",
2503 "-DXNN_TEST_MODE=1",
2504 ],
2505 gcc_copts = xnnpack_gcc_std_copts(),
2506 msvc_copts = xnnpack_msvc_std_copts(),
2507 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002508 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002509 deps = [
2510 ":tables",
2511 "@FP16",
2512 "@FXdiv",
2513 "@pthreadpool",
2514 ],
2515)
2516
2517xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002518 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002519 hdrs = INTERNAL_HDRS,
2520 aarch32_copts = [
2521 "-marm",
2522 "-mfpu=neon",
2523 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002524 gcc_copts = xnnpack_gcc_std_copts(),
2525 gcc_x86_copts = ["-msse2"],
2526 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002527 optimized_copts = [
2528 "-O3",
2529 "-ffast-math",
2530 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002531 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002532 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002533 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002534 "@FP16",
2535 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002536 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002537 ],
2538)
2539
2540xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002541 name = "psimd_fastmath_ukernels_test_mode",
2542 hdrs = INTERNAL_HDRS,
2543 aarch32_copts = [
2544 "-marm",
2545 "-mfpu=neon",
2546 ],
2547 copts = [
2548 "-UNDEBUG",
2549 "-DXNN_TEST_MODE=1",
2550 ],
2551 gcc_copts = xnnpack_gcc_std_copts(),
2552 gcc_x86_copts = ["-msse2"],
2553 msvc_copts = xnnpack_msvc_std_copts(),
2554 optimized_copts = [
2555 "-O3",
2556 "-ffast-math",
2557 ],
2558 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2559 deps = [
2560 ":tables",
2561 "@FP16",
2562 "@psimd",
2563 "@pthreadpool",
2564 ],
2565)
2566
2567xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002568 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002569 hdrs = INTERNAL_HDRS,
2570 aarch32_copts = [
2571 "-marm",
2572 "-mfpu=neon",
2573 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002574 gcc_copts = xnnpack_gcc_std_copts(),
2575 gcc_x86_copts = ["-msse2"],
2576 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002577 optimized_copts = [
2578 "-O3",
2579 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002580 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002581 deps = [
2582 ":tables",
2583 "@FP16",
2584 "@psimd",
2585 "@pthreadpool",
2586 ],
2587)
2588
2589xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002590 name = "psimd_accmath_ukernels_test_mode",
2591 hdrs = INTERNAL_HDRS,
2592 aarch32_copts = [
2593 "-marm",
2594 "-mfpu=neon",
2595 ],
2596 copts = [
2597 "-UNDEBUG",
2598 "-DXNN_TEST_MODE=1",
2599 ],
2600 gcc_copts = xnnpack_gcc_std_copts(),
2601 gcc_x86_copts = ["-msse2"],
2602 msvc_copts = xnnpack_msvc_std_copts(),
2603 optimized_copts = [
2604 "-O3",
2605 ],
2606 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2607 deps = [
2608 ":tables",
2609 "@FP16",
2610 "@psimd",
2611 "@pthreadpool",
2612 ],
2613)
2614
2615xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002616 name = "neon_ukernels",
2617 hdrs = INTERNAL_HDRS,
2618 aarch32_copts = [
2619 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002620 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002621 "-mfpu=neon",
2622 ],
2623 aarch32_srcs = NEON_UKERNELS,
2624 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002625 gcc_copts = xnnpack_gcc_std_copts(),
2626 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002627 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002628 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002629 "@FP16",
2630 "@pthreadpool",
2631 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002632)
2633
2634xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002635 name = "neon_ukernels_test_mode",
2636 hdrs = INTERNAL_HDRS,
2637 aarch32_copts = [
2638 "-marm",
2639 "-march=armv7-a",
2640 "-mfpu=neon",
2641 ],
2642 aarch32_srcs = NEON_UKERNELS,
2643 aarch64_srcs = NEON_UKERNELS,
2644 copts = [
2645 "-UNDEBUG",
2646 "-DXNN_TEST_MODE=1",
2647 ],
2648 gcc_copts = xnnpack_gcc_std_copts(),
2649 msvc_copts = xnnpack_msvc_std_copts(),
2650 deps = [
2651 ":tables",
2652 "@FP16",
2653 "@pthreadpool",
2654 ],
2655)
2656
2657xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002658 name = "neonfma_ukernels",
2659 hdrs = INTERNAL_HDRS,
2660 aarch32_copts = [
2661 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002662 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002663 "-mfpu=neon-vfpv4",
2664 ],
2665 aarch32_srcs = NEONFMA_UKERNELS,
2666 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002667 apple_aarch32_copts = [
2668 "-mcpu=swift",
2669 "-mtune=generic",
2670 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002671 gcc_copts = xnnpack_gcc_std_copts(),
2672 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002673 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002674 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002675 "@FP16",
2676 "@pthreadpool",
2677 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002678)
2679
2680xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002681 name = "neonfma_ukernels_test_mode",
2682 hdrs = INTERNAL_HDRS,
2683 aarch32_copts = [
2684 "-marm",
2685 "-march=armv7-a",
2686 "-mfpu=neon-vfpv4",
2687 ],
2688 aarch32_srcs = NEONFMA_UKERNELS,
2689 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002690 apple_aarch32_copts = [
2691 "-mcpu=swift",
2692 "-mtune=generic",
2693 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002694 copts = [
2695 "-UNDEBUG",
2696 "-DXNN_TEST_MODE=1",
2697 ],
2698 gcc_copts = xnnpack_gcc_std_copts(),
2699 msvc_copts = xnnpack_msvc_std_copts(),
2700 deps = [
2701 ":tables",
2702 "@FP16",
2703 "@pthreadpool",
2704 ],
2705)
2706
2707xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002708 name = "neonv8_ukernels",
2709 hdrs = INTERNAL_HDRS,
2710 aarch32_copts = [
2711 "-marm",
2712 "-march=armv8-a",
2713 "-mfpu=neon-fp-armv8",
2714 ],
2715 aarch32_srcs = NEONV8_UKERNELS,
2716 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002717 apple_aarch32_copts = [
2718 "-mcpu=cyclone",
2719 "-mtune=generic",
2720 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002721 gcc_copts = xnnpack_gcc_std_copts(),
2722 msvc_copts = xnnpack_msvc_std_copts(),
2723 deps = [
2724 ":tables",
2725 "@FP16",
2726 "@pthreadpool",
2727 ],
2728)
2729
2730xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002731 name = "neonv8_ukernels_test_mode",
2732 hdrs = INTERNAL_HDRS,
2733 aarch32_copts = [
2734 "-marm",
2735 "-march=armv8-a",
2736 "-mfpu=neon-fp-armv8",
2737 ],
2738 aarch32_srcs = NEONV8_UKERNELS,
2739 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002740 apple_aarch32_copts = [
2741 "-mcpu=cyclone",
2742 "-mtune=generic",
2743 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002744 copts = [
2745 "-UNDEBUG",
2746 "-DXNN_TEST_MODE=1",
2747 ],
2748 gcc_copts = xnnpack_gcc_std_copts(),
2749 msvc_copts = xnnpack_msvc_std_copts(),
2750 deps = [
2751 ":tables",
2752 "@FP16",
2753 "@pthreadpool",
2754 ],
2755)
2756
2757xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002758 name = "neonfp16arith_ukernels",
2759 hdrs = INTERNAL_HDRS,
2760 aarch64_copts = ["-march=armv8.2-a+fp16"],
2761 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002762 gcc_copts = xnnpack_gcc_std_copts(),
2763 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002764 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002765 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002766 "@FP16",
2767 "@pthreadpool",
2768 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002769)
2770
2771xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002772 name = "neonfp16arith_ukernels_test_mode",
2773 hdrs = INTERNAL_HDRS,
2774 aarch64_copts = ["-march=armv8.2-a+fp16"],
2775 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2776 copts = [
2777 "-UNDEBUG",
2778 "-DXNN_TEST_MODE=1",
2779 ],
2780 gcc_copts = xnnpack_gcc_std_copts(),
2781 msvc_copts = xnnpack_msvc_std_copts(),
2782 deps = [
2783 ":tables",
2784 "@FP16",
2785 "@pthreadpool",
2786 ],
2787)
2788
2789xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002790 name = "sse2_ukernels",
2791 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002792 gcc_copts = xnnpack_gcc_std_copts(),
2793 gcc_x86_copts = ["-msse2"],
2794 msvc_copts = xnnpack_msvc_std_copts(),
2795 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002796 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002797 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002798 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002799 "@FP16",
2800 "@pthreadpool",
2801 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002802)
2803
2804xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002805 name = "sse2_ukernels_test_mode",
2806 hdrs = INTERNAL_HDRS,
2807 copts = [
2808 "-UNDEBUG",
2809 "-DXNN_TEST_MODE=1",
2810 ],
2811 gcc_copts = xnnpack_gcc_std_copts(),
2812 gcc_x86_copts = ["-msse2"],
2813 msvc_copts = xnnpack_msvc_std_copts(),
2814 msvc_x86_32_copts = ["/arch:SSE2"],
2815 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2816 deps = [
2817 ":tables",
2818 "@FP16",
2819 "@pthreadpool",
2820 ],
2821)
2822
2823xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002824 name = "ssse3_ukernels",
2825 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002826 gcc_copts = xnnpack_gcc_std_copts(),
2827 gcc_x86_copts = ["-mssse3"],
2828 msvc_copts = xnnpack_msvc_std_copts(),
2829 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002830 x86_srcs = SSSE3_UKERNELS,
2831 deps = [
2832 ":tables",
2833 "@FP16",
2834 "@pthreadpool",
2835 ],
2836)
2837
2838xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002839 name = "ssse3_ukernels_test_mode",
2840 hdrs = INTERNAL_HDRS,
2841 copts = [
2842 "-UNDEBUG",
2843 "-DXNN_TEST_MODE=1",
2844 ],
2845 gcc_copts = xnnpack_gcc_std_copts(),
2846 gcc_x86_copts = ["-mssse3"],
2847 msvc_copts = xnnpack_msvc_std_copts(),
2848 msvc_x86_32_copts = ["/arch:SSE2"],
2849 x86_srcs = SSSE3_UKERNELS,
2850 deps = [
2851 ":tables",
2852 "@FP16",
2853 "@pthreadpool",
2854 ],
2855)
2856
2857xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002858 name = "sse41_ukernels",
2859 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002860 gcc_copts = xnnpack_gcc_std_copts(),
2861 gcc_x86_copts = ["-msse4.1"],
2862 msvc_copts = xnnpack_msvc_std_copts(),
2863 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002864 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002865 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002866 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002867 "@FP16",
2868 "@pthreadpool",
2869 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002870)
2871
2872xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002873 name = "sse41_ukernels_test_mode",
2874 hdrs = INTERNAL_HDRS,
2875 copts = [
2876 "-UNDEBUG",
2877 "-DXNN_TEST_MODE=1",
2878 ],
2879 gcc_copts = xnnpack_gcc_std_copts(),
2880 gcc_x86_copts = ["-msse4.1"],
2881 msvc_copts = xnnpack_msvc_std_copts(),
2882 msvc_x86_32_copts = ["/arch:SSE2"],
2883 x86_srcs = SSE41_UKERNELS,
2884 deps = [
2885 ":tables",
2886 "@FP16",
2887 "@pthreadpool",
2888 ],
2889)
2890
2891xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002892 name = "avx_ukernels",
2893 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002894 gcc_copts = xnnpack_gcc_std_copts(),
2895 gcc_x86_copts = ["-mavx"],
2896 msvc_copts = xnnpack_msvc_std_copts(),
2897 msvc_x86_32_copts = ["/arch:AVX"],
2898 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002899 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002900 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002901 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002902 "@FP16",
2903 "@pthreadpool",
2904 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002905)
2906
2907xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002908 name = "avx_ukernels_test_mode",
2909 hdrs = INTERNAL_HDRS,
2910 copts = [
2911 "-UNDEBUG",
2912 "-DXNN_TEST_MODE=1",
2913 ],
2914 gcc_copts = xnnpack_gcc_std_copts(),
2915 gcc_x86_copts = ["-mavx"],
2916 msvc_copts = xnnpack_msvc_std_copts(),
2917 msvc_x86_32_copts = ["/arch:AVX"],
2918 msvc_x86_64_copts = ["/arch:AVX"],
2919 x86_srcs = AVX_UKERNELS,
2920 deps = [
2921 ":tables",
2922 "@FP16",
2923 "@pthreadpool",
2924 ],
2925)
2926
2927xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002928 name = "fma3_ukernels",
2929 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002930 gcc_copts = xnnpack_gcc_std_copts(),
2931 gcc_x86_copts = ["-mfma"],
2932 msvc_copts = xnnpack_msvc_std_copts(),
2933 msvc_x86_32_copts = ["/arch:AVX"],
2934 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002935 x86_srcs = FMA3_UKERNELS,
2936 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002937 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002938 "@FP16",
2939 "@pthreadpool",
2940 ],
2941)
2942
2943xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002944 name = "fma3_ukernels_test_mode",
2945 hdrs = INTERNAL_HDRS,
2946 copts = [
2947 "-UNDEBUG",
2948 "-DXNN_TEST_MODE=1",
2949 ],
2950 gcc_copts = xnnpack_gcc_std_copts(),
2951 gcc_x86_copts = ["-mfma"],
2952 msvc_copts = xnnpack_msvc_std_copts(),
2953 msvc_x86_32_copts = ["/arch:AVX"],
2954 msvc_x86_64_copts = ["/arch:AVX"],
2955 x86_srcs = FMA3_UKERNELS,
2956 deps = [
2957 ":tables",
2958 "@FP16",
2959 "@pthreadpool",
2960 ],
2961)
2962
2963xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002964 name = "avx2_ukernels",
2965 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002966 gcc_copts = xnnpack_gcc_std_copts(),
2967 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002968 "-mfma",
2969 "-mavx2",
2970 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002971 msvc_copts = xnnpack_msvc_std_copts(),
2972 msvc_x86_32_copts = ["/arch:AVX2"],
2973 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002974 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002975 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002976 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002977 "@FP16",
2978 "@pthreadpool",
2979 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002980)
2981
2982xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002983 name = "avx2_ukernels_test_mode",
2984 hdrs = INTERNAL_HDRS,
2985 copts = [
2986 "-UNDEBUG",
2987 "-DXNN_TEST_MODE=1",
2988 ],
2989 gcc_copts = xnnpack_gcc_std_copts(),
2990 gcc_x86_copts = [
2991 "-mfma",
2992 "-mavx2",
2993 ],
2994 msvc_copts = xnnpack_msvc_std_copts(),
2995 msvc_x86_32_copts = ["/arch:AVX2"],
2996 msvc_x86_64_copts = ["/arch:AVX2"],
2997 x86_srcs = AVX2_UKERNELS,
2998 deps = [
2999 ":tables",
3000 "@FP16",
3001 "@pthreadpool",
3002 ],
3003)
3004
3005xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003006 name = "avx512f_ukernels",
3007 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003008 gcc_copts = xnnpack_gcc_std_copts(),
3009 gcc_x86_copts = ["-mavx512f"],
3010 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3011 msvc_copts = xnnpack_msvc_std_copts(),
3012 msvc_x86_32_copts = ["/arch:AVX512"],
3013 msvc_x86_64_copts = ["/arch:AVX512"],
3014 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003015 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003016 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003017 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003018 "@FP16",
3019 "@pthreadpool",
3020 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021)
3022
3023xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003024 name = "avx512f_ukernels_test_mode",
3025 hdrs = INTERNAL_HDRS,
3026 copts = [
3027 "-UNDEBUG",
3028 "-DXNN_TEST_MODE=1",
3029 ],
3030 gcc_copts = xnnpack_gcc_std_copts(),
3031 gcc_x86_copts = ["-mavx512f"],
3032 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3033 msvc_copts = xnnpack_msvc_std_copts(),
3034 msvc_x86_32_copts = ["/arch:AVX512"],
3035 msvc_x86_64_copts = ["/arch:AVX512"],
3036 msys_copts = ["-fno-asynchronous-unwind-tables"],
3037 x86_srcs = AVX512F_UKERNELS,
3038 deps = [
3039 ":tables",
3040 "@FP16",
3041 "@pthreadpool",
3042 ],
3043)
3044
3045xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003046 name = "asm_ukernels",
3047 hdrs = ["src/xnnpack/assembly.h"],
3048 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003049 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003050 aarch64_srcs = AARCH64_ASM_UKERNELS,
3051)
3052
Marat Dukhan3b59de22020-06-03 20:15:19 -07003053xnnpack_cc_library(
3054 name = "logging_utils",
3055 srcs = LOGGING_SRCS,
3056 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3057 copts = LOGGING_COPTS + [
3058 "-Isrc",
3059 "-Iinclude",
3060 ] + select({
3061 ":debug_build": [],
3062 "//conditions:default": xnnpack_min_size_copts(),
3063 }),
3064 gcc_copts = xnnpack_gcc_std_copts(),
3065 msvc_copts = xnnpack_msvc_std_copts(),
3066 visibility = xnnpack_visibility(),
3067 deps = [
3068 "@FP16",
3069 "@clog",
3070 "@pthreadpool",
3071 ],
3072)
3073
Marat Dukhan08c4a432019-10-03 09:29:21 -07003074xnnpack_aggregate_library(
3075 name = "ukernels",
3076 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003077 ":neon_ukernels",
3078 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003079 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003080 ":asm_ukernels",
3081 ],
3082 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003083 ":neon_ukernels",
3084 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003085 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 ":neonfp16arith_ukernels",
3087 ":asm_ukernels",
3088 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003089 generic_deps = [
3090 ":scalar_ukernels",
3091 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003092 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003093 ":psimd_fastmath_ukernels",
3094 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003096 wasm_deps = [
3097 ":wasm_ukernels",
3098 ],
3099 wasmsimd_deps = [
3100 ":wasm_ukernels",
3101 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003102 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003103 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003104 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003105 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003107 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003108 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003109 ":avx512f_ukernels",
3110 ],
3111)
3112
Marat Dukhan33fcf782020-05-24 14:27:15 -07003113xnnpack_aggregate_library(
3114 name = "ukernels_test_mode",
3115 aarch32_deps = [
3116 ":neon_ukernels_test_mode",
3117 ":neonfma_ukernels_test_mode",
3118 ":neonv8_ukernels_test_mode",
3119 ":asm_ukernels",
3120 ],
3121 aarch64_deps = [
3122 ":neon_ukernels_test_mode",
3123 ":neonfma_ukernels_test_mode",
3124 ":neonv8_ukernels_test_mode",
3125 ":neonfp16arith_ukernels_test_mode",
3126 ":asm_ukernels",
3127 ],
3128 generic_deps = [
3129 ":scalar_ukernels_test_mode",
3130 ],
3131 psimd_deps = [
3132 ":psimd_fastmath_ukernels_test_mode",
3133 ":psimd_accmath_ukernels_test_mode",
3134 ],
3135 wasm_deps = [
3136 ":wasm_ukernels_test_mode",
3137 ],
3138 wasmsimd_deps = [
3139 ":wasm_ukernels_test_mode",
3140 ],
3141 x86_deps = [
3142 ":sse2_ukernels_test_mode",
3143 ":ssse3_ukernels_test_mode",
3144 ":sse41_ukernels_test_mode",
3145 ":avx_ukernels_test_mode",
3146 ":fma3_ukernels_test_mode",
3147 ":avx2_ukernels_test_mode",
3148 ":avx512f_ukernels_test_mode",
3149 ],
3150)
3151
Marat Dukhan86df3842020-06-17 14:03:31 -07003152# This target must be compatible with Android NDK r17 and compile with Clang 6.
3153# It is similar to :ukernels target, but excludes microkernels which require
3154# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003155xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003156 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003157 aarch32_deps = [
3158 ":neon_ukernels",
3159 ":neonfma_ukernels",
3160 ":neonv8_ukernels",
3161 ":asm_ukernels",
3162 ],
3163 aarch64_deps = [
3164 ":neon_ukernels",
3165 ":neonfma_ukernels",
3166 ":neonv8_ukernels",
3167 ":asm_ukernels",
3168 ],
3169 generic_deps = [
3170 ":scalar_ukernels",
3171 ],
3172 psimd_deps = [
3173 ":psimd_fastmath_ukernels",
3174 ":psimd_accmath_ukernels",
3175 ],
3176 wasm_deps = [
3177 ":wasm_ukernels",
3178 ],
3179 wasmsimd_deps = [
3180 ":wasm_ukernels",
3181 ],
3182 x86_deps = [
3183 ":sse2_ukernels",
3184 ":ssse3_ukernels",
3185 ":sse41_ukernels",
3186 ":avx_ukernels",
3187 ":fma3_ukernels",
3188 ":avx2_ukernels",
3189 ":avx512f_ukernels",
3190 ],
3191)
3192
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193xnnpack_cc_library(
3194 name = "im2col",
3195 srcs = ["src/im2col.c"],
3196 hdrs = [
3197 "src/xnnpack/common.h",
3198 "src/xnnpack/im2col.h",
3199 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003200 gcc_copts = xnnpack_gcc_std_copts(),
3201 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003202)
3203
3204xnnpack_cc_library(
3205 name = "indirection",
3206 srcs = ["src/indirection.c"],
3207 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003208 gcc_copts = xnnpack_gcc_std_copts(),
3209 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210 deps = [
3211 "@FP16",
3212 "@FXdiv",
3213 "@pthreadpool",
3214 ],
3215)
3216
3217xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003218 name = "indirection_test_mode",
3219 srcs = ["src/indirection.c"],
3220 hdrs = INTERNAL_HDRS,
3221 copts = [
3222 "-UNDEBUG",
3223 "-DXNN_TEST_MODE=1",
3224 ],
3225 gcc_copts = xnnpack_gcc_std_copts(),
3226 msvc_copts = xnnpack_msvc_std_copts(),
3227 deps = [
3228 "@FP16",
3229 "@FXdiv",
3230 "@pthreadpool",
3231 ],
3232)
3233
3234xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003235 name = "packing",
3236 srcs = ["src/packing.c"],
3237 hdrs = INTERNAL_HDRS,
3238 gcc_copts = xnnpack_gcc_std_copts(),
3239 msvc_copts = xnnpack_msvc_std_copts(),
3240 deps = [
3241 "@FP16",
3242 "@FXdiv",
3243 "@pthreadpool",
3244 ],
3245)
3246
3247xnnpack_cc_library(
3248 name = "packing_test_mode",
3249 srcs = ["src/packing.c"],
3250 hdrs = INTERNAL_HDRS,
3251 copts = [
3252 "-UNDEBUG",
3253 "-DXNN_TEST_MODE=1",
3254 ],
3255 gcc_copts = xnnpack_gcc_std_copts(),
3256 msvc_copts = xnnpack_msvc_std_copts(),
3257 deps = [
3258 "@FP16",
3259 "@FXdiv",
3260 "@pthreadpool",
3261 ],
3262)
3263
3264xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265 name = "operator_run",
3266 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003267 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003268 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003269 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3270 "//conditions:default": [],
3271 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003272 gcc_copts = xnnpack_gcc_std_copts(),
3273 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003274 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003275 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003276 "@FP16",
3277 "@FXdiv",
3278 "@clog",
3279 "@pthreadpool",
3280 ],
3281)
3282
Chao Mei6ddfc602020-05-13 22:29:36 -07003283xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003284 name = "operator_run_test_mode",
3285 srcs = ["src/operator-run.c"],
3286 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3287 copts = LOGGING_COPTS + [
3288 "-UNDEBUG",
3289 "-DXNN_TEST_MODE=1",
3290 ] + select({
3291 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3292 "//conditions:default": [],
3293 }),
3294 gcc_copts = xnnpack_gcc_std_copts(),
3295 msvc_copts = xnnpack_msvc_std_copts(),
3296 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003297 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003298 "@FP16",
3299 "@FXdiv",
3300 "@clog",
3301 "@pthreadpool",
3302 ],
3303)
3304
3305xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003306 name = "memory_planner",
3307 srcs = ["src/memory-planner.c"],
3308 hdrs = INTERNAL_HDRS,
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",
Chao Mei6ddfc602020-05-13 22:29:36 -07003318 "@pthreadpool",
3319 ],
3320)
3321
Marat Dukhan33fcf782020-05-24 14:27:15 -07003322xnnpack_cc_library(
3323 name = "memory_planner_test_mode",
3324 srcs = ["src/memory-planner.c"],
3325 hdrs = INTERNAL_HDRS,
3326 copts = [
3327 "-UNDEBUG",
3328 "-DXNN_TEST_MODE=1",
3329 ],
3330 defines = select({
3331 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3332 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3333 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3334 }),
3335 gcc_copts = xnnpack_gcc_std_copts(),
3336 msvc_copts = xnnpack_msvc_std_copts(),
3337 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003338 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003339 "@pthreadpool",
3340 ],
3341)
3342
Marat Dukhan08c4a432019-10-03 09:29:21 -07003343cc_library(
3344 name = "enable_assembly",
3345 defines = select({
3346 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3347 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003348 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003349 }),
3350)
3351
Marat Dukhan9de90e02020-06-18 16:04:12 -07003352cc_library(
3353 name = "enable_sparse",
3354 defines = select({
3355 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3356 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003357 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003358 }),
3359)
3360
Marat Dukhancf056b22019-10-07 10:26:29 -07003361xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003362 name = "operators",
3363 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003364 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003365 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003366 ],
3367 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003368 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003369 "-Isrc",
3370 "-Iinclude",
3371 ] + select({
3372 ":debug_build": [],
3373 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003374 }) + select({
3375 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3376 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003378 gcc_copts = xnnpack_gcc_std_copts(),
3379 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003380 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003381 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003382 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003383 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003384 "@FP16",
3385 "@FXdiv",
3386 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003387 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003388 ],
3389)
3390
Marat Dukhan10a38082020-04-17 03:58:35 -07003391xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003392 name = "operators_test_mode",
3393 srcs = OPERATOR_SRCS + [
3394 "src/memory.c",
3395 "src/operator-delete.c",
3396 ],
3397 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3398 copts = LOGGING_COPTS + [
3399 "-Isrc",
3400 "-Iinclude",
3401 "-UNDEBUG",
3402 "-DXNN_TEST_MODE=1",
3403 ] + select({
3404 ":debug_build": [],
3405 "//conditions:default": xnnpack_min_size_copts(),
3406 }) + select({
3407 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3408 "//conditions:default": [],
3409 }),
3410 gcc_copts = xnnpack_gcc_std_copts(),
3411 msvc_copts = xnnpack_msvc_std_copts(),
3412 deps = [
3413 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003414 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003415 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003416 "@FP16",
3417 "@FXdiv",
3418 "@clog",
3419 "@pthreadpool",
3420 ],
3421)
3422
3423xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003424 name = "XNNPACK",
3425 srcs = [
3426 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003427 "src/runtime.c",
3428 "src/subgraph.c",
3429 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003430 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003431 hdrs = ["include/xnnpack.h"],
3432 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003433 "-Isrc",
3434 "-Iinclude",
3435 ] + select({
3436 ":debug_build": [],
3437 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003438 }) + select({
3439 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3440 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003441 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003442 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003443 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003444 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003445 visibility = xnnpack_visibility(),
3446 deps = [
3447 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003448 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003449 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003450 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003451 ":operator_run",
3452 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003453 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003454 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003455 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003456 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003457 ] + select({
3458 ":emscripten": [],
3459 "//conditions:default": ["@cpuinfo"],
3460 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003461)
3462
Marat Dukhan10a38082020-04-17 03:58:35 -07003463xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003464 name = "XNNPACK_test_mode",
3465 srcs = [
3466 "src/init.c",
3467 "src/runtime.c",
3468 "src/subgraph.c",
3469 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003470 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003471 hdrs = ["include/xnnpack.h"],
3472 copts = LOGGING_COPTS + [
3473 "-Isrc",
3474 "-Iinclude",
3475 "-UNDEBUG",
3476 "-DXNN_TEST_MODE=1",
3477 ] + select({
3478 ":debug_build": [],
3479 "//conditions:default": xnnpack_min_size_copts(),
3480 }) + select({
3481 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3482 "//conditions:default": [],
3483 }),
3484 gcc_copts = xnnpack_gcc_std_copts(),
3485 includes = ["include"],
3486 msvc_copts = xnnpack_msvc_std_copts(),
3487 visibility = xnnpack_visibility(),
3488 deps = [
3489 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003490 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003491 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003492 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003493 ":operator_run_test_mode",
3494 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003495 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003496 "@clog",
3497 "@FP16",
3498 "@pthreadpool",
3499 ] + select({
3500 ":emscripten": [],
3501 "//conditions:default": ["@cpuinfo"],
3502 }),
3503)
3504
3505xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003506 name = "xnnpack_f32",
3507 srcs = [
3508 "src/init.c",
3509 "src/runtime.c",
3510 "src/subgraph.c",
3511 "src/tensor.c",
3512 ] + SUBGRAPH_SRCS,
3513 hdrs = ["include/xnnpack.h"],
3514 copts = LOGGING_COPTS + [
3515 "-Isrc",
3516 "-Iinclude",
3517 ] + select({
3518 ":debug_build": [],
3519 "//conditions:default": xnnpack_min_size_copts(),
3520 }) + select({
3521 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3522 "//conditions:default": [],
3523 }),
3524 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003525 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003526 "XNN_NO_U8_OPERATORS",
3527 "XNN_NO_X8_OPERATORS",
3528 "XNN_NO_F16_OPERATORS",
3529 "XNN_NO_X16_OPERATORS",
3530 ],
3531 gcc_copts = xnnpack_gcc_std_copts(),
3532 includes = ["include"],
3533 msvc_copts = xnnpack_msvc_std_copts(),
3534 visibility = xnnpack_visibility(),
3535 deps = [
3536 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003537 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003538 ":logging_utils",
3539 ":memory_planner",
3540 ":operator_run",
3541 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003542 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003543 "@clog",
3544 "@FP16",
3545 "@pthreadpool",
3546 ] + select({
3547 ":emscripten": [],
3548 "//conditions:default": ["@cpuinfo"],
3549 }),
3550)
3551
3552xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003553 name = "xnnpack_operators_nhwc_f32",
3554 srcs = [
3555 "src/init.c",
3556 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003557 hdrs = ["include/xnnpack.h"],
3558 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003559 "-Isrc",
3560 "-Iinclude",
3561 ] + select({
3562 ":debug_build": [],
3563 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003564 }) + select({
3565 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3566 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003567 }),
3568 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003569 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003570 "XNN_NO_U8_OPERATORS",
3571 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003572 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003573 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003574 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003575 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003576 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003577 visibility = xnnpack_visibility(),
3578 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003579 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003580 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003581 ":operator_run",
3582 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003583 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003584 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003586 ] + select({
3587 ":emscripten": [],
3588 "//conditions:default": ["@cpuinfo"],
3589 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003590)
3591
Marat Dukhancf056b22019-10-07 10:26:29 -07003592xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593 name = "bench_utils",
3594 srcs = ["bench/utils.cc"],
3595 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003596 deps = [
3597 "@com_google_benchmark//:benchmark",
3598 "@cpuinfo",
3599 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600)
3601
Frank Barchard7e955972019-10-11 10:34:25 -07003602######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603
3604xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003605 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003606 srcs = [
3607 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003608 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003609 "src/xnnpack/AlignedAllocator.h",
3610 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003611 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003612 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003613)
3614
3615xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003616 name = "f16_igemm_bench",
3617 srcs = [
3618 "bench/f16-igemm.cc",
3619 "bench/conv.h",
3620 "bench/google/conv.h",
3621 "src/xnnpack/AlignedAllocator.h",
3622 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003623 deps = MICROKERNEL_BENCHMARK_DEPS + [
3624 ":indirection",
3625 ":packing",
3626 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003627)
3628
3629xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003630 name = "f16_gemm_bench",
3631 srcs = [
3632 "bench/f16-gemm.cc",
3633 "bench/gemm.h",
3634 "src/xnnpack/AlignedAllocator.h",
3635 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003636 deps = MICROKERNEL_BENCHMARK_DEPS + [
3637 ":packing",
3638 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003639)
3640
3641xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003642 name = "f16_spmm_bench",
3643 srcs = [
3644 "bench/f16-spmm.cc",
3645 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003646 "src/xnnpack/AlignedAllocator.h",
3647 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003648 deps = MICROKERNEL_BENCHMARK_DEPS,
3649)
3650
3651xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652 name = "f32_igemm_bench",
3653 srcs = [
3654 "bench/f32-igemm.cc",
3655 "bench/conv.h",
3656 "src/xnnpack/AlignedAllocator.h",
3657 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003658 deps = MICROKERNEL_BENCHMARK_DEPS + [
3659 ":indirection",
3660 ":packing",
3661 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662)
3663
3664xnnpack_benchmark(
3665 name = "f32_conv_hwc_bench",
3666 srcs = [
3667 "bench/f32-conv-hwc.cc",
3668 "bench/dconv.h",
3669 "src/xnnpack/AlignedAllocator.h",
3670 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003671 deps = MICROKERNEL_BENCHMARK_DEPS + [
3672 ":packing",
3673 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003674)
3675
3676xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003677 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003678 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003679 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003680 "bench/dconv.h",
3681 "src/xnnpack/AlignedAllocator.h",
3682 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003683 deps = MICROKERNEL_BENCHMARK_DEPS + [
3684 ":packing",
3685 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003686)
3687
3688xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003689 name = "f16_dwconv_bench",
3690 srcs = [
3691 "bench/f16-dwconv.cc",
3692 "bench/dwconv.h",
3693 "bench/google/dwconv.h",
3694 "src/xnnpack/AlignedAllocator.h",
3695 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003696 deps = MICROKERNEL_BENCHMARK_DEPS + [
3697 ":indirection",
3698 ":packing",
3699 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003700)
3701
3702xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703 name = "f32_dwconv_bench",
3704 srcs = [
3705 "bench/f32-dwconv.cc",
3706 "bench/dwconv.h",
3707 "src/xnnpack/AlignedAllocator.h",
3708 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003709 deps = MICROKERNEL_BENCHMARK_DEPS + [
3710 ":indirection",
3711 ":packing",
3712 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003713)
3714
3715xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003716 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003717 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003718 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719 "bench/dwconv.h",
3720 "src/xnnpack/AlignedAllocator.h",
3721 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003722 deps = MICROKERNEL_BENCHMARK_DEPS + [
3723 ":indirection",
3724 ":packing",
3725 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003726)
3727
3728xnnpack_benchmark(
3729 name = "f32_gemm_bench",
3730 srcs = [
3731 "bench/f32-gemm.cc",
3732 "bench/gemm.h",
3733 "src/xnnpack/AlignedAllocator.h",
3734 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003735 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003736 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737)
3738
3739xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003740 name = "f32_hswish_bench",
3741 srcs = [
3742 "bench/f32-hswish.cc",
3743 "src/xnnpack/AlignedAllocator.h",
3744 ] + MICROKERNEL_BENCHMARK_HDRS,
3745 deps = MICROKERNEL_BENCHMARK_DEPS,
3746)
3747
3748xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003749 name = "f32_raddexpminusmax_bench",
3750 srcs = [
3751 "bench/f32-raddexpminusmax.cc",
3752 "src/xnnpack/AlignedAllocator.h",
3753 ] + MICROKERNEL_BENCHMARK_HDRS,
3754 deps = MICROKERNEL_BENCHMARK_DEPS,
3755)
3756
3757xnnpack_benchmark(
3758 name = "f32_raddextexp_bench",
3759 srcs = [
3760 "bench/f32-raddextexp.cc",
3761 "src/xnnpack/AlignedAllocator.h",
3762 ] + MICROKERNEL_BENCHMARK_HDRS,
3763 deps = MICROKERNEL_BENCHMARK_DEPS,
3764)
3765
3766xnnpack_benchmark(
3767 name = "f32_raddstoreexpminusmax_bench",
3768 srcs = [
3769 "bench/f32-raddstoreexpminusmax.cc",
3770 "src/xnnpack/AlignedAllocator.h",
3771 ] + MICROKERNEL_BENCHMARK_HDRS,
3772 deps = MICROKERNEL_BENCHMARK_DEPS,
3773)
3774
3775xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003776 name = "f32_rmax_bench",
3777 srcs = [
3778 "bench/f32-rmax.cc",
3779 "src/xnnpack/AlignedAllocator.h",
3780 ] + MICROKERNEL_BENCHMARK_HDRS,
3781 deps = MICROKERNEL_BENCHMARK_DEPS,
3782)
3783
3784xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003785 name = "f32_sigmoid_bench",
3786 srcs = [
3787 "bench/f32-sigmoid.cc",
3788 "src/xnnpack/AlignedAllocator.h",
3789 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003790 deps = MICROKERNEL_BENCHMARK_DEPS,
3791)
3792
3793xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794 name = "f32_spmm_bench",
3795 srcs = [
3796 "bench/f32-spmm.cc",
3797 "bench/gemm.h",
3798 "src/xnnpack/AlignedAllocator.h",
3799 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003800 deps = MICROKERNEL_BENCHMARK_DEPS,
3801)
3802
3803xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003804 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003805 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003806 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003807 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003808 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003809 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003810)
3811
3812xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003813 name = "f32_vscaleexpminusmax_bench",
3814 srcs = [
3815 "bench/f32-vscaleexpminusmax.cc",
3816 "src/xnnpack/AlignedAllocator.h",
3817 ] + MICROKERNEL_BENCHMARK_HDRS,
3818 deps = MICROKERNEL_BENCHMARK_DEPS,
3819)
3820
3821xnnpack_benchmark(
3822 name = "f32_vscaleextexp_bench",
3823 srcs = [
3824 "bench/f32-vscaleextexp.cc",
3825 "src/xnnpack/AlignedAllocator.h",
3826 ] + MICROKERNEL_BENCHMARK_HDRS,
3827 deps = MICROKERNEL_BENCHMARK_DEPS,
3828)
3829
3830xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003831 name = "f32_vsqrt_bench",
3832 srcs = [
3833 "bench/f32-vsqrt.cc",
3834 "src/xnnpack/AlignedAllocator.h",
3835 ] + MICROKERNEL_BENCHMARK_HDRS,
3836 deps = MICROKERNEL_BENCHMARK_DEPS,
3837)
3838
3839xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003840 name = "f32_im2col_gemm_bench",
3841 srcs = [
3842 "bench/f32-im2col-gemm.cc",
3843 "bench/conv.h",
3844 "src/xnnpack/AlignedAllocator.h",
3845 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003846 deps = MICROKERNEL_BENCHMARK_DEPS + [
3847 ":im2col",
3848 ":packing",
3849 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003850)
3851
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003852xnnpack_benchmark(
3853 name = "requantization_bench",
3854 srcs = [
3855 "bench/requantization.cc",
3856 "src/xnnpack/requantization-stubs.h",
3857 "src/xnnpack/AlignedAllocator.h",
3858 ] + MICROKERNEL_BENCHMARK_HDRS,
3859 deps = MICROKERNEL_BENCHMARK_DEPS,
3860)
3861
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003862xnnpack_benchmark(
3863 name = "rounding_bench",
3864 srcs = [
3865 "bench/rounding.cc",
3866 "src/xnnpack/math-stubs.h",
3867 "src/xnnpack/AlignedAllocator.h",
3868 ] + MICROKERNEL_BENCHMARK_HDRS,
3869 deps = MICROKERNEL_BENCHMARK_DEPS,
3870)
3871
Marat Dukhan08c4a432019-10-03 09:29:21 -07003872########################### Benchmarks for operators ###########################
3873
3874xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003875 name = "average_pooling_bench",
3876 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003877 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003878 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003879 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003880)
3881
3882xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003883 name = "bankers_rounding_bench",
3884 srcs = ["bench/bankers-rounding.cc"],
3885 copts = xnnpack_optional_tflite_copts(),
3886 tags = ["nowin32"],
3887 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3888)
3889
3890xnnpack_benchmark(
3891 name = "ceiling_bench",
3892 srcs = ["bench/ceiling.cc"],
3893 copts = xnnpack_optional_tflite_copts(),
3894 tags = ["nowin32"],
3895 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3896)
3897
3898xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003899 name = "channel_shuffle_bench",
3900 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003901 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003902)
3903
3904xnnpack_benchmark(
3905 name = "convolution_bench",
3906 srcs = ["bench/convolution.cc"],
3907 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003908 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003909 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003910)
3911
3912xnnpack_benchmark(
3913 name = "deconvolution_bench",
3914 srcs = ["bench/deconvolution.cc"],
3915 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003916 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003917 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003918)
3919
3920xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003921 name = "floor_bench",
3922 srcs = ["bench/floor.cc"],
3923 copts = xnnpack_optional_tflite_copts(),
3924 tags = ["nowin32"],
3925 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3926)
3927
3928xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003929 name = "global_average_pooling_bench",
3930 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003931 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003932)
3933
3934xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003935 name = "hardswish_bench",
3936 srcs = ["bench/hardswish.cc"],
3937 copts = xnnpack_optional_tflite_copts(),
3938 tags = ["nowin32"],
3939 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3940)
3941
3942xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003943 name = "max_pooling_bench",
3944 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003945 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003946)
3947
3948xnnpack_benchmark(
3949 name = "sigmoid_bench",
3950 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003951 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003952 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003953 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954)
3955
3956xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003957 name = "prelu_bench",
3958 srcs = ["bench/prelu.cc"],
3959 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003960 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003961 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003962)
3963
3964xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003965 name = "softmax_bench",
3966 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003967 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003968 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003969 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003970)
3971
Marat Dukhan87727142020-06-24 15:24:10 -07003972xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003973 name = "square_root_bench",
3974 srcs = ["bench/square-root.cc"],
3975 copts = xnnpack_optional_tflite_copts(),
3976 tags = ["nowin32"],
3977 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3978)
3979
3980xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003981 name = "truncation_bench",
3982 srcs = ["bench/truncation.cc"],
3983 deps = OPERATOR_BENCHMARK_DEPS,
3984)
3985
Marat Dukhanc068bb62019-10-04 13:24:39 -07003986############################# End-to-end benchmarks ############################
3987
3988cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003989 name = "fp32_mobilenet_v1",
3990 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003991 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003992 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003993 linkstatic = True,
3994 deps = [
3995 ":XNNPACK",
3996 "@pthreadpool",
3997 ],
3998)
3999
4000cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004001 name = "fp16_mobilenet_v1",
4002 srcs = ["models/fp16-mobilenet-v1.cc"],
4003 hdrs = ["models/models.h"],
4004 copts = xnnpack_std_cxxopts(),
4005 linkstatic = True,
4006 deps = [
4007 ":XNNPACK",
4008 "@FP16",
4009 "@pthreadpool",
4010 ],
4011)
4012
4013cc_library(
4014 name = "fp32_mobilenet_v2",
4015 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004016 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004017 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004018 linkstatic = True,
4019 deps = [
4020 ":XNNPACK",
4021 "@pthreadpool",
4022 ],
4023)
4024
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004025cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004026 name = "fp16_mobilenet_v2",
4027 srcs = ["models/fp16-mobilenet-v2.cc"],
4028 hdrs = ["models/models.h"],
4029 copts = xnnpack_std_cxxopts(),
4030 linkstatic = True,
4031 deps = [
4032 ":XNNPACK",
4033 "@FP16",
4034 "@pthreadpool",
4035 ],
4036)
4037
4038cc_library(
4039 name = "fp32_mobilenet_v3_large",
4040 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004041 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004042 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004043 linkstatic = True,
4044 deps = [
4045 ":XNNPACK",
4046 "@pthreadpool",
4047 ],
4048)
4049
4050cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004051 name = "fp32_mobilenet_v3_small",
4052 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004053 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004054 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004055 linkstatic = True,
4056 deps = [
4057 ":XNNPACK",
4058 "@pthreadpool",
4059 ],
4060)
4061
Marat Dukhanc068bb62019-10-04 13:24:39 -07004062xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004063 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004064 srcs = [
4065 "bench/f32-dwconv-e2e.cc",
4066 "bench/end2end.h",
4067 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004068 deps = MICROKERNEL_BENCHMARK_DEPS + [
4069 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004070 ":fp32_mobilenet_v1",
4071 ":fp32_mobilenet_v2",
4072 ":fp32_mobilenet_v3_large",
4073 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004074 ],
4075)
4076
4077xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004078 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004079 srcs = [
4080 "bench/f32-gemm-e2e.cc",
4081 "bench/end2end.h",
4082 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004083 deps = MICROKERNEL_BENCHMARK_DEPS + [
4084 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004085 ":fp32_mobilenet_v1",
4086 ":fp32_mobilenet_v2",
4087 ":fp32_mobilenet_v3_large",
4088 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004089 ],
4090)
4091
4092xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004093 name = "end2end_bench",
4094 srcs = ["bench/end2end.cc"],
4095 deps = [
4096 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004097 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004098 ":fp16_mobilenet_v1",
4099 ":fp16_mobilenet_v2",
4100 ":fp32_mobilenet_v1",
4101 ":fp32_mobilenet_v2",
4102 ":fp32_mobilenet_v3_large",
4103 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004104 "@pthreadpool",
4105 ],
4106)
4107
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004108#################### Accuracy evaluation for math functions ####################
4109
4110xnnpack_benchmark(
4111 name = "f32_exp_eval",
4112 srcs = [
4113 "eval/f32-exp.cc",
4114 "src/xnnpack/AlignedAllocator.h",
4115 ] + ACCURACY_EVAL_HDRS,
4116 deps = ACCURACY_EVAL_DEPS,
4117)
4118
Marat Dukhan515c9772019-10-17 18:07:57 -07004119xnnpack_benchmark(
4120 name = "f32_expminus_eval",
4121 srcs = [
4122 "eval/f32-expminus.cc",
4123 "src/xnnpack/AlignedAllocator.h",
4124 ] + ACCURACY_EVAL_HDRS,
4125 deps = ACCURACY_EVAL_DEPS,
4126)
4127
Marat Dukhan98ba4412019-10-23 02:14:28 -07004128xnnpack_benchmark(
4129 name = "f32_extexp_eval",
4130 srcs = [
4131 "eval/f32-extexp.cc",
4132 "src/xnnpack/AlignedAllocator.h",
4133 ] + ACCURACY_EVAL_HDRS,
4134 deps = ACCURACY_EVAL_DEPS,
4135)
4136
Marat Dukhan8853b822020-05-07 12:19:01 -07004137xnnpack_unit_test(
4138 name = "f32_roundne_eval",
4139 srcs = [
4140 "eval/f32-roundne.cc",
4141 "src/xnnpack/AlignedAllocator.h",
4142 "src/xnnpack/math-stubs.h",
4143 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004144 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004145 deps = MICROKERNEL_TEST_DEPS,
4146)
4147
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004148xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004149 name = "f32_roundd_eval",
4150 srcs = [
4151 "eval/f32-roundd.cc",
4152 "src/xnnpack/AlignedAllocator.h",
4153 "src/xnnpack/math-stubs.h",
4154 ] + MICROKERNEL_TEST_HDRS,
4155 automatic = False,
4156 deps = MICROKERNEL_TEST_DEPS,
4157)
4158
4159xnnpack_unit_test(
4160 name = "f32_roundu_eval",
4161 srcs = [
4162 "eval/f32-roundu.cc",
4163 "src/xnnpack/AlignedAllocator.h",
4164 "src/xnnpack/math-stubs.h",
4165 ] + MICROKERNEL_TEST_HDRS,
4166 automatic = False,
4167 deps = MICROKERNEL_TEST_DEPS,
4168)
4169
4170xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004171 name = "f32_roundz_eval",
4172 srcs = [
4173 "eval/f32-roundz.cc",
4174 "src/xnnpack/AlignedAllocator.h",
4175 "src/xnnpack/math-stubs.h",
4176 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004177 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004178 deps = MICROKERNEL_TEST_DEPS,
4179)
4180
Marat Dukhan346a9e52019-11-15 09:06:30 -08004181xnnpack_benchmark(
4182 name = "f32_sigmoid_eval",
4183 srcs = [
4184 "eval/f32-sigmoid.cc",
4185 "src/xnnpack/AlignedAllocator.h",
4186 ] + ACCURACY_EVAL_HDRS,
4187 deps = ACCURACY_EVAL_DEPS,
4188)
4189
Marat Dukhan84000762020-06-29 18:38:43 -07004190xnnpack_benchmark(
4191 name = "f32_sqrt_eval",
4192 srcs = [
4193 "eval/f32-sqrt.cc",
4194 "src/xnnpack/AlignedAllocator.h",
4195 ] + ACCURACY_EVAL_HDRS,
4196 deps = ACCURACY_EVAL_DEPS + [
4197 ":bench_utils",
4198 ],
4199)
4200
Marat Dukhan08c4a432019-10-03 09:29:21 -07004201######################### Unit tests for micro-kernels #########################
4202
4203xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004204 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004206 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004207 "test/gemm-microkernel-tester.h",
4208 "src/xnnpack/AlignedAllocator.h",
4209 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004210 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004211)
4212
4213xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004214 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004215 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004216 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004217 "test/spmm-microkernel-tester.h",
4218 "src/xnnpack/AlignedAllocator.h",
4219 ] + MICROKERNEL_TEST_HDRS,
4220 deps = MICROKERNEL_TEST_DEPS,
4221)
4222
4223xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004224 name = "f32_argmaxpool_test",
4225 srcs = [
4226 "test/f32-argmaxpool.cc",
4227 "test/argmaxpool-microkernel-tester.h",
4228 "src/xnnpack/AlignedAllocator.h",
4229 ] + MICROKERNEL_TEST_HDRS,
4230 deps = MICROKERNEL_TEST_DEPS,
4231)
4232
4233xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004234 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004235 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004236 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237 "test/avgpool-microkernel-tester.h",
4238 "src/xnnpack/AlignedAllocator.h",
4239 ] + MICROKERNEL_TEST_HDRS,
4240 deps = MICROKERNEL_TEST_DEPS,
4241)
4242
4243xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004244 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004245 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004246 "test/f32-ibilinear.cc",
4247 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004248 "src/xnnpack/AlignedAllocator.h",
4249 ] + MICROKERNEL_TEST_HDRS,
4250 deps = MICROKERNEL_TEST_DEPS,
4251)
4252
4253xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004254 name = "f16_clamp_test",
4255 srcs = [
4256 "test/f16-clamp.cc",
4257 "test/clamp-microkernel-tester.h",
4258 ] + MICROKERNEL_TEST_HDRS,
4259 deps = MICROKERNEL_TEST_DEPS,
4260)
4261
4262xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263 name = "f32_clamp_test",
4264 srcs = [
4265 "test/f32-clamp.cc",
4266 "test/clamp-microkernel-tester.h",
4267 ] + MICROKERNEL_TEST_HDRS,
4268 deps = MICROKERNEL_TEST_DEPS,
4269)
4270
4271xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004272 name = "f32_igemm_test",
4273 srcs = [
4274 "test/f32-igemm.cc",
4275 "test/gemm-microkernel-tester.h",
4276 "src/xnnpack/AlignedAllocator.h",
4277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004278 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004279)
4280
4281xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004282 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004284 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285 "test/gemm-microkernel-tester.h",
4286 "src/xnnpack/AlignedAllocator.h",
4287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004288 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289)
4290
4291xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004292 name = "f32_igemm_minmax_test",
4293 srcs = [
4294 "test/f32-igemm-minmax.cc",
4295 "test/gemm-microkernel-tester.h",
4296 "src/xnnpack/AlignedAllocator.h",
4297 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004298 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004299)
4300
4301xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004302 name = "f16_igemm_minmax_test",
4303 srcs = [
4304 "test/f16-igemm-minmax.cc",
4305 "test/gemm-microkernel-tester.h",
4306 "src/xnnpack/AlignedAllocator.h",
4307 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004308 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004309)
4310
4311xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004312 name = "f32_conv_hwc_test",
4313 srcs = [
4314 "test/f32-conv-hwc.cc",
4315 "test/conv-hwc-microkernel-tester.h",
4316 "src/xnnpack/AlignedAllocator.h",
4317 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004318 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319)
4320
4321xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004322 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004323 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004324 "test/f32-conv-hwc2chw.cc",
4325 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326 "src/xnnpack/AlignedAllocator.h",
4327 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004328 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329)
4330
4331xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004332 name = "f32_dwconv_test",
4333 srcs = [
4334 "test/f32-dwconv.cc",
4335 "test/dwconv-microkernel-tester.h",
4336 "src/xnnpack/AlignedAllocator.h",
4337 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004338 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004339)
4340
4341xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004342 name = "f16_dwconv_minmax_test",
4343 srcs = [
4344 "test/f16-dwconv-minmax.cc",
4345 "test/dwconv-microkernel-tester.h",
4346 "src/xnnpack/AlignedAllocator.h",
4347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004348 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004349)
4350
4351xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004352 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004353 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004354 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 "test/dwconv-microkernel-tester.h",
4356 "src/xnnpack/AlignedAllocator.h",
4357 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004358 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004359)
4360
4361xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004362 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004363 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004364 "test/f32-dwconv-chw.cc",
4365 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004366 "src/xnnpack/AlignedAllocator.h",
4367 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004368 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004369)
4370
4371xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004372 name = "f16_gavgpool_minmax_test",
4373 srcs = [
4374 "test/f16-gavgpool-minmax.cc",
4375 "test/gavgpool-microkernel-tester.h",
4376 "src/xnnpack/AlignedAllocator.h",
4377 ] + MICROKERNEL_TEST_HDRS,
4378 deps = MICROKERNEL_TEST_DEPS,
4379)
4380
4381xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004382 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004384 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004385 "test/gavgpool-microkernel-tester.h",
4386 "src/xnnpack/AlignedAllocator.h",
4387 ] + MICROKERNEL_TEST_HDRS,
4388 deps = MICROKERNEL_TEST_DEPS,
4389)
4390
4391xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004392 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004394 "test/f32-gavgpool-cw.cc",
4395 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 "src/xnnpack/AlignedAllocator.h",
4397 ] + MICROKERNEL_TEST_HDRS,
4398 deps = MICROKERNEL_TEST_DEPS,
4399)
4400
4401xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004402 name = "f32_gemm_test",
4403 srcs = [
4404 "test/f32-gemm.cc",
4405 "test/gemm-microkernel-tester.h",
4406 "src/xnnpack/AlignedAllocator.h",
4407 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004408 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004409)
4410
4411xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004412 name = "f32_gemm_relu_test",
4413 srcs = [
4414 "test/f32-gemm-relu.cc",
4415 "test/gemm-microkernel-tester.h",
4416 "src/xnnpack/AlignedAllocator.h",
4417 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004418 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004419)
4420
4421xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004422 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004423 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004424 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004425 "test/gemm-microkernel-tester.h",
4426 "src/xnnpack/AlignedAllocator.h",
4427 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004428 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004429)
4430
4431xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004432 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004433 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004434 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004435 "test/gemm-microkernel-tester.h",
4436 "src/xnnpack/AlignedAllocator.h",
4437 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004438 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004439)
4440
4441xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004442 name = "f16_hswish_test",
4443 srcs = [
4444 "test/f16-hswish.cc",
4445 "test/hswish-microkernel-tester.h",
4446 ] + MICROKERNEL_TEST_HDRS,
4447 deps = MICROKERNEL_TEST_DEPS,
4448)
4449
4450xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 name = "f32_hswish_test",
4452 srcs = [
4453 "test/f32-hswish.cc",
4454 "test/hswish-microkernel-tester.h",
4455 ] + MICROKERNEL_TEST_HDRS,
4456 deps = MICROKERNEL_TEST_DEPS,
4457)
4458
4459xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004460 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004462 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004463 "test/maxpool-microkernel-tester.h",
4464 ] + MICROKERNEL_TEST_HDRS,
4465 deps = MICROKERNEL_TEST_DEPS,
4466)
4467
4468xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004469 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004471 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004472 "test/avgpool-microkernel-tester.h",
4473 "src/xnnpack/AlignedAllocator.h",
4474 ] + MICROKERNEL_TEST_HDRS,
4475 deps = MICROKERNEL_TEST_DEPS,
4476)
4477
4478xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004479 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004481 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004482 "test/gemm-microkernel-tester.h",
4483 "src/xnnpack/AlignedAllocator.h",
4484 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004485 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486)
4487
4488xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004489 name = "f16_prelu_test",
4490 srcs = [
4491 "test/f16-prelu.cc",
4492 "test/prelu-microkernel-tester.h",
4493 "src/xnnpack/AlignedAllocator.h",
4494 ] + MICROKERNEL_TEST_HDRS,
4495 deps = MICROKERNEL_TEST_DEPS,
4496)
4497
4498xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499 name = "f32_prelu_test",
4500 srcs = [
4501 "test/f32-prelu.cc",
4502 "test/prelu-microkernel-tester.h",
4503 "src/xnnpack/AlignedAllocator.h",
4504 ] + MICROKERNEL_TEST_HDRS,
4505 deps = MICROKERNEL_TEST_DEPS,
4506)
4507
4508xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004509 name = "f32_raddexpminusmax_test",
4510 srcs = [
4511 "test/f32-raddexpminusmax.cc",
4512 "test/raddexpminusmax-microkernel-tester.h",
4513 ] + MICROKERNEL_TEST_HDRS,
4514 deps = MICROKERNEL_TEST_DEPS,
4515)
4516
4517xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004518 name = "f32_raddextexp_test",
4519 srcs = [
4520 "test/f32-raddextexp.cc",
4521 "test/raddextexp-microkernel-tester.h",
4522 ] + MICROKERNEL_TEST_HDRS,
4523 deps = MICROKERNEL_TEST_DEPS,
4524)
4525
4526xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004527 name = "f32_raddstoreexpminusmax_test",
4528 srcs = [
4529 "test/f32-raddstoreexpminusmax.cc",
4530 "test/raddstoreexpminusmax-microkernel-tester.h",
4531 ] + MICROKERNEL_TEST_HDRS,
4532 deps = MICROKERNEL_TEST_DEPS,
4533)
4534
4535xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004536 name = "f32_relu_test",
4537 srcs = [
4538 "test/f32-relu.cc",
4539 "test/vunary-microkernel-tester.h",
4540 ] + MICROKERNEL_TEST_HDRS,
4541 deps = MICROKERNEL_TEST_DEPS,
4542)
4543
4544xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004545 name = "f32_rmax_test",
4546 srcs = [
4547 "test/f32-rmax.cc",
4548 "test/rmax-microkernel-tester.h",
4549 ] + MICROKERNEL_TEST_HDRS,
4550 deps = MICROKERNEL_TEST_DEPS,
4551)
4552
4553xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004554 name = "f32_sigmoid_test",
4555 srcs = [
4556 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004557 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004558 ] + MICROKERNEL_TEST_HDRS,
4559 deps = MICROKERNEL_TEST_DEPS,
4560)
4561
4562xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004563 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004564 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004565 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004566 "test/spmm-microkernel-tester.h",
4567 "src/xnnpack/AlignedAllocator.h",
4568 ] + MICROKERNEL_TEST_HDRS,
4569 deps = MICROKERNEL_TEST_DEPS,
4570)
4571
4572xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004573 name = "f16_vadd_minmax_test",
4574 srcs = [
4575 "test/f16-vadd-minmax.cc",
4576 "test/vbinary-microkernel-tester.h",
4577 ] + MICROKERNEL_TEST_HDRS,
4578 deps = MICROKERNEL_TEST_DEPS,
4579)
4580
4581xnnpack_unit_test(
4582 name = "f16_vaddc_minmax_test",
4583 srcs = [
4584 "test/f16-vaddc-minmax.cc",
4585 "test/vbinaryc-microkernel-tester.h",
4586 ] + MICROKERNEL_TEST_HDRS,
4587 deps = MICROKERNEL_TEST_DEPS,
4588)
4589
4590xnnpack_unit_test(
4591 name = "f16_vdiv_minmax_test",
4592 srcs = [
4593 "test/f16-vdiv-minmax.cc",
4594 "test/vbinary-microkernel-tester.h",
4595 ] + MICROKERNEL_TEST_HDRS,
4596 deps = MICROKERNEL_TEST_DEPS,
4597)
4598
4599xnnpack_unit_test(
4600 name = "f16_vdivc_minmax_test",
4601 srcs = [
4602 "test/f16-vdivc-minmax.cc",
4603 "test/vbinaryc-microkernel-tester.h",
4604 ] + MICROKERNEL_TEST_HDRS,
4605 deps = MICROKERNEL_TEST_DEPS,
4606)
4607
4608xnnpack_unit_test(
4609 name = "f16_vrdivc_minmax_test",
4610 srcs = [
4611 "test/f16-vrdivc-minmax.cc",
4612 "test/vbinaryc-microkernel-tester.h",
4613 ] + MICROKERNEL_TEST_HDRS,
4614 deps = MICROKERNEL_TEST_DEPS,
4615)
4616
4617xnnpack_unit_test(
4618 name = "f16_vmax_test",
4619 srcs = [
4620 "test/f16-vmax.cc",
4621 "test/vbinary-microkernel-tester.h",
4622 ] + MICROKERNEL_TEST_HDRS,
4623 deps = MICROKERNEL_TEST_DEPS,
4624)
4625
4626xnnpack_unit_test(
4627 name = "f16_vmaxc_test",
4628 srcs = [
4629 "test/f16-vmaxc.cc",
4630 "test/vbinaryc-microkernel-tester.h",
4631 ] + MICROKERNEL_TEST_HDRS,
4632 deps = MICROKERNEL_TEST_DEPS,
4633)
4634
4635xnnpack_unit_test(
4636 name = "f16_vmin_test",
4637 srcs = [
4638 "test/f16-vmin.cc",
4639 "test/vbinary-microkernel-tester.h",
4640 ] + MICROKERNEL_TEST_HDRS,
4641 deps = MICROKERNEL_TEST_DEPS,
4642)
4643
4644xnnpack_unit_test(
4645 name = "f16_vminc_test",
4646 srcs = [
4647 "test/f16-vminc.cc",
4648 "test/vbinaryc-microkernel-tester.h",
4649 ] + MICROKERNEL_TEST_HDRS,
4650 deps = MICROKERNEL_TEST_DEPS,
4651)
4652
4653xnnpack_unit_test(
4654 name = "f16_vmul_minmax_test",
4655 srcs = [
4656 "test/f16-vmul-minmax.cc",
4657 "test/vbinary-microkernel-tester.h",
4658 ] + MICROKERNEL_TEST_HDRS,
4659 deps = MICROKERNEL_TEST_DEPS,
4660)
4661
4662xnnpack_unit_test(
4663 name = "f16_vmulc_minmax_test",
4664 srcs = [
4665 "test/f16-vmulc-minmax.cc",
4666 "test/vbinaryc-microkernel-tester.h",
4667 ] + MICROKERNEL_TEST_HDRS,
4668 deps = MICROKERNEL_TEST_DEPS,
4669)
4670
4671xnnpack_unit_test(
4672 name = "f16_vsub_minmax_test",
4673 srcs = [
4674 "test/f16-vsub-minmax.cc",
4675 "test/vbinary-microkernel-tester.h",
4676 ] + MICROKERNEL_TEST_HDRS,
4677 deps = MICROKERNEL_TEST_DEPS,
4678)
4679
4680xnnpack_unit_test(
4681 name = "f16_vsubc_minmax_test",
4682 srcs = [
4683 "test/f16-vsubc-minmax.cc",
4684 "test/vbinaryc-microkernel-tester.h",
4685 ] + MICROKERNEL_TEST_HDRS,
4686 deps = MICROKERNEL_TEST_DEPS,
4687)
4688
4689xnnpack_unit_test(
4690 name = "f16_vrsubc_minmax_test",
4691 srcs = [
4692 "test/f16-vrsubc-minmax.cc",
4693 "test/vbinaryc-microkernel-tester.h",
4694 ] + MICROKERNEL_TEST_HDRS,
4695 deps = MICROKERNEL_TEST_DEPS,
4696)
4697
4698xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004699 name = "f32_vabs_test",
4700 srcs = [
4701 "test/f32-vabs.cc",
4702 "test/vunary-microkernel-tester.h",
4703 ] + MICROKERNEL_TEST_HDRS,
4704 deps = MICROKERNEL_TEST_DEPS,
4705)
4706
4707xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004708 name = "f32_vadd_test",
4709 srcs = [
4710 "test/f32-vadd.cc",
4711 "test/vbinary-microkernel-tester.h",
4712 ] + MICROKERNEL_TEST_HDRS,
4713 deps = MICROKERNEL_TEST_DEPS,
4714)
4715
4716xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004717 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004718 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004719 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004720 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004721 ] + MICROKERNEL_TEST_HDRS,
4722 deps = MICROKERNEL_TEST_DEPS,
4723)
4724
4725xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004726 name = "f32_vaddc_test",
4727 srcs = [
4728 "test/f32-vaddc.cc",
4729 "test/vbinaryc-microkernel-tester.h",
4730 ] + MICROKERNEL_TEST_HDRS,
4731 deps = MICROKERNEL_TEST_DEPS,
4732)
4733
4734xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004735 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004736 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004737 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004738 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004739 ] + MICROKERNEL_TEST_HDRS,
4740 deps = MICROKERNEL_TEST_DEPS,
4741)
4742
4743xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004744 name = "f32_vdiv_test",
4745 srcs = [
4746 "test/f32-vdiv.cc",
4747 "test/vbinary-microkernel-tester.h",
4748 ] + MICROKERNEL_TEST_HDRS,
4749 deps = MICROKERNEL_TEST_DEPS,
4750)
4751
4752xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004753 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004754 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004755 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004756 "test/vbinary-microkernel-tester.h",
4757 ] + MICROKERNEL_TEST_HDRS,
4758 deps = MICROKERNEL_TEST_DEPS,
4759)
4760
4761xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004762 name = "f32_vdivc_test",
4763 srcs = [
4764 "test/f32-vdivc.cc",
4765 "test/vbinaryc-microkernel-tester.h",
4766 ] + MICROKERNEL_TEST_HDRS,
4767 deps = MICROKERNEL_TEST_DEPS,
4768)
4769
4770xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004771 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004772 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004773 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004774 "test/vbinaryc-microkernel-tester.h",
4775 ] + MICROKERNEL_TEST_HDRS,
4776 deps = MICROKERNEL_TEST_DEPS,
4777)
4778
4779xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004780 name = "f32_vrdivc_test",
4781 srcs = [
4782 "test/f32-vrdivc.cc",
4783 "test/vbinaryc-microkernel-tester.h",
4784 ] + MICROKERNEL_TEST_HDRS,
4785 deps = MICROKERNEL_TEST_DEPS,
4786)
4787
4788xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004789 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004790 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004791 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004792 "test/vbinaryc-microkernel-tester.h",
4793 ] + MICROKERNEL_TEST_HDRS,
4794 deps = MICROKERNEL_TEST_DEPS,
4795)
4796
4797xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004798 name = "f32_vmax_test",
4799 srcs = [
4800 "test/f32-vmax.cc",
4801 "test/vbinary-microkernel-tester.h",
4802 ] + MICROKERNEL_TEST_HDRS,
4803 deps = MICROKERNEL_TEST_DEPS,
4804)
4805
4806xnnpack_unit_test(
4807 name = "f32_vmaxc_test",
4808 srcs = [
4809 "test/f32-vmaxc.cc",
4810 "test/vbinaryc-microkernel-tester.h",
4811 ] + MICROKERNEL_TEST_HDRS,
4812 deps = MICROKERNEL_TEST_DEPS,
4813)
4814
4815xnnpack_unit_test(
4816 name = "f32_vmin_test",
4817 srcs = [
4818 "test/f32-vmin.cc",
4819 "test/vbinary-microkernel-tester.h",
4820 ] + MICROKERNEL_TEST_HDRS,
4821 deps = MICROKERNEL_TEST_DEPS,
4822)
4823
4824xnnpack_unit_test(
4825 name = "f32_vminc_test",
4826 srcs = [
4827 "test/f32-vminc.cc",
4828 "test/vbinaryc-microkernel-tester.h",
4829 ] + MICROKERNEL_TEST_HDRS,
4830 deps = MICROKERNEL_TEST_DEPS,
4831)
4832
4833xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004834 name = "f32_vmul_test",
4835 srcs = [
4836 "test/f32-vmul.cc",
4837 "test/vbinary-microkernel-tester.h",
4838 ] + MICROKERNEL_TEST_HDRS,
4839 deps = MICROKERNEL_TEST_DEPS,
4840)
4841
4842xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004843 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004844 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004845 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004846 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004847 ] + MICROKERNEL_TEST_HDRS,
4848 deps = MICROKERNEL_TEST_DEPS,
4849)
4850
4851xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004852 name = "f32_vmulc_test",
4853 srcs = [
4854 "test/f32-vmulc.cc",
4855 "test/vbinaryc-microkernel-tester.h",
4856 ] + MICROKERNEL_TEST_HDRS,
4857 deps = MICROKERNEL_TEST_DEPS,
4858)
4859
4860xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004861 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004862 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004863 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004864 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004865 ] + MICROKERNEL_TEST_HDRS,
4866 deps = MICROKERNEL_TEST_DEPS,
4867)
4868
4869xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004870 name = "f16_vmulcaddc_minmax_test",
4871 srcs = [
4872 "test/f16-vmulcaddc-minmax.cc",
4873 "test/vmulcaddc-microkernel-tester.h",
4874 "src/xnnpack/AlignedAllocator.h",
4875 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004876 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004877)
4878
4879xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004880 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004881 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004882 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883 "test/vmulcaddc-microkernel-tester.h",
4884 "src/xnnpack/AlignedAllocator.h",
4885 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004886 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004887)
4888
4889xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004890 name = "f32_vlrelu_test",
4891 srcs = [
4892 "test/f32-vlrelu.cc",
4893 "test/vunary-microkernel-tester.h",
4894 ] + MICROKERNEL_TEST_HDRS,
4895 deps = MICROKERNEL_TEST_DEPS,
4896)
4897
4898xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004899 name = "f32_vneg_test",
4900 srcs = [
4901 "test/f32-vneg.cc",
4902 "test/vunary-microkernel-tester.h",
4903 ] + MICROKERNEL_TEST_HDRS,
4904 deps = MICROKERNEL_TEST_DEPS,
4905)
4906
4907xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004908 name = "f32_vrndne_test",
4909 srcs = [
4910 "test/f32-vrndne.cc",
4911 "test/vunary-microkernel-tester.h",
4912 ] + MICROKERNEL_TEST_HDRS,
4913 deps = MICROKERNEL_TEST_DEPS,
4914)
4915
4916xnnpack_unit_test(
4917 name = "f32_vrndz_test",
4918 srcs = [
4919 "test/f32-vrndz.cc",
4920 "test/vunary-microkernel-tester.h",
4921 ] + MICROKERNEL_TEST_HDRS,
4922 deps = MICROKERNEL_TEST_DEPS,
4923)
4924
4925xnnpack_unit_test(
4926 name = "f32_vrndu_test",
4927 srcs = [
4928 "test/f32-vrndu.cc",
4929 "test/vunary-microkernel-tester.h",
4930 ] + MICROKERNEL_TEST_HDRS,
4931 deps = MICROKERNEL_TEST_DEPS,
4932)
4933
4934xnnpack_unit_test(
4935 name = "f32_vrndd_test",
4936 srcs = [
4937 "test/f32-vrndd.cc",
4938 "test/vunary-microkernel-tester.h",
4939 ] + MICROKERNEL_TEST_HDRS,
4940 deps = MICROKERNEL_TEST_DEPS,
4941)
4942
4943xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004944 name = "f32_vscale_test",
4945 srcs = [
4946 "test/f32-vscale.cc",
4947 "test/vscale-microkernel-tester.h",
4948 ] + MICROKERNEL_TEST_HDRS,
4949 deps = MICROKERNEL_TEST_DEPS,
4950)
4951
4952xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004953 name = "f32_vscaleexpminusmax_test",
4954 srcs = [
4955 "test/f32-vscaleexpminusmax.cc",
4956 "test/vscaleexpminusmax-microkernel-tester.h",
4957 ] + MICROKERNEL_TEST_HDRS,
4958 deps = MICROKERNEL_TEST_DEPS,
4959)
4960
4961xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004962 name = "f32_vscaleextexp_test",
4963 srcs = [
4964 "test/f32-vscaleextexp.cc",
4965 "test/vscaleextexp-microkernel-tester.h",
4966 ] + MICROKERNEL_TEST_HDRS,
4967 deps = MICROKERNEL_TEST_DEPS,
4968)
4969
4970xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004971 name = "f32_vsqr_test",
4972 srcs = [
4973 "test/f32-vsqr.cc",
4974 "test/vunary-microkernel-tester.h",
4975 ] + MICROKERNEL_TEST_HDRS,
4976 deps = MICROKERNEL_TEST_DEPS,
4977)
4978
4979xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004980 name = "f32_vsqrdiff_test",
4981 srcs = [
4982 "test/f32-vsqrdiff.cc",
4983 "test/vbinary-microkernel-tester.h",
4984 ] + MICROKERNEL_TEST_HDRS,
4985 deps = MICROKERNEL_TEST_DEPS,
4986)
4987
4988xnnpack_unit_test(
4989 name = "f32_vsqrdiffc_test",
4990 srcs = [
4991 "test/f32-vsqrdiffc.cc",
4992 "test/vbinaryc-microkernel-tester.h",
4993 ] + MICROKERNEL_TEST_HDRS,
4994 deps = MICROKERNEL_TEST_DEPS,
4995)
4996
4997xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004998 name = "f32_vsqrt_test",
4999 srcs = [
5000 "test/f32-vsqrt.cc",
5001 "test/vunary-microkernel-tester.h",
5002 ] + MICROKERNEL_TEST_HDRS,
5003 deps = MICROKERNEL_TEST_DEPS,
5004)
5005
5006xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005007 name = "f32_vsub_test",
5008 srcs = [
5009 "test/f32-vsub.cc",
5010 "test/vbinary-microkernel-tester.h",
5011 ] + MICROKERNEL_TEST_HDRS,
5012 deps = MICROKERNEL_TEST_DEPS,
5013)
5014
5015xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005016 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005017 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005018 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005019 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005020 ] + MICROKERNEL_TEST_HDRS,
5021 deps = MICROKERNEL_TEST_DEPS,
5022)
5023
5024xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005025 name = "f32_vsubc_test",
5026 srcs = [
5027 "test/f32-vsubc.cc",
5028 "test/vbinaryc-microkernel-tester.h",
5029 ] + MICROKERNEL_TEST_HDRS,
5030 deps = MICROKERNEL_TEST_DEPS,
5031)
5032
5033xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005034 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005035 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005036 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005037 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005038 ] + MICROKERNEL_TEST_HDRS,
5039 deps = MICROKERNEL_TEST_DEPS,
5040)
5041
5042xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005043 name = "f32_vrsubc_test",
5044 srcs = [
5045 "test/f32-vrsubc.cc",
5046 "test/vbinaryc-microkernel-tester.h",
5047 ] + MICROKERNEL_TEST_HDRS,
5048 deps = MICROKERNEL_TEST_DEPS,
5049)
5050
5051xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005052 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005053 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005054 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005055 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005056 ] + MICROKERNEL_TEST_HDRS,
5057 deps = MICROKERNEL_TEST_DEPS,
5058)
5059
5060xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005061 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005062 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005063 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005064 "test/avgpool-microkernel-tester.h",
5065 "src/xnnpack/AlignedAllocator.h",
5066 ] + MICROKERNEL_TEST_HDRS,
5067 deps = MICROKERNEL_TEST_DEPS,
5068)
5069
5070xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005071 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005072 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005073 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005074 "test/gemm-microkernel-tester.h",
5075 "src/xnnpack/AlignedAllocator.h",
5076 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005077 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005078)
5079
5080xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005081 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005082 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005083 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084 "test/dwconv-microkernel-tester.h",
5085 "src/xnnpack/AlignedAllocator.h",
5086 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005087 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005088)
5089
5090xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005091 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005093 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005094 "test/gavgpool-microkernel-tester.h",
5095 "src/xnnpack/AlignedAllocator.h",
5096 ] + MICROKERNEL_TEST_HDRS,
5097 deps = MICROKERNEL_TEST_DEPS,
5098)
5099
5100xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005101 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005102 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005103 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005104 "test/gemm-microkernel-tester.h",
5105 "src/xnnpack/AlignedAllocator.h",
5106 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005107 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005108)
5109
5110xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005111 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005112 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005113 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005114 "test/vadd-microkernel-tester.h",
5115 ] + MICROKERNEL_TEST_HDRS,
5116 deps = MICROKERNEL_TEST_DEPS,
5117)
5118
5119xnnpack_unit_test(
5120 name = "u8_clamp_test",
5121 srcs = [
5122 "test/u8-clamp.cc",
5123 "test/clamp-microkernel-tester.h",
5124 ] + MICROKERNEL_TEST_HDRS,
5125 deps = MICROKERNEL_TEST_DEPS,
5126)
5127
5128xnnpack_unit_test(
5129 name = "u8_lut32norm_test",
5130 srcs = [
5131 "test/u8-lut32norm.cc",
5132 "test/lut-norm-microkernel-tester.h",
5133 ] + MICROKERNEL_TEST_HDRS,
5134 deps = MICROKERNEL_TEST_DEPS,
5135)
5136
5137xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005138 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005139 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005140 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005141 "test/maxpool-microkernel-tester.h",
5142 ] + MICROKERNEL_TEST_HDRS,
5143 deps = MICROKERNEL_TEST_DEPS,
5144)
5145
5146xnnpack_unit_test(
5147 name = "u8_rmax_test",
5148 srcs = [
5149 "test/u8-rmax.cc",
5150 "test/rmax-microkernel-tester.h",
5151 ] + MICROKERNEL_TEST_HDRS,
5152 deps = MICROKERNEL_TEST_DEPS,
5153)
5154
5155xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005156 name = "x32_fill_test",
5157 srcs = [
5158 "test/x32-fill.cc",
5159 "test/fill-microkernel-tester.h",
5160 ] + MICROKERNEL_TEST_HDRS,
5161 deps = MICROKERNEL_TEST_DEPS,
5162)
5163
5164xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165 name = "x32_packx_test",
5166 srcs = [
5167 "test/x32-packx.cc",
5168 "test/pack-microkernel-tester.h",
5169 "src/xnnpack/AlignedAllocator.h",
5170 ] + MICROKERNEL_TEST_HDRS,
5171 deps = MICROKERNEL_TEST_DEPS,
5172)
5173
5174xnnpack_unit_test(
5175 name = "x32_pad_test",
5176 srcs = [
5177 "test/x32-pad.cc",
5178 "test/pad-microkernel-tester.h",
5179 ] + MICROKERNEL_TEST_HDRS,
5180 deps = MICROKERNEL_TEST_DEPS,
5181)
5182
5183xnnpack_unit_test(
5184 name = "x32_unpool_test",
5185 srcs = [
5186 "test/x32-unpool.cc",
5187 "test/unpool-microkernel-tester.h",
5188 ] + MICROKERNEL_TEST_HDRS,
5189 deps = MICROKERNEL_TEST_DEPS,
5190)
5191
5192xnnpack_unit_test(
5193 name = "x32_zip_test",
5194 srcs = [
5195 "test/x32-zip.cc",
5196 "test/zip-microkernel-tester.h",
5197 ] + MICROKERNEL_TEST_HDRS,
5198 deps = MICROKERNEL_TEST_DEPS,
5199)
5200
5201xnnpack_unit_test(
5202 name = "x8_lut_test",
5203 srcs = [
5204 "test/x8-lut.cc",
5205 "test/lut-microkernel-tester.h",
5206 ] + MICROKERNEL_TEST_HDRS,
5207 deps = MICROKERNEL_TEST_DEPS,
5208)
5209
5210xnnpack_unit_test(
5211 name = "x8_zip_test",
5212 srcs = [
5213 "test/x8-zip.cc",
5214 "test/zip-microkernel-tester.h",
5215 ] + MICROKERNEL_TEST_HDRS,
5216 deps = MICROKERNEL_TEST_DEPS,
5217)
5218
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005219xnnpack_unit_test(
5220 name = "requantization_test",
5221 srcs = [
5222 "src/xnnpack/requantization-stubs.h",
5223 "test/requantization.cc",
5224 "test/requantization-tester.h",
5225 ] + MICROKERNEL_TEST_HDRS,
5226 deps = MICROKERNEL_TEST_DEPS,
5227)
5228
Marat Dukhan20c3b922020-03-10 03:45:06 -07005229########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230
5231xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005232 name = "operator_size_test",
5233 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005234 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235)
5236
Marat Dukhan20c3b922020-03-10 03:45:06 -07005237xnnpack_binary(
5238 name = "subgraph_size_test",
5239 srcs = ["test/subgraph-size.c"],
5240 deps = [":XNNPACK"],
5241)
5242
5243########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244
5245xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005246 name = "abs_nc_test",
5247 srcs = [
5248 "test/abs-nc.cc",
5249 "test/abs-operator-tester.h",
5250 ],
5251 deps = OPERATOR_TEST_DEPS,
5252)
5253
5254xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005255 name = "add_nd_test",
5256 srcs = [
5257 "test/add-nd.cc",
5258 "test/binary-elementwise-operator-tester.h",
5259 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005260 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005261)
5262
5263xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005264 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005266 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267 "test/argmax-pooling-operator-tester.h",
5268 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005269 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005270)
5271
5272xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005273 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005274 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005275 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005276 "test/average-pooling-operator-tester.h",
5277 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005278 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005279)
5280
5281xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005282 name = "bankers_rounding_nc_test",
5283 srcs = [
5284 "test/bankers-rounding-nc.cc",
5285 "test/bankers-rounding-operator-tester.h",
5286 ],
5287 deps = OPERATOR_TEST_DEPS,
5288)
5289
5290xnnpack_unit_test(
5291 name = "ceiling_nc_test",
5292 srcs = [
5293 "test/ceiling-nc.cc",
5294 "test/ceiling-operator-tester.h",
5295 ],
5296 deps = OPERATOR_TEST_DEPS,
5297)
5298
5299xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005300 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005301 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005302 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303 "test/channel-shuffle-operator-tester.h",
5304 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005305 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306)
5307
5308xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005309 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005311 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312 "test/clamp-operator-tester.h",
5313 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005314 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005315)
5316
5317xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005318 name = "constant_pad_nd_test",
5319 srcs = [
5320 "test/constant-pad-nd.cc",
5321 "test/constant-pad-operator-tester.h",
5322 ],
5323 deps = OPERATOR_TEST_DEPS,
5324)
5325
5326xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005327 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005328 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005329 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005330 "test/convolution-operator-tester.h",
5331 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005332 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333)
5334
5335xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005336 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005337 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005338 "test/convolution-nchw.cc",
5339 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005340 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005341 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005342)
5343
5344xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005345 name = "copy_nc_test",
5346 srcs = [
5347 "test/copy-nc.cc",
5348 "test/copy-operator-tester.h",
5349 ],
5350 deps = OPERATOR_TEST_DEPS,
5351)
5352
5353xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005354 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005355 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005356 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005357 "test/deconvolution-operator-tester.h",
5358 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005359 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005360)
5361
5362xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005363 name = "divide_nd_test",
5364 srcs = [
5365 "test/binary-elementwise-operator-tester.h",
5366 "test/divide-nd.cc",
5367 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005368 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005369)
5370
5371xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005372 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005373 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005374 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005375 "test/fully-connected-operator-tester.h",
5376 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005377 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005378)
5379
5380xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005381 name = "floor_nc_test",
5382 srcs = [
5383 "test/floor-nc.cc",
5384 "test/floor-operator-tester.h",
5385 ],
5386 deps = OPERATOR_TEST_DEPS,
5387)
5388
5389xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005390 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005391 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005392 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005393 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005394 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005395 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005396)
5397
5398xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005399 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005400 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005401 "test/global-average-pooling-ncw.cc",
5402 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005404 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005405)
5406
5407xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005408 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005409 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005410 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 "test/hardswish-operator-tester.h",
5412 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005413 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005414)
5415
5416xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005417 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005418 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005419 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005420 "test/leaky-relu-operator-tester.h",
5421 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005422 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005423)
5424
5425xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005426 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005427 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005428 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005429 "test/max-pooling-operator-tester.h",
5430 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005431 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005432)
5433
5434xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005435 name = "maximum_nd_test",
5436 srcs = [
5437 "test/binary-elementwise-operator-tester.h",
5438 "test/maximum-nd.cc",
5439 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005440 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005441)
5442
5443xnnpack_unit_test(
5444 name = "minimum_nd_test",
5445 srcs = [
5446 "test/binary-elementwise-operator-tester.h",
5447 "test/minimum-nd.cc",
5448 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005449 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005450)
5451
5452xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005453 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005454 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005455 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005456 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005457 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005458 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005459)
5460
5461xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005462 name = "negate_nc_test",
5463 srcs = [
5464 "test/negate-nc.cc",
5465 "test/negate-operator-tester.h",
5466 ],
5467 deps = OPERATOR_TEST_DEPS,
5468)
5469
5470xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005471 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005472 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005473 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005474 "test/prelu-operator-tester.h",
5475 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005476 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005477)
5478
5479xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005480 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005481 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005482 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005483 "test/resize-bilinear-operator-tester.h",
5484 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005485 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005486)
5487
5488xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005489 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005491 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005492 "test/sigmoid-operator-tester.h",
5493 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005494 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005495)
5496
5497xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005498 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005499 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005500 "test/softmax-nc.cc",
5501 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005502 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005503 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005504)
5505
5506xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005507 name = "square_nc_test",
5508 srcs = [
5509 "test/square-nc.cc",
5510 "test/square-operator-tester.h",
5511 ],
5512 deps = OPERATOR_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005516 name = "square_root_nc_test",
5517 srcs = [
5518 "test/square-root-nc.cc",
5519 "test/square-root-operator-tester.h",
5520 ],
5521 deps = OPERATOR_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005525 name = "squared_difference_nd_test",
5526 srcs = [
5527 "test/binary-elementwise-operator-tester.h",
5528 "test/squared-difference-nd.cc",
5529 ],
5530 deps = OPERATOR_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005534 name = "subtract_nd_test",
5535 srcs = [
5536 "test/binary-elementwise-operator-tester.h",
5537 "test/subtract-nd.cc",
5538 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005539 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005540)
5541
5542xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005543 name = "truncation_nc_test",
5544 srcs = [
5545 "test/truncation-nc.cc",
5546 "test/truncation-operator-tester.h",
5547 ],
5548 deps = OPERATOR_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005552 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005553 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005554 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005555 "test/unpooling-operator-tester.h",
5556 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005557 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005558)
5559
Chao Mei6ddfc602020-05-13 22:29:36 -07005560############################### Misc unit tests ###############################
5561
5562xnnpack_unit_test(
5563 name = "memory_planner_test",
5564 srcs = [
5565 "test/memory-planner-test.cc",
5566 ],
5567 deps = [
5568 ":XNNPACK",
5569 ":memory_planner",
5570 ],
5571)
5572
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573############################# Build configurations #############################
5574
Marat Dukhanb8642352019-10-30 15:43:02 -07005575# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005576config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005577 name = "xnn_enable_assembly_explicit_true",
5578 define_values = {"xnn_enable_assembly": "true"},
5579)
5580
5581# Disables usage of assembly kernels.
5582config_setting(
5583 name = "xnn_enable_assembly_explicit_false",
5584 define_values = {"xnn_enable_assembly": "false"},
5585)
5586
Marat Dukhan9de90e02020-06-18 16:04:12 -07005587# Enables usage of sparse inference.
5588config_setting(
5589 name = "xnn_enable_sparse_explicit_true",
5590 define_values = {"xnn_enable_sparse": "true"},
5591)
5592
5593# Disables usage of sparse inference.
5594config_setting(
5595 name = "xnn_enable_sparse_explicit_false",
5596 define_values = {"xnn_enable_sparse": "false"},
5597)
5598
Marat Dukhan05702cf2020-03-26 15:41:33 -07005599# Disables usage of HMP-aware optimizations.
5600config_setting(
5601 name = "xnn_enable_hmp_explicit_false",
5602 define_values = {"xnn_enable_hmp": "false"},
5603)
5604
Chao Mei6ddfc602020-05-13 22:29:36 -07005605# Enable usage of optimized memory allocation
5606config_setting(
5607 name = "xnn_enable_memopt_explicit_true",
5608 define_values = {"xnnpack_enable_memopt": "true"},
5609)
5610
5611# Disable usage of optimized memory allocation
5612config_setting(
5613 name = "xnn_enable_memopt_explicit_false",
5614 define_values = {"xnnpack_enable_memopt": "false"},
5615)
5616
Marat Dukhanb8642352019-10-30 15:43:02 -07005617# Builds with -c dbg
5618config_setting(
5619 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005620 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005621 "compilation_mode": "dbg",
5622 },
5623)
5624
5625# Builds with -c opt
5626config_setting(
5627 name = "optimized_build",
5628 values = {
5629 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005630 },
5631)
5632
5633config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005634 name = "linux_k8",
5635 values = {"cpu": "k8"},
5636)
5637
5638config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005639 name = "linux_arm",
5640 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005641)
5642
5643config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005644 name = "linux_armeabi",
5645 values = {"cpu": "armeabi"},
5646)
5647
5648config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005649 name = "linux_armhf",
5650 values = {"cpu": "armhf"},
5651)
5652
5653config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005654 name = "linux_armv7a",
5655 values = {"cpu": "armv7a"},
5656)
5657
5658config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005659 name = "linux_aarch64",
5660 values = {"cpu": "aarch64"},
5661)
5662
5663config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005664 name = "android",
5665 values = {"crosstool_top": "//external:android/crosstool"},
5666)
5667
5668config_setting(
5669 name = "android_armv7",
5670 values = {
5671 "crosstool_top": "//external:android/crosstool",
5672 "cpu": "armeabi-v7a",
5673 },
5674)
5675
5676config_setting(
5677 name = "android_arm64",
5678 values = {
5679 "crosstool_top": "//external:android/crosstool",
5680 "cpu": "arm64-v8a",
5681 },
5682)
5683
5684config_setting(
5685 name = "android_x86",
5686 values = {
5687 "crosstool_top": "//external:android/crosstool",
5688 "cpu": "x86",
5689 },
5690)
5691
5692config_setting(
5693 name = "android_x86_64",
5694 values = {
5695 "crosstool_top": "//external:android/crosstool",
5696 "cpu": "x86_64",
5697 },
5698)
5699
5700config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005701 name = "windows_x86_64",
5702 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005703)
5704
5705config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005706 name = "windows_x86_64_clang",
5707 values = {
5708 "compiler": "clang-cl",
5709 "cpu": "x64_windows",
5710 },
5711)
5712
5713config_setting(
5714 name = "windows_x86_64_mingw",
5715 values = {
5716 "compiler": "mingw-gcc",
5717 "cpu": "x64_windows",
5718 },
5719)
5720
5721config_setting(
5722 name = "windows_x86_64_msys",
5723 values = {
5724 "compiler": "msys-gcc",
5725 "cpu": "x64_windows",
5726 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005727)
5728
5729config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005730 name = "macos_x86_64",
5731 values = {
5732 "apple_platform_type": "macos",
5733 "cpu": "darwin",
5734 },
5735)
5736
5737config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005738 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005739 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005740)
5741
5742config_setting(
5743 name = "emscripten_wasm",
5744 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005745 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005746 "cpu": "wasm",
5747 },
5748)
5749
5750config_setting(
5751 name = "emscripten_wasmsimd",
5752 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005753 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005754 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005755 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005756 },
5757)
5758
5759config_setting(
5760 name = "emscripten_asmjs",
5761 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005762 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005763 "cpu": "asmjs",
5764 },
5765)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005766
5767config_setting(
5768 name = "ios_armv7",
5769 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005770 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005771 "cpu": "ios_armv7",
5772 },
5773)
5774
5775config_setting(
5776 name = "ios_arm64",
5777 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005778 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005779 "cpu": "ios_arm64",
5780 },
5781)
5782
5783config_setting(
5784 name = "ios_arm64e",
5785 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005786 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005787 "cpu": "ios_arm64e",
5788 },
5789)
5790
5791config_setting(
5792 name = "ios_x86",
5793 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005794 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005795 "cpu": "ios_i386",
5796 },
5797)
5798
5799config_setting(
5800 name = "ios_x86_64",
5801 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005802 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005803 "cpu": "ios_x86_64",
5804 },
5805)
5806
5807config_setting(
5808 name = "watchos_armv7k",
5809 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005810 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005811 "cpu": "watchos_armv7k",
5812 },
5813)
5814
5815config_setting(
5816 name = "watchos_arm64_32",
5817 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005818 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005819 "cpu": "watchos_arm64_32",
5820 },
5821)
5822
5823config_setting(
5824 name = "watchos_x86",
5825 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005826 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005827 "cpu": "watchos_i386",
5828 },
5829)
5830
5831config_setting(
5832 name = "watchos_x86_64",
5833 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005834 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005835 "cpu": "watchos_x86_64",
5836 },
5837)
5838
5839config_setting(
5840 name = "tvos_arm64",
5841 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005842 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005843 "cpu": "tvos_arm64",
5844 },
5845)
5846
5847config_setting(
5848 name = "tvos_x86_64",
5849 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005850 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005851 "cpu": "tvos_x86_64",
5852 },
5853)