blob: 28bf3295b5b77cf62c831b04a4fb208024573d21 [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 Dukhan3b7432d2020-07-16 17:46:32 -0700516 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
517 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
518 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
519 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700520 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
521 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
522 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
523 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700524 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
525 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
526 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
527 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
528 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
529 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700530 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
531 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
532 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
533 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
534 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
535 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
536 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
537 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
538 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
539 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
540 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
541 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
542 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
543 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
544 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
545 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
546 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
547 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
548 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
549 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
550 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
551 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
552 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
553 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700554 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
555 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
556 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700557 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700558 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
559 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
560 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
561 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700562 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
563 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
564 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
565 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
566 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
567 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
568 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
569 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
570 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
571 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
572 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
573 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
574 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
578 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
579 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
580 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
581 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
582 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
583 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
584 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
585 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
586 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
587 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
588 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
589 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700592 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
593 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700594 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
595 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
596 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
597 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
598 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
599 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
600 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
601 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
602 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
603 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
604 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
605 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
606 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
610 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
611 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
612 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
613 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
614 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
615 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
616 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
617 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
618 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
619 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
620 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
621 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700624 "src/f32-ibilinear/gen/wasmsimd-c4.c",
625 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700626 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
627 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
628 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700629 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700630 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
631 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
632 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
633 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700634 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
635 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
636 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
637 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
638 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
639 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
640 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
641 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
642 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
643 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
644 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
645 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
646 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
650 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
651 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
652 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
653 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
654 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
655 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
656 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
657 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
658 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
659 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
660 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
661 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700664 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
665 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700666 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
667 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
668 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
669 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700670 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
671 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
672 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700673 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
674 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700675 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
676 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
677 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
678 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700679 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
680 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
681 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
682 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700683 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
684 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
685 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
686 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
687 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
688 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
689 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
690 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
691 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
692 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
693 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
694 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700695 "src/f32-relu/gen/wasmsimd-x4.c",
696 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700697 "src/f32-rmax/wasmsimd-arm.c",
698 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700699 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
700 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
701 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
702 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
703 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
704 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700705 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
706 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
707 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
708 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700709 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
710 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700711 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
712 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
713 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
714 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700715 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
716 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700717 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
718 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
719 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
720 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700721 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
722 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700723 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
724 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
725 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
726 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700727 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
728 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700729 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
730 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
731 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
732 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700733 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
734 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
735 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
736 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700737 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
738 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
739 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
740 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700741 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700745 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
746 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
747 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
748 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700749 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
750 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700751 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
752 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
753 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
754 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700755 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
756 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700757 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
758 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
759 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
760 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700761 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
762 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700763 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
764 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
765 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
766 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700767 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
768 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700769 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
770 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700771 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
772 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700773 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700777 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
778 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700779 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
780 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
781 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
782 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700783 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
784 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700785 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
786 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
787 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
788 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700789 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
790 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
791 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
792 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700793 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
794 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
795 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
796 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
797 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
798 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
799 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
800 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
801 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
802 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
803 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
804 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
805 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
806 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700807 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
808 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700809 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
810 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
811 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
812 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
813 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
814 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700815 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700816 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700817 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700818 "src/x32-unpool/wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700819 "src/math/roundne-wasmsimd-addsub.c",
820 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700821 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700822 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700823 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700824 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700825 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700826 "src/requantization/fp32-wasmsimd.c",
827]
828
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800829PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800830 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800831 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800832 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700833 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
834 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700835 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700836 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700837 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700838 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700839 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700840 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
841 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
842 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
843 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
844 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
845 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
846 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
847 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
848 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
849 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
850 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
851 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700852 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700853 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
854 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700855 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
856 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
857 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
858 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
859 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
860 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
861 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
862 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
863 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
864 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
865 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
866 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
867 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
868 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
869 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
870 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
871 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
872 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
873 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700874 "src/f32-ibilinear/gen/psimd-c4.c",
875 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700876 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
877 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
878 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
879 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
880 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
881 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
882 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
883 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
884 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
885 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700886 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700887 "src/f32-spmm/gen/4x1-minmax-psimd.c",
888 "src/f32-spmm/gen/8x1-minmax-psimd.c",
889 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700890 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
891 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700892 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800893 "src/f32-prelu/gen/psimd-2x4.c",
894 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800895 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700896 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
897 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
898 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
899 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
900 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
901 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
902 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
903 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800904 "src/f32-vbinary/gen/vmax-psimd-x4.c",
905 "src/f32-vbinary/gen/vmax-psimd-x8.c",
906 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
907 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
908 "src/f32-vbinary/gen/vmin-psimd-x4.c",
909 "src/f32-vbinary/gen/vmin-psimd-x8.c",
910 "src/f32-vbinary/gen/vminc-psimd-x4.c",
911 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700912 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
913 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
914 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
915 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
916 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
917 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
918 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
919 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700920 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
921 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
922 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
923 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700924 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
925 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
926 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
927 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700928 "src/f32-vlrelu/gen/vlrelu-psimd-x4.c",
929 "src/f32-vlrelu/gen/vlrelu-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700930 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
931 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700932 "src/f32-vunary/gen/vabs-psimd-x4.c",
933 "src/f32-vunary/gen/vabs-psimd-x8.c",
934 "src/f32-vunary/gen/vneg-psimd-x4.c",
935 "src/f32-vunary/gen/vneg-psimd-x8.c",
936 "src/f32-vunary/gen/vsqr-psimd-x4.c",
937 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700938 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700939 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700940 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700941 "src/x32-unpool/psimd.c",
942 "src/x32-zip/x2-psimd.c",
943 "src/x32-zip/x3-psimd.c",
944 "src/x32-zip/x4-psimd.c",
945 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700946 "src/requantization/precise-psimd.c",
947 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700948]
949
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800950PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800951 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
952 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
953 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
954 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
955 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
956 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
957 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
958 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
959 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
960 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
961 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
962 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800963 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
964 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
965 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
966 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
967 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
968 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700969 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
970 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
971 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
972 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
973 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
974 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
975 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
976 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700977 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700978 "src/math/roundd-psimd-addsub.c",
979 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700980 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800981 "src/math/sigmoid-psimd-p5-div.c",
982]
983
Marat Dukhan08c4a432019-10-03 09:29:21 -0700984# ISA-specific micro-kernels
985NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700986 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
987 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700988 "src/f32-clamp/gen/neon-x4.c",
989 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700990 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
991 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700992 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
993 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700994 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
995 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700996 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
997 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700998 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
999 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1000 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1001 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001002 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1003 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1004 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1005 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001006 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1007 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1008 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1009 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001010 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001011 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1012 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001013 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1014 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1015 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1016 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1017 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1018 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1019 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1020 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1021 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1022 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1023 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1024 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1025 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1026 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1027 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1028 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1029 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1030 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1031 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1032 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1033 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1034 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1035 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1036 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1037 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1038 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1039 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1040 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1041 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1042 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1043 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001044 "src/f32-hswish/gen/hswish-neon-x4.c",
1045 "src/f32-hswish/gen/hswish-neon-x8.c",
1046 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001047 "src/f32-ibilinear/gen/neon-c4.c",
1048 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001049 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1050 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1051 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1052 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1053 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1054 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1055 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1056 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1057 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1058 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1059 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1060 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1061 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1062 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1063 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1064 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001065 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1066 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1067 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001068 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1069 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001070 "src/f32-prelu/gen/neon-2x4.c",
1071 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001072 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1073 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1074 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1075 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1076 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1077 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1078 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1079 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1080 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1081 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1082 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1083 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1084 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1085 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1086 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1087 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1088 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1090 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1091 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1092 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1093 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1094 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1095 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001096 "src/f32-relu/gen/neon-x4.c",
1097 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001098 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001099 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001100 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1101 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1102 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1103 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1104 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1105 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1106 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1107 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1108 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1109 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1110 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1111 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1112 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1113 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1114 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1115 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1116 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1117 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001118 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1119 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1120 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1121 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001122 "src/f32-vbinary/gen/vmax-neon-x4.c",
1123 "src/f32-vbinary/gen/vmax-neon-x8.c",
1124 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1125 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1126 "src/f32-vbinary/gen/vmin-neon-x4.c",
1127 "src/f32-vbinary/gen/vmin-neon-x8.c",
1128 "src/f32-vbinary/gen/vminc-neon-x4.c",
1129 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001130 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1131 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1132 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1133 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1134 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1135 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001136 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1137 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1138 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1139 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001140 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1141 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1142 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1143 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001144 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1145 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001146 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1147 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001148 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1149 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1150 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1151 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1152 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1153 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1154 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1155 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001156 "src/f32-vunary/gen/vabs-neon-x4.c",
1157 "src/f32-vunary/gen/vabs-neon-x8.c",
1158 "src/f32-vunary/gen/vneg-neon-x4.c",
1159 "src/f32-vunary/gen/vneg-neon-x8.c",
1160 "src/f32-vunary/gen/vsqr-neon-x4.c",
1161 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001162 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1163 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1164 "src/qu8-dwconv/up8x9-minmax-neon.c",
1165 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1166 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1167 "src/qu8-gemm/4x8-minmax-neon.c",
1168 "src/qu8-gemm/8x8-minmax-neon.c",
1169 "src/qu8-igemm/4x8-minmax-neon.c",
1170 "src/qu8-igemm/8x8-minmax-neon.c",
1171 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001172 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001173 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001174 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001175 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001176 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001177 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001178 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001179 "src/x32-zip/x2-neon.c",
1180 "src/x32-zip/x3-neon.c",
1181 "src/x32-zip/x4-neon.c",
1182 "src/x32-zip/xm-neon.c",
1183 "src/x8-zip/x2-neon.c",
1184 "src/x8-zip/x3-neon.c",
1185 "src/x8-zip/x4-neon.c",
1186 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001187 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001188 "src/math/roundd-neon-addsub.c",
1189 "src/math/roundd-neon-cvt.c",
1190 "src/math/roundu-neon-addsub.c",
1191 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001192 "src/math/roundz-neon-addsub.c",
1193 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001194 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001195 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1196 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1197 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1198 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1199 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1200 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001201 "src/math/sqrt-neon-nr1rsqrts.c",
1202 "src/math/sqrt-neon-nr2rsqrts.c",
1203 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001204 "src/requantization/precise-neon.c",
1205 "src/requantization/fp32-neon.c",
1206 "src/requantization/q31-neon.c",
1207 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001208]
1209
1210NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001211 "src/f32-ibilinear/gen/neonfma-c4.c",
1212 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001213 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1214 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1215 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1216 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1217 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1218 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1219 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1220 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1221 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001222 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1223 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1224 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1225 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001226 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1227 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1228 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1229 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001230 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1231 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1232 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1233 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001234 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1235 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1236 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1237 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1238 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1239 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1240 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1241 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1242 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1243 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1244 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1245 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1246 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1247 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1248 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1249 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1250 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1251 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001252 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1253 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001254 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1255 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1256 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1257 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1258 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1259 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1260 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1261 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1262 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1263 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1264 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1265 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1266 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1267 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1268 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1269 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1270 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1271 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1272 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1273 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1274 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1275 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1276 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1277 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001278 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1279 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1280 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1281 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1282 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1283 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1284 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1285 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1286 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1287 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1288 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1289 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1290 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1291 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1292 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1293 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1294 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1295 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1296 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1297 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1298 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1299 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1300 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1301 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1302 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1303 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1304 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1305 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1306 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1307 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1308 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1309 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1310 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1311 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1312 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1313 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1314 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1315 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1316 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1318 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1319 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1320 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1321 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1322 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1323 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1325 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1326 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1327 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001332 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1333 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001334 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1335 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1336 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1337 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1338 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1339 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1340 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1341 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1342 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1343 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1344 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1345 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1346 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1347 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1348 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1349 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1350 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1351 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1352 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1353 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001354 "src/math/exp-neonfma-lut64-p2.c",
1355 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001356 "src/math/expminus-neonfma-lut2048-p1.c",
1357 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001358 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001359 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1360 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1361 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1362 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1363 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1364 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1365 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1366 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1367 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1368 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1369 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1370 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1371 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1372 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1373 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1374 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1375 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1376 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001377 "src/math/sqrt-neonfma-nr1fma.c",
1378 "src/math/sqrt-neonfma-nr2fma.c",
1379 "src/math/sqrt-neonfma-nr3fma.c",
1380 "src/math/sqrt-neonfma-nr2fma1adj.c",
1381 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001382]
1383
1384AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001385 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1386 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1387 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1388 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1389 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1390 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001391 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1392 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1393 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1394 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1395 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1396 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1397 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1398 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1399 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1400 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1401 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1402 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1403 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1404 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1405 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1406 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1407 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1408 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1409 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1410 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001411 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1412 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001413 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1414 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001415 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1416 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001417 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1418 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001419 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1420 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1421 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1422 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1423 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001424 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1425 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1426 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1427 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1428 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1429 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1430 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1431 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1432 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1433 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1434 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1435 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1436 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1437 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1438 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1439 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1440 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1441 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001442 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1443 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1444 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1445 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1446 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1447 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1448 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1449 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1450 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1451 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1452 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1453 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1454 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1455 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1456 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1457 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1458 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1459 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001460 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1461 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001462 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1463 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1464 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1465 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1466 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1467 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001468]
1469
Marat Dukhan8853b822020-05-07 12:19:01 -07001470NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001471 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1472 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1473 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1474 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1475 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1476 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1477 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1478 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001479 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001480 "src/math/roundd-neonv8.c",
1481 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001482 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001483]
1484
Marat Dukhan08c4a432019-10-03 09:29:21 -07001485AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001486 "src/f16-clamp/gen/neonfp16arith-x16.c",
1487 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001488 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1489 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1490 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1491 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1492 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1493 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1494 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1495 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1496 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1497 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1498 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1499 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001500 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1501 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001502 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1503 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001504 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1505 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001506 "src/f16-relu/gen/neonfp16arith-x16.c",
1507 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001508 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1509 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1510 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1511 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1512 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1513 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1514 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1515 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1516 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1517 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1518 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1519 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1520 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1521 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1522 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1523 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1524 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1525 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1526 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1527 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1528 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1529 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1530 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1531 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1532 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1533 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1534 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1535 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001536 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1537 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001538 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1539 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1540 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1541 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1542 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1543 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1544 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1545 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001546 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1547 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1548 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1549 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001550 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1551 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1552 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1553 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1554 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1555 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1556 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1557 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1558 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1559 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1560 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1561 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001562 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1563 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1564 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1565 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1566 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1567 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1568 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1569 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001570]
1571
1572SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001573 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1574 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001575 "src/f32-clamp/gen/sse-x4.c",
1576 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001577 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1578 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001579 "src/f32-dwconv-chw/3x3p1-sse.c",
1580 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001581 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1582 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1583 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1584 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1585 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1586 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1587 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1588 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1589 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1590 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1591 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1592 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001593 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001594 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1595 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001596 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1597 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1598 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1599 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1600 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1601 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1602 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1603 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1604 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1605 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1606 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1607 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1608 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001609 "src/f32-hswish/gen/hswish-sse-x4.c",
1610 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001611 "src/f32-ibilinear/gen/sse-c4.c",
1612 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001613 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1614 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1615 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1616 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1617 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1618 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1619 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001620 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1621 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1622 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001623 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001624 "src/f32-prelu/gen/sse-2x4.c",
1625 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001626 "src/f32-relu/gen/sse-x4.c",
1627 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001628 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001629 "src/f32-spmm/gen/4x1-minmax-sse.c",
1630 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001631 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001632 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1633 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1634 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1635 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1636 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1637 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1638 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1639 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001640 "src/f32-vbinary/gen/vmax-sse-x4.c",
1641 "src/f32-vbinary/gen/vmax-sse-x8.c",
1642 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1643 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1644 "src/f32-vbinary/gen/vmin-sse-x4.c",
1645 "src/f32-vbinary/gen/vmin-sse-x8.c",
1646 "src/f32-vbinary/gen/vminc-sse-x4.c",
1647 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001648 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1649 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1650 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1651 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1652 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1653 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1654 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1655 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001656 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1657 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1658 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1659 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001660 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1661 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1662 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1663 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001664 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1665 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001666 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1667 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001668 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1669 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001670 "src/f32-vunary/gen/vabs-sse-x4.c",
1671 "src/f32-vunary/gen/vabs-sse-x8.c",
1672 "src/f32-vunary/gen/vneg-sse-x4.c",
1673 "src/f32-vunary/gen/vneg-sse-x8.c",
1674 "src/f32-vunary/gen/vsqr-sse-x4.c",
1675 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001676 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001677 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001678 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001679 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001680 "src/math/roundd-sse-addsub.c",
1681 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001682 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001683 "src/math/sqrt-sse-hh1mac.c",
1684 "src/math/sqrt-sse-nr1mac.c",
1685 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001686]
1687
1688SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001689 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1690 "src/f32-argmaxpool/4x-sse2-c4.c",
1691 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001692 "src/f32-prelu/gen/sse2-2x4.c",
1693 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001694 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1695 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1696 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1697 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1698 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1699 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1700 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1701 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1702 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1703 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1704 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1705 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001706 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001707 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001708 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001709 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001710 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1711 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001712 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1713 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001714 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1715 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1716 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1717 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1718 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1719 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1720 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1721 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001722 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1723 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1724 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1725 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1726 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1727 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1728 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1729 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1730 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001731 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001732 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001733 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001734 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001735 "src/x32-zip/x2-sse2.c",
1736 "src/x32-zip/x3-sse2.c",
1737 "src/x32-zip/x4-sse2.c",
1738 "src/x32-zip/xm-sse2.c",
1739 "src/x8-zip/x2-sse2.c",
1740 "src/x8-zip/x3-sse2.c",
1741 "src/x8-zip/x4-sse2.c",
1742 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001743 "src/math/exp-sse2-p5.c",
1744 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001745 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001746 "src/math/roundd-sse2-cvt.c",
1747 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001748 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001749 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001750 "src/requantization/precise-sse2.c",
1751 "src/requantization/fp32-sse2.c",
1752 "src/requantization/q31-sse2.c",
1753 "src/requantization/gemmlowp-sse2.c",
1754]
1755
1756SSSE3_UKERNELS = [
1757 "src/requantization/precise-ssse3.c",
1758 "src/requantization/q31-ssse3.c",
1759 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001760]
1761
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001762SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001763 "src/f32-prelu/gen/sse41-2x4.c",
1764 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001765 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001766 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001767 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001768 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001769 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1770 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001771 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1772 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001773 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1774 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1775 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1776 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1777 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1778 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1779 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1780 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001781 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001782 "src/math/roundd-sse41.c",
1783 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001784 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001785 "src/requantization/precise-sse4.c",
1786 "src/requantization/q31-sse4.c",
1787 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001788]
1789
Marat Dukhan08c4a432019-10-03 09:29:21 -07001790AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001791 "src/f32-clamp/gen/avx-x8.c",
1792 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001793 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1794 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1795 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1796 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1797 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1798 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1799 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1800 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1801 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1802 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1803 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1804 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1805 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1806 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1807 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1808 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1809 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1810 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1811 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1812 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1813 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1814 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1815 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1816 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1817 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1818 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1819 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1820 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1821 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1822 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001823 "src/f32-hswish/gen/hswish-avx-x8.c",
1824 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001825 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1826 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1827 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1828 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1829 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1830 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1831 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1832 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1833 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001834 "src/f32-prelu/gen/avx-2x8.c",
1835 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001836 "src/f32-relu/gen/avx-x16.c",
1837 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001839 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1840 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1841 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1842 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1843 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1844 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1845 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1846 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001847 "src/f32-vbinary/gen/vmax-avx-x8.c",
1848 "src/f32-vbinary/gen/vmax-avx-x16.c",
1849 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1850 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1851 "src/f32-vbinary/gen/vmin-avx-x8.c",
1852 "src/f32-vbinary/gen/vmin-avx-x16.c",
1853 "src/f32-vbinary/gen/vminc-avx-x8.c",
1854 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001855 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1856 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1857 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1858 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1859 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1860 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1861 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1862 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001863 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1864 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1865 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1866 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001867 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1868 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1869 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1870 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001871 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1872 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001873 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001874 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1875 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1876 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1877 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1878 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1879 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1880 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1881 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001882 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1883 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001884 "src/f32-vunary/gen/vabs-avx-x8.c",
1885 "src/f32-vunary/gen/vabs-avx-x16.c",
1886 "src/f32-vunary/gen/vneg-avx-x8.c",
1887 "src/f32-vunary/gen/vneg-avx-x16.c",
1888 "src/f32-vunary/gen/vsqr-avx-x8.c",
1889 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001890]
1891
Marat Dukhanfda12b82019-11-21 12:27:59 -08001892FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001893 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1894 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1895 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1896 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1897 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1898 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1899 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1900 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1901 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1902 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1903 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1904 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1905 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1906 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1907 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1908 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1909 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1910 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1911 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1912 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1913 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1914 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1915 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1916 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1917 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1918 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1919 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1920 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1921 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1922 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1923 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1924 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1925 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1926 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1927 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1928 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1929 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1930 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1931 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1932 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001933 "src/f32-hswish/gen/hswish-fma3-x8.c",
1934 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001935 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1936 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1937 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1938 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1939 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1940 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1941 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1942 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1943 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1944 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1945 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1946 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1947 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1948 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001949 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1950 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1951 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1952 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1953 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1954 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1955 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1956 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001957 "src/math/sqrt-fma3-nr1fma.c",
1958 "src/math/sqrt-fma3-nr2fma.c",
1959 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001960]
1961
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001962AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001963 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1964 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1965 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1966 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1967 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1968 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1969 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1970 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1971 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1972 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1973 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1974 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1975 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1976 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1977 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1978 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1979 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1980 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1981 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1982 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1983 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1984 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1985 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1986 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1987 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1988 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1989 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1990 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1991 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1992 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1993 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1994 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1995 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1996 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1997 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1998 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001999 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2000 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2001 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2002 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2003 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2004 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2005 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2006 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2007 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2008 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2009 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2010 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2011 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2012 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2013 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2014 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2015 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2016 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2017 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2018 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2019 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2020 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2021 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2022 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2023 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2024 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2025 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2026 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2027 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2028 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002029 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2030 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2031 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2032 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2033 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2034 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2035 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2036 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2037 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2038 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2039 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2040 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2041 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2042 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2043 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2044 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2045 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2046 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2047 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2048 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2049 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2050 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2051 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2052 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002053 "src/math/exp-avx2-p5.c",
2054 "src/math/exp-avx2-perm-p3.c",
2055 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002056 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002057 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002058 "src/math/sigmoid-avx2-rr2-p5-div.c",
2059 "src/math/sigmoid-avx2-rr1-p5-div.c",
2060 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2061 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2062 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2063 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002064]
2065
Marat Dukhan08c4a432019-10-03 09:29:21 -07002066AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002067 "src/f32-clamp/gen/avx512f-x16.c",
2068 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002069 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2070 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2071 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2072 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2073 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2074 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2075 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2076 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2077 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2078 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2079 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2080 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2081 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2082 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2083 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2084 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2085 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2086 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2087 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2088 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2089 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2090 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2091 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2092 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002093 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2094 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002095 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2096 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2097 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2098 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2099 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2100 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002101 "src/f32-prelu/gen/avx512f-2x16.c",
2102 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002103 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2104 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2105 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2106 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2107 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2108 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2109 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2110 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2111 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2112 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2113 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2114 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2115 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2116 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2117 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2118 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2119 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2120 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2121 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2122 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2123 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2124 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2125 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2126 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2127 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2128 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2129 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2130 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2131 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2132 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2133 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2134 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2135 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2136 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2137 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2138 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002139 "src/f32-relu/gen/avx512f-x16.c",
2140 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002141 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002142 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2143 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2144 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2145 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2146 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2147 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2148 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2149 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002150 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2151 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2152 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2153 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2154 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2155 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2156 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2157 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002158 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2159 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2160 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2161 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2162 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2163 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2164 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2165 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002166 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2167 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2168 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2169 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002170 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2171 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2172 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2173 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002174 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2175 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002176 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002177 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2178 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2179 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2180 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2181 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2182 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2183 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2184 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2185 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2186 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2187 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2188 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2189 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2190 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2191 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2192 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2193 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2194 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2195 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2196 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2197 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2198 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2199 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2200 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002201 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2202 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2203 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2204 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2205 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2206 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2207 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2208 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002209 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2210 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2211 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2212 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2213 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2214 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2215 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2216 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002217 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2218 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2219 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2220 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2221 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2222 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002223 "src/math/exp-avx512f-p5-scalef.c",
2224 "src/math/exp-avx512f-p5.c",
2225 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002226 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002227 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002228 "src/math/sqrt-avx512f-nr1fma.c",
2229 "src/math/sqrt-avx512f-nr2fma.c",
2230 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002231]
2232
2233AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002234 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002235 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002236 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2237 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002238 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002239 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002240 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002241 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002242 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2243 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002244 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2245 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2246 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2247 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002248]
2249
2250AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002251 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002252 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2253 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002254 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002255 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2256 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002257 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2258 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2259 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002260 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002261 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2262 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2263 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002264 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002265 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2266 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002267 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002268 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2269 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2270 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2271 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2272 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2273 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2274 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2275 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2276 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2277 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2278 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2279 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2280 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2281 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2282 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2283 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2284 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2285 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002286 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2287 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002288 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002289 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2290 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2291 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2292 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2293 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2294 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2295 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2296 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2297 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2298 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2299 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2300 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2301 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2302 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2303 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2304 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2305 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2306 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002307 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2308 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2309 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2310 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2311 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2312 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2313 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2314 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2315 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2316 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2317 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2318 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2319 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2320 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2321 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2322 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2323 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2324 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325]
2326
Marat Dukhan1b354632020-03-23 12:50:22 -07002327INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002328 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002329 "src/xnnpack/argmaxpool.h",
2330 "src/xnnpack/avgpool.h",
2331 "src/xnnpack/clamp.h",
2332 "src/xnnpack/common.h",
2333 "src/xnnpack/conv.h",
2334 "src/xnnpack/dwconv.h",
2335 "src/xnnpack/gavgpool.h",
2336 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002337 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002338 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002339 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002341 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002342 "src/xnnpack/lut.h",
2343 "src/xnnpack/math.h",
2344 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002345 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002346 "src/xnnpack/packx.h",
2347 "src/xnnpack/pad.h",
2348 "src/xnnpack/params.h",
2349 "src/xnnpack/pavgpool.h",
2350 "src/xnnpack/ppmm.h",
2351 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002352 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002353 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002354 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002355 "src/xnnpack/rmax.h",
2356 "src/xnnpack/scalar-utils.h",
2357 "src/xnnpack/spmm.h",
2358 "src/xnnpack/unpool.h",
2359 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002360 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002361 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002362 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002363 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002364 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002365 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002366 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002367]
2368
2369INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002370 "include/xnnpack.h",
2371 "src/xnnpack/allocator.h",
2372 "src/xnnpack/compute.h",
2373 "src/xnnpack/im2col.h",
2374 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002375 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002376 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002377 "src/xnnpack/operator.h",
2378 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002379 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002380 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002381 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002382 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002383]
2384
Marat Dukhan1b354632020-03-23 12:50:22 -07002385ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002386 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002387]
2388
Marat Dukhan1b354632020-03-23 12:50:22 -07002389MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002390 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391 "include/xnnpack.h",
2392]
2393
Marat Dukhan1b354632020-03-23 12:50:22 -07002394MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002395 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002396 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002397 "src/xnnpack/requantization.h",
2398 "include/xnnpack.h",
2399]
2400
2401OPERATOR_TEST_PARAMS_HDRS = [
2402 "src/xnnpack/params.h",
2403 "src/xnnpack/common.h",
2404]
2405
2406WEIGHTS_PACK_HDRS = [
2407 "src/xnnpack/pack.h",
2408 "src/xnnpack/operator.h",
2409 "src/xnnpack/compute.h",
2410]
2411
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002412LOGGING_COPTS = select({
2413 # No logging in optimized mode
2414 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2415 # Full logging in debug mode
2416 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2417 # Error-only logging in default (fastbuild) mode
2418 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2419})
2420
Marat Dukhan3b59de22020-06-03 20:15:19 -07002421LOGGING_SRCS = select({
2422 # No logging in optimized mode
2423 ":optimized_build": [],
2424 "//conditions:default": [
2425 "src/operator-strings.c",
2426 "src/subgraph-strings.c",
2427 ],
2428})
2429
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002430LOGGING_HDRS = [
2431 "src/xnnpack/log.h",
2432]
2433
Marat Dukhan08c4a432019-10-03 09:29:21 -07002434xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002435 name = "tables",
2436 srcs = TABLE_SRCS,
2437 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002438 gcc_copts = xnnpack_gcc_std_copts(),
2439 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002440)
2441
2442xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002443 name = "scalar_ukernels",
2444 srcs = SCALAR_UKERNELS,
2445 hdrs = INTERNAL_HDRS,
2446 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002447 gcc_copts = xnnpack_gcc_std_copts(),
2448 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002449 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002450 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002451 "@FP16",
2452 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002453 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002454 ],
2455)
2456
2457xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002458 name = "scalar_ukernels_test_mode",
2459 srcs = SCALAR_UKERNELS,
2460 hdrs = INTERNAL_HDRS,
2461 aarch32_copts = ["-marm"],
2462 copts = [
2463 "-UNDEBUG",
2464 "-DXNN_TEST_MODE=1",
2465 ],
2466 gcc_copts = xnnpack_gcc_std_copts(),
2467 msvc_copts = xnnpack_msvc_std_copts(),
2468 deps = [
2469 ":tables",
2470 "@FP16",
2471 "@FXdiv",
2472 "@pthreadpool",
2473 ],
2474)
2475
2476xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002477 name = "wasm_ukernels",
2478 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002479 gcc_copts = xnnpack_gcc_std_copts(),
2480 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002481 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002482 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002483 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002484 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002485 "@FP16",
2486 "@FXdiv",
2487 "@pthreadpool",
2488 ],
2489)
2490
2491xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002492 name = "wasm_ukernels_test_mode",
2493 hdrs = INTERNAL_HDRS,
2494 copts = [
2495 "-UNDEBUG",
2496 "-DXNN_TEST_MODE=1",
2497 ],
2498 gcc_copts = xnnpack_gcc_std_copts(),
2499 msvc_copts = xnnpack_msvc_std_copts(),
2500 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002501 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002502 deps = [
2503 ":tables",
2504 "@FP16",
2505 "@FXdiv",
2506 "@pthreadpool",
2507 ],
2508)
2509
2510xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002511 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002512 hdrs = INTERNAL_HDRS,
2513 aarch32_copts = [
2514 "-marm",
2515 "-mfpu=neon",
2516 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002517 gcc_copts = xnnpack_gcc_std_copts(),
2518 gcc_x86_copts = ["-msse2"],
2519 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002520 optimized_copts = [
2521 "-O3",
2522 "-ffast-math",
2523 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002524 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002525 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002526 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002527 "@FP16",
2528 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002529 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002530 ],
2531)
2532
2533xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002534 name = "psimd_fastmath_ukernels_test_mode",
2535 hdrs = INTERNAL_HDRS,
2536 aarch32_copts = [
2537 "-marm",
2538 "-mfpu=neon",
2539 ],
2540 copts = [
2541 "-UNDEBUG",
2542 "-DXNN_TEST_MODE=1",
2543 ],
2544 gcc_copts = xnnpack_gcc_std_copts(),
2545 gcc_x86_copts = ["-msse2"],
2546 msvc_copts = xnnpack_msvc_std_copts(),
2547 optimized_copts = [
2548 "-O3",
2549 "-ffast-math",
2550 ],
2551 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2552 deps = [
2553 ":tables",
2554 "@FP16",
2555 "@psimd",
2556 "@pthreadpool",
2557 ],
2558)
2559
2560xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002561 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002562 hdrs = INTERNAL_HDRS,
2563 aarch32_copts = [
2564 "-marm",
2565 "-mfpu=neon",
2566 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002567 gcc_copts = xnnpack_gcc_std_copts(),
2568 gcc_x86_copts = ["-msse2"],
2569 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002570 optimized_copts = [
2571 "-O3",
2572 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002573 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002574 deps = [
2575 ":tables",
2576 "@FP16",
2577 "@psimd",
2578 "@pthreadpool",
2579 ],
2580)
2581
2582xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002583 name = "psimd_accmath_ukernels_test_mode",
2584 hdrs = INTERNAL_HDRS,
2585 aarch32_copts = [
2586 "-marm",
2587 "-mfpu=neon",
2588 ],
2589 copts = [
2590 "-UNDEBUG",
2591 "-DXNN_TEST_MODE=1",
2592 ],
2593 gcc_copts = xnnpack_gcc_std_copts(),
2594 gcc_x86_copts = ["-msse2"],
2595 msvc_copts = xnnpack_msvc_std_copts(),
2596 optimized_copts = [
2597 "-O3",
2598 ],
2599 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2600 deps = [
2601 ":tables",
2602 "@FP16",
2603 "@psimd",
2604 "@pthreadpool",
2605 ],
2606)
2607
2608xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002609 name = "neon_ukernels",
2610 hdrs = INTERNAL_HDRS,
2611 aarch32_copts = [
2612 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002613 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002614 "-mfpu=neon",
2615 ],
2616 aarch32_srcs = NEON_UKERNELS,
2617 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002618 gcc_copts = xnnpack_gcc_std_copts(),
2619 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002620 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002621 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002622 "@FP16",
2623 "@pthreadpool",
2624 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002625)
2626
2627xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002628 name = "neon_ukernels_test_mode",
2629 hdrs = INTERNAL_HDRS,
2630 aarch32_copts = [
2631 "-marm",
2632 "-march=armv7-a",
2633 "-mfpu=neon",
2634 ],
2635 aarch32_srcs = NEON_UKERNELS,
2636 aarch64_srcs = NEON_UKERNELS,
2637 copts = [
2638 "-UNDEBUG",
2639 "-DXNN_TEST_MODE=1",
2640 ],
2641 gcc_copts = xnnpack_gcc_std_copts(),
2642 msvc_copts = xnnpack_msvc_std_copts(),
2643 deps = [
2644 ":tables",
2645 "@FP16",
2646 "@pthreadpool",
2647 ],
2648)
2649
2650xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002651 name = "neonfma_ukernels",
2652 hdrs = INTERNAL_HDRS,
2653 aarch32_copts = [
2654 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002655 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002656 "-mfpu=neon-vfpv4",
2657 ],
2658 aarch32_srcs = NEONFMA_UKERNELS,
2659 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002660 apple_aarch32_copts = [
2661 "-mcpu=swift",
2662 "-mtune=generic",
2663 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002664 gcc_copts = xnnpack_gcc_std_copts(),
2665 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002666 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002667 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002668 "@FP16",
2669 "@pthreadpool",
2670 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002671)
2672
2673xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002674 name = "neonfma_ukernels_test_mode",
2675 hdrs = INTERNAL_HDRS,
2676 aarch32_copts = [
2677 "-marm",
2678 "-march=armv7-a",
2679 "-mfpu=neon-vfpv4",
2680 ],
2681 aarch32_srcs = NEONFMA_UKERNELS,
2682 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002683 apple_aarch32_copts = [
2684 "-mcpu=swift",
2685 "-mtune=generic",
2686 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002687 copts = [
2688 "-UNDEBUG",
2689 "-DXNN_TEST_MODE=1",
2690 ],
2691 gcc_copts = xnnpack_gcc_std_copts(),
2692 msvc_copts = xnnpack_msvc_std_copts(),
2693 deps = [
2694 ":tables",
2695 "@FP16",
2696 "@pthreadpool",
2697 ],
2698)
2699
2700xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002701 name = "neonv8_ukernels",
2702 hdrs = INTERNAL_HDRS,
2703 aarch32_copts = [
2704 "-marm",
2705 "-march=armv8-a",
2706 "-mfpu=neon-fp-armv8",
2707 ],
2708 aarch32_srcs = NEONV8_UKERNELS,
2709 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002710 apple_aarch32_copts = [
2711 "-mcpu=cyclone",
2712 "-mtune=generic",
2713 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002714 gcc_copts = xnnpack_gcc_std_copts(),
2715 msvc_copts = xnnpack_msvc_std_copts(),
2716 deps = [
2717 ":tables",
2718 "@FP16",
2719 "@pthreadpool",
2720 ],
2721)
2722
2723xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002724 name = "neonv8_ukernels_test_mode",
2725 hdrs = INTERNAL_HDRS,
2726 aarch32_copts = [
2727 "-marm",
2728 "-march=armv8-a",
2729 "-mfpu=neon-fp-armv8",
2730 ],
2731 aarch32_srcs = NEONV8_UKERNELS,
2732 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002733 apple_aarch32_copts = [
2734 "-mcpu=cyclone",
2735 "-mtune=generic",
2736 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002737 copts = [
2738 "-UNDEBUG",
2739 "-DXNN_TEST_MODE=1",
2740 ],
2741 gcc_copts = xnnpack_gcc_std_copts(),
2742 msvc_copts = xnnpack_msvc_std_copts(),
2743 deps = [
2744 ":tables",
2745 "@FP16",
2746 "@pthreadpool",
2747 ],
2748)
2749
2750xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002751 name = "neonfp16arith_ukernels",
2752 hdrs = INTERNAL_HDRS,
2753 aarch64_copts = ["-march=armv8.2-a+fp16"],
2754 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002755 gcc_copts = xnnpack_gcc_std_copts(),
2756 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002757 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002758 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002759 "@FP16",
2760 "@pthreadpool",
2761 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002762)
2763
2764xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002765 name = "neonfp16arith_ukernels_test_mode",
2766 hdrs = INTERNAL_HDRS,
2767 aarch64_copts = ["-march=armv8.2-a+fp16"],
2768 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2769 copts = [
2770 "-UNDEBUG",
2771 "-DXNN_TEST_MODE=1",
2772 ],
2773 gcc_copts = xnnpack_gcc_std_copts(),
2774 msvc_copts = xnnpack_msvc_std_copts(),
2775 deps = [
2776 ":tables",
2777 "@FP16",
2778 "@pthreadpool",
2779 ],
2780)
2781
2782xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002783 name = "sse2_ukernels",
2784 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002785 gcc_copts = xnnpack_gcc_std_copts(),
2786 gcc_x86_copts = ["-msse2"],
2787 msvc_copts = xnnpack_msvc_std_copts(),
2788 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002789 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002790 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002791 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002792 "@FP16",
2793 "@pthreadpool",
2794 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002795)
2796
2797xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002798 name = "sse2_ukernels_test_mode",
2799 hdrs = INTERNAL_HDRS,
2800 copts = [
2801 "-UNDEBUG",
2802 "-DXNN_TEST_MODE=1",
2803 ],
2804 gcc_copts = xnnpack_gcc_std_copts(),
2805 gcc_x86_copts = ["-msse2"],
2806 msvc_copts = xnnpack_msvc_std_copts(),
2807 msvc_x86_32_copts = ["/arch:SSE2"],
2808 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2809 deps = [
2810 ":tables",
2811 "@FP16",
2812 "@pthreadpool",
2813 ],
2814)
2815
2816xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002817 name = "ssse3_ukernels",
2818 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002819 gcc_copts = xnnpack_gcc_std_copts(),
2820 gcc_x86_copts = ["-mssse3"],
2821 msvc_copts = xnnpack_msvc_std_copts(),
2822 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002823 x86_srcs = SSSE3_UKERNELS,
2824 deps = [
2825 ":tables",
2826 "@FP16",
2827 "@pthreadpool",
2828 ],
2829)
2830
2831xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002832 name = "ssse3_ukernels_test_mode",
2833 hdrs = INTERNAL_HDRS,
2834 copts = [
2835 "-UNDEBUG",
2836 "-DXNN_TEST_MODE=1",
2837 ],
2838 gcc_copts = xnnpack_gcc_std_copts(),
2839 gcc_x86_copts = ["-mssse3"],
2840 msvc_copts = xnnpack_msvc_std_copts(),
2841 msvc_x86_32_copts = ["/arch:SSE2"],
2842 x86_srcs = SSSE3_UKERNELS,
2843 deps = [
2844 ":tables",
2845 "@FP16",
2846 "@pthreadpool",
2847 ],
2848)
2849
2850xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002851 name = "sse41_ukernels",
2852 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002853 gcc_copts = xnnpack_gcc_std_copts(),
2854 gcc_x86_copts = ["-msse4.1"],
2855 msvc_copts = xnnpack_msvc_std_copts(),
2856 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002857 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002858 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002859 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002860 "@FP16",
2861 "@pthreadpool",
2862 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002863)
2864
2865xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002866 name = "sse41_ukernels_test_mode",
2867 hdrs = INTERNAL_HDRS,
2868 copts = [
2869 "-UNDEBUG",
2870 "-DXNN_TEST_MODE=1",
2871 ],
2872 gcc_copts = xnnpack_gcc_std_copts(),
2873 gcc_x86_copts = ["-msse4.1"],
2874 msvc_copts = xnnpack_msvc_std_copts(),
2875 msvc_x86_32_copts = ["/arch:SSE2"],
2876 x86_srcs = SSE41_UKERNELS,
2877 deps = [
2878 ":tables",
2879 "@FP16",
2880 "@pthreadpool",
2881 ],
2882)
2883
2884xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 name = "avx_ukernels",
2886 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002887 gcc_copts = xnnpack_gcc_std_copts(),
2888 gcc_x86_copts = ["-mavx"],
2889 msvc_copts = xnnpack_msvc_std_copts(),
2890 msvc_x86_32_copts = ["/arch:AVX"],
2891 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002892 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002893 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002894 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002895 "@FP16",
2896 "@pthreadpool",
2897 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898)
2899
2900xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002901 name = "avx_ukernels_test_mode",
2902 hdrs = INTERNAL_HDRS,
2903 copts = [
2904 "-UNDEBUG",
2905 "-DXNN_TEST_MODE=1",
2906 ],
2907 gcc_copts = xnnpack_gcc_std_copts(),
2908 gcc_x86_copts = ["-mavx"],
2909 msvc_copts = xnnpack_msvc_std_copts(),
2910 msvc_x86_32_copts = ["/arch:AVX"],
2911 msvc_x86_64_copts = ["/arch:AVX"],
2912 x86_srcs = AVX_UKERNELS,
2913 deps = [
2914 ":tables",
2915 "@FP16",
2916 "@pthreadpool",
2917 ],
2918)
2919
2920xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002921 name = "fma3_ukernels",
2922 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002923 gcc_copts = xnnpack_gcc_std_copts(),
2924 gcc_x86_copts = ["-mfma"],
2925 msvc_copts = xnnpack_msvc_std_copts(),
2926 msvc_x86_32_copts = ["/arch:AVX"],
2927 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002928 x86_srcs = FMA3_UKERNELS,
2929 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002930 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002931 "@FP16",
2932 "@pthreadpool",
2933 ],
2934)
2935
2936xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002937 name = "fma3_ukernels_test_mode",
2938 hdrs = INTERNAL_HDRS,
2939 copts = [
2940 "-UNDEBUG",
2941 "-DXNN_TEST_MODE=1",
2942 ],
2943 gcc_copts = xnnpack_gcc_std_copts(),
2944 gcc_x86_copts = ["-mfma"],
2945 msvc_copts = xnnpack_msvc_std_copts(),
2946 msvc_x86_32_copts = ["/arch:AVX"],
2947 msvc_x86_64_copts = ["/arch:AVX"],
2948 x86_srcs = FMA3_UKERNELS,
2949 deps = [
2950 ":tables",
2951 "@FP16",
2952 "@pthreadpool",
2953 ],
2954)
2955
2956xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002957 name = "avx2_ukernels",
2958 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002959 gcc_copts = xnnpack_gcc_std_copts(),
2960 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002961 "-mfma",
2962 "-mavx2",
2963 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002964 msvc_copts = xnnpack_msvc_std_copts(),
2965 msvc_x86_32_copts = ["/arch:AVX2"],
2966 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002967 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002968 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002969 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002970 "@FP16",
2971 "@pthreadpool",
2972 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002973)
2974
2975xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002976 name = "avx2_ukernels_test_mode",
2977 hdrs = INTERNAL_HDRS,
2978 copts = [
2979 "-UNDEBUG",
2980 "-DXNN_TEST_MODE=1",
2981 ],
2982 gcc_copts = xnnpack_gcc_std_copts(),
2983 gcc_x86_copts = [
2984 "-mfma",
2985 "-mavx2",
2986 ],
2987 msvc_copts = xnnpack_msvc_std_copts(),
2988 msvc_x86_32_copts = ["/arch:AVX2"],
2989 msvc_x86_64_copts = ["/arch:AVX2"],
2990 x86_srcs = AVX2_UKERNELS,
2991 deps = [
2992 ":tables",
2993 "@FP16",
2994 "@pthreadpool",
2995 ],
2996)
2997
2998xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 name = "avx512f_ukernels",
3000 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003001 gcc_copts = xnnpack_gcc_std_copts(),
3002 gcc_x86_copts = ["-mavx512f"],
3003 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3004 msvc_copts = xnnpack_msvc_std_copts(),
3005 msvc_x86_32_copts = ["/arch:AVX512"],
3006 msvc_x86_64_copts = ["/arch:AVX512"],
3007 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003009 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003010 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003011 "@FP16",
3012 "@pthreadpool",
3013 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003014)
3015
3016xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003017 name = "avx512f_ukernels_test_mode",
3018 hdrs = INTERNAL_HDRS,
3019 copts = [
3020 "-UNDEBUG",
3021 "-DXNN_TEST_MODE=1",
3022 ],
3023 gcc_copts = xnnpack_gcc_std_copts(),
3024 gcc_x86_copts = ["-mavx512f"],
3025 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3026 msvc_copts = xnnpack_msvc_std_copts(),
3027 msvc_x86_32_copts = ["/arch:AVX512"],
3028 msvc_x86_64_copts = ["/arch:AVX512"],
3029 msys_copts = ["-fno-asynchronous-unwind-tables"],
3030 x86_srcs = AVX512F_UKERNELS,
3031 deps = [
3032 ":tables",
3033 "@FP16",
3034 "@pthreadpool",
3035 ],
3036)
3037
3038xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039 name = "asm_ukernels",
3040 hdrs = ["src/xnnpack/assembly.h"],
3041 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003042 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043 aarch64_srcs = AARCH64_ASM_UKERNELS,
3044)
3045
Marat Dukhan3b59de22020-06-03 20:15:19 -07003046xnnpack_cc_library(
3047 name = "logging_utils",
3048 srcs = LOGGING_SRCS,
3049 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3050 copts = LOGGING_COPTS + [
3051 "-Isrc",
3052 "-Iinclude",
3053 ] + select({
3054 ":debug_build": [],
3055 "//conditions:default": xnnpack_min_size_copts(),
3056 }),
3057 gcc_copts = xnnpack_gcc_std_copts(),
3058 msvc_copts = xnnpack_msvc_std_copts(),
3059 visibility = xnnpack_visibility(),
3060 deps = [
3061 "@FP16",
3062 "@clog",
3063 "@pthreadpool",
3064 ],
3065)
3066
Marat Dukhan08c4a432019-10-03 09:29:21 -07003067xnnpack_aggregate_library(
3068 name = "ukernels",
3069 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003070 ":neon_ukernels",
3071 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003072 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003073 ":asm_ukernels",
3074 ],
3075 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003076 ":neon_ukernels",
3077 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003078 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003079 ":neonfp16arith_ukernels",
3080 ":asm_ukernels",
3081 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003082 generic_deps = [
3083 ":scalar_ukernels",
3084 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003085 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003086 ":psimd_fastmath_ukernels",
3087 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003088 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003089 wasm_deps = [
3090 ":wasm_ukernels",
3091 ],
3092 wasmsimd_deps = [
3093 ":wasm_ukernels",
3094 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003096 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003097 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003098 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003099 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003100 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003101 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003102 ":avx512f_ukernels",
3103 ],
3104)
3105
Marat Dukhan33fcf782020-05-24 14:27:15 -07003106xnnpack_aggregate_library(
3107 name = "ukernels_test_mode",
3108 aarch32_deps = [
3109 ":neon_ukernels_test_mode",
3110 ":neonfma_ukernels_test_mode",
3111 ":neonv8_ukernels_test_mode",
3112 ":asm_ukernels",
3113 ],
3114 aarch64_deps = [
3115 ":neon_ukernels_test_mode",
3116 ":neonfma_ukernels_test_mode",
3117 ":neonv8_ukernels_test_mode",
3118 ":neonfp16arith_ukernels_test_mode",
3119 ":asm_ukernels",
3120 ],
3121 generic_deps = [
3122 ":scalar_ukernels_test_mode",
3123 ],
3124 psimd_deps = [
3125 ":psimd_fastmath_ukernels_test_mode",
3126 ":psimd_accmath_ukernels_test_mode",
3127 ],
3128 wasm_deps = [
3129 ":wasm_ukernels_test_mode",
3130 ],
3131 wasmsimd_deps = [
3132 ":wasm_ukernels_test_mode",
3133 ],
3134 x86_deps = [
3135 ":sse2_ukernels_test_mode",
3136 ":ssse3_ukernels_test_mode",
3137 ":sse41_ukernels_test_mode",
3138 ":avx_ukernels_test_mode",
3139 ":fma3_ukernels_test_mode",
3140 ":avx2_ukernels_test_mode",
3141 ":avx512f_ukernels_test_mode",
3142 ],
3143)
3144
Marat Dukhan86df3842020-06-17 14:03:31 -07003145# This target must be compatible with Android NDK r17 and compile with Clang 6.
3146# It is similar to :ukernels target, but excludes microkernels which require
3147# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003148xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003149 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003150 aarch32_deps = [
3151 ":neon_ukernels",
3152 ":neonfma_ukernels",
3153 ":neonv8_ukernels",
3154 ":asm_ukernels",
3155 ],
3156 aarch64_deps = [
3157 ":neon_ukernels",
3158 ":neonfma_ukernels",
3159 ":neonv8_ukernels",
3160 ":asm_ukernels",
3161 ],
3162 generic_deps = [
3163 ":scalar_ukernels",
3164 ],
3165 psimd_deps = [
3166 ":psimd_fastmath_ukernels",
3167 ":psimd_accmath_ukernels",
3168 ],
3169 wasm_deps = [
3170 ":wasm_ukernels",
3171 ],
3172 wasmsimd_deps = [
3173 ":wasm_ukernels",
3174 ],
3175 x86_deps = [
3176 ":sse2_ukernels",
3177 ":ssse3_ukernels",
3178 ":sse41_ukernels",
3179 ":avx_ukernels",
3180 ":fma3_ukernels",
3181 ":avx2_ukernels",
3182 ":avx512f_ukernels",
3183 ],
3184)
3185
Marat Dukhan08c4a432019-10-03 09:29:21 -07003186xnnpack_cc_library(
3187 name = "im2col",
3188 srcs = ["src/im2col.c"],
3189 hdrs = [
3190 "src/xnnpack/common.h",
3191 "src/xnnpack/im2col.h",
3192 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003193 gcc_copts = xnnpack_gcc_std_copts(),
3194 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003195)
3196
3197xnnpack_cc_library(
3198 name = "indirection",
3199 srcs = ["src/indirection.c"],
3200 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003201 gcc_copts = xnnpack_gcc_std_copts(),
3202 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 deps = [
3204 "@FP16",
3205 "@FXdiv",
3206 "@pthreadpool",
3207 ],
3208)
3209
3210xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003211 name = "indirection_test_mode",
3212 srcs = ["src/indirection.c"],
3213 hdrs = INTERNAL_HDRS,
3214 copts = [
3215 "-UNDEBUG",
3216 "-DXNN_TEST_MODE=1",
3217 ],
3218 gcc_copts = xnnpack_gcc_std_copts(),
3219 msvc_copts = xnnpack_msvc_std_copts(),
3220 deps = [
3221 "@FP16",
3222 "@FXdiv",
3223 "@pthreadpool",
3224 ],
3225)
3226
3227xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003228 name = "packing",
3229 srcs = ["src/packing.c"],
3230 hdrs = INTERNAL_HDRS,
3231 gcc_copts = xnnpack_gcc_std_copts(),
3232 msvc_copts = xnnpack_msvc_std_copts(),
3233 deps = [
3234 "@FP16",
3235 "@FXdiv",
3236 "@pthreadpool",
3237 ],
3238)
3239
3240xnnpack_cc_library(
3241 name = "packing_test_mode",
3242 srcs = ["src/packing.c"],
3243 hdrs = INTERNAL_HDRS,
3244 copts = [
3245 "-UNDEBUG",
3246 "-DXNN_TEST_MODE=1",
3247 ],
3248 gcc_copts = xnnpack_gcc_std_copts(),
3249 msvc_copts = xnnpack_msvc_std_copts(),
3250 deps = [
3251 "@FP16",
3252 "@FXdiv",
3253 "@pthreadpool",
3254 ],
3255)
3256
3257xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003258 name = "operator_run",
3259 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003260 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003261 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003262 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3263 "//conditions:default": [],
3264 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003265 gcc_copts = xnnpack_gcc_std_copts(),
3266 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003267 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003268 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003269 "@FP16",
3270 "@FXdiv",
3271 "@clog",
3272 "@pthreadpool",
3273 ],
3274)
3275
Chao Mei6ddfc602020-05-13 22:29:36 -07003276xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003277 name = "operator_run_test_mode",
3278 srcs = ["src/operator-run.c"],
3279 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3280 copts = LOGGING_COPTS + [
3281 "-UNDEBUG",
3282 "-DXNN_TEST_MODE=1",
3283 ] + select({
3284 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3285 "//conditions:default": [],
3286 }),
3287 gcc_copts = xnnpack_gcc_std_copts(),
3288 msvc_copts = xnnpack_msvc_std_copts(),
3289 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003290 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003291 "@FP16",
3292 "@FXdiv",
3293 "@clog",
3294 "@pthreadpool",
3295 ],
3296)
3297
3298xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003299 name = "memory_planner",
3300 srcs = ["src/memory-planner.c"],
3301 hdrs = INTERNAL_HDRS,
3302 defines = select({
3303 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3304 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3305 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3306 }),
3307 gcc_copts = xnnpack_gcc_std_copts(),
3308 msvc_copts = xnnpack_msvc_std_copts(),
3309 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003310 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003311 "@pthreadpool",
3312 ],
3313)
3314
Marat Dukhan33fcf782020-05-24 14:27:15 -07003315xnnpack_cc_library(
3316 name = "memory_planner_test_mode",
3317 srcs = ["src/memory-planner.c"],
3318 hdrs = INTERNAL_HDRS,
3319 copts = [
3320 "-UNDEBUG",
3321 "-DXNN_TEST_MODE=1",
3322 ],
3323 defines = select({
3324 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3325 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3326 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3327 }),
3328 gcc_copts = xnnpack_gcc_std_copts(),
3329 msvc_copts = xnnpack_msvc_std_copts(),
3330 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003331 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003332 "@pthreadpool",
3333 ],
3334)
3335
Marat Dukhan08c4a432019-10-03 09:29:21 -07003336cc_library(
3337 name = "enable_assembly",
3338 defines = select({
3339 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3340 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003341 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003342 }),
3343)
3344
Marat Dukhan9de90e02020-06-18 16:04:12 -07003345cc_library(
3346 name = "enable_sparse",
3347 defines = select({
3348 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3349 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003350 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003351 }),
3352)
3353
Marat Dukhancf056b22019-10-07 10:26:29 -07003354xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003355 name = "operators",
3356 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003357 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003358 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003359 ],
3360 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003361 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003362 "-Isrc",
3363 "-Iinclude",
3364 ] + select({
3365 ":debug_build": [],
3366 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003367 }) + select({
3368 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3369 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003370 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003371 gcc_copts = xnnpack_gcc_std_copts(),
3372 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003373 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003374 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003375 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003376 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377 "@FP16",
3378 "@FXdiv",
3379 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003380 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003381 ],
3382)
3383
Marat Dukhan10a38082020-04-17 03:58:35 -07003384xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003385 name = "operators_test_mode",
3386 srcs = OPERATOR_SRCS + [
3387 "src/memory.c",
3388 "src/operator-delete.c",
3389 ],
3390 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3391 copts = LOGGING_COPTS + [
3392 "-Isrc",
3393 "-Iinclude",
3394 "-UNDEBUG",
3395 "-DXNN_TEST_MODE=1",
3396 ] + select({
3397 ":debug_build": [],
3398 "//conditions:default": xnnpack_min_size_copts(),
3399 }) + select({
3400 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3401 "//conditions:default": [],
3402 }),
3403 gcc_copts = xnnpack_gcc_std_copts(),
3404 msvc_copts = xnnpack_msvc_std_copts(),
3405 deps = [
3406 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003407 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003408 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003409 "@FP16",
3410 "@FXdiv",
3411 "@clog",
3412 "@pthreadpool",
3413 ],
3414)
3415
3416xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003417 name = "XNNPACK",
3418 srcs = [
3419 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003420 "src/runtime.c",
3421 "src/subgraph.c",
3422 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003423 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003424 hdrs = ["include/xnnpack.h"],
3425 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003426 "-Isrc",
3427 "-Iinclude",
3428 ] + select({
3429 ":debug_build": [],
3430 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003431 }) + select({
3432 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3433 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003434 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003435 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003436 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003437 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003438 visibility = xnnpack_visibility(),
3439 deps = [
3440 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003441 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003442 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003443 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003444 ":operator_run",
3445 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003446 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003447 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003448 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003449 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003450 ] + select({
3451 ":emscripten": [],
3452 "//conditions:default": ["@cpuinfo"],
3453 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003454)
3455
Marat Dukhan10a38082020-04-17 03:58:35 -07003456xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003457 name = "XNNPACK_test_mode",
3458 srcs = [
3459 "src/init.c",
3460 "src/runtime.c",
3461 "src/subgraph.c",
3462 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003463 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003464 hdrs = ["include/xnnpack.h"],
3465 copts = LOGGING_COPTS + [
3466 "-Isrc",
3467 "-Iinclude",
3468 "-UNDEBUG",
3469 "-DXNN_TEST_MODE=1",
3470 ] + select({
3471 ":debug_build": [],
3472 "//conditions:default": xnnpack_min_size_copts(),
3473 }) + select({
3474 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3475 "//conditions:default": [],
3476 }),
3477 gcc_copts = xnnpack_gcc_std_copts(),
3478 includes = ["include"],
3479 msvc_copts = xnnpack_msvc_std_copts(),
3480 visibility = xnnpack_visibility(),
3481 deps = [
3482 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003483 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003484 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003485 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003486 ":operator_run_test_mode",
3487 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003488 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003489 "@clog",
3490 "@FP16",
3491 "@pthreadpool",
3492 ] + select({
3493 ":emscripten": [],
3494 "//conditions:default": ["@cpuinfo"],
3495 }),
3496)
3497
3498xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003499 name = "xnnpack_f32",
3500 srcs = [
3501 "src/init.c",
3502 "src/runtime.c",
3503 "src/subgraph.c",
3504 "src/tensor.c",
3505 ] + SUBGRAPH_SRCS,
3506 hdrs = ["include/xnnpack.h"],
3507 copts = LOGGING_COPTS + [
3508 "-Isrc",
3509 "-Iinclude",
3510 ] + select({
3511 ":debug_build": [],
3512 "//conditions:default": xnnpack_min_size_copts(),
3513 }) + select({
3514 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3515 "//conditions:default": [],
3516 }),
3517 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003518 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003519 "XNN_NO_U8_OPERATORS",
3520 "XNN_NO_X8_OPERATORS",
3521 "XNN_NO_F16_OPERATORS",
3522 "XNN_NO_X16_OPERATORS",
3523 ],
3524 gcc_copts = xnnpack_gcc_std_copts(),
3525 includes = ["include"],
3526 msvc_copts = xnnpack_msvc_std_copts(),
3527 visibility = xnnpack_visibility(),
3528 deps = [
3529 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003530 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003531 ":logging_utils",
3532 ":memory_planner",
3533 ":operator_run",
3534 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003535 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003536 "@clog",
3537 "@FP16",
3538 "@pthreadpool",
3539 ] + select({
3540 ":emscripten": [],
3541 "//conditions:default": ["@cpuinfo"],
3542 }),
3543)
3544
3545xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003546 name = "xnnpack_operators_nhwc_f32",
3547 srcs = [
3548 "src/init.c",
3549 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003550 hdrs = ["include/xnnpack.h"],
3551 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003552 "-Isrc",
3553 "-Iinclude",
3554 ] + select({
3555 ":debug_build": [],
3556 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003557 }) + select({
3558 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3559 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003560 }),
3561 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003562 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003563 "XNN_NO_U8_OPERATORS",
3564 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003565 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003566 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003567 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003569 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003570 visibility = xnnpack_visibility(),
3571 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003572 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003573 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 ":operator_run",
3575 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003576 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003577 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003578 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003579 ] + select({
3580 ":emscripten": [],
3581 "//conditions:default": ["@cpuinfo"],
3582 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583)
3584
Marat Dukhancf056b22019-10-07 10:26:29 -07003585xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 name = "bench_utils",
3587 srcs = ["bench/utils.cc"],
3588 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003589 deps = [
3590 "@com_google_benchmark//:benchmark",
3591 "@cpuinfo",
3592 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593)
3594
Frank Barchard7e955972019-10-11 10:34:25 -07003595######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003596
3597xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003598 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003599 srcs = [
3600 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003601 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 "src/xnnpack/AlignedAllocator.h",
3603 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003604 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003605 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003606)
3607
3608xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003609 name = "f16_igemm_bench",
3610 srcs = [
3611 "bench/f16-igemm.cc",
3612 "bench/conv.h",
3613 "bench/google/conv.h",
3614 "src/xnnpack/AlignedAllocator.h",
3615 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003616 deps = MICROKERNEL_BENCHMARK_DEPS + [
3617 ":indirection",
3618 ":packing",
3619 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003620)
3621
3622xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003623 name = "f16_gemm_bench",
3624 srcs = [
3625 "bench/f16-gemm.cc",
3626 "bench/gemm.h",
3627 "src/xnnpack/AlignedAllocator.h",
3628 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003629 deps = MICROKERNEL_BENCHMARK_DEPS + [
3630 ":packing",
3631 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632)
3633
3634xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003635 name = "f16_spmm_bench",
3636 srcs = [
3637 "bench/f16-spmm.cc",
3638 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003639 "src/xnnpack/AlignedAllocator.h",
3640 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003641 deps = MICROKERNEL_BENCHMARK_DEPS,
3642)
3643
3644xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 name = "f32_igemm_bench",
3646 srcs = [
3647 "bench/f32-igemm.cc",
3648 "bench/conv.h",
3649 "src/xnnpack/AlignedAllocator.h",
3650 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003651 deps = MICROKERNEL_BENCHMARK_DEPS + [
3652 ":indirection",
3653 ":packing",
3654 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003655)
3656
3657xnnpack_benchmark(
3658 name = "f32_conv_hwc_bench",
3659 srcs = [
3660 "bench/f32-conv-hwc.cc",
3661 "bench/dconv.h",
3662 "src/xnnpack/AlignedAllocator.h",
3663 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003664 deps = MICROKERNEL_BENCHMARK_DEPS + [
3665 ":packing",
3666 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003667)
3668
3669xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003670 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003671 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003672 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003673 "bench/dconv.h",
3674 "src/xnnpack/AlignedAllocator.h",
3675 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003676 deps = MICROKERNEL_BENCHMARK_DEPS + [
3677 ":packing",
3678 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003679)
3680
3681xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003682 name = "f16_dwconv_bench",
3683 srcs = [
3684 "bench/f16-dwconv.cc",
3685 "bench/dwconv.h",
3686 "bench/google/dwconv.h",
3687 "src/xnnpack/AlignedAllocator.h",
3688 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003689 deps = MICROKERNEL_BENCHMARK_DEPS + [
3690 ":indirection",
3691 ":packing",
3692 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003693)
3694
3695xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 name = "f32_dwconv_bench",
3697 srcs = [
3698 "bench/f32-dwconv.cc",
3699 "bench/dwconv.h",
3700 "src/xnnpack/AlignedAllocator.h",
3701 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003702 deps = MICROKERNEL_BENCHMARK_DEPS + [
3703 ":indirection",
3704 ":packing",
3705 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003706)
3707
3708xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003709 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003710 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003711 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003712 "bench/dwconv.h",
3713 "src/xnnpack/AlignedAllocator.h",
3714 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003715 deps = MICROKERNEL_BENCHMARK_DEPS + [
3716 ":indirection",
3717 ":packing",
3718 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719)
3720
3721xnnpack_benchmark(
3722 name = "f32_gemm_bench",
3723 srcs = [
3724 "bench/f32-gemm.cc",
3725 "bench/gemm.h",
3726 "src/xnnpack/AlignedAllocator.h",
3727 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003728 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003729 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003730)
3731
3732xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003733 name = "f32_hswish_bench",
3734 srcs = [
3735 "bench/f32-hswish.cc",
3736 "src/xnnpack/AlignedAllocator.h",
3737 ] + MICROKERNEL_BENCHMARK_HDRS,
3738 deps = MICROKERNEL_BENCHMARK_DEPS,
3739)
3740
3741xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003742 name = "f32_raddexpminusmax_bench",
3743 srcs = [
3744 "bench/f32-raddexpminusmax.cc",
3745 "src/xnnpack/AlignedAllocator.h",
3746 ] + MICROKERNEL_BENCHMARK_HDRS,
3747 deps = MICROKERNEL_BENCHMARK_DEPS,
3748)
3749
3750xnnpack_benchmark(
3751 name = "f32_raddextexp_bench",
3752 srcs = [
3753 "bench/f32-raddextexp.cc",
3754 "src/xnnpack/AlignedAllocator.h",
3755 ] + MICROKERNEL_BENCHMARK_HDRS,
3756 deps = MICROKERNEL_BENCHMARK_DEPS,
3757)
3758
3759xnnpack_benchmark(
3760 name = "f32_raddstoreexpminusmax_bench",
3761 srcs = [
3762 "bench/f32-raddstoreexpminusmax.cc",
3763 "src/xnnpack/AlignedAllocator.h",
3764 ] + MICROKERNEL_BENCHMARK_HDRS,
3765 deps = MICROKERNEL_BENCHMARK_DEPS,
3766)
3767
3768xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 name = "f32_rmax_bench",
3770 srcs = [
3771 "bench/f32-rmax.cc",
3772 "src/xnnpack/AlignedAllocator.h",
3773 ] + MICROKERNEL_BENCHMARK_HDRS,
3774 deps = MICROKERNEL_BENCHMARK_DEPS,
3775)
3776
3777xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003778 name = "f32_sigmoid_bench",
3779 srcs = [
3780 "bench/f32-sigmoid.cc",
3781 "src/xnnpack/AlignedAllocator.h",
3782 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003783 deps = MICROKERNEL_BENCHMARK_DEPS,
3784)
3785
3786xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003787 name = "f32_spmm_bench",
3788 srcs = [
3789 "bench/f32-spmm.cc",
3790 "bench/gemm.h",
3791 "src/xnnpack/AlignedAllocator.h",
3792 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003793 deps = MICROKERNEL_BENCHMARK_DEPS,
3794)
3795
3796xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003797 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003798 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003799 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003800 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003801 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003802 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003803)
3804
3805xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003806 name = "f32_vscaleexpminusmax_bench",
3807 srcs = [
3808 "bench/f32-vscaleexpminusmax.cc",
3809 "src/xnnpack/AlignedAllocator.h",
3810 ] + MICROKERNEL_BENCHMARK_HDRS,
3811 deps = MICROKERNEL_BENCHMARK_DEPS,
3812)
3813
3814xnnpack_benchmark(
3815 name = "f32_vscaleextexp_bench",
3816 srcs = [
3817 "bench/f32-vscaleextexp.cc",
3818 "src/xnnpack/AlignedAllocator.h",
3819 ] + MICROKERNEL_BENCHMARK_HDRS,
3820 deps = MICROKERNEL_BENCHMARK_DEPS,
3821)
3822
3823xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003824 name = "f32_vsqrt_bench",
3825 srcs = [
3826 "bench/f32-vsqrt.cc",
3827 "src/xnnpack/AlignedAllocator.h",
3828 ] + MICROKERNEL_BENCHMARK_HDRS,
3829 deps = MICROKERNEL_BENCHMARK_DEPS,
3830)
3831
3832xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003833 name = "f32_im2col_gemm_bench",
3834 srcs = [
3835 "bench/f32-im2col-gemm.cc",
3836 "bench/conv.h",
3837 "src/xnnpack/AlignedAllocator.h",
3838 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003839 deps = MICROKERNEL_BENCHMARK_DEPS + [
3840 ":im2col",
3841 ":packing",
3842 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003843)
3844
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003845xnnpack_benchmark(
3846 name = "requantization_bench",
3847 srcs = [
3848 "bench/requantization.cc",
3849 "src/xnnpack/requantization-stubs.h",
3850 "src/xnnpack/AlignedAllocator.h",
3851 ] + MICROKERNEL_BENCHMARK_HDRS,
3852 deps = MICROKERNEL_BENCHMARK_DEPS,
3853)
3854
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003855xnnpack_benchmark(
3856 name = "rounding_bench",
3857 srcs = [
3858 "bench/rounding.cc",
3859 "src/xnnpack/math-stubs.h",
3860 "src/xnnpack/AlignedAllocator.h",
3861 ] + MICROKERNEL_BENCHMARK_HDRS,
3862 deps = MICROKERNEL_BENCHMARK_DEPS,
3863)
3864
Marat Dukhan08c4a432019-10-03 09:29:21 -07003865########################### Benchmarks for operators ###########################
3866
3867xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003868 name = "average_pooling_bench",
3869 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003870 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003871 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003872 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003873)
3874
3875xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003876 name = "bankers_rounding_bench",
3877 srcs = ["bench/bankers-rounding.cc"],
3878 copts = xnnpack_optional_tflite_copts(),
3879 tags = ["nowin32"],
3880 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3881)
3882
3883xnnpack_benchmark(
3884 name = "ceiling_bench",
3885 srcs = ["bench/ceiling.cc"],
3886 copts = xnnpack_optional_tflite_copts(),
3887 tags = ["nowin32"],
3888 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3889)
3890
3891xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003892 name = "channel_shuffle_bench",
3893 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003894 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003895)
3896
3897xnnpack_benchmark(
3898 name = "convolution_bench",
3899 srcs = ["bench/convolution.cc"],
3900 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003901 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003902 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003903)
3904
3905xnnpack_benchmark(
3906 name = "deconvolution_bench",
3907 srcs = ["bench/deconvolution.cc"],
3908 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003909 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003910 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003911)
3912
3913xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003914 name = "floor_bench",
3915 srcs = ["bench/floor.cc"],
3916 copts = xnnpack_optional_tflite_copts(),
3917 tags = ["nowin32"],
3918 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3919)
3920
3921xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003922 name = "global_average_pooling_bench",
3923 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003924 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003925)
3926
3927xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003928 name = "hardswish_bench",
3929 srcs = ["bench/hardswish.cc"],
3930 copts = xnnpack_optional_tflite_copts(),
3931 tags = ["nowin32"],
3932 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3933)
3934
3935xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003936 name = "max_pooling_bench",
3937 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003938 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003939)
3940
3941xnnpack_benchmark(
3942 name = "sigmoid_bench",
3943 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003944 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003945 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003946 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003947)
3948
3949xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003950 name = "prelu_bench",
3951 srcs = ["bench/prelu.cc"],
3952 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003953 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003954 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003955)
3956
3957xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003958 name = "softmax_bench",
3959 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003960 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003961 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003962 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003963)
3964
Marat Dukhan87727142020-06-24 15:24:10 -07003965xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003966 name = "square_root_bench",
3967 srcs = ["bench/square-root.cc"],
3968 copts = xnnpack_optional_tflite_copts(),
3969 tags = ["nowin32"],
3970 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3971)
3972
3973xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003974 name = "truncation_bench",
3975 srcs = ["bench/truncation.cc"],
3976 deps = OPERATOR_BENCHMARK_DEPS,
3977)
3978
Marat Dukhanc068bb62019-10-04 13:24:39 -07003979############################# End-to-end benchmarks ############################
3980
3981cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003982 name = "fp32_mobilenet_v1",
3983 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003984 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003985 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003986 linkstatic = True,
3987 deps = [
3988 ":XNNPACK",
3989 "@pthreadpool",
3990 ],
3991)
3992
3993cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003994 name = "fp16_mobilenet_v1",
3995 srcs = ["models/fp16-mobilenet-v1.cc"],
3996 hdrs = ["models/models.h"],
3997 copts = xnnpack_std_cxxopts(),
3998 linkstatic = True,
3999 deps = [
4000 ":XNNPACK",
4001 "@FP16",
4002 "@pthreadpool",
4003 ],
4004)
4005
4006cc_library(
4007 name = "fp32_mobilenet_v2",
4008 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004009 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004010 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004011 linkstatic = True,
4012 deps = [
4013 ":XNNPACK",
4014 "@pthreadpool",
4015 ],
4016)
4017
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004018cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004019 name = "fp16_mobilenet_v2",
4020 srcs = ["models/fp16-mobilenet-v2.cc"],
4021 hdrs = ["models/models.h"],
4022 copts = xnnpack_std_cxxopts(),
4023 linkstatic = True,
4024 deps = [
4025 ":XNNPACK",
4026 "@FP16",
4027 "@pthreadpool",
4028 ],
4029)
4030
4031cc_library(
4032 name = "fp32_mobilenet_v3_large",
4033 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004034 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004035 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004036 linkstatic = True,
4037 deps = [
4038 ":XNNPACK",
4039 "@pthreadpool",
4040 ],
4041)
4042
4043cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004044 name = "fp32_mobilenet_v3_small",
4045 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004046 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004047 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004048 linkstatic = True,
4049 deps = [
4050 ":XNNPACK",
4051 "@pthreadpool",
4052 ],
4053)
4054
Marat Dukhanc068bb62019-10-04 13:24:39 -07004055xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004056 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004057 srcs = [
4058 "bench/f32-dwconv-e2e.cc",
4059 "bench/end2end.h",
4060 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004061 deps = MICROKERNEL_BENCHMARK_DEPS + [
4062 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004063 ":fp32_mobilenet_v1",
4064 ":fp32_mobilenet_v2",
4065 ":fp32_mobilenet_v3_large",
4066 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004067 ],
4068)
4069
4070xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004071 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004072 srcs = [
4073 "bench/f32-gemm-e2e.cc",
4074 "bench/end2end.h",
4075 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004076 deps = MICROKERNEL_BENCHMARK_DEPS + [
4077 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004078 ":fp32_mobilenet_v1",
4079 ":fp32_mobilenet_v2",
4080 ":fp32_mobilenet_v3_large",
4081 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004082 ],
4083)
4084
4085xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004086 name = "end2end_bench",
4087 srcs = ["bench/end2end.cc"],
4088 deps = [
4089 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004090 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004091 ":fp16_mobilenet_v1",
4092 ":fp16_mobilenet_v2",
4093 ":fp32_mobilenet_v1",
4094 ":fp32_mobilenet_v2",
4095 ":fp32_mobilenet_v3_large",
4096 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004097 "@pthreadpool",
4098 ],
4099)
4100
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004101#################### Accuracy evaluation for math functions ####################
4102
4103xnnpack_benchmark(
4104 name = "f32_exp_eval",
4105 srcs = [
4106 "eval/f32-exp.cc",
4107 "src/xnnpack/AlignedAllocator.h",
4108 ] + ACCURACY_EVAL_HDRS,
4109 deps = ACCURACY_EVAL_DEPS,
4110)
4111
Marat Dukhan515c9772019-10-17 18:07:57 -07004112xnnpack_benchmark(
4113 name = "f32_expminus_eval",
4114 srcs = [
4115 "eval/f32-expminus.cc",
4116 "src/xnnpack/AlignedAllocator.h",
4117 ] + ACCURACY_EVAL_HDRS,
4118 deps = ACCURACY_EVAL_DEPS,
4119)
4120
Marat Dukhan98ba4412019-10-23 02:14:28 -07004121xnnpack_benchmark(
4122 name = "f32_extexp_eval",
4123 srcs = [
4124 "eval/f32-extexp.cc",
4125 "src/xnnpack/AlignedAllocator.h",
4126 ] + ACCURACY_EVAL_HDRS,
4127 deps = ACCURACY_EVAL_DEPS,
4128)
4129
Marat Dukhan8853b822020-05-07 12:19:01 -07004130xnnpack_unit_test(
4131 name = "f32_roundne_eval",
4132 srcs = [
4133 "eval/f32-roundne.cc",
4134 "src/xnnpack/AlignedAllocator.h",
4135 "src/xnnpack/math-stubs.h",
4136 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004137 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004138 deps = MICROKERNEL_TEST_DEPS,
4139)
4140
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004141xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004142 name = "f32_roundd_eval",
4143 srcs = [
4144 "eval/f32-roundd.cc",
4145 "src/xnnpack/AlignedAllocator.h",
4146 "src/xnnpack/math-stubs.h",
4147 ] + MICROKERNEL_TEST_HDRS,
4148 automatic = False,
4149 deps = MICROKERNEL_TEST_DEPS,
4150)
4151
4152xnnpack_unit_test(
4153 name = "f32_roundu_eval",
4154 srcs = [
4155 "eval/f32-roundu.cc",
4156 "src/xnnpack/AlignedAllocator.h",
4157 "src/xnnpack/math-stubs.h",
4158 ] + MICROKERNEL_TEST_HDRS,
4159 automatic = False,
4160 deps = MICROKERNEL_TEST_DEPS,
4161)
4162
4163xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004164 name = "f32_roundz_eval",
4165 srcs = [
4166 "eval/f32-roundz.cc",
4167 "src/xnnpack/AlignedAllocator.h",
4168 "src/xnnpack/math-stubs.h",
4169 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004170 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004171 deps = MICROKERNEL_TEST_DEPS,
4172)
4173
Marat Dukhan346a9e52019-11-15 09:06:30 -08004174xnnpack_benchmark(
4175 name = "f32_sigmoid_eval",
4176 srcs = [
4177 "eval/f32-sigmoid.cc",
4178 "src/xnnpack/AlignedAllocator.h",
4179 ] + ACCURACY_EVAL_HDRS,
4180 deps = ACCURACY_EVAL_DEPS,
4181)
4182
Marat Dukhan84000762020-06-29 18:38:43 -07004183xnnpack_benchmark(
4184 name = "f32_sqrt_eval",
4185 srcs = [
4186 "eval/f32-sqrt.cc",
4187 "src/xnnpack/AlignedAllocator.h",
4188 ] + ACCURACY_EVAL_HDRS,
4189 deps = ACCURACY_EVAL_DEPS + [
4190 ":bench_utils",
4191 ],
4192)
4193
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194######################### Unit tests for micro-kernels #########################
4195
4196xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004197 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004199 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004200 "test/gemm-microkernel-tester.h",
4201 "src/xnnpack/AlignedAllocator.h",
4202 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004203 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004204)
4205
4206xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004207 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004208 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004209 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004210 "test/spmm-microkernel-tester.h",
4211 "src/xnnpack/AlignedAllocator.h",
4212 ] + MICROKERNEL_TEST_HDRS,
4213 deps = MICROKERNEL_TEST_DEPS,
4214)
4215
4216xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004217 name = "f32_argmaxpool_test",
4218 srcs = [
4219 "test/f32-argmaxpool.cc",
4220 "test/argmaxpool-microkernel-tester.h",
4221 "src/xnnpack/AlignedAllocator.h",
4222 ] + MICROKERNEL_TEST_HDRS,
4223 deps = MICROKERNEL_TEST_DEPS,
4224)
4225
4226xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004227 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004229 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004230 "test/avgpool-microkernel-tester.h",
4231 "src/xnnpack/AlignedAllocator.h",
4232 ] + MICROKERNEL_TEST_HDRS,
4233 deps = MICROKERNEL_TEST_DEPS,
4234)
4235
4236xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004237 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004238 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004239 "test/f32-ibilinear.cc",
4240 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004241 "src/xnnpack/AlignedAllocator.h",
4242 ] + MICROKERNEL_TEST_HDRS,
4243 deps = MICROKERNEL_TEST_DEPS,
4244)
4245
4246xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004247 name = "f16_clamp_test",
4248 srcs = [
4249 "test/f16-clamp.cc",
4250 "test/clamp-microkernel-tester.h",
4251 ] + MICROKERNEL_TEST_HDRS,
4252 deps = MICROKERNEL_TEST_DEPS,
4253)
4254
4255xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004256 name = "f32_clamp_test",
4257 srcs = [
4258 "test/f32-clamp.cc",
4259 "test/clamp-microkernel-tester.h",
4260 ] + MICROKERNEL_TEST_HDRS,
4261 deps = MICROKERNEL_TEST_DEPS,
4262)
4263
4264xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004265 name = "f32_igemm_test",
4266 srcs = [
4267 "test/f32-igemm.cc",
4268 "test/gemm-microkernel-tester.h",
4269 "src/xnnpack/AlignedAllocator.h",
4270 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004271 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004272)
4273
4274xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004275 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004276 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004277 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278 "test/gemm-microkernel-tester.h",
4279 "src/xnnpack/AlignedAllocator.h",
4280 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004281 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004282)
4283
4284xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004285 name = "f32_igemm_minmax_test",
4286 srcs = [
4287 "test/f32-igemm-minmax.cc",
4288 "test/gemm-microkernel-tester.h",
4289 "src/xnnpack/AlignedAllocator.h",
4290 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004291 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004292)
4293
4294xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004295 name = "f16_igemm_minmax_test",
4296 srcs = [
4297 "test/f16-igemm-minmax.cc",
4298 "test/gemm-microkernel-tester.h",
4299 "src/xnnpack/AlignedAllocator.h",
4300 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004301 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004302)
4303
4304xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305 name = "f32_conv_hwc_test",
4306 srcs = [
4307 "test/f32-conv-hwc.cc",
4308 "test/conv-hwc-microkernel-tester.h",
4309 "src/xnnpack/AlignedAllocator.h",
4310 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004311 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004312)
4313
4314xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004315 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004316 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004317 "test/f32-conv-hwc2chw.cc",
4318 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319 "src/xnnpack/AlignedAllocator.h",
4320 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004321 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004322)
4323
4324xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004325 name = "f32_dwconv_test",
4326 srcs = [
4327 "test/f32-dwconv.cc",
4328 "test/dwconv-microkernel-tester.h",
4329 "src/xnnpack/AlignedAllocator.h",
4330 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004331 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004332)
4333
4334xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004335 name = "f16_dwconv_minmax_test",
4336 srcs = [
4337 "test/f16-dwconv-minmax.cc",
4338 "test/dwconv-microkernel-tester.h",
4339 "src/xnnpack/AlignedAllocator.h",
4340 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004341 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004342)
4343
4344xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004345 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004346 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004347 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348 "test/dwconv-microkernel-tester.h",
4349 "src/xnnpack/AlignedAllocator.h",
4350 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004351 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352)
4353
4354xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004355 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004356 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004357 "test/f32-dwconv-chw.cc",
4358 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004359 "src/xnnpack/AlignedAllocator.h",
4360 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004361 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004362)
4363
4364xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004365 name = "f16_gavgpool_minmax_test",
4366 srcs = [
4367 "test/f16-gavgpool-minmax.cc",
4368 "test/gavgpool-microkernel-tester.h",
4369 "src/xnnpack/AlignedAllocator.h",
4370 ] + MICROKERNEL_TEST_HDRS,
4371 deps = MICROKERNEL_TEST_DEPS,
4372)
4373
4374xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004375 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004376 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004377 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004378 "test/gavgpool-microkernel-tester.h",
4379 "src/xnnpack/AlignedAllocator.h",
4380 ] + MICROKERNEL_TEST_HDRS,
4381 deps = MICROKERNEL_TEST_DEPS,
4382)
4383
4384xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004385 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004387 "test/f32-gavgpool-cw.cc",
4388 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004389 "src/xnnpack/AlignedAllocator.h",
4390 ] + MICROKERNEL_TEST_HDRS,
4391 deps = MICROKERNEL_TEST_DEPS,
4392)
4393
4394xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004395 name = "f32_gemm_test",
4396 srcs = [
4397 "test/f32-gemm.cc",
4398 "test/gemm-microkernel-tester.h",
4399 "src/xnnpack/AlignedAllocator.h",
4400 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004401 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004402)
4403
4404xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004405 name = "f32_gemm_relu_test",
4406 srcs = [
4407 "test/f32-gemm-relu.cc",
4408 "test/gemm-microkernel-tester.h",
4409 "src/xnnpack/AlignedAllocator.h",
4410 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004411 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004412)
4413
4414xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004415 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004417 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004418 "test/gemm-microkernel-tester.h",
4419 "src/xnnpack/AlignedAllocator.h",
4420 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004421 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004422)
4423
4424xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004425 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004426 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004427 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004428 "test/gemm-microkernel-tester.h",
4429 "src/xnnpack/AlignedAllocator.h",
4430 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004431 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004432)
4433
4434xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004435 name = "f16_hswish_test",
4436 srcs = [
4437 "test/f16-hswish.cc",
4438 "test/hswish-microkernel-tester.h",
4439 ] + MICROKERNEL_TEST_HDRS,
4440 deps = MICROKERNEL_TEST_DEPS,
4441)
4442
4443xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004444 name = "f32_hswish_test",
4445 srcs = [
4446 "test/f32-hswish.cc",
4447 "test/hswish-microkernel-tester.h",
4448 ] + MICROKERNEL_TEST_HDRS,
4449 deps = MICROKERNEL_TEST_DEPS,
4450)
4451
4452xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004453 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004454 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004455 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004456 "test/maxpool-microkernel-tester.h",
4457 ] + MICROKERNEL_TEST_HDRS,
4458 deps = MICROKERNEL_TEST_DEPS,
4459)
4460
4461xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004462 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004463 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004464 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465 "test/avgpool-microkernel-tester.h",
4466 "src/xnnpack/AlignedAllocator.h",
4467 ] + MICROKERNEL_TEST_HDRS,
4468 deps = MICROKERNEL_TEST_DEPS,
4469)
4470
4471xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004472 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004474 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004475 "test/gemm-microkernel-tester.h",
4476 "src/xnnpack/AlignedAllocator.h",
4477 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004478 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479)
4480
4481xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004482 name = "f16_prelu_test",
4483 srcs = [
4484 "test/f16-prelu.cc",
4485 "test/prelu-microkernel-tester.h",
4486 "src/xnnpack/AlignedAllocator.h",
4487 ] + MICROKERNEL_TEST_HDRS,
4488 deps = MICROKERNEL_TEST_DEPS,
4489)
4490
4491xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004492 name = "f32_prelu_test",
4493 srcs = [
4494 "test/f32-prelu.cc",
4495 "test/prelu-microkernel-tester.h",
4496 "src/xnnpack/AlignedAllocator.h",
4497 ] + MICROKERNEL_TEST_HDRS,
4498 deps = MICROKERNEL_TEST_DEPS,
4499)
4500
4501xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004502 name = "f32_raddexpminusmax_test",
4503 srcs = [
4504 "test/f32-raddexpminusmax.cc",
4505 "test/raddexpminusmax-microkernel-tester.h",
4506 ] + MICROKERNEL_TEST_HDRS,
4507 deps = MICROKERNEL_TEST_DEPS,
4508)
4509
4510xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004511 name = "f32_raddextexp_test",
4512 srcs = [
4513 "test/f32-raddextexp.cc",
4514 "test/raddextexp-microkernel-tester.h",
4515 ] + MICROKERNEL_TEST_HDRS,
4516 deps = MICROKERNEL_TEST_DEPS,
4517)
4518
4519xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004520 name = "f32_raddstoreexpminusmax_test",
4521 srcs = [
4522 "test/f32-raddstoreexpminusmax.cc",
4523 "test/raddstoreexpminusmax-microkernel-tester.h",
4524 ] + MICROKERNEL_TEST_HDRS,
4525 deps = MICROKERNEL_TEST_DEPS,
4526)
4527
4528xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004529 name = "f32_relu_test",
4530 srcs = [
4531 "test/f32-relu.cc",
4532 "test/vunary-microkernel-tester.h",
4533 ] + MICROKERNEL_TEST_HDRS,
4534 deps = MICROKERNEL_TEST_DEPS,
4535)
4536
4537xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004538 name = "f32_rmax_test",
4539 srcs = [
4540 "test/f32-rmax.cc",
4541 "test/rmax-microkernel-tester.h",
4542 ] + MICROKERNEL_TEST_HDRS,
4543 deps = MICROKERNEL_TEST_DEPS,
4544)
4545
4546xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004547 name = "f32_sigmoid_test",
4548 srcs = [
4549 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004550 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004551 ] + MICROKERNEL_TEST_HDRS,
4552 deps = MICROKERNEL_TEST_DEPS,
4553)
4554
4555xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004556 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004557 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004558 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004559 "test/spmm-microkernel-tester.h",
4560 "src/xnnpack/AlignedAllocator.h",
4561 ] + MICROKERNEL_TEST_HDRS,
4562 deps = MICROKERNEL_TEST_DEPS,
4563)
4564
4565xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004566 name = "f16_vadd_minmax_test",
4567 srcs = [
4568 "test/f16-vadd-minmax.cc",
4569 "test/vbinary-microkernel-tester.h",
4570 ] + MICROKERNEL_TEST_HDRS,
4571 deps = MICROKERNEL_TEST_DEPS,
4572)
4573
4574xnnpack_unit_test(
4575 name = "f16_vaddc_minmax_test",
4576 srcs = [
4577 "test/f16-vaddc-minmax.cc",
4578 "test/vbinaryc-microkernel-tester.h",
4579 ] + MICROKERNEL_TEST_HDRS,
4580 deps = MICROKERNEL_TEST_DEPS,
4581)
4582
4583xnnpack_unit_test(
4584 name = "f16_vdiv_minmax_test",
4585 srcs = [
4586 "test/f16-vdiv-minmax.cc",
4587 "test/vbinary-microkernel-tester.h",
4588 ] + MICROKERNEL_TEST_HDRS,
4589 deps = MICROKERNEL_TEST_DEPS,
4590)
4591
4592xnnpack_unit_test(
4593 name = "f16_vdivc_minmax_test",
4594 srcs = [
4595 "test/f16-vdivc-minmax.cc",
4596 "test/vbinaryc-microkernel-tester.h",
4597 ] + MICROKERNEL_TEST_HDRS,
4598 deps = MICROKERNEL_TEST_DEPS,
4599)
4600
4601xnnpack_unit_test(
4602 name = "f16_vrdivc_minmax_test",
4603 srcs = [
4604 "test/f16-vrdivc-minmax.cc",
4605 "test/vbinaryc-microkernel-tester.h",
4606 ] + MICROKERNEL_TEST_HDRS,
4607 deps = MICROKERNEL_TEST_DEPS,
4608)
4609
4610xnnpack_unit_test(
4611 name = "f16_vmax_test",
4612 srcs = [
4613 "test/f16-vmax.cc",
4614 "test/vbinary-microkernel-tester.h",
4615 ] + MICROKERNEL_TEST_HDRS,
4616 deps = MICROKERNEL_TEST_DEPS,
4617)
4618
4619xnnpack_unit_test(
4620 name = "f16_vmaxc_test",
4621 srcs = [
4622 "test/f16-vmaxc.cc",
4623 "test/vbinaryc-microkernel-tester.h",
4624 ] + MICROKERNEL_TEST_HDRS,
4625 deps = MICROKERNEL_TEST_DEPS,
4626)
4627
4628xnnpack_unit_test(
4629 name = "f16_vmin_test",
4630 srcs = [
4631 "test/f16-vmin.cc",
4632 "test/vbinary-microkernel-tester.h",
4633 ] + MICROKERNEL_TEST_HDRS,
4634 deps = MICROKERNEL_TEST_DEPS,
4635)
4636
4637xnnpack_unit_test(
4638 name = "f16_vminc_test",
4639 srcs = [
4640 "test/f16-vminc.cc",
4641 "test/vbinaryc-microkernel-tester.h",
4642 ] + MICROKERNEL_TEST_HDRS,
4643 deps = MICROKERNEL_TEST_DEPS,
4644)
4645
4646xnnpack_unit_test(
4647 name = "f16_vmul_minmax_test",
4648 srcs = [
4649 "test/f16-vmul-minmax.cc",
4650 "test/vbinary-microkernel-tester.h",
4651 ] + MICROKERNEL_TEST_HDRS,
4652 deps = MICROKERNEL_TEST_DEPS,
4653)
4654
4655xnnpack_unit_test(
4656 name = "f16_vmulc_minmax_test",
4657 srcs = [
4658 "test/f16-vmulc-minmax.cc",
4659 "test/vbinaryc-microkernel-tester.h",
4660 ] + MICROKERNEL_TEST_HDRS,
4661 deps = MICROKERNEL_TEST_DEPS,
4662)
4663
4664xnnpack_unit_test(
4665 name = "f16_vsub_minmax_test",
4666 srcs = [
4667 "test/f16-vsub-minmax.cc",
4668 "test/vbinary-microkernel-tester.h",
4669 ] + MICROKERNEL_TEST_HDRS,
4670 deps = MICROKERNEL_TEST_DEPS,
4671)
4672
4673xnnpack_unit_test(
4674 name = "f16_vsubc_minmax_test",
4675 srcs = [
4676 "test/f16-vsubc-minmax.cc",
4677 "test/vbinaryc-microkernel-tester.h",
4678 ] + MICROKERNEL_TEST_HDRS,
4679 deps = MICROKERNEL_TEST_DEPS,
4680)
4681
4682xnnpack_unit_test(
4683 name = "f16_vrsubc_minmax_test",
4684 srcs = [
4685 "test/f16-vrsubc-minmax.cc",
4686 "test/vbinaryc-microkernel-tester.h",
4687 ] + MICROKERNEL_TEST_HDRS,
4688 deps = MICROKERNEL_TEST_DEPS,
4689)
4690
4691xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004692 name = "f32_vabs_test",
4693 srcs = [
4694 "test/f32-vabs.cc",
4695 "test/vunary-microkernel-tester.h",
4696 ] + MICROKERNEL_TEST_HDRS,
4697 deps = MICROKERNEL_TEST_DEPS,
4698)
4699
4700xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004701 name = "f32_vadd_test",
4702 srcs = [
4703 "test/f32-vadd.cc",
4704 "test/vbinary-microkernel-tester.h",
4705 ] + MICROKERNEL_TEST_HDRS,
4706 deps = MICROKERNEL_TEST_DEPS,
4707)
4708
4709xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004710 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004711 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004712 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004713 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004714 ] + MICROKERNEL_TEST_HDRS,
4715 deps = MICROKERNEL_TEST_DEPS,
4716)
4717
4718xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004719 name = "f32_vaddc_test",
4720 srcs = [
4721 "test/f32-vaddc.cc",
4722 "test/vbinaryc-microkernel-tester.h",
4723 ] + MICROKERNEL_TEST_HDRS,
4724 deps = MICROKERNEL_TEST_DEPS,
4725)
4726
4727xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004728 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004729 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004730 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004731 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004732 ] + MICROKERNEL_TEST_HDRS,
4733 deps = MICROKERNEL_TEST_DEPS,
4734)
4735
4736xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004737 name = "f32_vdiv_test",
4738 srcs = [
4739 "test/f32-vdiv.cc",
4740 "test/vbinary-microkernel-tester.h",
4741 ] + MICROKERNEL_TEST_HDRS,
4742 deps = MICROKERNEL_TEST_DEPS,
4743)
4744
4745xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004746 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004747 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004748 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004749 "test/vbinary-microkernel-tester.h",
4750 ] + MICROKERNEL_TEST_HDRS,
4751 deps = MICROKERNEL_TEST_DEPS,
4752)
4753
4754xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004755 name = "f32_vdivc_test",
4756 srcs = [
4757 "test/f32-vdivc.cc",
4758 "test/vbinaryc-microkernel-tester.h",
4759 ] + MICROKERNEL_TEST_HDRS,
4760 deps = MICROKERNEL_TEST_DEPS,
4761)
4762
4763xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004764 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004765 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004766 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004767 "test/vbinaryc-microkernel-tester.h",
4768 ] + MICROKERNEL_TEST_HDRS,
4769 deps = MICROKERNEL_TEST_DEPS,
4770)
4771
4772xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004773 name = "f32_vrdivc_test",
4774 srcs = [
4775 "test/f32-vrdivc.cc",
4776 "test/vbinaryc-microkernel-tester.h",
4777 ] + MICROKERNEL_TEST_HDRS,
4778 deps = MICROKERNEL_TEST_DEPS,
4779)
4780
4781xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004782 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004783 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004784 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004785 "test/vbinaryc-microkernel-tester.h",
4786 ] + MICROKERNEL_TEST_HDRS,
4787 deps = MICROKERNEL_TEST_DEPS,
4788)
4789
4790xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004791 name = "f32_vmax_test",
4792 srcs = [
4793 "test/f32-vmax.cc",
4794 "test/vbinary-microkernel-tester.h",
4795 ] + MICROKERNEL_TEST_HDRS,
4796 deps = MICROKERNEL_TEST_DEPS,
4797)
4798
4799xnnpack_unit_test(
4800 name = "f32_vmaxc_test",
4801 srcs = [
4802 "test/f32-vmaxc.cc",
4803 "test/vbinaryc-microkernel-tester.h",
4804 ] + MICROKERNEL_TEST_HDRS,
4805 deps = MICROKERNEL_TEST_DEPS,
4806)
4807
4808xnnpack_unit_test(
4809 name = "f32_vmin_test",
4810 srcs = [
4811 "test/f32-vmin.cc",
4812 "test/vbinary-microkernel-tester.h",
4813 ] + MICROKERNEL_TEST_HDRS,
4814 deps = MICROKERNEL_TEST_DEPS,
4815)
4816
4817xnnpack_unit_test(
4818 name = "f32_vminc_test",
4819 srcs = [
4820 "test/f32-vminc.cc",
4821 "test/vbinaryc-microkernel-tester.h",
4822 ] + MICROKERNEL_TEST_HDRS,
4823 deps = MICROKERNEL_TEST_DEPS,
4824)
4825
4826xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004827 name = "f32_vmul_test",
4828 srcs = [
4829 "test/f32-vmul.cc",
4830 "test/vbinary-microkernel-tester.h",
4831 ] + MICROKERNEL_TEST_HDRS,
4832 deps = MICROKERNEL_TEST_DEPS,
4833)
4834
4835xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004836 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004837 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004838 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004839 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004840 ] + MICROKERNEL_TEST_HDRS,
4841 deps = MICROKERNEL_TEST_DEPS,
4842)
4843
4844xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004845 name = "f32_vmulc_test",
4846 srcs = [
4847 "test/f32-vmulc.cc",
4848 "test/vbinaryc-microkernel-tester.h",
4849 ] + MICROKERNEL_TEST_HDRS,
4850 deps = MICROKERNEL_TEST_DEPS,
4851)
4852
4853xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004854 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004855 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004856 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004857 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858 ] + MICROKERNEL_TEST_HDRS,
4859 deps = MICROKERNEL_TEST_DEPS,
4860)
4861
4862xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004863 name = "f16_vmulcaddc_minmax_test",
4864 srcs = [
4865 "test/f16-vmulcaddc-minmax.cc",
4866 "test/vmulcaddc-microkernel-tester.h",
4867 "src/xnnpack/AlignedAllocator.h",
4868 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004869 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004870)
4871
4872xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004873 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004874 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004875 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004876 "test/vmulcaddc-microkernel-tester.h",
4877 "src/xnnpack/AlignedAllocator.h",
4878 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004879 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004880)
4881
4882xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004883 name = "f32_vlrelu_test",
4884 srcs = [
4885 "test/f32-vlrelu.cc",
4886 "test/vunary-microkernel-tester.h",
4887 ] + MICROKERNEL_TEST_HDRS,
4888 deps = MICROKERNEL_TEST_DEPS,
4889)
4890
4891xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004892 name = "f32_vneg_test",
4893 srcs = [
4894 "test/f32-vneg.cc",
4895 "test/vunary-microkernel-tester.h",
4896 ] + MICROKERNEL_TEST_HDRS,
4897 deps = MICROKERNEL_TEST_DEPS,
4898)
4899
4900xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004901 name = "f32_vrndne_test",
4902 srcs = [
4903 "test/f32-vrndne.cc",
4904 "test/vunary-microkernel-tester.h",
4905 ] + MICROKERNEL_TEST_HDRS,
4906 deps = MICROKERNEL_TEST_DEPS,
4907)
4908
4909xnnpack_unit_test(
4910 name = "f32_vrndz_test",
4911 srcs = [
4912 "test/f32-vrndz.cc",
4913 "test/vunary-microkernel-tester.h",
4914 ] + MICROKERNEL_TEST_HDRS,
4915 deps = MICROKERNEL_TEST_DEPS,
4916)
4917
4918xnnpack_unit_test(
4919 name = "f32_vrndu_test",
4920 srcs = [
4921 "test/f32-vrndu.cc",
4922 "test/vunary-microkernel-tester.h",
4923 ] + MICROKERNEL_TEST_HDRS,
4924 deps = MICROKERNEL_TEST_DEPS,
4925)
4926
4927xnnpack_unit_test(
4928 name = "f32_vrndd_test",
4929 srcs = [
4930 "test/f32-vrndd.cc",
4931 "test/vunary-microkernel-tester.h",
4932 ] + MICROKERNEL_TEST_HDRS,
4933 deps = MICROKERNEL_TEST_DEPS,
4934)
4935
4936xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004937 name = "f32_vscale_test",
4938 srcs = [
4939 "test/f32-vscale.cc",
4940 "test/vscale-microkernel-tester.h",
4941 ] + MICROKERNEL_TEST_HDRS,
4942 deps = MICROKERNEL_TEST_DEPS,
4943)
4944
4945xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004946 name = "f32_vscaleexpminusmax_test",
4947 srcs = [
4948 "test/f32-vscaleexpminusmax.cc",
4949 "test/vscaleexpminusmax-microkernel-tester.h",
4950 ] + MICROKERNEL_TEST_HDRS,
4951 deps = MICROKERNEL_TEST_DEPS,
4952)
4953
4954xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004955 name = "f32_vscaleextexp_test",
4956 srcs = [
4957 "test/f32-vscaleextexp.cc",
4958 "test/vscaleextexp-microkernel-tester.h",
4959 ] + MICROKERNEL_TEST_HDRS,
4960 deps = MICROKERNEL_TEST_DEPS,
4961)
4962
4963xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004964 name = "f32_vsqr_test",
4965 srcs = [
4966 "test/f32-vsqr.cc",
4967 "test/vunary-microkernel-tester.h",
4968 ] + MICROKERNEL_TEST_HDRS,
4969 deps = MICROKERNEL_TEST_DEPS,
4970)
4971
4972xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004973 name = "f32_vsqrdiff_test",
4974 srcs = [
4975 "test/f32-vsqrdiff.cc",
4976 "test/vbinary-microkernel-tester.h",
4977 ] + MICROKERNEL_TEST_HDRS,
4978 deps = MICROKERNEL_TEST_DEPS,
4979)
4980
4981xnnpack_unit_test(
4982 name = "f32_vsqrdiffc_test",
4983 srcs = [
4984 "test/f32-vsqrdiffc.cc",
4985 "test/vbinaryc-microkernel-tester.h",
4986 ] + MICROKERNEL_TEST_HDRS,
4987 deps = MICROKERNEL_TEST_DEPS,
4988)
4989
4990xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004991 name = "f32_vsqrt_test",
4992 srcs = [
4993 "test/f32-vsqrt.cc",
4994 "test/vunary-microkernel-tester.h",
4995 ] + MICROKERNEL_TEST_HDRS,
4996 deps = MICROKERNEL_TEST_DEPS,
4997)
4998
4999xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005000 name = "f32_vsub_test",
5001 srcs = [
5002 "test/f32-vsub.cc",
5003 "test/vbinary-microkernel-tester.h",
5004 ] + MICROKERNEL_TEST_HDRS,
5005 deps = MICROKERNEL_TEST_DEPS,
5006)
5007
5008xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005009 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005010 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005011 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005012 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005013 ] + MICROKERNEL_TEST_HDRS,
5014 deps = MICROKERNEL_TEST_DEPS,
5015)
5016
5017xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005018 name = "f32_vsubc_test",
5019 srcs = [
5020 "test/f32-vsubc.cc",
5021 "test/vbinaryc-microkernel-tester.h",
5022 ] + MICROKERNEL_TEST_HDRS,
5023 deps = MICROKERNEL_TEST_DEPS,
5024)
5025
5026xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005027 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005028 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005029 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005030 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005031 ] + MICROKERNEL_TEST_HDRS,
5032 deps = MICROKERNEL_TEST_DEPS,
5033)
5034
5035xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005036 name = "f32_vrsubc_test",
5037 srcs = [
5038 "test/f32-vrsubc.cc",
5039 "test/vbinaryc-microkernel-tester.h",
5040 ] + MICROKERNEL_TEST_HDRS,
5041 deps = MICROKERNEL_TEST_DEPS,
5042)
5043
5044xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005045 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005046 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005047 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005048 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005049 ] + MICROKERNEL_TEST_HDRS,
5050 deps = MICROKERNEL_TEST_DEPS,
5051)
5052
5053xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005054 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005055 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005056 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057 "test/avgpool-microkernel-tester.h",
5058 "src/xnnpack/AlignedAllocator.h",
5059 ] + MICROKERNEL_TEST_HDRS,
5060 deps = MICROKERNEL_TEST_DEPS,
5061)
5062
5063xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005064 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005065 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005066 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005067 "test/gemm-microkernel-tester.h",
5068 "src/xnnpack/AlignedAllocator.h",
5069 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005070 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005071)
5072
5073xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005074 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005075 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005076 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005077 "test/dwconv-microkernel-tester.h",
5078 "src/xnnpack/AlignedAllocator.h",
5079 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005080 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005081)
5082
5083xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005084 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005086 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005087 "test/gavgpool-microkernel-tester.h",
5088 "src/xnnpack/AlignedAllocator.h",
5089 ] + MICROKERNEL_TEST_HDRS,
5090 deps = MICROKERNEL_TEST_DEPS,
5091)
5092
5093xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005094 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005096 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005097 "test/gemm-microkernel-tester.h",
5098 "src/xnnpack/AlignedAllocator.h",
5099 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005100 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005101)
5102
5103xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005104 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005105 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005106 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005107 "test/vadd-microkernel-tester.h",
5108 ] + MICROKERNEL_TEST_HDRS,
5109 deps = MICROKERNEL_TEST_DEPS,
5110)
5111
5112xnnpack_unit_test(
5113 name = "u8_clamp_test",
5114 srcs = [
5115 "test/u8-clamp.cc",
5116 "test/clamp-microkernel-tester.h",
5117 ] + MICROKERNEL_TEST_HDRS,
5118 deps = MICROKERNEL_TEST_DEPS,
5119)
5120
5121xnnpack_unit_test(
5122 name = "u8_lut32norm_test",
5123 srcs = [
5124 "test/u8-lut32norm.cc",
5125 "test/lut-norm-microkernel-tester.h",
5126 ] + MICROKERNEL_TEST_HDRS,
5127 deps = MICROKERNEL_TEST_DEPS,
5128)
5129
5130xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005131 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005132 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005133 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134 "test/maxpool-microkernel-tester.h",
5135 ] + MICROKERNEL_TEST_HDRS,
5136 deps = MICROKERNEL_TEST_DEPS,
5137)
5138
5139xnnpack_unit_test(
5140 name = "u8_rmax_test",
5141 srcs = [
5142 "test/u8-rmax.cc",
5143 "test/rmax-microkernel-tester.h",
5144 ] + MICROKERNEL_TEST_HDRS,
5145 deps = MICROKERNEL_TEST_DEPS,
5146)
5147
5148xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005149 name = "x32_fill_test",
5150 srcs = [
5151 "test/x32-fill.cc",
5152 "test/fill-microkernel-tester.h",
5153 ] + MICROKERNEL_TEST_HDRS,
5154 deps = MICROKERNEL_TEST_DEPS,
5155)
5156
5157xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005158 name = "x32_packx_test",
5159 srcs = [
5160 "test/x32-packx.cc",
5161 "test/pack-microkernel-tester.h",
5162 "src/xnnpack/AlignedAllocator.h",
5163 ] + MICROKERNEL_TEST_HDRS,
5164 deps = MICROKERNEL_TEST_DEPS,
5165)
5166
5167xnnpack_unit_test(
5168 name = "x32_pad_test",
5169 srcs = [
5170 "test/x32-pad.cc",
5171 "test/pad-microkernel-tester.h",
5172 ] + MICROKERNEL_TEST_HDRS,
5173 deps = MICROKERNEL_TEST_DEPS,
5174)
5175
5176xnnpack_unit_test(
5177 name = "x32_unpool_test",
5178 srcs = [
5179 "test/x32-unpool.cc",
5180 "test/unpool-microkernel-tester.h",
5181 ] + MICROKERNEL_TEST_HDRS,
5182 deps = MICROKERNEL_TEST_DEPS,
5183)
5184
5185xnnpack_unit_test(
5186 name = "x32_zip_test",
5187 srcs = [
5188 "test/x32-zip.cc",
5189 "test/zip-microkernel-tester.h",
5190 ] + MICROKERNEL_TEST_HDRS,
5191 deps = MICROKERNEL_TEST_DEPS,
5192)
5193
5194xnnpack_unit_test(
5195 name = "x8_lut_test",
5196 srcs = [
5197 "test/x8-lut.cc",
5198 "test/lut-microkernel-tester.h",
5199 ] + MICROKERNEL_TEST_HDRS,
5200 deps = MICROKERNEL_TEST_DEPS,
5201)
5202
5203xnnpack_unit_test(
5204 name = "x8_zip_test",
5205 srcs = [
5206 "test/x8-zip.cc",
5207 "test/zip-microkernel-tester.h",
5208 ] + MICROKERNEL_TEST_HDRS,
5209 deps = MICROKERNEL_TEST_DEPS,
5210)
5211
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005212xnnpack_unit_test(
5213 name = "requantization_test",
5214 srcs = [
5215 "src/xnnpack/requantization-stubs.h",
5216 "test/requantization.cc",
5217 "test/requantization-tester.h",
5218 ] + MICROKERNEL_TEST_HDRS,
5219 deps = MICROKERNEL_TEST_DEPS,
5220)
5221
Marat Dukhan20c3b922020-03-10 03:45:06 -07005222########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223
5224xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005225 name = "operator_size_test",
5226 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005227 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005228)
5229
Marat Dukhan20c3b922020-03-10 03:45:06 -07005230xnnpack_binary(
5231 name = "subgraph_size_test",
5232 srcs = ["test/subgraph-size.c"],
5233 deps = [":XNNPACK"],
5234)
5235
5236########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237
5238xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005239 name = "abs_nc_test",
5240 srcs = [
5241 "test/abs-nc.cc",
5242 "test/abs-operator-tester.h",
5243 ],
5244 deps = OPERATOR_TEST_DEPS,
5245)
5246
5247xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005248 name = "add_nd_test",
5249 srcs = [
5250 "test/add-nd.cc",
5251 "test/binary-elementwise-operator-tester.h",
5252 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005253 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005254)
5255
5256xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005257 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005259 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260 "test/argmax-pooling-operator-tester.h",
5261 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005262 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263)
5264
5265xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005266 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005268 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005269 "test/average-pooling-operator-tester.h",
5270 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005271 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272)
5273
5274xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005275 name = "bankers_rounding_nc_test",
5276 srcs = [
5277 "test/bankers-rounding-nc.cc",
5278 "test/bankers-rounding-operator-tester.h",
5279 ],
5280 deps = OPERATOR_TEST_DEPS,
5281)
5282
5283xnnpack_unit_test(
5284 name = "ceiling_nc_test",
5285 srcs = [
5286 "test/ceiling-nc.cc",
5287 "test/ceiling-operator-tester.h",
5288 ],
5289 deps = OPERATOR_TEST_DEPS,
5290)
5291
5292xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005293 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005295 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005296 "test/channel-shuffle-operator-tester.h",
5297 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005299)
5300
5301xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005302 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005304 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305 "test/clamp-operator-tester.h",
5306 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005307 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005308)
5309
5310xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005311 name = "constant_pad_nd_test",
5312 srcs = [
5313 "test/constant-pad-nd.cc",
5314 "test/constant-pad-operator-tester.h",
5315 ],
5316 deps = OPERATOR_TEST_DEPS,
5317)
5318
5319xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005320 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005322 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323 "test/convolution-operator-tester.h",
5324 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005325 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005326)
5327
5328xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005329 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005330 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005331 "test/convolution-nchw.cc",
5332 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005334 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005335)
5336
5337xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005338 name = "copy_nc_test",
5339 srcs = [
5340 "test/copy-nc.cc",
5341 "test/copy-operator-tester.h",
5342 ],
5343 deps = OPERATOR_TEST_DEPS,
5344)
5345
5346xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005347 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005348 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005349 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005350 "test/deconvolution-operator-tester.h",
5351 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005352 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005353)
5354
5355xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005356 name = "divide_nd_test",
5357 srcs = [
5358 "test/binary-elementwise-operator-tester.h",
5359 "test/divide-nd.cc",
5360 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005361 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005362)
5363
5364xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005365 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005367 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005368 "test/fully-connected-operator-tester.h",
5369 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005371)
5372
5373xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005374 name = "floor_nc_test",
5375 srcs = [
5376 "test/floor-nc.cc",
5377 "test/floor-operator-tester.h",
5378 ],
5379 deps = OPERATOR_TEST_DEPS,
5380)
5381
5382xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005383 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005384 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005385 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005386 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005387 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005388 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005389)
5390
5391xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005392 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005393 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005394 "test/global-average-pooling-ncw.cc",
5395 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005396 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005397 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005398)
5399
5400xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005401 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005402 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005403 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005404 "test/hardswish-operator-tester.h",
5405 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005406 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005407)
5408
5409xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005410 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005412 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 "test/leaky-relu-operator-tester.h",
5414 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005415 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005416)
5417
5418xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005419 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005420 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005421 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005422 "test/max-pooling-operator-tester.h",
5423 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005424 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005425)
5426
5427xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005428 name = "maximum_nd_test",
5429 srcs = [
5430 "test/binary-elementwise-operator-tester.h",
5431 "test/maximum-nd.cc",
5432 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005433 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005434)
5435
5436xnnpack_unit_test(
5437 name = "minimum_nd_test",
5438 srcs = [
5439 "test/binary-elementwise-operator-tester.h",
5440 "test/minimum-nd.cc",
5441 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005442 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005443)
5444
5445xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005446 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005447 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005448 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005449 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005450 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005451 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005452)
5453
5454xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005455 name = "negate_nc_test",
5456 srcs = [
5457 "test/negate-nc.cc",
5458 "test/negate-operator-tester.h",
5459 ],
5460 deps = OPERATOR_TEST_DEPS,
5461)
5462
5463xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005464 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005466 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467 "test/prelu-operator-tester.h",
5468 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005469 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005470)
5471
5472xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005473 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005474 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005475 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005476 "test/resize-bilinear-operator-tester.h",
5477 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005478 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005479)
5480
5481xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005482 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005483 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005484 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005485 "test/sigmoid-operator-tester.h",
5486 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005487 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005488)
5489
5490xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005491 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005492 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005493 "test/softmax-nc.cc",
5494 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005495 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005496 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005497)
5498
5499xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005500 name = "square_nc_test",
5501 srcs = [
5502 "test/square-nc.cc",
5503 "test/square-operator-tester.h",
5504 ],
5505 deps = OPERATOR_TEST_DEPS,
5506)
5507
5508xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005509 name = "square_root_nc_test",
5510 srcs = [
5511 "test/square-root-nc.cc",
5512 "test/square-root-operator-tester.h",
5513 ],
5514 deps = OPERATOR_TEST_DEPS,
5515)
5516
5517xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005518 name = "squared_difference_nd_test",
5519 srcs = [
5520 "test/binary-elementwise-operator-tester.h",
5521 "test/squared-difference-nd.cc",
5522 ],
5523 deps = OPERATOR_TEST_DEPS,
5524)
5525
5526xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005527 name = "subtract_nd_test",
5528 srcs = [
5529 "test/binary-elementwise-operator-tester.h",
5530 "test/subtract-nd.cc",
5531 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005532 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005533)
5534
5535xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005536 name = "truncation_nc_test",
5537 srcs = [
5538 "test/truncation-nc.cc",
5539 "test/truncation-operator-tester.h",
5540 ],
5541 deps = OPERATOR_TEST_DEPS,
5542)
5543
5544xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005545 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005546 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005547 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005548 "test/unpooling-operator-tester.h",
5549 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005550 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005551)
5552
Chao Mei6ddfc602020-05-13 22:29:36 -07005553############################### Misc unit tests ###############################
5554
5555xnnpack_unit_test(
5556 name = "memory_planner_test",
5557 srcs = [
5558 "test/memory-planner-test.cc",
5559 ],
5560 deps = [
5561 ":XNNPACK",
5562 ":memory_planner",
5563 ],
5564)
5565
Marat Dukhan08c4a432019-10-03 09:29:21 -07005566############################# Build configurations #############################
5567
Marat Dukhanb8642352019-10-30 15:43:02 -07005568# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005569config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005570 name = "xnn_enable_assembly_explicit_true",
5571 define_values = {"xnn_enable_assembly": "true"},
5572)
5573
5574# Disables usage of assembly kernels.
5575config_setting(
5576 name = "xnn_enable_assembly_explicit_false",
5577 define_values = {"xnn_enable_assembly": "false"},
5578)
5579
Marat Dukhan9de90e02020-06-18 16:04:12 -07005580# Enables usage of sparse inference.
5581config_setting(
5582 name = "xnn_enable_sparse_explicit_true",
5583 define_values = {"xnn_enable_sparse": "true"},
5584)
5585
5586# Disables usage of sparse inference.
5587config_setting(
5588 name = "xnn_enable_sparse_explicit_false",
5589 define_values = {"xnn_enable_sparse": "false"},
5590)
5591
Marat Dukhan05702cf2020-03-26 15:41:33 -07005592# Disables usage of HMP-aware optimizations.
5593config_setting(
5594 name = "xnn_enable_hmp_explicit_false",
5595 define_values = {"xnn_enable_hmp": "false"},
5596)
5597
Chao Mei6ddfc602020-05-13 22:29:36 -07005598# Enable usage of optimized memory allocation
5599config_setting(
5600 name = "xnn_enable_memopt_explicit_true",
5601 define_values = {"xnnpack_enable_memopt": "true"},
5602)
5603
5604# Disable usage of optimized memory allocation
5605config_setting(
5606 name = "xnn_enable_memopt_explicit_false",
5607 define_values = {"xnnpack_enable_memopt": "false"},
5608)
5609
Marat Dukhanb8642352019-10-30 15:43:02 -07005610# Builds with -c dbg
5611config_setting(
5612 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005613 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005614 "compilation_mode": "dbg",
5615 },
5616)
5617
5618# Builds with -c opt
5619config_setting(
5620 name = "optimized_build",
5621 values = {
5622 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005623 },
5624)
5625
5626config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005627 name = "linux_k8",
5628 values = {"cpu": "k8"},
5629)
5630
5631config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005632 name = "linux_arm",
5633 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005634)
5635
5636config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005637 name = "linux_armeabi",
5638 values = {"cpu": "armeabi"},
5639)
5640
5641config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005642 name = "linux_armhf",
5643 values = {"cpu": "armhf"},
5644)
5645
5646config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005647 name = "linux_armv7a",
5648 values = {"cpu": "armv7a"},
5649)
5650
5651config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005652 name = "linux_aarch64",
5653 values = {"cpu": "aarch64"},
5654)
5655
5656config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005657 name = "android",
5658 values = {"crosstool_top": "//external:android/crosstool"},
5659)
5660
5661config_setting(
5662 name = "android_armv7",
5663 values = {
5664 "crosstool_top": "//external:android/crosstool",
5665 "cpu": "armeabi-v7a",
5666 },
5667)
5668
5669config_setting(
5670 name = "android_arm64",
5671 values = {
5672 "crosstool_top": "//external:android/crosstool",
5673 "cpu": "arm64-v8a",
5674 },
5675)
5676
5677config_setting(
5678 name = "android_x86",
5679 values = {
5680 "crosstool_top": "//external:android/crosstool",
5681 "cpu": "x86",
5682 },
5683)
5684
5685config_setting(
5686 name = "android_x86_64",
5687 values = {
5688 "crosstool_top": "//external:android/crosstool",
5689 "cpu": "x86_64",
5690 },
5691)
5692
5693config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005694 name = "windows_x86_64",
5695 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005696)
5697
5698config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005699 name = "windows_x86_64_clang",
5700 values = {
5701 "compiler": "clang-cl",
5702 "cpu": "x64_windows",
5703 },
5704)
5705
5706config_setting(
5707 name = "windows_x86_64_mingw",
5708 values = {
5709 "compiler": "mingw-gcc",
5710 "cpu": "x64_windows",
5711 },
5712)
5713
5714config_setting(
5715 name = "windows_x86_64_msys",
5716 values = {
5717 "compiler": "msys-gcc",
5718 "cpu": "x64_windows",
5719 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005720)
5721
5722config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005723 name = "macos_x86_64",
5724 values = {
5725 "apple_platform_type": "macos",
5726 "cpu": "darwin",
5727 },
5728)
5729
5730config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005731 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005732 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005733)
5734
5735config_setting(
5736 name = "emscripten_wasm",
5737 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005738 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005739 "cpu": "wasm",
5740 },
5741)
5742
5743config_setting(
5744 name = "emscripten_wasmsimd",
5745 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005746 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005747 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005748 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005749 },
5750)
5751
5752config_setting(
5753 name = "emscripten_asmjs",
5754 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005755 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005756 "cpu": "asmjs",
5757 },
5758)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005759
5760config_setting(
5761 name = "ios_armv7",
5762 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005763 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005764 "cpu": "ios_armv7",
5765 },
5766)
5767
5768config_setting(
5769 name = "ios_arm64",
5770 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005771 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005772 "cpu": "ios_arm64",
5773 },
5774)
5775
5776config_setting(
5777 name = "ios_arm64e",
5778 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005779 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005780 "cpu": "ios_arm64e",
5781 },
5782)
5783
5784config_setting(
5785 name = "ios_x86",
5786 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005787 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005788 "cpu": "ios_i386",
5789 },
5790)
5791
5792config_setting(
5793 name = "ios_x86_64",
5794 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005795 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005796 "cpu": "ios_x86_64",
5797 },
5798)
5799
5800config_setting(
5801 name = "watchos_armv7k",
5802 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005803 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005804 "cpu": "watchos_armv7k",
5805 },
5806)
5807
5808config_setting(
5809 name = "watchos_arm64_32",
5810 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005811 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005812 "cpu": "watchos_arm64_32",
5813 },
5814)
5815
5816config_setting(
5817 name = "watchos_x86",
5818 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005819 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005820 "cpu": "watchos_i386",
5821 },
5822)
5823
5824config_setting(
5825 name = "watchos_x86_64",
5826 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005827 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005828 "cpu": "watchos_x86_64",
5829 },
5830)
5831
5832config_setting(
5833 name = "tvos_arm64",
5834 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005835 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005836 "cpu": "tvos_arm64",
5837 },
5838)
5839
5840config_setting(
5841 name = "tvos_x86_64",
5842 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005843 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005844 "cpu": "tvos_x86_64",
5845 },
5846)