blob: c610f091198cb4493d5eeb106c39586cb62c19cc [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 Dukhanefaac272020-07-16 16:58:07 -0700666 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
667 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
668 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700669 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
670 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700671 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
672 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
673 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
674 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700675 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
676 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
677 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
678 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700679 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
680 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
681 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
682 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
683 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
684 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
685 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
686 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
687 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
688 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
689 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
690 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700691 "src/f32-relu/gen/wasmsimd-x4.c",
692 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700693 "src/f32-rmax/wasmsimd-arm.c",
694 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700695 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
696 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
697 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
698 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
699 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
700 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700701 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
702 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
703 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
704 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700705 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
706 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700707 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
708 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
709 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
710 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700711 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
712 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700713 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
714 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
715 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
716 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700717 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
718 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700719 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
720 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
721 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
722 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700723 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
724 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700725 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
726 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
727 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
728 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700729 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
730 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
731 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
732 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700733 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
734 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
735 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
736 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700737 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
738 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
739 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
740 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700741 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700747 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700751 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
752 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700753 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
754 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
755 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
756 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700757 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700763 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
764 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700765 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
766 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700767 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
768 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700769 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
770 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
771 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
772 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700773 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
774 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700775 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
776 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
777 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
778 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700779 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
780 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700781 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
782 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
783 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
784 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700785 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
786 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
787 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
788 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700789 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
790 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
791 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
792 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
793 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
794 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
795 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
796 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
797 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
798 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
799 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
800 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
801 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
802 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700803 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
804 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700805 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
806 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
807 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
808 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
809 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
810 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700811 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700812 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700813 "src/x32-pad/wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700814 "src/math/roundne-wasmsimd-addsub.c",
815 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700816 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700817 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700818 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700819 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700820 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700821 "src/requantization/fp32-wasmsimd.c",
822]
823
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800824PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800825 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800826 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800827 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700828 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
829 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700830 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700831 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700832 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700833 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700834 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700835 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
836 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
837 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
838 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
839 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
840 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
841 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
842 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
843 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
844 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
845 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
846 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700847 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700848 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
849 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700850 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
851 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
852 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
853 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
854 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
855 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
856 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
857 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
858 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
859 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
860 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
861 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
862 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
863 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
864 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
865 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
866 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
867 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
868 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700869 "src/f32-ibilinear/gen/psimd-c4.c",
870 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700871 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
872 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
873 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
874 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
875 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
876 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
877 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
878 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
879 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
880 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700881 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700882 "src/f32-spmm/gen/4x1-minmax-psimd.c",
883 "src/f32-spmm/gen/8x1-minmax-psimd.c",
884 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700885 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
886 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700887 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800888 "src/f32-prelu/gen/psimd-2x4.c",
889 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800890 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700891 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
892 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
893 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
894 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
895 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
896 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
897 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
898 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800899 "src/f32-vbinary/gen/vmax-psimd-x4.c",
900 "src/f32-vbinary/gen/vmax-psimd-x8.c",
901 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
902 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
903 "src/f32-vbinary/gen/vmin-psimd-x4.c",
904 "src/f32-vbinary/gen/vmin-psimd-x8.c",
905 "src/f32-vbinary/gen/vminc-psimd-x4.c",
906 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700907 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
908 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
909 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
910 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
911 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
912 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
913 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
914 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700915 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
916 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
917 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
918 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700919 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
920 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
921 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
922 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700923 "src/f32-vlrelu/gen/vlrelu-psimd-x4.c",
924 "src/f32-vlrelu/gen/vlrelu-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700925 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
926 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700927 "src/f32-vunary/gen/vabs-psimd-x4.c",
928 "src/f32-vunary/gen/vabs-psimd-x8.c",
929 "src/f32-vunary/gen/vneg-psimd-x4.c",
930 "src/f32-vunary/gen/vneg-psimd-x8.c",
931 "src/f32-vunary/gen/vsqr-psimd-x4.c",
932 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700933 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700934 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700935 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700936 "src/x32-unpool/psimd.c",
937 "src/x32-zip/x2-psimd.c",
938 "src/x32-zip/x3-psimd.c",
939 "src/x32-zip/x4-psimd.c",
940 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700941 "src/requantization/precise-psimd.c",
942 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700943]
944
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800945PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800946 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
947 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
948 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
949 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
950 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
951 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
952 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
953 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
954 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
955 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
956 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
957 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800958 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
959 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
960 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
961 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
962 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
963 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700964 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
965 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
966 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
967 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
968 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
969 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
970 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
971 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700972 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700973 "src/math/roundd-psimd-addsub.c",
974 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700975 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800976 "src/math/sigmoid-psimd-p5-div.c",
977]
978
Marat Dukhan08c4a432019-10-03 09:29:21 -0700979# ISA-specific micro-kernels
980NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700981 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
982 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700983 "src/f32-clamp/gen/neon-x4.c",
984 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700985 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
986 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700987 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
988 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700989 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
990 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700991 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
992 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700993 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
994 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
995 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
996 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700997 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
998 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
999 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1000 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001001 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1002 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1003 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1004 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001005 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001006 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1007 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001008 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1009 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1010 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1011 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1012 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1013 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1014 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1015 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1016 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1017 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1018 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1019 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1020 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1021 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1022 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1023 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1024 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1025 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1026 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1027 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1028 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1029 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1030 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1031 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1032 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1033 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1034 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1035 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1036 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1037 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1038 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001039 "src/f32-hswish/gen/hswish-neon-x4.c",
1040 "src/f32-hswish/gen/hswish-neon-x8.c",
1041 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001042 "src/f32-ibilinear/gen/neon-c4.c",
1043 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001044 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1045 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1046 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1047 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1048 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1049 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1050 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1051 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1052 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1053 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1054 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1055 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1056 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1057 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1058 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1059 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001060 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1061 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1062 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001063 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1064 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001065 "src/f32-prelu/gen/neon-2x4.c",
1066 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001067 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1068 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1069 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1070 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1071 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1072 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1073 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1074 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1075 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1076 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1077 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1078 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1079 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1080 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1081 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1082 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1083 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1084 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1085 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1086 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1087 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1088 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1089 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1090 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001091 "src/f32-relu/gen/neon-x4.c",
1092 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001093 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001094 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001095 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1096 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1097 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1098 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1099 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1100 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1101 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1102 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1103 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1104 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1105 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1106 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1107 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1108 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1109 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1110 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1111 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1112 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001113 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1114 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1115 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1116 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001117 "src/f32-vbinary/gen/vmax-neon-x4.c",
1118 "src/f32-vbinary/gen/vmax-neon-x8.c",
1119 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1120 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1121 "src/f32-vbinary/gen/vmin-neon-x4.c",
1122 "src/f32-vbinary/gen/vmin-neon-x8.c",
1123 "src/f32-vbinary/gen/vminc-neon-x4.c",
1124 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001125 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1126 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1127 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1128 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1129 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1130 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001131 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1132 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1133 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1134 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001135 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1136 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1137 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1138 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001139 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1140 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001141 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1142 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001143 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1144 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1145 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1146 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1147 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1148 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1149 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1150 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001151 "src/f32-vunary/gen/vabs-neon-x4.c",
1152 "src/f32-vunary/gen/vabs-neon-x8.c",
1153 "src/f32-vunary/gen/vneg-neon-x4.c",
1154 "src/f32-vunary/gen/vneg-neon-x8.c",
1155 "src/f32-vunary/gen/vsqr-neon-x4.c",
1156 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001157 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1158 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1159 "src/qu8-dwconv/up8x9-minmax-neon.c",
1160 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1161 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1162 "src/qu8-gemm/4x8-minmax-neon.c",
1163 "src/qu8-gemm/8x8-minmax-neon.c",
1164 "src/qu8-igemm/4x8-minmax-neon.c",
1165 "src/qu8-igemm/8x8-minmax-neon.c",
1166 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001167 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001168 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001169 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001170 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001171 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001172 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001173 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001174 "src/x32-zip/x2-neon.c",
1175 "src/x32-zip/x3-neon.c",
1176 "src/x32-zip/x4-neon.c",
1177 "src/x32-zip/xm-neon.c",
1178 "src/x8-zip/x2-neon.c",
1179 "src/x8-zip/x3-neon.c",
1180 "src/x8-zip/x4-neon.c",
1181 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001182 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001183 "src/math/roundd-neon-addsub.c",
1184 "src/math/roundd-neon-cvt.c",
1185 "src/math/roundu-neon-addsub.c",
1186 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001187 "src/math/roundz-neon-addsub.c",
1188 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001189 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001190 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1191 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1192 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1193 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1194 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1195 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001196 "src/math/sqrt-neon-nr1rsqrts.c",
1197 "src/math/sqrt-neon-nr2rsqrts.c",
1198 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001199 "src/requantization/precise-neon.c",
1200 "src/requantization/fp32-neon.c",
1201 "src/requantization/q31-neon.c",
1202 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001203]
1204
1205NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001206 "src/f32-ibilinear/gen/neonfma-c4.c",
1207 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001208 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1209 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1210 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1211 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1212 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1213 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1214 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1215 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1216 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001217 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1218 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1219 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1220 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001221 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1222 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1223 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1224 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001225 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1226 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1227 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1228 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001229 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1230 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1231 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1232 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1233 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1234 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1235 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1236 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1237 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1238 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1239 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1240 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1241 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1242 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1243 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1244 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1245 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1246 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001247 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1248 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001249 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1250 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1251 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1252 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1253 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1254 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1255 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1256 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1257 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1258 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1259 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1260 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1261 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1262 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1263 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1264 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1265 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1266 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1267 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1268 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1269 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1270 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1271 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1272 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001273 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1275 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1276 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1277 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1278 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1279 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1280 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1281 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1282 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1283 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1284 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1285 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1286 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1287 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1288 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1289 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1290 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1291 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1292 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1293 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1294 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1295 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1296 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1297 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1298 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1299 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1300 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1301 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1302 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1303 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1304 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1305 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1306 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1307 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1308 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1309 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1310 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1311 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1312 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1313 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1314 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1315 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1316 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1318 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1319 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1320 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1321 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1322 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1323 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1325 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1326 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001327 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1328 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001329 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1330 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1331 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1332 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1333 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1334 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1335 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1336 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1337 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1338 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1339 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1340 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1341 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1342 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1343 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1344 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1345 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1346 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1347 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1348 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001349 "src/math/exp-neonfma-lut64-p2.c",
1350 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001351 "src/math/expminus-neonfma-lut2048-p1.c",
1352 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001353 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001354 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1355 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1356 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1357 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1358 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1359 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1360 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1361 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1362 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1363 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1364 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1365 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1366 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1367 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1368 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1369 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1370 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1371 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001372 "src/math/sqrt-neonfma-nr1fma.c",
1373 "src/math/sqrt-neonfma-nr2fma.c",
1374 "src/math/sqrt-neonfma-nr3fma.c",
1375 "src/math/sqrt-neonfma-nr2fma1adj.c",
1376 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001377]
1378
1379AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001380 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1381 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1382 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1383 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1384 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1385 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001386 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1387 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1388 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1389 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1390 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1391 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1392 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1393 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1394 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1395 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1396 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1397 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1398 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1399 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1400 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1401 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1402 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1403 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1404 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1405 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001406 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1407 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001408 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1409 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001410 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1411 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001412 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1413 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001414 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1415 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1416 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1417 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1418 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001419 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1420 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1421 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1422 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1423 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1424 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1425 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1426 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1427 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1428 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1429 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1430 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1431 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1432 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1433 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1434 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1435 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1436 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001437 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1438 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1439 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1440 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1441 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1442 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1443 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1444 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1445 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1446 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1447 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1448 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1449 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1450 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1451 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1452 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1453 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1454 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001455 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1456 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001457 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1458 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1459 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1460 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1461 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1462 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001463]
1464
Marat Dukhan8853b822020-05-07 12:19:01 -07001465NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001466 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1467 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1468 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1469 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1470 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1471 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1472 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1473 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001474 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001475 "src/math/roundd-neonv8.c",
1476 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001477 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001478]
1479
Marat Dukhan08c4a432019-10-03 09:29:21 -07001480AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001481 "src/f16-clamp/gen/neonfp16arith-x16.c",
1482 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001483 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1484 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1485 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1486 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1487 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1488 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1489 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1490 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1491 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1492 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1493 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1494 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001495 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1496 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001497 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1498 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001499 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1500 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001501 "src/f16-relu/gen/neonfp16arith-x16.c",
1502 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001503 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1504 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1505 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1506 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1507 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1508 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1509 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1510 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1511 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1512 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1513 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1514 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1515 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1516 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1517 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1518 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1519 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1520 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1521 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1522 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1523 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1524 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1525 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1526 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1527 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1528 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1529 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1530 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001531 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1532 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001533 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1534 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1535 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1536 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1537 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1538 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1539 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1540 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001541 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1542 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1543 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1544 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001545 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1546 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1547 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1548 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1549 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1550 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1551 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1552 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1553 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1554 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1555 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1556 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001557 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1558 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1559 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1560 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1561 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1562 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1563 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1564 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001565]
1566
1567SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001568 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1569 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001570 "src/f32-clamp/gen/sse-x4.c",
1571 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001572 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1573 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001574 "src/f32-dwconv-chw/3x3p1-sse.c",
1575 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001576 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1577 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1578 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1579 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1580 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1581 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1582 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1583 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1584 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1585 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1586 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1587 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001588 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001589 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1590 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001591 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1592 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1593 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1594 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1595 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1596 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1597 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1598 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1599 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1600 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1601 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1602 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1603 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001604 "src/f32-hswish/gen/hswish-sse-x4.c",
1605 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001606 "src/f32-ibilinear/gen/sse-c4.c",
1607 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001608 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1609 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1610 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1611 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1612 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1613 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1614 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001615 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1616 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1617 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001618 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001619 "src/f32-prelu/gen/sse-2x4.c",
1620 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001621 "src/f32-relu/gen/sse-x4.c",
1622 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001623 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001624 "src/f32-spmm/gen/4x1-minmax-sse.c",
1625 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001626 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001627 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1628 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1629 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1630 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1631 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1632 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1633 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1634 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001635 "src/f32-vbinary/gen/vmax-sse-x4.c",
1636 "src/f32-vbinary/gen/vmax-sse-x8.c",
1637 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1638 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1639 "src/f32-vbinary/gen/vmin-sse-x4.c",
1640 "src/f32-vbinary/gen/vmin-sse-x8.c",
1641 "src/f32-vbinary/gen/vminc-sse-x4.c",
1642 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001643 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1644 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1645 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1646 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1647 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1648 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1649 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1650 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001651 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1652 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1653 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1654 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001655 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1656 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1657 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1658 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001659 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1660 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001661 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1662 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001663 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1664 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001665 "src/f32-vunary/gen/vabs-sse-x4.c",
1666 "src/f32-vunary/gen/vabs-sse-x8.c",
1667 "src/f32-vunary/gen/vneg-sse-x4.c",
1668 "src/f32-vunary/gen/vneg-sse-x8.c",
1669 "src/f32-vunary/gen/vsqr-sse-x4.c",
1670 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001671 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001672 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001673 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001674 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001675 "src/math/roundd-sse-addsub.c",
1676 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001677 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001678 "src/math/sqrt-sse-hh1mac.c",
1679 "src/math/sqrt-sse-nr1mac.c",
1680 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001681]
1682
1683SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001684 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1685 "src/f32-argmaxpool/4x-sse2-c4.c",
1686 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001687 "src/f32-prelu/gen/sse2-2x4.c",
1688 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001689 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1690 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1691 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1692 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1693 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1694 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1695 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1696 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1697 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1698 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1699 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1700 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001701 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001702 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001703 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001704 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001705 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1706 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001707 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1708 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001709 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1710 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1711 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1712 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1713 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1714 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1715 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1716 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001717 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1718 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1719 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1720 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1721 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1722 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1723 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1724 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1725 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001726 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001727 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001728 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001729 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730 "src/x32-zip/x2-sse2.c",
1731 "src/x32-zip/x3-sse2.c",
1732 "src/x32-zip/x4-sse2.c",
1733 "src/x32-zip/xm-sse2.c",
1734 "src/x8-zip/x2-sse2.c",
1735 "src/x8-zip/x3-sse2.c",
1736 "src/x8-zip/x4-sse2.c",
1737 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001738 "src/math/exp-sse2-p5.c",
1739 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001740 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001741 "src/math/roundd-sse2-cvt.c",
1742 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001743 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001744 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001745 "src/requantization/precise-sse2.c",
1746 "src/requantization/fp32-sse2.c",
1747 "src/requantization/q31-sse2.c",
1748 "src/requantization/gemmlowp-sse2.c",
1749]
1750
1751SSSE3_UKERNELS = [
1752 "src/requantization/precise-ssse3.c",
1753 "src/requantization/q31-ssse3.c",
1754 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001755]
1756
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001757SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001758 "src/f32-prelu/gen/sse41-2x4.c",
1759 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001760 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001761 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001762 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001763 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001764 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1765 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001766 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1767 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001768 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1769 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1770 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1771 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1772 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1773 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1774 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1775 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001776 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001777 "src/math/roundd-sse41.c",
1778 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001779 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001780 "src/requantization/precise-sse4.c",
1781 "src/requantization/q31-sse4.c",
1782 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001783]
1784
Marat Dukhan08c4a432019-10-03 09:29:21 -07001785AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001786 "src/f32-clamp/gen/avx-x8.c",
1787 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001788 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1789 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1790 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1791 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1792 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1793 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1794 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1795 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1796 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1797 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1798 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1799 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1800 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1801 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1802 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1803 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1804 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1805 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1806 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1807 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1808 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1809 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1810 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1811 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1812 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1813 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1814 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1815 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1816 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1817 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001818 "src/f32-hswish/gen/hswish-avx-x8.c",
1819 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001820 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1821 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1822 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1823 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1824 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1825 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1826 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1827 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1828 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001829 "src/f32-prelu/gen/avx-2x8.c",
1830 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001831 "src/f32-relu/gen/avx-x16.c",
1832 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001833 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001834 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1835 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1836 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1837 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1838 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1839 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1840 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1841 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001842 "src/f32-vbinary/gen/vmax-avx-x8.c",
1843 "src/f32-vbinary/gen/vmax-avx-x16.c",
1844 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1845 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1846 "src/f32-vbinary/gen/vmin-avx-x8.c",
1847 "src/f32-vbinary/gen/vmin-avx-x16.c",
1848 "src/f32-vbinary/gen/vminc-avx-x8.c",
1849 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001850 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1851 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1852 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1853 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1854 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1855 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1856 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1857 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001858 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1859 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1860 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1861 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001862 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1863 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1864 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1865 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001866 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1867 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001868 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001869 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1870 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1871 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1872 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1873 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1874 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1875 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1876 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001877 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1878 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001879 "src/f32-vunary/gen/vabs-avx-x8.c",
1880 "src/f32-vunary/gen/vabs-avx-x16.c",
1881 "src/f32-vunary/gen/vneg-avx-x8.c",
1882 "src/f32-vunary/gen/vneg-avx-x16.c",
1883 "src/f32-vunary/gen/vsqr-avx-x8.c",
1884 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001885]
1886
Marat Dukhanfda12b82019-11-21 12:27:59 -08001887FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001888 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1889 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1890 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1891 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1892 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1893 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1894 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1895 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1896 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1897 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1898 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1899 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1900 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1901 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1902 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1903 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1904 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1905 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1906 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1907 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1908 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1909 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1910 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1911 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1912 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1913 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1914 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1915 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1916 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1917 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1918 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1919 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1920 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1921 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1922 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1923 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1924 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1925 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1926 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1927 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001928 "src/f32-hswish/gen/hswish-fma3-x8.c",
1929 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001930 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1931 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1932 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1933 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1934 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1935 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1936 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1937 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1938 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1939 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1940 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1941 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1942 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1943 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001944 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1945 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1946 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1947 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1948 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1949 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1950 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1951 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001952 "src/math/sqrt-fma3-nr1fma.c",
1953 "src/math/sqrt-fma3-nr2fma.c",
1954 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001955]
1956
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001957AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001958 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1959 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1960 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1961 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1962 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1963 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1964 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1965 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1966 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1967 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1968 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1969 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1970 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1971 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1972 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1973 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1974 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1975 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1976 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1977 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1978 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1979 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1980 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1981 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1982 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1983 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1984 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1985 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1986 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1987 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1988 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1989 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1990 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1991 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1992 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1993 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001994 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1995 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1996 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1997 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1998 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1999 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2000 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2001 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2002 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2003 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2004 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2005 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2006 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2007 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2008 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2009 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2010 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2011 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2012 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2013 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2014 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2015 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2016 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2017 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2018 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2019 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2020 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2021 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2022 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2023 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002024 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2025 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2026 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2027 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2028 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2029 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2030 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2031 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2032 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2033 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2034 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2035 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2036 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2037 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2038 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2039 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2040 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2041 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2042 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2043 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2044 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2045 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2046 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2047 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002048 "src/math/exp-avx2-p5.c",
2049 "src/math/exp-avx2-perm-p3.c",
2050 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002051 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002052 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002053 "src/math/sigmoid-avx2-rr2-p5-div.c",
2054 "src/math/sigmoid-avx2-rr1-p5-div.c",
2055 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2056 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2057 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2058 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002059]
2060
Marat Dukhan08c4a432019-10-03 09:29:21 -07002061AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002062 "src/f32-clamp/gen/avx512f-x16.c",
2063 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002064 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2065 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2066 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2067 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2068 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2069 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2070 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2071 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2072 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2073 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2074 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2075 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2076 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2077 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2078 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2079 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2080 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2081 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2082 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2083 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2084 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2085 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2086 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2087 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002088 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2089 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002090 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2091 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2092 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2093 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2094 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2095 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002096 "src/f32-prelu/gen/avx512f-2x16.c",
2097 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002098 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2099 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2100 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2101 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2102 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2103 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2104 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2105 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2106 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2107 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2108 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2109 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2110 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2111 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2112 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2113 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2114 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2115 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2116 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2117 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2118 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2119 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2120 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2121 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2122 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2123 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2124 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2125 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2126 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2127 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2128 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2129 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2130 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2131 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2132 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2133 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002134 "src/f32-relu/gen/avx512f-x16.c",
2135 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002136 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002137 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2138 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2139 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2140 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2141 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2142 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2143 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2144 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002145 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2146 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2147 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2148 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2149 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2150 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2151 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2152 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002153 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2154 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2155 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2156 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2157 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2158 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2159 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2160 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002161 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2162 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2163 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2164 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002165 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2166 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2167 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2168 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002169 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2170 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002171 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002172 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2173 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2174 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2175 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2176 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2177 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2178 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2179 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2180 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2181 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2182 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2183 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2184 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2185 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2186 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2187 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2188 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2189 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2190 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2191 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2192 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2193 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2194 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2195 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002196 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2197 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2198 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2199 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2200 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2201 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2202 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2203 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002204 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2205 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2206 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2207 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2208 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2209 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2210 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2211 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002212 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2213 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2214 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2215 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2216 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2217 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002218 "src/math/exp-avx512f-p5-scalef.c",
2219 "src/math/exp-avx512f-p5.c",
2220 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002221 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002222 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002223 "src/math/sqrt-avx512f-nr1fma.c",
2224 "src/math/sqrt-avx512f-nr2fma.c",
2225 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002226]
2227
2228AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002229 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002230 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002231 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2232 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2233 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2234 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002235 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2236 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2237 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2238 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002239 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2240 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2241 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2242 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243]
2244
2245AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002246 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002247 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2248 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002249 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002250 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2251 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002252 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2253 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2254 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002255 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002256 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2257 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2258 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002259 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002260 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2261 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002262 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002263 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2264 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2265 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2266 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2267 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2268 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2269 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2270 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2271 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2272 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2273 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2274 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2275 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2276 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2277 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2278 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2279 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2280 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002281 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2282 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002283 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002284 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2285 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2286 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2287 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2288 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2289 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2290 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2291 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2292 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2293 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2294 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2295 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2296 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2297 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2298 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2299 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2300 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2301 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002302 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2303 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2304 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2305 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2306 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2307 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2308 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2309 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2310 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2311 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2312 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2313 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2314 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2315 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2316 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2317 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2318 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2319 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002320]
2321
Marat Dukhan1b354632020-03-23 12:50:22 -07002322INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002323 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002324 "src/xnnpack/argmaxpool.h",
2325 "src/xnnpack/avgpool.h",
2326 "src/xnnpack/clamp.h",
2327 "src/xnnpack/common.h",
2328 "src/xnnpack/conv.h",
2329 "src/xnnpack/dwconv.h",
2330 "src/xnnpack/gavgpool.h",
2331 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002332 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002333 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002334 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002335 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002336 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002337 "src/xnnpack/lut.h",
2338 "src/xnnpack/math.h",
2339 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002340 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002341 "src/xnnpack/packx.h",
2342 "src/xnnpack/pad.h",
2343 "src/xnnpack/params.h",
2344 "src/xnnpack/pavgpool.h",
2345 "src/xnnpack/ppmm.h",
2346 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002347 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002348 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002349 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002350 "src/xnnpack/rmax.h",
2351 "src/xnnpack/scalar-utils.h",
2352 "src/xnnpack/spmm.h",
2353 "src/xnnpack/unpool.h",
2354 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002355 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002356 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002357 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002358 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002359 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002360 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002361 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002362]
2363
2364INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002365 "include/xnnpack.h",
2366 "src/xnnpack/allocator.h",
2367 "src/xnnpack/compute.h",
2368 "src/xnnpack/im2col.h",
2369 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002370 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002371 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002372 "src/xnnpack/operator.h",
2373 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002374 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002375 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002376 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002377 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002378]
2379
Marat Dukhan1b354632020-03-23 12:50:22 -07002380ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002381 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002382]
2383
Marat Dukhan1b354632020-03-23 12:50:22 -07002384MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002385 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002386 "include/xnnpack.h",
2387]
2388
Marat Dukhan1b354632020-03-23 12:50:22 -07002389MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002390 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002391 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002392 "src/xnnpack/requantization.h",
2393 "include/xnnpack.h",
2394]
2395
2396OPERATOR_TEST_PARAMS_HDRS = [
2397 "src/xnnpack/params.h",
2398 "src/xnnpack/common.h",
2399]
2400
2401WEIGHTS_PACK_HDRS = [
2402 "src/xnnpack/pack.h",
2403 "src/xnnpack/operator.h",
2404 "src/xnnpack/compute.h",
2405]
2406
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002407LOGGING_COPTS = select({
2408 # No logging in optimized mode
2409 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2410 # Full logging in debug mode
2411 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2412 # Error-only logging in default (fastbuild) mode
2413 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2414})
2415
Marat Dukhan3b59de22020-06-03 20:15:19 -07002416LOGGING_SRCS = select({
2417 # No logging in optimized mode
2418 ":optimized_build": [],
2419 "//conditions:default": [
2420 "src/operator-strings.c",
2421 "src/subgraph-strings.c",
2422 ],
2423})
2424
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002425LOGGING_HDRS = [
2426 "src/xnnpack/log.h",
2427]
2428
Marat Dukhan08c4a432019-10-03 09:29:21 -07002429xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002430 name = "tables",
2431 srcs = TABLE_SRCS,
2432 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002433 gcc_copts = xnnpack_gcc_std_copts(),
2434 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002435)
2436
2437xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002438 name = "scalar_ukernels",
2439 srcs = SCALAR_UKERNELS,
2440 hdrs = INTERNAL_HDRS,
2441 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002442 gcc_copts = xnnpack_gcc_std_copts(),
2443 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002444 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002445 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002446 "@FP16",
2447 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002448 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002449 ],
2450)
2451
2452xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002453 name = "scalar_ukernels_test_mode",
2454 srcs = SCALAR_UKERNELS,
2455 hdrs = INTERNAL_HDRS,
2456 aarch32_copts = ["-marm"],
2457 copts = [
2458 "-UNDEBUG",
2459 "-DXNN_TEST_MODE=1",
2460 ],
2461 gcc_copts = xnnpack_gcc_std_copts(),
2462 msvc_copts = xnnpack_msvc_std_copts(),
2463 deps = [
2464 ":tables",
2465 "@FP16",
2466 "@FXdiv",
2467 "@pthreadpool",
2468 ],
2469)
2470
2471xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002472 name = "wasm_ukernels",
2473 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002474 gcc_copts = xnnpack_gcc_std_copts(),
2475 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002476 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002477 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002478 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002479 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002480 "@FP16",
2481 "@FXdiv",
2482 "@pthreadpool",
2483 ],
2484)
2485
2486xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002487 name = "wasm_ukernels_test_mode",
2488 hdrs = INTERNAL_HDRS,
2489 copts = [
2490 "-UNDEBUG",
2491 "-DXNN_TEST_MODE=1",
2492 ],
2493 gcc_copts = xnnpack_gcc_std_copts(),
2494 msvc_copts = xnnpack_msvc_std_copts(),
2495 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002496 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002497 deps = [
2498 ":tables",
2499 "@FP16",
2500 "@FXdiv",
2501 "@pthreadpool",
2502 ],
2503)
2504
2505xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002506 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002507 hdrs = INTERNAL_HDRS,
2508 aarch32_copts = [
2509 "-marm",
2510 "-mfpu=neon",
2511 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002512 gcc_copts = xnnpack_gcc_std_copts(),
2513 gcc_x86_copts = ["-msse2"],
2514 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002515 optimized_copts = [
2516 "-O3",
2517 "-ffast-math",
2518 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002519 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002520 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002521 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002522 "@FP16",
2523 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002524 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002525 ],
2526)
2527
2528xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002529 name = "psimd_fastmath_ukernels_test_mode",
2530 hdrs = INTERNAL_HDRS,
2531 aarch32_copts = [
2532 "-marm",
2533 "-mfpu=neon",
2534 ],
2535 copts = [
2536 "-UNDEBUG",
2537 "-DXNN_TEST_MODE=1",
2538 ],
2539 gcc_copts = xnnpack_gcc_std_copts(),
2540 gcc_x86_copts = ["-msse2"],
2541 msvc_copts = xnnpack_msvc_std_copts(),
2542 optimized_copts = [
2543 "-O3",
2544 "-ffast-math",
2545 ],
2546 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2547 deps = [
2548 ":tables",
2549 "@FP16",
2550 "@psimd",
2551 "@pthreadpool",
2552 ],
2553)
2554
2555xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002556 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002557 hdrs = INTERNAL_HDRS,
2558 aarch32_copts = [
2559 "-marm",
2560 "-mfpu=neon",
2561 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002562 gcc_copts = xnnpack_gcc_std_copts(),
2563 gcc_x86_copts = ["-msse2"],
2564 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002565 optimized_copts = [
2566 "-O3",
2567 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002568 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002569 deps = [
2570 ":tables",
2571 "@FP16",
2572 "@psimd",
2573 "@pthreadpool",
2574 ],
2575)
2576
2577xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002578 name = "psimd_accmath_ukernels_test_mode",
2579 hdrs = INTERNAL_HDRS,
2580 aarch32_copts = [
2581 "-marm",
2582 "-mfpu=neon",
2583 ],
2584 copts = [
2585 "-UNDEBUG",
2586 "-DXNN_TEST_MODE=1",
2587 ],
2588 gcc_copts = xnnpack_gcc_std_copts(),
2589 gcc_x86_copts = ["-msse2"],
2590 msvc_copts = xnnpack_msvc_std_copts(),
2591 optimized_copts = [
2592 "-O3",
2593 ],
2594 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2595 deps = [
2596 ":tables",
2597 "@FP16",
2598 "@psimd",
2599 "@pthreadpool",
2600 ],
2601)
2602
2603xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002604 name = "neon_ukernels",
2605 hdrs = INTERNAL_HDRS,
2606 aarch32_copts = [
2607 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002608 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002609 "-mfpu=neon",
2610 ],
2611 aarch32_srcs = NEON_UKERNELS,
2612 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002613 gcc_copts = xnnpack_gcc_std_copts(),
2614 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002615 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002616 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002617 "@FP16",
2618 "@pthreadpool",
2619 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002620)
2621
2622xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002623 name = "neon_ukernels_test_mode",
2624 hdrs = INTERNAL_HDRS,
2625 aarch32_copts = [
2626 "-marm",
2627 "-march=armv7-a",
2628 "-mfpu=neon",
2629 ],
2630 aarch32_srcs = NEON_UKERNELS,
2631 aarch64_srcs = NEON_UKERNELS,
2632 copts = [
2633 "-UNDEBUG",
2634 "-DXNN_TEST_MODE=1",
2635 ],
2636 gcc_copts = xnnpack_gcc_std_copts(),
2637 msvc_copts = xnnpack_msvc_std_copts(),
2638 deps = [
2639 ":tables",
2640 "@FP16",
2641 "@pthreadpool",
2642 ],
2643)
2644
2645xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002646 name = "neonfma_ukernels",
2647 hdrs = INTERNAL_HDRS,
2648 aarch32_copts = [
2649 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002650 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002651 "-mfpu=neon-vfpv4",
2652 ],
2653 aarch32_srcs = NEONFMA_UKERNELS,
2654 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002655 apple_aarch32_copts = [
2656 "-mcpu=swift",
2657 "-mtune=generic",
2658 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002659 gcc_copts = xnnpack_gcc_std_copts(),
2660 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002661 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002662 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002663 "@FP16",
2664 "@pthreadpool",
2665 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002666)
2667
2668xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002669 name = "neonfma_ukernels_test_mode",
2670 hdrs = INTERNAL_HDRS,
2671 aarch32_copts = [
2672 "-marm",
2673 "-march=armv7-a",
2674 "-mfpu=neon-vfpv4",
2675 ],
2676 aarch32_srcs = NEONFMA_UKERNELS,
2677 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002678 apple_aarch32_copts = [
2679 "-mcpu=swift",
2680 "-mtune=generic",
2681 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002682 copts = [
2683 "-UNDEBUG",
2684 "-DXNN_TEST_MODE=1",
2685 ],
2686 gcc_copts = xnnpack_gcc_std_copts(),
2687 msvc_copts = xnnpack_msvc_std_copts(),
2688 deps = [
2689 ":tables",
2690 "@FP16",
2691 "@pthreadpool",
2692 ],
2693)
2694
2695xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002696 name = "neonv8_ukernels",
2697 hdrs = INTERNAL_HDRS,
2698 aarch32_copts = [
2699 "-marm",
2700 "-march=armv8-a",
2701 "-mfpu=neon-fp-armv8",
2702 ],
2703 aarch32_srcs = NEONV8_UKERNELS,
2704 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002705 apple_aarch32_copts = [
2706 "-mcpu=cyclone",
2707 "-mtune=generic",
2708 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002709 gcc_copts = xnnpack_gcc_std_copts(),
2710 msvc_copts = xnnpack_msvc_std_copts(),
2711 deps = [
2712 ":tables",
2713 "@FP16",
2714 "@pthreadpool",
2715 ],
2716)
2717
2718xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002719 name = "neonv8_ukernels_test_mode",
2720 hdrs = INTERNAL_HDRS,
2721 aarch32_copts = [
2722 "-marm",
2723 "-march=armv8-a",
2724 "-mfpu=neon-fp-armv8",
2725 ],
2726 aarch32_srcs = NEONV8_UKERNELS,
2727 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002728 apple_aarch32_copts = [
2729 "-mcpu=cyclone",
2730 "-mtune=generic",
2731 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002732 copts = [
2733 "-UNDEBUG",
2734 "-DXNN_TEST_MODE=1",
2735 ],
2736 gcc_copts = xnnpack_gcc_std_copts(),
2737 msvc_copts = xnnpack_msvc_std_copts(),
2738 deps = [
2739 ":tables",
2740 "@FP16",
2741 "@pthreadpool",
2742 ],
2743)
2744
2745xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002746 name = "neonfp16arith_ukernels",
2747 hdrs = INTERNAL_HDRS,
2748 aarch64_copts = ["-march=armv8.2-a+fp16"],
2749 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002750 gcc_copts = xnnpack_gcc_std_copts(),
2751 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002752 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002753 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002754 "@FP16",
2755 "@pthreadpool",
2756 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002757)
2758
2759xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002760 name = "neonfp16arith_ukernels_test_mode",
2761 hdrs = INTERNAL_HDRS,
2762 aarch64_copts = ["-march=armv8.2-a+fp16"],
2763 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2764 copts = [
2765 "-UNDEBUG",
2766 "-DXNN_TEST_MODE=1",
2767 ],
2768 gcc_copts = xnnpack_gcc_std_copts(),
2769 msvc_copts = xnnpack_msvc_std_copts(),
2770 deps = [
2771 ":tables",
2772 "@FP16",
2773 "@pthreadpool",
2774 ],
2775)
2776
2777xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002778 name = "sse2_ukernels",
2779 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002780 gcc_copts = xnnpack_gcc_std_copts(),
2781 gcc_x86_copts = ["-msse2"],
2782 msvc_copts = xnnpack_msvc_std_copts(),
2783 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002784 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002785 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002786 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002787 "@FP16",
2788 "@pthreadpool",
2789 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002790)
2791
2792xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002793 name = "sse2_ukernels_test_mode",
2794 hdrs = INTERNAL_HDRS,
2795 copts = [
2796 "-UNDEBUG",
2797 "-DXNN_TEST_MODE=1",
2798 ],
2799 gcc_copts = xnnpack_gcc_std_copts(),
2800 gcc_x86_copts = ["-msse2"],
2801 msvc_copts = xnnpack_msvc_std_copts(),
2802 msvc_x86_32_copts = ["/arch:SSE2"],
2803 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2804 deps = [
2805 ":tables",
2806 "@FP16",
2807 "@pthreadpool",
2808 ],
2809)
2810
2811xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002812 name = "ssse3_ukernels",
2813 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002814 gcc_copts = xnnpack_gcc_std_copts(),
2815 gcc_x86_copts = ["-mssse3"],
2816 msvc_copts = xnnpack_msvc_std_copts(),
2817 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002818 x86_srcs = SSSE3_UKERNELS,
2819 deps = [
2820 ":tables",
2821 "@FP16",
2822 "@pthreadpool",
2823 ],
2824)
2825
2826xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002827 name = "ssse3_ukernels_test_mode",
2828 hdrs = INTERNAL_HDRS,
2829 copts = [
2830 "-UNDEBUG",
2831 "-DXNN_TEST_MODE=1",
2832 ],
2833 gcc_copts = xnnpack_gcc_std_copts(),
2834 gcc_x86_copts = ["-mssse3"],
2835 msvc_copts = xnnpack_msvc_std_copts(),
2836 msvc_x86_32_copts = ["/arch:SSE2"],
2837 x86_srcs = SSSE3_UKERNELS,
2838 deps = [
2839 ":tables",
2840 "@FP16",
2841 "@pthreadpool",
2842 ],
2843)
2844
2845xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002846 name = "sse41_ukernels",
2847 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002848 gcc_copts = xnnpack_gcc_std_copts(),
2849 gcc_x86_copts = ["-msse4.1"],
2850 msvc_copts = xnnpack_msvc_std_copts(),
2851 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002852 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002853 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002854 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002855 "@FP16",
2856 "@pthreadpool",
2857 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002858)
2859
2860xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002861 name = "sse41_ukernels_test_mode",
2862 hdrs = INTERNAL_HDRS,
2863 copts = [
2864 "-UNDEBUG",
2865 "-DXNN_TEST_MODE=1",
2866 ],
2867 gcc_copts = xnnpack_gcc_std_copts(),
2868 gcc_x86_copts = ["-msse4.1"],
2869 msvc_copts = xnnpack_msvc_std_copts(),
2870 msvc_x86_32_copts = ["/arch:SSE2"],
2871 x86_srcs = SSE41_UKERNELS,
2872 deps = [
2873 ":tables",
2874 "@FP16",
2875 "@pthreadpool",
2876 ],
2877)
2878
2879xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002880 name = "avx_ukernels",
2881 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002882 gcc_copts = xnnpack_gcc_std_copts(),
2883 gcc_x86_copts = ["-mavx"],
2884 msvc_copts = xnnpack_msvc_std_copts(),
2885 msvc_x86_32_copts = ["/arch:AVX"],
2886 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002887 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002888 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002889 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002890 "@FP16",
2891 "@pthreadpool",
2892 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002893)
2894
2895xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002896 name = "avx_ukernels_test_mode",
2897 hdrs = INTERNAL_HDRS,
2898 copts = [
2899 "-UNDEBUG",
2900 "-DXNN_TEST_MODE=1",
2901 ],
2902 gcc_copts = xnnpack_gcc_std_copts(),
2903 gcc_x86_copts = ["-mavx"],
2904 msvc_copts = xnnpack_msvc_std_copts(),
2905 msvc_x86_32_copts = ["/arch:AVX"],
2906 msvc_x86_64_copts = ["/arch:AVX"],
2907 x86_srcs = AVX_UKERNELS,
2908 deps = [
2909 ":tables",
2910 "@FP16",
2911 "@pthreadpool",
2912 ],
2913)
2914
2915xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002916 name = "fma3_ukernels",
2917 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002918 gcc_copts = xnnpack_gcc_std_copts(),
2919 gcc_x86_copts = ["-mfma"],
2920 msvc_copts = xnnpack_msvc_std_copts(),
2921 msvc_x86_32_copts = ["/arch:AVX"],
2922 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002923 x86_srcs = FMA3_UKERNELS,
2924 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002925 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002926 "@FP16",
2927 "@pthreadpool",
2928 ],
2929)
2930
2931xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002932 name = "fma3_ukernels_test_mode",
2933 hdrs = INTERNAL_HDRS,
2934 copts = [
2935 "-UNDEBUG",
2936 "-DXNN_TEST_MODE=1",
2937 ],
2938 gcc_copts = xnnpack_gcc_std_copts(),
2939 gcc_x86_copts = ["-mfma"],
2940 msvc_copts = xnnpack_msvc_std_copts(),
2941 msvc_x86_32_copts = ["/arch:AVX"],
2942 msvc_x86_64_copts = ["/arch:AVX"],
2943 x86_srcs = FMA3_UKERNELS,
2944 deps = [
2945 ":tables",
2946 "@FP16",
2947 "@pthreadpool",
2948 ],
2949)
2950
2951xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002952 name = "avx2_ukernels",
2953 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002954 gcc_copts = xnnpack_gcc_std_copts(),
2955 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002956 "-mfma",
2957 "-mavx2",
2958 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002959 msvc_copts = xnnpack_msvc_std_copts(),
2960 msvc_x86_32_copts = ["/arch:AVX2"],
2961 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002962 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002963 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002964 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002965 "@FP16",
2966 "@pthreadpool",
2967 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002968)
2969
2970xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002971 name = "avx2_ukernels_test_mode",
2972 hdrs = INTERNAL_HDRS,
2973 copts = [
2974 "-UNDEBUG",
2975 "-DXNN_TEST_MODE=1",
2976 ],
2977 gcc_copts = xnnpack_gcc_std_copts(),
2978 gcc_x86_copts = [
2979 "-mfma",
2980 "-mavx2",
2981 ],
2982 msvc_copts = xnnpack_msvc_std_copts(),
2983 msvc_x86_32_copts = ["/arch:AVX2"],
2984 msvc_x86_64_copts = ["/arch:AVX2"],
2985 x86_srcs = AVX2_UKERNELS,
2986 deps = [
2987 ":tables",
2988 "@FP16",
2989 "@pthreadpool",
2990 ],
2991)
2992
2993xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002994 name = "avx512f_ukernels",
2995 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002996 gcc_copts = xnnpack_gcc_std_copts(),
2997 gcc_x86_copts = ["-mavx512f"],
2998 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2999 msvc_copts = xnnpack_msvc_std_copts(),
3000 msvc_x86_32_copts = ["/arch:AVX512"],
3001 msvc_x86_64_copts = ["/arch:AVX512"],
3002 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003003 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003004 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003005 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003006 "@FP16",
3007 "@pthreadpool",
3008 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009)
3010
3011xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003012 name = "avx512f_ukernels_test_mode",
3013 hdrs = INTERNAL_HDRS,
3014 copts = [
3015 "-UNDEBUG",
3016 "-DXNN_TEST_MODE=1",
3017 ],
3018 gcc_copts = xnnpack_gcc_std_copts(),
3019 gcc_x86_copts = ["-mavx512f"],
3020 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3021 msvc_copts = xnnpack_msvc_std_copts(),
3022 msvc_x86_32_copts = ["/arch:AVX512"],
3023 msvc_x86_64_copts = ["/arch:AVX512"],
3024 msys_copts = ["-fno-asynchronous-unwind-tables"],
3025 x86_srcs = AVX512F_UKERNELS,
3026 deps = [
3027 ":tables",
3028 "@FP16",
3029 "@pthreadpool",
3030 ],
3031)
3032
3033xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003034 name = "asm_ukernels",
3035 hdrs = ["src/xnnpack/assembly.h"],
3036 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003037 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003038 aarch64_srcs = AARCH64_ASM_UKERNELS,
3039)
3040
Marat Dukhan3b59de22020-06-03 20:15:19 -07003041xnnpack_cc_library(
3042 name = "logging_utils",
3043 srcs = LOGGING_SRCS,
3044 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3045 copts = LOGGING_COPTS + [
3046 "-Isrc",
3047 "-Iinclude",
3048 ] + select({
3049 ":debug_build": [],
3050 "//conditions:default": xnnpack_min_size_copts(),
3051 }),
3052 gcc_copts = xnnpack_gcc_std_copts(),
3053 msvc_copts = xnnpack_msvc_std_copts(),
3054 visibility = xnnpack_visibility(),
3055 deps = [
3056 "@FP16",
3057 "@clog",
3058 "@pthreadpool",
3059 ],
3060)
3061
Marat Dukhan08c4a432019-10-03 09:29:21 -07003062xnnpack_aggregate_library(
3063 name = "ukernels",
3064 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003065 ":neon_ukernels",
3066 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003067 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003068 ":asm_ukernels",
3069 ],
3070 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003071 ":neon_ukernels",
3072 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003073 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003074 ":neonfp16arith_ukernels",
3075 ":asm_ukernels",
3076 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003077 generic_deps = [
3078 ":scalar_ukernels",
3079 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003080 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003081 ":psimd_fastmath_ukernels",
3082 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003083 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003084 wasm_deps = [
3085 ":wasm_ukernels",
3086 ],
3087 wasmsimd_deps = [
3088 ":wasm_ukernels",
3089 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003090 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003091 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003092 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003093 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003094 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003095 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003096 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003097 ":avx512f_ukernels",
3098 ],
3099)
3100
Marat Dukhan33fcf782020-05-24 14:27:15 -07003101xnnpack_aggregate_library(
3102 name = "ukernels_test_mode",
3103 aarch32_deps = [
3104 ":neon_ukernels_test_mode",
3105 ":neonfma_ukernels_test_mode",
3106 ":neonv8_ukernels_test_mode",
3107 ":asm_ukernels",
3108 ],
3109 aarch64_deps = [
3110 ":neon_ukernels_test_mode",
3111 ":neonfma_ukernels_test_mode",
3112 ":neonv8_ukernels_test_mode",
3113 ":neonfp16arith_ukernels_test_mode",
3114 ":asm_ukernels",
3115 ],
3116 generic_deps = [
3117 ":scalar_ukernels_test_mode",
3118 ],
3119 psimd_deps = [
3120 ":psimd_fastmath_ukernels_test_mode",
3121 ":psimd_accmath_ukernels_test_mode",
3122 ],
3123 wasm_deps = [
3124 ":wasm_ukernels_test_mode",
3125 ],
3126 wasmsimd_deps = [
3127 ":wasm_ukernels_test_mode",
3128 ],
3129 x86_deps = [
3130 ":sse2_ukernels_test_mode",
3131 ":ssse3_ukernels_test_mode",
3132 ":sse41_ukernels_test_mode",
3133 ":avx_ukernels_test_mode",
3134 ":fma3_ukernels_test_mode",
3135 ":avx2_ukernels_test_mode",
3136 ":avx512f_ukernels_test_mode",
3137 ],
3138)
3139
Marat Dukhan86df3842020-06-17 14:03:31 -07003140# This target must be compatible with Android NDK r17 and compile with Clang 6.
3141# It is similar to :ukernels target, but excludes microkernels which require
3142# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003143xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003144 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003145 aarch32_deps = [
3146 ":neon_ukernels",
3147 ":neonfma_ukernels",
3148 ":neonv8_ukernels",
3149 ":asm_ukernels",
3150 ],
3151 aarch64_deps = [
3152 ":neon_ukernels",
3153 ":neonfma_ukernels",
3154 ":neonv8_ukernels",
3155 ":asm_ukernels",
3156 ],
3157 generic_deps = [
3158 ":scalar_ukernels",
3159 ],
3160 psimd_deps = [
3161 ":psimd_fastmath_ukernels",
3162 ":psimd_accmath_ukernels",
3163 ],
3164 wasm_deps = [
3165 ":wasm_ukernels",
3166 ],
3167 wasmsimd_deps = [
3168 ":wasm_ukernels",
3169 ],
3170 x86_deps = [
3171 ":sse2_ukernels",
3172 ":ssse3_ukernels",
3173 ":sse41_ukernels",
3174 ":avx_ukernels",
3175 ":fma3_ukernels",
3176 ":avx2_ukernels",
3177 ":avx512f_ukernels",
3178 ],
3179)
3180
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181xnnpack_cc_library(
3182 name = "im2col",
3183 srcs = ["src/im2col.c"],
3184 hdrs = [
3185 "src/xnnpack/common.h",
3186 "src/xnnpack/im2col.h",
3187 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003188 gcc_copts = xnnpack_gcc_std_copts(),
3189 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190)
3191
3192xnnpack_cc_library(
3193 name = "indirection",
3194 srcs = ["src/indirection.c"],
3195 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003196 gcc_copts = xnnpack_gcc_std_copts(),
3197 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003198 deps = [
3199 "@FP16",
3200 "@FXdiv",
3201 "@pthreadpool",
3202 ],
3203)
3204
3205xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003206 name = "indirection_test_mode",
3207 srcs = ["src/indirection.c"],
3208 hdrs = INTERNAL_HDRS,
3209 copts = [
3210 "-UNDEBUG",
3211 "-DXNN_TEST_MODE=1",
3212 ],
3213 gcc_copts = xnnpack_gcc_std_copts(),
3214 msvc_copts = xnnpack_msvc_std_copts(),
3215 deps = [
3216 "@FP16",
3217 "@FXdiv",
3218 "@pthreadpool",
3219 ],
3220)
3221
3222xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003223 name = "packing",
3224 srcs = ["src/packing.c"],
3225 hdrs = INTERNAL_HDRS,
3226 gcc_copts = xnnpack_gcc_std_copts(),
3227 msvc_copts = xnnpack_msvc_std_copts(),
3228 deps = [
3229 "@FP16",
3230 "@FXdiv",
3231 "@pthreadpool",
3232 ],
3233)
3234
3235xnnpack_cc_library(
3236 name = "packing_test_mode",
3237 srcs = ["src/packing.c"],
3238 hdrs = INTERNAL_HDRS,
3239 copts = [
3240 "-UNDEBUG",
3241 "-DXNN_TEST_MODE=1",
3242 ],
3243 gcc_copts = xnnpack_gcc_std_copts(),
3244 msvc_copts = xnnpack_msvc_std_copts(),
3245 deps = [
3246 "@FP16",
3247 "@FXdiv",
3248 "@pthreadpool",
3249 ],
3250)
3251
3252xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 name = "operator_run",
3254 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003255 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003256 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003257 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3258 "//conditions:default": [],
3259 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003260 gcc_copts = xnnpack_gcc_std_copts(),
3261 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003262 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003263 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003264 "@FP16",
3265 "@FXdiv",
3266 "@clog",
3267 "@pthreadpool",
3268 ],
3269)
3270
Chao Mei6ddfc602020-05-13 22:29:36 -07003271xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003272 name = "operator_run_test_mode",
3273 srcs = ["src/operator-run.c"],
3274 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3275 copts = LOGGING_COPTS + [
3276 "-UNDEBUG",
3277 "-DXNN_TEST_MODE=1",
3278 ] + select({
3279 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3280 "//conditions:default": [],
3281 }),
3282 gcc_copts = xnnpack_gcc_std_copts(),
3283 msvc_copts = xnnpack_msvc_std_copts(),
3284 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003285 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003286 "@FP16",
3287 "@FXdiv",
3288 "@clog",
3289 "@pthreadpool",
3290 ],
3291)
3292
3293xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003294 name = "memory_planner",
3295 srcs = ["src/memory-planner.c"],
3296 hdrs = INTERNAL_HDRS,
3297 defines = select({
3298 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3299 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3300 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3301 }),
3302 gcc_copts = xnnpack_gcc_std_copts(),
3303 msvc_copts = xnnpack_msvc_std_copts(),
3304 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003305 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003306 "@pthreadpool",
3307 ],
3308)
3309
Marat Dukhan33fcf782020-05-24 14:27:15 -07003310xnnpack_cc_library(
3311 name = "memory_planner_test_mode",
3312 srcs = ["src/memory-planner.c"],
3313 hdrs = INTERNAL_HDRS,
3314 copts = [
3315 "-UNDEBUG",
3316 "-DXNN_TEST_MODE=1",
3317 ],
3318 defines = select({
3319 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3320 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3321 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3322 }),
3323 gcc_copts = xnnpack_gcc_std_copts(),
3324 msvc_copts = xnnpack_msvc_std_copts(),
3325 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003326 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003327 "@pthreadpool",
3328 ],
3329)
3330
Marat Dukhan08c4a432019-10-03 09:29:21 -07003331cc_library(
3332 name = "enable_assembly",
3333 defines = select({
3334 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3335 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003336 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003337 }),
3338)
3339
Marat Dukhan9de90e02020-06-18 16:04:12 -07003340cc_library(
3341 name = "enable_sparse",
3342 defines = select({
3343 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3344 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003345 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003346 }),
3347)
3348
Marat Dukhancf056b22019-10-07 10:26:29 -07003349xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350 name = "operators",
3351 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003352 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003353 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003354 ],
3355 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003356 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003357 "-Isrc",
3358 "-Iinclude",
3359 ] + select({
3360 ":debug_build": [],
3361 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003362 }) + select({
3363 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3364 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003365 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003366 gcc_copts = xnnpack_gcc_std_copts(),
3367 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003368 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003369 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003370 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003371 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 "@FP16",
3373 "@FXdiv",
3374 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003376 ],
3377)
3378
Marat Dukhan10a38082020-04-17 03:58:35 -07003379xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003380 name = "operators_test_mode",
3381 srcs = OPERATOR_SRCS + [
3382 "src/memory.c",
3383 "src/operator-delete.c",
3384 ],
3385 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3386 copts = LOGGING_COPTS + [
3387 "-Isrc",
3388 "-Iinclude",
3389 "-UNDEBUG",
3390 "-DXNN_TEST_MODE=1",
3391 ] + select({
3392 ":debug_build": [],
3393 "//conditions:default": xnnpack_min_size_copts(),
3394 }) + select({
3395 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3396 "//conditions:default": [],
3397 }),
3398 gcc_copts = xnnpack_gcc_std_copts(),
3399 msvc_copts = xnnpack_msvc_std_copts(),
3400 deps = [
3401 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003402 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003403 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003404 "@FP16",
3405 "@FXdiv",
3406 "@clog",
3407 "@pthreadpool",
3408 ],
3409)
3410
3411xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003412 name = "XNNPACK",
3413 srcs = [
3414 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003415 "src/runtime.c",
3416 "src/subgraph.c",
3417 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003418 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003419 hdrs = ["include/xnnpack.h"],
3420 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003421 "-Isrc",
3422 "-Iinclude",
3423 ] + select({
3424 ":debug_build": [],
3425 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003426 }) + select({
3427 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3428 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003429 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003430 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003431 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003432 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003433 visibility = xnnpack_visibility(),
3434 deps = [
3435 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003436 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003437 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003438 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003439 ":operator_run",
3440 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003441 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003442 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003443 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003444 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003445 ] + select({
3446 ":emscripten": [],
3447 "//conditions:default": ["@cpuinfo"],
3448 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003449)
3450
Marat Dukhan10a38082020-04-17 03:58:35 -07003451xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003452 name = "XNNPACK_test_mode",
3453 srcs = [
3454 "src/init.c",
3455 "src/runtime.c",
3456 "src/subgraph.c",
3457 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003458 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003459 hdrs = ["include/xnnpack.h"],
3460 copts = LOGGING_COPTS + [
3461 "-Isrc",
3462 "-Iinclude",
3463 "-UNDEBUG",
3464 "-DXNN_TEST_MODE=1",
3465 ] + select({
3466 ":debug_build": [],
3467 "//conditions:default": xnnpack_min_size_copts(),
3468 }) + select({
3469 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3470 "//conditions:default": [],
3471 }),
3472 gcc_copts = xnnpack_gcc_std_copts(),
3473 includes = ["include"],
3474 msvc_copts = xnnpack_msvc_std_copts(),
3475 visibility = xnnpack_visibility(),
3476 deps = [
3477 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003478 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003479 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003480 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003481 ":operator_run_test_mode",
3482 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003483 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003484 "@clog",
3485 "@FP16",
3486 "@pthreadpool",
3487 ] + select({
3488 ":emscripten": [],
3489 "//conditions:default": ["@cpuinfo"],
3490 }),
3491)
3492
3493xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003494 name = "xnnpack_f32",
3495 srcs = [
3496 "src/init.c",
3497 "src/runtime.c",
3498 "src/subgraph.c",
3499 "src/tensor.c",
3500 ] + SUBGRAPH_SRCS,
3501 hdrs = ["include/xnnpack.h"],
3502 copts = LOGGING_COPTS + [
3503 "-Isrc",
3504 "-Iinclude",
3505 ] + select({
3506 ":debug_build": [],
3507 "//conditions:default": xnnpack_min_size_copts(),
3508 }) + select({
3509 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3510 "//conditions:default": [],
3511 }),
3512 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003513 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003514 "XNN_NO_U8_OPERATORS",
3515 "XNN_NO_X8_OPERATORS",
3516 "XNN_NO_F16_OPERATORS",
3517 "XNN_NO_X16_OPERATORS",
3518 ],
3519 gcc_copts = xnnpack_gcc_std_copts(),
3520 includes = ["include"],
3521 msvc_copts = xnnpack_msvc_std_copts(),
3522 visibility = xnnpack_visibility(),
3523 deps = [
3524 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003525 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003526 ":logging_utils",
3527 ":memory_planner",
3528 ":operator_run",
3529 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003530 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003531 "@clog",
3532 "@FP16",
3533 "@pthreadpool",
3534 ] + select({
3535 ":emscripten": [],
3536 "//conditions:default": ["@cpuinfo"],
3537 }),
3538)
3539
3540xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003541 name = "xnnpack_operators_nhwc_f32",
3542 srcs = [
3543 "src/init.c",
3544 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003545 hdrs = ["include/xnnpack.h"],
3546 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003547 "-Isrc",
3548 "-Iinclude",
3549 ] + select({
3550 ":debug_build": [],
3551 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003552 }) + select({
3553 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3554 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003555 }),
3556 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003557 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003558 "XNN_NO_U8_OPERATORS",
3559 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003560 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003561 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003562 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003563 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003564 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003565 visibility = xnnpack_visibility(),
3566 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003567 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003568 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003569 ":operator_run",
3570 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003571 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003572 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003573 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003574 ] + select({
3575 ":emscripten": [],
3576 "//conditions:default": ["@cpuinfo"],
3577 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003578)
3579
Marat Dukhancf056b22019-10-07 10:26:29 -07003580xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003581 name = "bench_utils",
3582 srcs = ["bench/utils.cc"],
3583 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003584 deps = [
3585 "@com_google_benchmark//:benchmark",
3586 "@cpuinfo",
3587 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588)
3589
Frank Barchard7e955972019-10-11 10:34:25 -07003590######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003591
3592xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003593 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594 srcs = [
3595 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003596 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003597 "src/xnnpack/AlignedAllocator.h",
3598 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003599 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003600 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601)
3602
3603xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003604 name = "f16_igemm_bench",
3605 srcs = [
3606 "bench/f16-igemm.cc",
3607 "bench/conv.h",
3608 "bench/google/conv.h",
3609 "src/xnnpack/AlignedAllocator.h",
3610 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003611 deps = MICROKERNEL_BENCHMARK_DEPS + [
3612 ":indirection",
3613 ":packing",
3614 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003615)
3616
3617xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618 name = "f16_gemm_bench",
3619 srcs = [
3620 "bench/f16-gemm.cc",
3621 "bench/gemm.h",
3622 "src/xnnpack/AlignedAllocator.h",
3623 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003624 deps = MICROKERNEL_BENCHMARK_DEPS + [
3625 ":packing",
3626 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003627)
3628
3629xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003630 name = "f16_spmm_bench",
3631 srcs = [
3632 "bench/f16-spmm.cc",
3633 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003634 "src/xnnpack/AlignedAllocator.h",
3635 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003636 deps = MICROKERNEL_BENCHMARK_DEPS,
3637)
3638
3639xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003640 name = "f32_igemm_bench",
3641 srcs = [
3642 "bench/f32-igemm.cc",
3643 "bench/conv.h",
3644 "src/xnnpack/AlignedAllocator.h",
3645 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003646 deps = MICROKERNEL_BENCHMARK_DEPS + [
3647 ":indirection",
3648 ":packing",
3649 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003650)
3651
3652xnnpack_benchmark(
3653 name = "f32_conv_hwc_bench",
3654 srcs = [
3655 "bench/f32-conv-hwc.cc",
3656 "bench/dconv.h",
3657 "src/xnnpack/AlignedAllocator.h",
3658 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003659 deps = MICROKERNEL_BENCHMARK_DEPS + [
3660 ":packing",
3661 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662)
3663
3664xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003665 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003666 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003667 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003668 "bench/dconv.h",
3669 "src/xnnpack/AlignedAllocator.h",
3670 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003671 deps = MICROKERNEL_BENCHMARK_DEPS + [
3672 ":packing",
3673 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003674)
3675
3676xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003677 name = "f16_dwconv_bench",
3678 srcs = [
3679 "bench/f16-dwconv.cc",
3680 "bench/dwconv.h",
3681 "bench/google/dwconv.h",
3682 "src/xnnpack/AlignedAllocator.h",
3683 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003684 deps = MICROKERNEL_BENCHMARK_DEPS + [
3685 ":indirection",
3686 ":packing",
3687 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003688)
3689
3690xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003691 name = "f32_dwconv_bench",
3692 srcs = [
3693 "bench/f32-dwconv.cc",
3694 "bench/dwconv.h",
3695 "src/xnnpack/AlignedAllocator.h",
3696 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003697 deps = MICROKERNEL_BENCHMARK_DEPS + [
3698 ":indirection",
3699 ":packing",
3700 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003701)
3702
3703xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003704 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003705 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003706 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003707 "bench/dwconv.h",
3708 "src/xnnpack/AlignedAllocator.h",
3709 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003710 deps = MICROKERNEL_BENCHMARK_DEPS + [
3711 ":indirection",
3712 ":packing",
3713 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003714)
3715
3716xnnpack_benchmark(
3717 name = "f32_gemm_bench",
3718 srcs = [
3719 "bench/f32-gemm.cc",
3720 "bench/gemm.h",
3721 "src/xnnpack/AlignedAllocator.h",
3722 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003723 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003724 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003725)
3726
3727xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003728 name = "f32_hswish_bench",
3729 srcs = [
3730 "bench/f32-hswish.cc",
3731 "src/xnnpack/AlignedAllocator.h",
3732 ] + MICROKERNEL_BENCHMARK_HDRS,
3733 deps = MICROKERNEL_BENCHMARK_DEPS,
3734)
3735
3736xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003737 name = "f32_raddexpminusmax_bench",
3738 srcs = [
3739 "bench/f32-raddexpminusmax.cc",
3740 "src/xnnpack/AlignedAllocator.h",
3741 ] + MICROKERNEL_BENCHMARK_HDRS,
3742 deps = MICROKERNEL_BENCHMARK_DEPS,
3743)
3744
3745xnnpack_benchmark(
3746 name = "f32_raddextexp_bench",
3747 srcs = [
3748 "bench/f32-raddextexp.cc",
3749 "src/xnnpack/AlignedAllocator.h",
3750 ] + MICROKERNEL_BENCHMARK_HDRS,
3751 deps = MICROKERNEL_BENCHMARK_DEPS,
3752)
3753
3754xnnpack_benchmark(
3755 name = "f32_raddstoreexpminusmax_bench",
3756 srcs = [
3757 "bench/f32-raddstoreexpminusmax.cc",
3758 "src/xnnpack/AlignedAllocator.h",
3759 ] + MICROKERNEL_BENCHMARK_HDRS,
3760 deps = MICROKERNEL_BENCHMARK_DEPS,
3761)
3762
3763xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003764 name = "f32_rmax_bench",
3765 srcs = [
3766 "bench/f32-rmax.cc",
3767 "src/xnnpack/AlignedAllocator.h",
3768 ] + MICROKERNEL_BENCHMARK_HDRS,
3769 deps = MICROKERNEL_BENCHMARK_DEPS,
3770)
3771
3772xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003773 name = "f32_sigmoid_bench",
3774 srcs = [
3775 "bench/f32-sigmoid.cc",
3776 "src/xnnpack/AlignedAllocator.h",
3777 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003778 deps = MICROKERNEL_BENCHMARK_DEPS,
3779)
3780
3781xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003782 name = "f32_spmm_bench",
3783 srcs = [
3784 "bench/f32-spmm.cc",
3785 "bench/gemm.h",
3786 "src/xnnpack/AlignedAllocator.h",
3787 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003788 deps = MICROKERNEL_BENCHMARK_DEPS,
3789)
3790
3791xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003792 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003793 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003794 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003795 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003796 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003797 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003798)
3799
3800xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003801 name = "f32_vscaleexpminusmax_bench",
3802 srcs = [
3803 "bench/f32-vscaleexpminusmax.cc",
3804 "src/xnnpack/AlignedAllocator.h",
3805 ] + MICROKERNEL_BENCHMARK_HDRS,
3806 deps = MICROKERNEL_BENCHMARK_DEPS,
3807)
3808
3809xnnpack_benchmark(
3810 name = "f32_vscaleextexp_bench",
3811 srcs = [
3812 "bench/f32-vscaleextexp.cc",
3813 "src/xnnpack/AlignedAllocator.h",
3814 ] + MICROKERNEL_BENCHMARK_HDRS,
3815 deps = MICROKERNEL_BENCHMARK_DEPS,
3816)
3817
3818xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003819 name = "f32_vsqrt_bench",
3820 srcs = [
3821 "bench/f32-vsqrt.cc",
3822 "src/xnnpack/AlignedAllocator.h",
3823 ] + MICROKERNEL_BENCHMARK_HDRS,
3824 deps = MICROKERNEL_BENCHMARK_DEPS,
3825)
3826
3827xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003828 name = "f32_im2col_gemm_bench",
3829 srcs = [
3830 "bench/f32-im2col-gemm.cc",
3831 "bench/conv.h",
3832 "src/xnnpack/AlignedAllocator.h",
3833 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003834 deps = MICROKERNEL_BENCHMARK_DEPS + [
3835 ":im2col",
3836 ":packing",
3837 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003838)
3839
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003840xnnpack_benchmark(
3841 name = "requantization_bench",
3842 srcs = [
3843 "bench/requantization.cc",
3844 "src/xnnpack/requantization-stubs.h",
3845 "src/xnnpack/AlignedAllocator.h",
3846 ] + MICROKERNEL_BENCHMARK_HDRS,
3847 deps = MICROKERNEL_BENCHMARK_DEPS,
3848)
3849
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003850xnnpack_benchmark(
3851 name = "rounding_bench",
3852 srcs = [
3853 "bench/rounding.cc",
3854 "src/xnnpack/math-stubs.h",
3855 "src/xnnpack/AlignedAllocator.h",
3856 ] + MICROKERNEL_BENCHMARK_HDRS,
3857 deps = MICROKERNEL_BENCHMARK_DEPS,
3858)
3859
Marat Dukhan08c4a432019-10-03 09:29:21 -07003860########################### Benchmarks for operators ###########################
3861
3862xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003863 name = "average_pooling_bench",
3864 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003865 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003866 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003867 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003868)
3869
3870xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003871 name = "bankers_rounding_bench",
3872 srcs = ["bench/bankers-rounding.cc"],
3873 copts = xnnpack_optional_tflite_copts(),
3874 tags = ["nowin32"],
3875 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3876)
3877
3878xnnpack_benchmark(
3879 name = "ceiling_bench",
3880 srcs = ["bench/ceiling.cc"],
3881 copts = xnnpack_optional_tflite_copts(),
3882 tags = ["nowin32"],
3883 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3884)
3885
3886xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003887 name = "channel_shuffle_bench",
3888 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003889 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890)
3891
3892xnnpack_benchmark(
3893 name = "convolution_bench",
3894 srcs = ["bench/convolution.cc"],
3895 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003896 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003897 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003898)
3899
3900xnnpack_benchmark(
3901 name = "deconvolution_bench",
3902 srcs = ["bench/deconvolution.cc"],
3903 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003904 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003905 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906)
3907
3908xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003909 name = "floor_bench",
3910 srcs = ["bench/floor.cc"],
3911 copts = xnnpack_optional_tflite_copts(),
3912 tags = ["nowin32"],
3913 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3914)
3915
3916xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003917 name = "global_average_pooling_bench",
3918 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003919 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003920)
3921
3922xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003923 name = "hardswish_bench",
3924 srcs = ["bench/hardswish.cc"],
3925 copts = xnnpack_optional_tflite_copts(),
3926 tags = ["nowin32"],
3927 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3928)
3929
3930xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003931 name = "max_pooling_bench",
3932 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003933 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003934)
3935
3936xnnpack_benchmark(
3937 name = "sigmoid_bench",
3938 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003939 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003940 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003941 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003942)
3943
3944xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003945 name = "prelu_bench",
3946 srcs = ["bench/prelu.cc"],
3947 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003948 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003949 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003950)
3951
3952xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003953 name = "softmax_bench",
3954 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003955 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003956 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003957 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003958)
3959
Marat Dukhan87727142020-06-24 15:24:10 -07003960xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003961 name = "square_root_bench",
3962 srcs = ["bench/square-root.cc"],
3963 copts = xnnpack_optional_tflite_copts(),
3964 tags = ["nowin32"],
3965 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3966)
3967
3968xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003969 name = "truncation_bench",
3970 srcs = ["bench/truncation.cc"],
3971 deps = OPERATOR_BENCHMARK_DEPS,
3972)
3973
Marat Dukhanc068bb62019-10-04 13:24:39 -07003974############################# End-to-end benchmarks ############################
3975
3976cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003977 name = "fp32_mobilenet_v1",
3978 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003979 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003980 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003981 linkstatic = True,
3982 deps = [
3983 ":XNNPACK",
3984 "@pthreadpool",
3985 ],
3986)
3987
3988cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003989 name = "fp16_mobilenet_v1",
3990 srcs = ["models/fp16-mobilenet-v1.cc"],
3991 hdrs = ["models/models.h"],
3992 copts = xnnpack_std_cxxopts(),
3993 linkstatic = True,
3994 deps = [
3995 ":XNNPACK",
3996 "@FP16",
3997 "@pthreadpool",
3998 ],
3999)
4000
4001cc_library(
4002 name = "fp32_mobilenet_v2",
4003 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004004 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004005 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004006 linkstatic = True,
4007 deps = [
4008 ":XNNPACK",
4009 "@pthreadpool",
4010 ],
4011)
4012
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004013cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004014 name = "fp16_mobilenet_v2",
4015 srcs = ["models/fp16-mobilenet-v2.cc"],
4016 hdrs = ["models/models.h"],
4017 copts = xnnpack_std_cxxopts(),
4018 linkstatic = True,
4019 deps = [
4020 ":XNNPACK",
4021 "@FP16",
4022 "@pthreadpool",
4023 ],
4024)
4025
4026cc_library(
4027 name = "fp32_mobilenet_v3_large",
4028 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004029 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004030 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004031 linkstatic = True,
4032 deps = [
4033 ":XNNPACK",
4034 "@pthreadpool",
4035 ],
4036)
4037
4038cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004039 name = "fp32_mobilenet_v3_small",
4040 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004041 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004042 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004043 linkstatic = True,
4044 deps = [
4045 ":XNNPACK",
4046 "@pthreadpool",
4047 ],
4048)
4049
Marat Dukhanc068bb62019-10-04 13:24:39 -07004050xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004051 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004052 srcs = [
4053 "bench/f32-dwconv-e2e.cc",
4054 "bench/end2end.h",
4055 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004056 deps = MICROKERNEL_BENCHMARK_DEPS + [
4057 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004058 ":fp32_mobilenet_v1",
4059 ":fp32_mobilenet_v2",
4060 ":fp32_mobilenet_v3_large",
4061 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004062 ],
4063)
4064
4065xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004066 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004067 srcs = [
4068 "bench/f32-gemm-e2e.cc",
4069 "bench/end2end.h",
4070 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004071 deps = MICROKERNEL_BENCHMARK_DEPS + [
4072 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004073 ":fp32_mobilenet_v1",
4074 ":fp32_mobilenet_v2",
4075 ":fp32_mobilenet_v3_large",
4076 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004077 ],
4078)
4079
4080xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004081 name = "end2end_bench",
4082 srcs = ["bench/end2end.cc"],
4083 deps = [
4084 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004085 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004086 ":fp16_mobilenet_v1",
4087 ":fp16_mobilenet_v2",
4088 ":fp32_mobilenet_v1",
4089 ":fp32_mobilenet_v2",
4090 ":fp32_mobilenet_v3_large",
4091 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004092 "@pthreadpool",
4093 ],
4094)
4095
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004096#################### Accuracy evaluation for math functions ####################
4097
4098xnnpack_benchmark(
4099 name = "f32_exp_eval",
4100 srcs = [
4101 "eval/f32-exp.cc",
4102 "src/xnnpack/AlignedAllocator.h",
4103 ] + ACCURACY_EVAL_HDRS,
4104 deps = ACCURACY_EVAL_DEPS,
4105)
4106
Marat Dukhan515c9772019-10-17 18:07:57 -07004107xnnpack_benchmark(
4108 name = "f32_expminus_eval",
4109 srcs = [
4110 "eval/f32-expminus.cc",
4111 "src/xnnpack/AlignedAllocator.h",
4112 ] + ACCURACY_EVAL_HDRS,
4113 deps = ACCURACY_EVAL_DEPS,
4114)
4115
Marat Dukhan98ba4412019-10-23 02:14:28 -07004116xnnpack_benchmark(
4117 name = "f32_extexp_eval",
4118 srcs = [
4119 "eval/f32-extexp.cc",
4120 "src/xnnpack/AlignedAllocator.h",
4121 ] + ACCURACY_EVAL_HDRS,
4122 deps = ACCURACY_EVAL_DEPS,
4123)
4124
Marat Dukhan8853b822020-05-07 12:19:01 -07004125xnnpack_unit_test(
4126 name = "f32_roundne_eval",
4127 srcs = [
4128 "eval/f32-roundne.cc",
4129 "src/xnnpack/AlignedAllocator.h",
4130 "src/xnnpack/math-stubs.h",
4131 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004132 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004133 deps = MICROKERNEL_TEST_DEPS,
4134)
4135
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004136xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004137 name = "f32_roundd_eval",
4138 srcs = [
4139 "eval/f32-roundd.cc",
4140 "src/xnnpack/AlignedAllocator.h",
4141 "src/xnnpack/math-stubs.h",
4142 ] + MICROKERNEL_TEST_HDRS,
4143 automatic = False,
4144 deps = MICROKERNEL_TEST_DEPS,
4145)
4146
4147xnnpack_unit_test(
4148 name = "f32_roundu_eval",
4149 srcs = [
4150 "eval/f32-roundu.cc",
4151 "src/xnnpack/AlignedAllocator.h",
4152 "src/xnnpack/math-stubs.h",
4153 ] + MICROKERNEL_TEST_HDRS,
4154 automatic = False,
4155 deps = MICROKERNEL_TEST_DEPS,
4156)
4157
4158xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004159 name = "f32_roundz_eval",
4160 srcs = [
4161 "eval/f32-roundz.cc",
4162 "src/xnnpack/AlignedAllocator.h",
4163 "src/xnnpack/math-stubs.h",
4164 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004165 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004166 deps = MICROKERNEL_TEST_DEPS,
4167)
4168
Marat Dukhan346a9e52019-11-15 09:06:30 -08004169xnnpack_benchmark(
4170 name = "f32_sigmoid_eval",
4171 srcs = [
4172 "eval/f32-sigmoid.cc",
4173 "src/xnnpack/AlignedAllocator.h",
4174 ] + ACCURACY_EVAL_HDRS,
4175 deps = ACCURACY_EVAL_DEPS,
4176)
4177
Marat Dukhan84000762020-06-29 18:38:43 -07004178xnnpack_benchmark(
4179 name = "f32_sqrt_eval",
4180 srcs = [
4181 "eval/f32-sqrt.cc",
4182 "src/xnnpack/AlignedAllocator.h",
4183 ] + ACCURACY_EVAL_HDRS,
4184 deps = ACCURACY_EVAL_DEPS + [
4185 ":bench_utils",
4186 ],
4187)
4188
Marat Dukhan08c4a432019-10-03 09:29:21 -07004189######################### Unit tests for micro-kernels #########################
4190
4191xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004192 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004193 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004194 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004195 "test/gemm-microkernel-tester.h",
4196 "src/xnnpack/AlignedAllocator.h",
4197 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004198 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004199)
4200
4201xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004202 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004203 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004204 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004205 "test/spmm-microkernel-tester.h",
4206 "src/xnnpack/AlignedAllocator.h",
4207 ] + MICROKERNEL_TEST_HDRS,
4208 deps = MICROKERNEL_TEST_DEPS,
4209)
4210
4211xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004212 name = "f32_argmaxpool_test",
4213 srcs = [
4214 "test/f32-argmaxpool.cc",
4215 "test/argmaxpool-microkernel-tester.h",
4216 "src/xnnpack/AlignedAllocator.h",
4217 ] + MICROKERNEL_TEST_HDRS,
4218 deps = MICROKERNEL_TEST_DEPS,
4219)
4220
4221xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004222 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004223 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004224 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004225 "test/avgpool-microkernel-tester.h",
4226 "src/xnnpack/AlignedAllocator.h",
4227 ] + MICROKERNEL_TEST_HDRS,
4228 deps = MICROKERNEL_TEST_DEPS,
4229)
4230
4231xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004232 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004233 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004234 "test/f32-ibilinear.cc",
4235 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004236 "src/xnnpack/AlignedAllocator.h",
4237 ] + MICROKERNEL_TEST_HDRS,
4238 deps = MICROKERNEL_TEST_DEPS,
4239)
4240
4241xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004242 name = "f16_clamp_test",
4243 srcs = [
4244 "test/f16-clamp.cc",
4245 "test/clamp-microkernel-tester.h",
4246 ] + MICROKERNEL_TEST_HDRS,
4247 deps = MICROKERNEL_TEST_DEPS,
4248)
4249
4250xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251 name = "f32_clamp_test",
4252 srcs = [
4253 "test/f32-clamp.cc",
4254 "test/clamp-microkernel-tester.h",
4255 ] + MICROKERNEL_TEST_HDRS,
4256 deps = MICROKERNEL_TEST_DEPS,
4257)
4258
4259xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004260 name = "f32_igemm_test",
4261 srcs = [
4262 "test/f32-igemm.cc",
4263 "test/gemm-microkernel-tester.h",
4264 "src/xnnpack/AlignedAllocator.h",
4265 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004266 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004267)
4268
4269xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004270 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004271 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004272 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004273 "test/gemm-microkernel-tester.h",
4274 "src/xnnpack/AlignedAllocator.h",
4275 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004276 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004277)
4278
4279xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004280 name = "f32_igemm_minmax_test",
4281 srcs = [
4282 "test/f32-igemm-minmax.cc",
4283 "test/gemm-microkernel-tester.h",
4284 "src/xnnpack/AlignedAllocator.h",
4285 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004286 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004287)
4288
4289xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004290 name = "f16_igemm_minmax_test",
4291 srcs = [
4292 "test/f16-igemm-minmax.cc",
4293 "test/gemm-microkernel-tester.h",
4294 "src/xnnpack/AlignedAllocator.h",
4295 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004296 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004297)
4298
4299xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004300 name = "f32_conv_hwc_test",
4301 srcs = [
4302 "test/f32-conv-hwc.cc",
4303 "test/conv-hwc-microkernel-tester.h",
4304 "src/xnnpack/AlignedAllocator.h",
4305 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004306 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307)
4308
4309xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004310 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004311 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004312 "test/f32-conv-hwc2chw.cc",
4313 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004314 "src/xnnpack/AlignedAllocator.h",
4315 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004316 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004317)
4318
4319xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004320 name = "f32_dwconv_test",
4321 srcs = [
4322 "test/f32-dwconv.cc",
4323 "test/dwconv-microkernel-tester.h",
4324 "src/xnnpack/AlignedAllocator.h",
4325 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004326 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004327)
4328
4329xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004330 name = "f16_dwconv_minmax_test",
4331 srcs = [
4332 "test/f16-dwconv-minmax.cc",
4333 "test/dwconv-microkernel-tester.h",
4334 "src/xnnpack/AlignedAllocator.h",
4335 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004336 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004337)
4338
4339xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004340 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004341 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004342 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004343 "test/dwconv-microkernel-tester.h",
4344 "src/xnnpack/AlignedAllocator.h",
4345 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004346 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004347)
4348
4349xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004350 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004351 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004352 "test/f32-dwconv-chw.cc",
4353 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004354 "src/xnnpack/AlignedAllocator.h",
4355 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004356 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357)
4358
4359xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004360 name = "f16_gavgpool_minmax_test",
4361 srcs = [
4362 "test/f16-gavgpool-minmax.cc",
4363 "test/gavgpool-microkernel-tester.h",
4364 "src/xnnpack/AlignedAllocator.h",
4365 ] + MICROKERNEL_TEST_HDRS,
4366 deps = MICROKERNEL_TEST_DEPS,
4367)
4368
4369xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004370 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004372 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004373 "test/gavgpool-microkernel-tester.h",
4374 "src/xnnpack/AlignedAllocator.h",
4375 ] + MICROKERNEL_TEST_HDRS,
4376 deps = MICROKERNEL_TEST_DEPS,
4377)
4378
4379xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004380 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004381 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004382 "test/f32-gavgpool-cw.cc",
4383 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 "src/xnnpack/AlignedAllocator.h",
4385 ] + MICROKERNEL_TEST_HDRS,
4386 deps = MICROKERNEL_TEST_DEPS,
4387)
4388
4389xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004390 name = "f32_gemm_test",
4391 srcs = [
4392 "test/f32-gemm.cc",
4393 "test/gemm-microkernel-tester.h",
4394 "src/xnnpack/AlignedAllocator.h",
4395 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004396 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004397)
4398
4399xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004400 name = "f32_gemm_relu_test",
4401 srcs = [
4402 "test/f32-gemm-relu.cc",
4403 "test/gemm-microkernel-tester.h",
4404 "src/xnnpack/AlignedAllocator.h",
4405 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004406 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004407)
4408
4409xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004410 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004411 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004412 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413 "test/gemm-microkernel-tester.h",
4414 "src/xnnpack/AlignedAllocator.h",
4415 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004416 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004417)
4418
4419xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004420 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004421 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004422 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004423 "test/gemm-microkernel-tester.h",
4424 "src/xnnpack/AlignedAllocator.h",
4425 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004426 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004427)
4428
4429xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004430 name = "f16_hswish_test",
4431 srcs = [
4432 "test/f16-hswish.cc",
4433 "test/hswish-microkernel-tester.h",
4434 ] + MICROKERNEL_TEST_HDRS,
4435 deps = MICROKERNEL_TEST_DEPS,
4436)
4437
4438xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004439 name = "f32_hswish_test",
4440 srcs = [
4441 "test/f32-hswish.cc",
4442 "test/hswish-microkernel-tester.h",
4443 ] + MICROKERNEL_TEST_HDRS,
4444 deps = MICROKERNEL_TEST_DEPS,
4445)
4446
4447xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004448 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004449 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004450 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 "test/maxpool-microkernel-tester.h",
4452 ] + MICROKERNEL_TEST_HDRS,
4453 deps = MICROKERNEL_TEST_DEPS,
4454)
4455
4456xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004457 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004459 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460 "test/avgpool-microkernel-tester.h",
4461 "src/xnnpack/AlignedAllocator.h",
4462 ] + MICROKERNEL_TEST_HDRS,
4463 deps = MICROKERNEL_TEST_DEPS,
4464)
4465
4466xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004467 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004468 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004469 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470 "test/gemm-microkernel-tester.h",
4471 "src/xnnpack/AlignedAllocator.h",
4472 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004473 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004474)
4475
4476xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004477 name = "f16_prelu_test",
4478 srcs = [
4479 "test/f16-prelu.cc",
4480 "test/prelu-microkernel-tester.h",
4481 "src/xnnpack/AlignedAllocator.h",
4482 ] + MICROKERNEL_TEST_HDRS,
4483 deps = MICROKERNEL_TEST_DEPS,
4484)
4485
4486xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004487 name = "f32_prelu_test",
4488 srcs = [
4489 "test/f32-prelu.cc",
4490 "test/prelu-microkernel-tester.h",
4491 "src/xnnpack/AlignedAllocator.h",
4492 ] + MICROKERNEL_TEST_HDRS,
4493 deps = MICROKERNEL_TEST_DEPS,
4494)
4495
4496xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004497 name = "f32_raddexpminusmax_test",
4498 srcs = [
4499 "test/f32-raddexpminusmax.cc",
4500 "test/raddexpminusmax-microkernel-tester.h",
4501 ] + MICROKERNEL_TEST_HDRS,
4502 deps = MICROKERNEL_TEST_DEPS,
4503)
4504
4505xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004506 name = "f32_raddextexp_test",
4507 srcs = [
4508 "test/f32-raddextexp.cc",
4509 "test/raddextexp-microkernel-tester.h",
4510 ] + MICROKERNEL_TEST_HDRS,
4511 deps = MICROKERNEL_TEST_DEPS,
4512)
4513
4514xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004515 name = "f32_raddstoreexpminusmax_test",
4516 srcs = [
4517 "test/f32-raddstoreexpminusmax.cc",
4518 "test/raddstoreexpminusmax-microkernel-tester.h",
4519 ] + MICROKERNEL_TEST_HDRS,
4520 deps = MICROKERNEL_TEST_DEPS,
4521)
4522
4523xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004524 name = "f32_relu_test",
4525 srcs = [
4526 "test/f32-relu.cc",
4527 "test/vunary-microkernel-tester.h",
4528 ] + MICROKERNEL_TEST_HDRS,
4529 deps = MICROKERNEL_TEST_DEPS,
4530)
4531
4532xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004533 name = "f32_rmax_test",
4534 srcs = [
4535 "test/f32-rmax.cc",
4536 "test/rmax-microkernel-tester.h",
4537 ] + MICROKERNEL_TEST_HDRS,
4538 deps = MICROKERNEL_TEST_DEPS,
4539)
4540
4541xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004542 name = "f32_sigmoid_test",
4543 srcs = [
4544 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004545 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004546 ] + MICROKERNEL_TEST_HDRS,
4547 deps = MICROKERNEL_TEST_DEPS,
4548)
4549
4550xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004551 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004552 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004553 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004554 "test/spmm-microkernel-tester.h",
4555 "src/xnnpack/AlignedAllocator.h",
4556 ] + MICROKERNEL_TEST_HDRS,
4557 deps = MICROKERNEL_TEST_DEPS,
4558)
4559
4560xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004561 name = "f16_vadd_minmax_test",
4562 srcs = [
4563 "test/f16-vadd-minmax.cc",
4564 "test/vbinary-microkernel-tester.h",
4565 ] + MICROKERNEL_TEST_HDRS,
4566 deps = MICROKERNEL_TEST_DEPS,
4567)
4568
4569xnnpack_unit_test(
4570 name = "f16_vaddc_minmax_test",
4571 srcs = [
4572 "test/f16-vaddc-minmax.cc",
4573 "test/vbinaryc-microkernel-tester.h",
4574 ] + MICROKERNEL_TEST_HDRS,
4575 deps = MICROKERNEL_TEST_DEPS,
4576)
4577
4578xnnpack_unit_test(
4579 name = "f16_vdiv_minmax_test",
4580 srcs = [
4581 "test/f16-vdiv-minmax.cc",
4582 "test/vbinary-microkernel-tester.h",
4583 ] + MICROKERNEL_TEST_HDRS,
4584 deps = MICROKERNEL_TEST_DEPS,
4585)
4586
4587xnnpack_unit_test(
4588 name = "f16_vdivc_minmax_test",
4589 srcs = [
4590 "test/f16-vdivc-minmax.cc",
4591 "test/vbinaryc-microkernel-tester.h",
4592 ] + MICROKERNEL_TEST_HDRS,
4593 deps = MICROKERNEL_TEST_DEPS,
4594)
4595
4596xnnpack_unit_test(
4597 name = "f16_vrdivc_minmax_test",
4598 srcs = [
4599 "test/f16-vrdivc-minmax.cc",
4600 "test/vbinaryc-microkernel-tester.h",
4601 ] + MICROKERNEL_TEST_HDRS,
4602 deps = MICROKERNEL_TEST_DEPS,
4603)
4604
4605xnnpack_unit_test(
4606 name = "f16_vmax_test",
4607 srcs = [
4608 "test/f16-vmax.cc",
4609 "test/vbinary-microkernel-tester.h",
4610 ] + MICROKERNEL_TEST_HDRS,
4611 deps = MICROKERNEL_TEST_DEPS,
4612)
4613
4614xnnpack_unit_test(
4615 name = "f16_vmaxc_test",
4616 srcs = [
4617 "test/f16-vmaxc.cc",
4618 "test/vbinaryc-microkernel-tester.h",
4619 ] + MICROKERNEL_TEST_HDRS,
4620 deps = MICROKERNEL_TEST_DEPS,
4621)
4622
4623xnnpack_unit_test(
4624 name = "f16_vmin_test",
4625 srcs = [
4626 "test/f16-vmin.cc",
4627 "test/vbinary-microkernel-tester.h",
4628 ] + MICROKERNEL_TEST_HDRS,
4629 deps = MICROKERNEL_TEST_DEPS,
4630)
4631
4632xnnpack_unit_test(
4633 name = "f16_vminc_test",
4634 srcs = [
4635 "test/f16-vminc.cc",
4636 "test/vbinaryc-microkernel-tester.h",
4637 ] + MICROKERNEL_TEST_HDRS,
4638 deps = MICROKERNEL_TEST_DEPS,
4639)
4640
4641xnnpack_unit_test(
4642 name = "f16_vmul_minmax_test",
4643 srcs = [
4644 "test/f16-vmul-minmax.cc",
4645 "test/vbinary-microkernel-tester.h",
4646 ] + MICROKERNEL_TEST_HDRS,
4647 deps = MICROKERNEL_TEST_DEPS,
4648)
4649
4650xnnpack_unit_test(
4651 name = "f16_vmulc_minmax_test",
4652 srcs = [
4653 "test/f16-vmulc-minmax.cc",
4654 "test/vbinaryc-microkernel-tester.h",
4655 ] + MICROKERNEL_TEST_HDRS,
4656 deps = MICROKERNEL_TEST_DEPS,
4657)
4658
4659xnnpack_unit_test(
4660 name = "f16_vsub_minmax_test",
4661 srcs = [
4662 "test/f16-vsub-minmax.cc",
4663 "test/vbinary-microkernel-tester.h",
4664 ] + MICROKERNEL_TEST_HDRS,
4665 deps = MICROKERNEL_TEST_DEPS,
4666)
4667
4668xnnpack_unit_test(
4669 name = "f16_vsubc_minmax_test",
4670 srcs = [
4671 "test/f16-vsubc-minmax.cc",
4672 "test/vbinaryc-microkernel-tester.h",
4673 ] + MICROKERNEL_TEST_HDRS,
4674 deps = MICROKERNEL_TEST_DEPS,
4675)
4676
4677xnnpack_unit_test(
4678 name = "f16_vrsubc_minmax_test",
4679 srcs = [
4680 "test/f16-vrsubc-minmax.cc",
4681 "test/vbinaryc-microkernel-tester.h",
4682 ] + MICROKERNEL_TEST_HDRS,
4683 deps = MICROKERNEL_TEST_DEPS,
4684)
4685
4686xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004687 name = "f32_vabs_test",
4688 srcs = [
4689 "test/f32-vabs.cc",
4690 "test/vunary-microkernel-tester.h",
4691 ] + MICROKERNEL_TEST_HDRS,
4692 deps = MICROKERNEL_TEST_DEPS,
4693)
4694
4695xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004696 name = "f32_vadd_test",
4697 srcs = [
4698 "test/f32-vadd.cc",
4699 "test/vbinary-microkernel-tester.h",
4700 ] + MICROKERNEL_TEST_HDRS,
4701 deps = MICROKERNEL_TEST_DEPS,
4702)
4703
4704xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004705 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004706 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004707 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004708 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004709 ] + MICROKERNEL_TEST_HDRS,
4710 deps = MICROKERNEL_TEST_DEPS,
4711)
4712
4713xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004714 name = "f32_vaddc_test",
4715 srcs = [
4716 "test/f32-vaddc.cc",
4717 "test/vbinaryc-microkernel-tester.h",
4718 ] + MICROKERNEL_TEST_HDRS,
4719 deps = MICROKERNEL_TEST_DEPS,
4720)
4721
4722xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004723 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004724 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004725 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004726 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004727 ] + MICROKERNEL_TEST_HDRS,
4728 deps = MICROKERNEL_TEST_DEPS,
4729)
4730
4731xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004732 name = "f32_vdiv_test",
4733 srcs = [
4734 "test/f32-vdiv.cc",
4735 "test/vbinary-microkernel-tester.h",
4736 ] + MICROKERNEL_TEST_HDRS,
4737 deps = MICROKERNEL_TEST_DEPS,
4738)
4739
4740xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004741 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004742 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004743 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004744 "test/vbinary-microkernel-tester.h",
4745 ] + MICROKERNEL_TEST_HDRS,
4746 deps = MICROKERNEL_TEST_DEPS,
4747)
4748
4749xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004750 name = "f32_vdivc_test",
4751 srcs = [
4752 "test/f32-vdivc.cc",
4753 "test/vbinaryc-microkernel-tester.h",
4754 ] + MICROKERNEL_TEST_HDRS,
4755 deps = MICROKERNEL_TEST_DEPS,
4756)
4757
4758xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004759 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004760 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004761 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004762 "test/vbinaryc-microkernel-tester.h",
4763 ] + MICROKERNEL_TEST_HDRS,
4764 deps = MICROKERNEL_TEST_DEPS,
4765)
4766
4767xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004768 name = "f32_vrdivc_test",
4769 srcs = [
4770 "test/f32-vrdivc.cc",
4771 "test/vbinaryc-microkernel-tester.h",
4772 ] + MICROKERNEL_TEST_HDRS,
4773 deps = MICROKERNEL_TEST_DEPS,
4774)
4775
4776xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004777 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004778 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004779 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004780 "test/vbinaryc-microkernel-tester.h",
4781 ] + MICROKERNEL_TEST_HDRS,
4782 deps = MICROKERNEL_TEST_DEPS,
4783)
4784
4785xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004786 name = "f32_vmax_test",
4787 srcs = [
4788 "test/f32-vmax.cc",
4789 "test/vbinary-microkernel-tester.h",
4790 ] + MICROKERNEL_TEST_HDRS,
4791 deps = MICROKERNEL_TEST_DEPS,
4792)
4793
4794xnnpack_unit_test(
4795 name = "f32_vmaxc_test",
4796 srcs = [
4797 "test/f32-vmaxc.cc",
4798 "test/vbinaryc-microkernel-tester.h",
4799 ] + MICROKERNEL_TEST_HDRS,
4800 deps = MICROKERNEL_TEST_DEPS,
4801)
4802
4803xnnpack_unit_test(
4804 name = "f32_vmin_test",
4805 srcs = [
4806 "test/f32-vmin.cc",
4807 "test/vbinary-microkernel-tester.h",
4808 ] + MICROKERNEL_TEST_HDRS,
4809 deps = MICROKERNEL_TEST_DEPS,
4810)
4811
4812xnnpack_unit_test(
4813 name = "f32_vminc_test",
4814 srcs = [
4815 "test/f32-vminc.cc",
4816 "test/vbinaryc-microkernel-tester.h",
4817 ] + MICROKERNEL_TEST_HDRS,
4818 deps = MICROKERNEL_TEST_DEPS,
4819)
4820
4821xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004822 name = "f32_vmul_test",
4823 srcs = [
4824 "test/f32-vmul.cc",
4825 "test/vbinary-microkernel-tester.h",
4826 ] + MICROKERNEL_TEST_HDRS,
4827 deps = MICROKERNEL_TEST_DEPS,
4828)
4829
4830xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004831 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004832 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004833 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004834 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004835 ] + MICROKERNEL_TEST_HDRS,
4836 deps = MICROKERNEL_TEST_DEPS,
4837)
4838
4839xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004840 name = "f32_vmulc_test",
4841 srcs = [
4842 "test/f32-vmulc.cc",
4843 "test/vbinaryc-microkernel-tester.h",
4844 ] + MICROKERNEL_TEST_HDRS,
4845 deps = MICROKERNEL_TEST_DEPS,
4846)
4847
4848xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004849 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004850 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004851 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004852 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004853 ] + MICROKERNEL_TEST_HDRS,
4854 deps = MICROKERNEL_TEST_DEPS,
4855)
4856
4857xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004858 name = "f16_vmulcaddc_minmax_test",
4859 srcs = [
4860 "test/f16-vmulcaddc-minmax.cc",
4861 "test/vmulcaddc-microkernel-tester.h",
4862 "src/xnnpack/AlignedAllocator.h",
4863 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004864 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004865)
4866
4867xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004868 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004870 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004871 "test/vmulcaddc-microkernel-tester.h",
4872 "src/xnnpack/AlignedAllocator.h",
4873 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004874 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875)
4876
4877xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004878 name = "f32_vlrelu_test",
4879 srcs = [
4880 "test/f32-vlrelu.cc",
4881 "test/vunary-microkernel-tester.h",
4882 ] + MICROKERNEL_TEST_HDRS,
4883 deps = MICROKERNEL_TEST_DEPS,
4884)
4885
4886xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004887 name = "f32_vneg_test",
4888 srcs = [
4889 "test/f32-vneg.cc",
4890 "test/vunary-microkernel-tester.h",
4891 ] + MICROKERNEL_TEST_HDRS,
4892 deps = MICROKERNEL_TEST_DEPS,
4893)
4894
4895xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004896 name = "f32_vrndne_test",
4897 srcs = [
4898 "test/f32-vrndne.cc",
4899 "test/vunary-microkernel-tester.h",
4900 ] + MICROKERNEL_TEST_HDRS,
4901 deps = MICROKERNEL_TEST_DEPS,
4902)
4903
4904xnnpack_unit_test(
4905 name = "f32_vrndz_test",
4906 srcs = [
4907 "test/f32-vrndz.cc",
4908 "test/vunary-microkernel-tester.h",
4909 ] + MICROKERNEL_TEST_HDRS,
4910 deps = MICROKERNEL_TEST_DEPS,
4911)
4912
4913xnnpack_unit_test(
4914 name = "f32_vrndu_test",
4915 srcs = [
4916 "test/f32-vrndu.cc",
4917 "test/vunary-microkernel-tester.h",
4918 ] + MICROKERNEL_TEST_HDRS,
4919 deps = MICROKERNEL_TEST_DEPS,
4920)
4921
4922xnnpack_unit_test(
4923 name = "f32_vrndd_test",
4924 srcs = [
4925 "test/f32-vrndd.cc",
4926 "test/vunary-microkernel-tester.h",
4927 ] + MICROKERNEL_TEST_HDRS,
4928 deps = MICROKERNEL_TEST_DEPS,
4929)
4930
4931xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004932 name = "f32_vscale_test",
4933 srcs = [
4934 "test/f32-vscale.cc",
4935 "test/vscale-microkernel-tester.h",
4936 ] + MICROKERNEL_TEST_HDRS,
4937 deps = MICROKERNEL_TEST_DEPS,
4938)
4939
4940xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004941 name = "f32_vscaleexpminusmax_test",
4942 srcs = [
4943 "test/f32-vscaleexpminusmax.cc",
4944 "test/vscaleexpminusmax-microkernel-tester.h",
4945 ] + MICROKERNEL_TEST_HDRS,
4946 deps = MICROKERNEL_TEST_DEPS,
4947)
4948
4949xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004950 name = "f32_vscaleextexp_test",
4951 srcs = [
4952 "test/f32-vscaleextexp.cc",
4953 "test/vscaleextexp-microkernel-tester.h",
4954 ] + MICROKERNEL_TEST_HDRS,
4955 deps = MICROKERNEL_TEST_DEPS,
4956)
4957
4958xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004959 name = "f32_vsqr_test",
4960 srcs = [
4961 "test/f32-vsqr.cc",
4962 "test/vunary-microkernel-tester.h",
4963 ] + MICROKERNEL_TEST_HDRS,
4964 deps = MICROKERNEL_TEST_DEPS,
4965)
4966
4967xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004968 name = "f32_vsqrdiff_test",
4969 srcs = [
4970 "test/f32-vsqrdiff.cc",
4971 "test/vbinary-microkernel-tester.h",
4972 ] + MICROKERNEL_TEST_HDRS,
4973 deps = MICROKERNEL_TEST_DEPS,
4974)
4975
4976xnnpack_unit_test(
4977 name = "f32_vsqrdiffc_test",
4978 srcs = [
4979 "test/f32-vsqrdiffc.cc",
4980 "test/vbinaryc-microkernel-tester.h",
4981 ] + MICROKERNEL_TEST_HDRS,
4982 deps = MICROKERNEL_TEST_DEPS,
4983)
4984
4985xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004986 name = "f32_vsqrt_test",
4987 srcs = [
4988 "test/f32-vsqrt.cc",
4989 "test/vunary-microkernel-tester.h",
4990 ] + MICROKERNEL_TEST_HDRS,
4991 deps = MICROKERNEL_TEST_DEPS,
4992)
4993
4994xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004995 name = "f32_vsub_test",
4996 srcs = [
4997 "test/f32-vsub.cc",
4998 "test/vbinary-microkernel-tester.h",
4999 ] + MICROKERNEL_TEST_HDRS,
5000 deps = MICROKERNEL_TEST_DEPS,
5001)
5002
5003xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005004 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005005 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005006 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005007 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005008 ] + MICROKERNEL_TEST_HDRS,
5009 deps = MICROKERNEL_TEST_DEPS,
5010)
5011
5012xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005013 name = "f32_vsubc_test",
5014 srcs = [
5015 "test/f32-vsubc.cc",
5016 "test/vbinaryc-microkernel-tester.h",
5017 ] + MICROKERNEL_TEST_HDRS,
5018 deps = MICROKERNEL_TEST_DEPS,
5019)
5020
5021xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005022 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005023 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005024 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005025 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005026 ] + MICROKERNEL_TEST_HDRS,
5027 deps = MICROKERNEL_TEST_DEPS,
5028)
5029
5030xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005031 name = "f32_vrsubc_test",
5032 srcs = [
5033 "test/f32-vrsubc.cc",
5034 "test/vbinaryc-microkernel-tester.h",
5035 ] + MICROKERNEL_TEST_HDRS,
5036 deps = MICROKERNEL_TEST_DEPS,
5037)
5038
5039xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005040 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005041 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005042 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005043 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005044 ] + MICROKERNEL_TEST_HDRS,
5045 deps = MICROKERNEL_TEST_DEPS,
5046)
5047
5048xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005049 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005051 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005052 "test/avgpool-microkernel-tester.h",
5053 "src/xnnpack/AlignedAllocator.h",
5054 ] + MICROKERNEL_TEST_HDRS,
5055 deps = MICROKERNEL_TEST_DEPS,
5056)
5057
5058xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005059 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005060 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005061 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005062 "test/gemm-microkernel-tester.h",
5063 "src/xnnpack/AlignedAllocator.h",
5064 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005065 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005066)
5067
5068xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005069 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005070 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005071 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005072 "test/dwconv-microkernel-tester.h",
5073 "src/xnnpack/AlignedAllocator.h",
5074 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005075 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005076)
5077
5078xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005079 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005080 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005081 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005082 "test/gavgpool-microkernel-tester.h",
5083 "src/xnnpack/AlignedAllocator.h",
5084 ] + MICROKERNEL_TEST_HDRS,
5085 deps = MICROKERNEL_TEST_DEPS,
5086)
5087
5088xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005089 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005090 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005091 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092 "test/gemm-microkernel-tester.h",
5093 "src/xnnpack/AlignedAllocator.h",
5094 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005095 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005096)
5097
5098xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005099 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005100 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005101 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005102 "test/vadd-microkernel-tester.h",
5103 ] + MICROKERNEL_TEST_HDRS,
5104 deps = MICROKERNEL_TEST_DEPS,
5105)
5106
5107xnnpack_unit_test(
5108 name = "u8_clamp_test",
5109 srcs = [
5110 "test/u8-clamp.cc",
5111 "test/clamp-microkernel-tester.h",
5112 ] + MICROKERNEL_TEST_HDRS,
5113 deps = MICROKERNEL_TEST_DEPS,
5114)
5115
5116xnnpack_unit_test(
5117 name = "u8_lut32norm_test",
5118 srcs = [
5119 "test/u8-lut32norm.cc",
5120 "test/lut-norm-microkernel-tester.h",
5121 ] + MICROKERNEL_TEST_HDRS,
5122 deps = MICROKERNEL_TEST_DEPS,
5123)
5124
5125xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005126 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005127 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005128 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005129 "test/maxpool-microkernel-tester.h",
5130 ] + MICROKERNEL_TEST_HDRS,
5131 deps = MICROKERNEL_TEST_DEPS,
5132)
5133
5134xnnpack_unit_test(
5135 name = "u8_rmax_test",
5136 srcs = [
5137 "test/u8-rmax.cc",
5138 "test/rmax-microkernel-tester.h",
5139 ] + MICROKERNEL_TEST_HDRS,
5140 deps = MICROKERNEL_TEST_DEPS,
5141)
5142
5143xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005144 name = "x32_fill_test",
5145 srcs = [
5146 "test/x32-fill.cc",
5147 "test/fill-microkernel-tester.h",
5148 ] + MICROKERNEL_TEST_HDRS,
5149 deps = MICROKERNEL_TEST_DEPS,
5150)
5151
5152xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005153 name = "x32_packx_test",
5154 srcs = [
5155 "test/x32-packx.cc",
5156 "test/pack-microkernel-tester.h",
5157 "src/xnnpack/AlignedAllocator.h",
5158 ] + MICROKERNEL_TEST_HDRS,
5159 deps = MICROKERNEL_TEST_DEPS,
5160)
5161
5162xnnpack_unit_test(
5163 name = "x32_pad_test",
5164 srcs = [
5165 "test/x32-pad.cc",
5166 "test/pad-microkernel-tester.h",
5167 ] + MICROKERNEL_TEST_HDRS,
5168 deps = MICROKERNEL_TEST_DEPS,
5169)
5170
5171xnnpack_unit_test(
5172 name = "x32_unpool_test",
5173 srcs = [
5174 "test/x32-unpool.cc",
5175 "test/unpool-microkernel-tester.h",
5176 ] + MICROKERNEL_TEST_HDRS,
5177 deps = MICROKERNEL_TEST_DEPS,
5178)
5179
5180xnnpack_unit_test(
5181 name = "x32_zip_test",
5182 srcs = [
5183 "test/x32-zip.cc",
5184 "test/zip-microkernel-tester.h",
5185 ] + MICROKERNEL_TEST_HDRS,
5186 deps = MICROKERNEL_TEST_DEPS,
5187)
5188
5189xnnpack_unit_test(
5190 name = "x8_lut_test",
5191 srcs = [
5192 "test/x8-lut.cc",
5193 "test/lut-microkernel-tester.h",
5194 ] + MICROKERNEL_TEST_HDRS,
5195 deps = MICROKERNEL_TEST_DEPS,
5196)
5197
5198xnnpack_unit_test(
5199 name = "x8_zip_test",
5200 srcs = [
5201 "test/x8-zip.cc",
5202 "test/zip-microkernel-tester.h",
5203 ] + MICROKERNEL_TEST_HDRS,
5204 deps = MICROKERNEL_TEST_DEPS,
5205)
5206
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005207xnnpack_unit_test(
5208 name = "requantization_test",
5209 srcs = [
5210 "src/xnnpack/requantization-stubs.h",
5211 "test/requantization.cc",
5212 "test/requantization-tester.h",
5213 ] + MICROKERNEL_TEST_HDRS,
5214 deps = MICROKERNEL_TEST_DEPS,
5215)
5216
Marat Dukhan20c3b922020-03-10 03:45:06 -07005217########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005218
5219xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005220 name = "operator_size_test",
5221 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005222 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223)
5224
Marat Dukhan20c3b922020-03-10 03:45:06 -07005225xnnpack_binary(
5226 name = "subgraph_size_test",
5227 srcs = ["test/subgraph-size.c"],
5228 deps = [":XNNPACK"],
5229)
5230
5231########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005232
5233xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005234 name = "abs_nc_test",
5235 srcs = [
5236 "test/abs-nc.cc",
5237 "test/abs-operator-tester.h",
5238 ],
5239 deps = OPERATOR_TEST_DEPS,
5240)
5241
5242xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005243 name = "add_nd_test",
5244 srcs = [
5245 "test/add-nd.cc",
5246 "test/binary-elementwise-operator-tester.h",
5247 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005248 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005249)
5250
5251xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005252 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005254 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005255 "test/argmax-pooling-operator-tester.h",
5256 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005257 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258)
5259
5260xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005261 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005262 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005263 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 "test/average-pooling-operator-tester.h",
5265 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005266 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267)
5268
5269xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005270 name = "bankers_rounding_nc_test",
5271 srcs = [
5272 "test/bankers-rounding-nc.cc",
5273 "test/bankers-rounding-operator-tester.h",
5274 ],
5275 deps = OPERATOR_TEST_DEPS,
5276)
5277
5278xnnpack_unit_test(
5279 name = "ceiling_nc_test",
5280 srcs = [
5281 "test/ceiling-nc.cc",
5282 "test/ceiling-operator-tester.h",
5283 ],
5284 deps = OPERATOR_TEST_DEPS,
5285)
5286
5287xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005288 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005289 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005290 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291 "test/channel-shuffle-operator-tester.h",
5292 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005293 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294)
5295
5296xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005297 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005298 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005299 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300 "test/clamp-operator-tester.h",
5301 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005302 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303)
5304
5305xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005306 name = "constant_pad_nd_test",
5307 srcs = [
5308 "test/constant-pad-nd.cc",
5309 "test/constant-pad-operator-tester.h",
5310 ],
5311 deps = OPERATOR_TEST_DEPS,
5312)
5313
5314xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005315 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005317 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318 "test/convolution-operator-tester.h",
5319 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005320 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321)
5322
5323xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005324 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005326 "test/convolution-nchw.cc",
5327 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005328 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005329 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005330)
5331
5332xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005333 name = "copy_nc_test",
5334 srcs = [
5335 "test/copy-nc.cc",
5336 "test/copy-operator-tester.h",
5337 ],
5338 deps = OPERATOR_TEST_DEPS,
5339)
5340
5341xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005342 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005343 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005344 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345 "test/deconvolution-operator-tester.h",
5346 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005347 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005348)
5349
5350xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005351 name = "divide_nd_test",
5352 srcs = [
5353 "test/binary-elementwise-operator-tester.h",
5354 "test/divide-nd.cc",
5355 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005356 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005357)
5358
5359xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005360 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005361 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005362 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363 "test/fully-connected-operator-tester.h",
5364 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005365 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366)
5367
5368xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005369 name = "floor_nc_test",
5370 srcs = [
5371 "test/floor-nc.cc",
5372 "test/floor-operator-tester.h",
5373 ],
5374 deps = OPERATOR_TEST_DEPS,
5375)
5376
5377xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005378 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005379 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005380 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005381 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005382 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005383 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005384)
5385
5386xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005387 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005388 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005389 "test/global-average-pooling-ncw.cc",
5390 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005391 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005392 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005393)
5394
5395xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005396 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005397 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005398 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399 "test/hardswish-operator-tester.h",
5400 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005401 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005402)
5403
5404xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005405 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005407 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005408 "test/leaky-relu-operator-tester.h",
5409 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005410 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411)
5412
5413xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005414 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005416 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417 "test/max-pooling-operator-tester.h",
5418 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005419 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005420)
5421
5422xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005423 name = "maximum_nd_test",
5424 srcs = [
5425 "test/binary-elementwise-operator-tester.h",
5426 "test/maximum-nd.cc",
5427 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005428 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005429)
5430
5431xnnpack_unit_test(
5432 name = "minimum_nd_test",
5433 srcs = [
5434 "test/binary-elementwise-operator-tester.h",
5435 "test/minimum-nd.cc",
5436 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005437 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005438)
5439
5440xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005441 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005442 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005443 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005444 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005445 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005446 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005447)
5448
5449xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005450 name = "negate_nc_test",
5451 srcs = [
5452 "test/negate-nc.cc",
5453 "test/negate-operator-tester.h",
5454 ],
5455 deps = OPERATOR_TEST_DEPS,
5456)
5457
5458xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005459 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005460 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005461 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462 "test/prelu-operator-tester.h",
5463 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005464 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465)
5466
5467xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005468 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005469 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005470 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005471 "test/resize-bilinear-operator-tester.h",
5472 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005473 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005474)
5475
5476xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005477 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005479 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005480 "test/sigmoid-operator-tester.h",
5481 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005482 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005483)
5484
5485xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005486 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005487 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005488 "test/softmax-nc.cc",
5489 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005491 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005492)
5493
5494xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005495 name = "square_nc_test",
5496 srcs = [
5497 "test/square-nc.cc",
5498 "test/square-operator-tester.h",
5499 ],
5500 deps = OPERATOR_TEST_DEPS,
5501)
5502
5503xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005504 name = "square_root_nc_test",
5505 srcs = [
5506 "test/square-root-nc.cc",
5507 "test/square-root-operator-tester.h",
5508 ],
5509 deps = OPERATOR_TEST_DEPS,
5510)
5511
5512xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005513 name = "squared_difference_nd_test",
5514 srcs = [
5515 "test/binary-elementwise-operator-tester.h",
5516 "test/squared-difference-nd.cc",
5517 ],
5518 deps = OPERATOR_TEST_DEPS,
5519)
5520
5521xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005522 name = "subtract_nd_test",
5523 srcs = [
5524 "test/binary-elementwise-operator-tester.h",
5525 "test/subtract-nd.cc",
5526 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005527 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005528)
5529
5530xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005531 name = "truncation_nc_test",
5532 srcs = [
5533 "test/truncation-nc.cc",
5534 "test/truncation-operator-tester.h",
5535 ],
5536 deps = OPERATOR_TEST_DEPS,
5537)
5538
5539xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005540 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005541 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005542 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005543 "test/unpooling-operator-tester.h",
5544 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005545 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005546)
5547
Chao Mei6ddfc602020-05-13 22:29:36 -07005548############################### Misc unit tests ###############################
5549
5550xnnpack_unit_test(
5551 name = "memory_planner_test",
5552 srcs = [
5553 "test/memory-planner-test.cc",
5554 ],
5555 deps = [
5556 ":XNNPACK",
5557 ":memory_planner",
5558 ],
5559)
5560
Marat Dukhan08c4a432019-10-03 09:29:21 -07005561############################# Build configurations #############################
5562
Marat Dukhanb8642352019-10-30 15:43:02 -07005563# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005564config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005565 name = "xnn_enable_assembly_explicit_true",
5566 define_values = {"xnn_enable_assembly": "true"},
5567)
5568
5569# Disables usage of assembly kernels.
5570config_setting(
5571 name = "xnn_enable_assembly_explicit_false",
5572 define_values = {"xnn_enable_assembly": "false"},
5573)
5574
Marat Dukhan9de90e02020-06-18 16:04:12 -07005575# Enables usage of sparse inference.
5576config_setting(
5577 name = "xnn_enable_sparse_explicit_true",
5578 define_values = {"xnn_enable_sparse": "true"},
5579)
5580
5581# Disables usage of sparse inference.
5582config_setting(
5583 name = "xnn_enable_sparse_explicit_false",
5584 define_values = {"xnn_enable_sparse": "false"},
5585)
5586
Marat Dukhan05702cf2020-03-26 15:41:33 -07005587# Disables usage of HMP-aware optimizations.
5588config_setting(
5589 name = "xnn_enable_hmp_explicit_false",
5590 define_values = {"xnn_enable_hmp": "false"},
5591)
5592
Chao Mei6ddfc602020-05-13 22:29:36 -07005593# Enable usage of optimized memory allocation
5594config_setting(
5595 name = "xnn_enable_memopt_explicit_true",
5596 define_values = {"xnnpack_enable_memopt": "true"},
5597)
5598
5599# Disable usage of optimized memory allocation
5600config_setting(
5601 name = "xnn_enable_memopt_explicit_false",
5602 define_values = {"xnnpack_enable_memopt": "false"},
5603)
5604
Marat Dukhanb8642352019-10-30 15:43:02 -07005605# Builds with -c dbg
5606config_setting(
5607 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005608 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005609 "compilation_mode": "dbg",
5610 },
5611)
5612
5613# Builds with -c opt
5614config_setting(
5615 name = "optimized_build",
5616 values = {
5617 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005618 },
5619)
5620
5621config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005622 name = "linux_k8",
5623 values = {"cpu": "k8"},
5624)
5625
5626config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005627 name = "linux_arm",
5628 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005629)
5630
5631config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005632 name = "linux_armeabi",
5633 values = {"cpu": "armeabi"},
5634)
5635
5636config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005637 name = "linux_armhf",
5638 values = {"cpu": "armhf"},
5639)
5640
5641config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005642 name = "linux_armv7a",
5643 values = {"cpu": "armv7a"},
5644)
5645
5646config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005647 name = "linux_aarch64",
5648 values = {"cpu": "aarch64"},
5649)
5650
5651config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005652 name = "android",
5653 values = {"crosstool_top": "//external:android/crosstool"},
5654)
5655
5656config_setting(
5657 name = "android_armv7",
5658 values = {
5659 "crosstool_top": "//external:android/crosstool",
5660 "cpu": "armeabi-v7a",
5661 },
5662)
5663
5664config_setting(
5665 name = "android_arm64",
5666 values = {
5667 "crosstool_top": "//external:android/crosstool",
5668 "cpu": "arm64-v8a",
5669 },
5670)
5671
5672config_setting(
5673 name = "android_x86",
5674 values = {
5675 "crosstool_top": "//external:android/crosstool",
5676 "cpu": "x86",
5677 },
5678)
5679
5680config_setting(
5681 name = "android_x86_64",
5682 values = {
5683 "crosstool_top": "//external:android/crosstool",
5684 "cpu": "x86_64",
5685 },
5686)
5687
5688config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005689 name = "windows_x86_64",
5690 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005691)
5692
5693config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005694 name = "windows_x86_64_clang",
5695 values = {
5696 "compiler": "clang-cl",
5697 "cpu": "x64_windows",
5698 },
5699)
5700
5701config_setting(
5702 name = "windows_x86_64_mingw",
5703 values = {
5704 "compiler": "mingw-gcc",
5705 "cpu": "x64_windows",
5706 },
5707)
5708
5709config_setting(
5710 name = "windows_x86_64_msys",
5711 values = {
5712 "compiler": "msys-gcc",
5713 "cpu": "x64_windows",
5714 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005715)
5716
5717config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005718 name = "macos_x86_64",
5719 values = {
5720 "apple_platform_type": "macos",
5721 "cpu": "darwin",
5722 },
5723)
5724
5725config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005726 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005727 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005728)
5729
5730config_setting(
5731 name = "emscripten_wasm",
5732 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005733 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005734 "cpu": "wasm",
5735 },
5736)
5737
5738config_setting(
5739 name = "emscripten_wasmsimd",
5740 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005741 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005742 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005743 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005744 },
5745)
5746
5747config_setting(
5748 name = "emscripten_asmjs",
5749 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005750 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005751 "cpu": "asmjs",
5752 },
5753)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005754
5755config_setting(
5756 name = "ios_armv7",
5757 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005758 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005759 "cpu": "ios_armv7",
5760 },
5761)
5762
5763config_setting(
5764 name = "ios_arm64",
5765 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005766 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005767 "cpu": "ios_arm64",
5768 },
5769)
5770
5771config_setting(
5772 name = "ios_arm64e",
5773 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005774 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005775 "cpu": "ios_arm64e",
5776 },
5777)
5778
5779config_setting(
5780 name = "ios_x86",
5781 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005782 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005783 "cpu": "ios_i386",
5784 },
5785)
5786
5787config_setting(
5788 name = "ios_x86_64",
5789 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005790 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005791 "cpu": "ios_x86_64",
5792 },
5793)
5794
5795config_setting(
5796 name = "watchos_armv7k",
5797 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005798 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005799 "cpu": "watchos_armv7k",
5800 },
5801)
5802
5803config_setting(
5804 name = "watchos_arm64_32",
5805 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005806 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005807 "cpu": "watchos_arm64_32",
5808 },
5809)
5810
5811config_setting(
5812 name = "watchos_x86",
5813 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005814 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005815 "cpu": "watchos_i386",
5816 },
5817)
5818
5819config_setting(
5820 name = "watchos_x86_64",
5821 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005822 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005823 "cpu": "watchos_x86_64",
5824 },
5825)
5826
5827config_setting(
5828 name = "tvos_arm64",
5829 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005830 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005831 "cpu": "tvos_arm64",
5832 },
5833)
5834
5835config_setting(
5836 name = "tvos_x86_64",
5837 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005838 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005839 "cpu": "tvos_x86_64",
5840 },
5841)