blob: 331f37d858d882c36591bf5331ebec7aadc95a35 [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 Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800124 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800125 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
133 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-scalar.c",
135 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-scalar.c",
137 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-scalar.c",
139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
141 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
146 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700158 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
160 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700161 "src/f32-gemm/gen/1x4-scalar.c",
162 "src/f32-gemm/gen/2x4-scalar.c",
163 "src/f32-gemm/gen/4x2-scalar.c",
164 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700165 "src/f32-gemm/gen/1x4-relu-scalar.c",
166 "src/f32-gemm/gen/2x4-relu-scalar.c",
167 "src/f32-gemm/gen/4x2-relu-scalar.c",
168 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700169 "src/f32-gemm/gen/1x4-minmax-scalar.c",
170 "src/f32-gemm/gen/2x4-minmax-scalar.c",
171 "src/f32-gemm/gen/4x2-minmax-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700191 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
193 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700194 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
195 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
197 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800198 "src/f32-prelu/gen/scalar-2x1.c",
199 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700212 "src/f32-relu/gen/scalar-x1.c",
213 "src/f32-relu/gen/scalar-x2.c",
214 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700215 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700225 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/1x1-minmax-scalar.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
228 "src/f32-spmm/gen/2x1-minmax-scalar.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/4x1-minmax-scalar.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/8x1-minmax-scalar.c",
233 "src/f32-spmm/gen/8x2-minmax-scalar.c",
234 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vadd-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
246 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
252 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700256 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
258 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800259 "src/f32-vbinary/gen/vmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
264 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
265 "src/f32-vbinary/gen/vmin-scalar-x1.c",
266 "src/f32-vbinary/gen/vmin-scalar-x2.c",
267 "src/f32-vbinary/gen/vmin-scalar-x4.c",
268 "src/f32-vbinary/gen/vminc-scalar-x1.c",
269 "src/f32-vbinary/gen/vminc-scalar-x2.c",
270 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmul-scalar-x1.c",
275 "src/f32-vbinary/gen/vmul-scalar-x2.c",
276 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
282 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
294 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700295 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
297 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
300 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsub-scalar-x1.c",
305 "src/f32-vbinary/gen/vsub-scalar-x2.c",
306 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
312 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700313 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
315 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
318 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700319 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
327 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
330 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700331 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
333 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700334 "src/f32-vunary/gen/vabs-scalar-x1.c",
335 "src/f32-vunary/gen/vabs-scalar-x2.c",
336 "src/f32-vunary/gen/vabs-scalar-x4.c",
337 "src/f32-vunary/gen/vneg-scalar-x1.c",
338 "src/f32-vunary/gen/vneg-scalar-x2.c",
339 "src/f32-vunary/gen/vneg-scalar-x4.c",
340 "src/f32-vunary/gen/vsqr-scalar-x1.c",
341 "src/f32-vunary/gen/vsqr-scalar-x2.c",
342 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800343 "src/math/expminus-scalar-lut2048-p1.c",
344 "src/math/expminus-scalar-lut64-p2.c",
345 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700346 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundne-scalar-nearbyint.c",
348 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700349 "src/math/roundd-scalar-addsub.c",
350 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700352 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700354 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700355 "src/math/roundz-scalar-addsub.c",
356 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700357 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800358 "src/math/sigmoid-scalar-lut2048-p1-div.c",
359 "src/math/sigmoid-scalar-lut64-p2-div.c",
360 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700361 "src/qs8-requantization/fp32-scalar-lrintf.c",
362 "src/qs8-requantization/fp32-scalar-magic.c",
363 "src/qs8-requantization/precise-scalar-signed64.c",
364 "src/qs8-requantization/precise-scalar-unsigned32.c",
365 "src/qs8-requantization/precise-scalar-unsigned64.c",
366 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700367 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
368 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
369 "src/qu8-dwconv/up1x9-minmax-scalar.c",
370 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
371 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
372 "src/qu8-gemm/2x2-minmax-scalar.c",
373 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700374 "src/qu8-requantization/fp32-scalar-lrintf.c",
375 "src/qu8-requantization/fp32-scalar-magic.c",
376 "src/qu8-requantization/precise-scalar-signed64.c",
377 "src/qu8-requantization/precise-scalar-unsigned32.c",
378 "src/qu8-requantization/precise-scalar-unsigned64.c",
379 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700380 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700381 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700382 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700384 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700385 "src/x32-fill/scalar-float.c",
386 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700387 "src/x32-packx/x2-scalar.c",
388 "src/x32-packx/x3-scalar.c",
389 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700390 "src/x32-pad/scalar-int.c",
391 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700392 "src/x32-unpool/scalar.c",
393 "src/x32-zip/x2-scalar.c",
394 "src/x32-zip/x3-scalar.c",
395 "src/x32-zip/x4-scalar.c",
396 "src/x32-zip/xm-scalar.c",
397 "src/x8-lut/scalar.c",
398 "src/x8-zip/x2-scalar.c",
399 "src/x8-zip/x3-scalar.c",
400 "src/x8-zip/x4-scalar.c",
401 "src/x8-zip/xm-scalar.c",
402]
403
Marat Dukhan436ebe62019-12-04 15:10:12 -0800404WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700405 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
406 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700407 "src/f32-clamp/gen/wasm-x1.c",
408 "src/f32-clamp/gen/wasm-x2.c",
409 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700410 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x4-wasm.c",
412 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x9-wasm.c",
414 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x25-wasm.c",
416 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x4-wasm.c",
418 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x9-wasm.c",
420 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
424 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700426 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
427 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700428 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
429 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
430 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
431 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700432 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
433 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700434 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
435 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700436 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
437 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
438 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700439 "src/f32-gemm/gen/1x4-wasm.c",
440 "src/f32-gemm/gen/2x4-wasm.c",
441 "src/f32-gemm/gen/4x2-wasm.c",
442 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700443 "src/f32-gemm/gen/1x4-relu-wasm.c",
444 "src/f32-gemm/gen/2x4-relu-wasm.c",
445 "src/f32-gemm/gen/4x2-relu-wasm.c",
446 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700447 "src/f32-gemm/gen/1x4-minmax-wasm.c",
448 "src/f32-gemm/gen/2x4-minmax-wasm.c",
449 "src/f32-gemm/gen/4x2-minmax-wasm.c",
450 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700451 "src/f32-hswish/gen/hswish-wasm-x1.c",
452 "src/f32-hswish/gen/hswish-wasm-x2.c",
453 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700454 "src/f32-igemm/gen/1x4-wasm.c",
455 "src/f32-igemm/gen/2x4-wasm.c",
456 "src/f32-igemm/gen/4x2-wasm.c",
457 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700458 "src/f32-igemm/gen/1x4-relu-wasm.c",
459 "src/f32-igemm/gen/2x4-relu-wasm.c",
460 "src/f32-igemm/gen/4x2-relu-wasm.c",
461 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700462 "src/f32-igemm/gen/1x4-minmax-wasm.c",
463 "src/f32-igemm/gen/2x4-minmax-wasm.c",
464 "src/f32-igemm/gen/4x2-minmax-wasm.c",
465 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700466 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
467 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
468 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700469 "src/f32-prelu/gen/wasm-2x1.c",
470 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700471 "src/f32-relu/gen/wasm-x1.c",
472 "src/f32-relu/gen/wasm-x2.c",
473 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700474 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
477 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
481 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
482 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
483 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
484 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
485 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800486 "src/f32-vbinary/gen/vmax-wasm-x1.c",
487 "src/f32-vbinary/gen/vmax-wasm-x2.c",
488 "src/f32-vbinary/gen/vmax-wasm-x4.c",
489 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
490 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
491 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
492 "src/f32-vbinary/gen/vmin-wasm-x1.c",
493 "src/f32-vbinary/gen/vmin-wasm-x2.c",
494 "src/f32-vbinary/gen/vmin-wasm-x4.c",
495 "src/f32-vbinary/gen/vminc-wasm-x1.c",
496 "src/f32-vbinary/gen/vminc-wasm-x2.c",
497 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700498 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
507 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
508 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
509 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
510 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
511 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
512 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
513 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
514 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
515 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700516 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
517 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
518 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
520 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
521 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800522]
523
Marat Dukhan290055c2020-06-09 12:24:29 -0700524WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700525 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
526 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
527 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700528 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
529 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
530 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
531 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700532 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
533 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
534 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
535 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700536 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
537 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
538 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
539 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
540 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
541 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700542 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
547 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
548 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
549 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
550 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
551 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
552 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
553 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
554 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
558 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
559 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
560 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
561 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
562 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
563 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
564 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
565 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700566 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
567 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
568 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700569 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700570 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
571 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
572 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
573 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700574 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
581 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
582 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
583 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
584 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
585 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
586 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
587 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
588 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
589 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
595 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
596 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
597 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
598 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
599 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
600 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
601 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
602 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
603 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700604 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
605 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700606 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
613 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
614 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
615 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
616 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
617 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
618 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
619 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
620 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
621 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
627 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
628 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
629 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
630 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
631 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
632 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
633 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
634 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
635 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700636 "src/f32-ibilinear/gen/wasmsimd-c4.c",
637 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700638 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
639 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
640 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700641 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700642 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
643 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
644 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
645 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700646 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
653 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
654 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
655 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
656 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
657 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
658 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
659 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
660 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
661 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
667 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
668 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
669 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
670 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
671 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
672 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
673 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
674 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
675 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700676 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
677 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700678 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700682 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
683 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700684 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
685 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
686 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700687 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
688 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700689 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
690 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
691 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
692 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700693 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
694 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700695 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
696 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
697 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700698 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
699 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700700 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
701 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
702 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
703 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
704 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
705 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
706 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700707 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
708 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700709 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
710 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
711 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
712 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700713 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
714 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
715 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
716 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
717 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
718 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
719 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
720 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
722 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
723 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700725 "src/f32-relu/gen/wasmsimd-x4.c",
726 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700727 "src/f32-rmax/wasmsimd-arm.c",
728 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700729 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
730 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
731 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
732 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
733 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
734 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700735 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
736 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
737 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
738 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
739 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
740 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700741 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700747 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700751 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
752 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700753 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
754 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
755 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
756 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700757 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700763 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
764 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700765 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
766 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
767 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
768 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700769 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
770 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
771 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
772 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700773 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700777 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
778 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
779 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
780 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700781 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
782 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
783 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
784 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700785 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
786 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700787 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
788 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
789 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
790 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700791 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
792 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700793 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
794 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
795 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
796 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700797 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
798 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700799 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
800 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
801 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
802 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700803 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
804 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700805 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
806 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700807 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
808 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700809 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
810 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
811 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
812 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700813 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
814 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700815 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
816 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
817 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
818 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700819 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
820 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700821 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
822 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
823 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
824 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700825 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
826 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
827 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
828 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700829 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
830 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
831 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
832 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
833 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
834 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
835 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
836 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
837 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
838 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
839 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
840 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
841 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
842 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700843 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
844 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700845 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
846 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
847 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
848 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
849 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
850 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700851 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700852 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700853 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700854 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700855 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700856 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700857 "src/x32-zip/x2-wasmsimd.c",
858 "src/x32-zip/x3-wasmsimd.c",
859 "src/x32-zip/x4-wasmsimd.c",
860 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700861 "src/math/roundne-wasmsimd-addsub.c",
862 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700863 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700864 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700865 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700866 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700867 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700868 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
869 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700870]
871
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800872PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700873 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700874 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700875 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700876 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700877 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700878 "src/f32-spmm/gen/4x1-minmax-psimd.c",
879 "src/f32-spmm/gen/8x1-minmax-psimd.c",
880 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700881]
882
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800883PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700884 "src/qs8-requantization/precise-psimd.c",
885 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700886 "src/qu8-requantization/precise-psimd.c",
887 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800888]
889
Marat Dukhan08c4a432019-10-03 09:29:21 -0700890# ISA-specific micro-kernels
891NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700892 "src/f32-argmaxpool/4x-neon-c4.c",
893 "src/f32-argmaxpool/9p8x-neon-c4.c",
894 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700895 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
896 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700897 "src/f32-clamp/gen/neon-x4.c",
898 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700899 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
900 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700903 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
904 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700905 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
906 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700907 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
908 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
909 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
910 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700911 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
912 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
913 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
914 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700915 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
916 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
917 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
918 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700919 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700920 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
921 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700922 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
923 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
924 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
925 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
926 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
927 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
928 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
929 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
930 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
931 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
932 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
933 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
934 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
935 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
936 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
937 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
938 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
939 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
940 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
941 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
942 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
944 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
945 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
946 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
947 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
948 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
949 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
950 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
951 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
952 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700953 "src/f32-hswish/gen/hswish-neon-x4.c",
954 "src/f32-hswish/gen/hswish-neon-x8.c",
955 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700956 "src/f32-ibilinear/gen/neon-c4.c",
957 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700958 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
959 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
960 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
961 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
962 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
963 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
964 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
965 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
966 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
967 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
968 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
969 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
970 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
971 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
972 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
973 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700974 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
975 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
976 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700977 "src/f32-ppmm/gen/4x8-minmax-neon.c",
978 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700979 "src/f32-prelu/gen/neon-1x4.c",
980 "src/f32-prelu/gen/neon-1x8.c",
981 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800982 "src/f32-prelu/gen/neon-2x4.c",
983 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700984 "src/f32-prelu/gen/neon-2x16.c",
985 "src/f32-prelu/gen/neon-4x4.c",
986 "src/f32-prelu/gen/neon-4x8.c",
987 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800988 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
989 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1005 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001012 "src/f32-relu/gen/neon-x4.c",
1013 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001014 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001015 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001016 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1017 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1018 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1019 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1020 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1021 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1022 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1023 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1026 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1027 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1028 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1029 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1030 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1031 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1032 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1033 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001034 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1035 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1036 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1037 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001038 "src/f32-vbinary/gen/vmax-neon-x4.c",
1039 "src/f32-vbinary/gen/vmax-neon-x8.c",
1040 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1041 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1042 "src/f32-vbinary/gen/vmin-neon-x4.c",
1043 "src/f32-vbinary/gen/vmin-neon-x8.c",
1044 "src/f32-vbinary/gen/vminc-neon-x4.c",
1045 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001046 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1047 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1048 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1049 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1050 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1051 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001052 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1053 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1054 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1055 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001056 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1057 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1058 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1059 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001060 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1061 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001062 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1063 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001064 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1065 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1066 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1067 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1068 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1069 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1070 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1071 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001072 "src/f32-vunary/gen/vabs-neon-x4.c",
1073 "src/f32-vunary/gen/vabs-neon-x8.c",
1074 "src/f32-vunary/gen/vneg-neon-x4.c",
1075 "src/f32-vunary/gen/vneg-neon-x8.c",
1076 "src/f32-vunary/gen/vsqr-neon-x4.c",
1077 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001078 "src/qs8-requantization/precise-neon.c",
1079 "src/qs8-requantization/fp32-neon.c",
1080 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001081 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1082 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1083 "src/qu8-dwconv/up8x9-minmax-neon.c",
1084 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1085 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1086 "src/qu8-gemm/4x8-minmax-neon.c",
1087 "src/qu8-gemm/8x8-minmax-neon.c",
1088 "src/qu8-igemm/4x8-minmax-neon.c",
1089 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001090 "src/qu8-requantization/precise-neon.c",
1091 "src/qu8-requantization/fp32-neon.c",
1092 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001093 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001094 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001095 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001096 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001097 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001098 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001099 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001100 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001101 "src/x32-zip/x2-neon.c",
1102 "src/x32-zip/x3-neon.c",
1103 "src/x32-zip/x4-neon.c",
1104 "src/x32-zip/xm-neon.c",
1105 "src/x8-zip/x2-neon.c",
1106 "src/x8-zip/x3-neon.c",
1107 "src/x8-zip/x4-neon.c",
1108 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001109 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001110 "src/math/roundd-neon-addsub.c",
1111 "src/math/roundd-neon-cvt.c",
1112 "src/math/roundu-neon-addsub.c",
1113 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001114 "src/math/roundz-neon-addsub.c",
1115 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001116 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001117 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1118 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1119 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1120 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1121 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1122 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001123 "src/math/sqrt-neon-nr1rsqrts.c",
1124 "src/math/sqrt-neon-nr2rsqrts.c",
1125 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001126]
1127
1128NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001129 "src/f32-ibilinear/gen/neonfma-c4.c",
1130 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001131 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1132 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1133 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1134 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1135 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1136 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1137 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1138 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1139 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001140 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1141 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1142 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1143 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001144 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1145 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1146 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1147 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001148 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1149 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1150 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1151 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001152 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1153 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1154 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1155 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1156 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1157 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1158 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1159 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1160 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1161 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1162 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1164 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1165 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1166 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1167 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1168 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1169 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001170 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1171 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001172 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1173 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1177 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1178 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1179 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1180 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1181 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1182 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1185 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1187 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1188 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1189 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1190 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1191 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001196 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001250 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1251 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001252 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1253 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1254 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1255 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1256 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1257 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1258 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1259 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1260 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1261 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1262 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1263 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1264 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1265 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1266 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1267 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1268 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1269 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1270 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1271 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001272 "src/math/exp-neonfma-lut64-p2.c",
1273 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001274 "src/math/expminus-neonfma-lut2048-p1.c",
1275 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001276 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001277 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1278 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1279 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1280 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1281 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1282 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1283 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1284 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1285 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1286 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1287 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1288 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1289 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1290 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1291 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1292 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1293 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1294 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001295 "src/math/sqrt-neonfma-nr1fma.c",
1296 "src/math/sqrt-neonfma-nr2fma.c",
1297 "src/math/sqrt-neonfma-nr3fma.c",
1298 "src/math/sqrt-neonfma-nr2fma1adj.c",
1299 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001300]
1301
1302AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001303 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1304 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1305 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1306 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1307 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1308 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001309 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1310 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1311 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1312 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1313 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1314 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1315 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1316 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1317 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1318 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1319 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1320 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1321 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1322 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1323 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1324 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1325 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1326 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1327 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1328 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001329 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1330 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001331 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1332 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001333 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1334 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001335 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1336 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001337 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1338 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1339 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1340 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1341 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001342 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001360 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1361 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1362 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1363 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1364 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1365 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1366 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1367 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1368 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1369 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1370 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1371 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1372 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1373 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1374 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1375 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1376 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1377 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001378 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1379 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001380 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1381 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1382 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1383 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1384 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1385 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001386]
1387
Marat Dukhan8853b822020-05-07 12:19:01 -07001388NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001389 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1390 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1391 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1392 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1393 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1394 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1395 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1396 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001397 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001398 "src/math/roundd-neonv8.c",
1399 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001400 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001401]
1402
Marat Dukhan08c4a432019-10-03 09:29:21 -07001403AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001404 "src/f16-clamp/gen/neonfp16arith-x16.c",
1405 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001406 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1407 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1408 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1409 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1410 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1411 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1412 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1413 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1414 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1415 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1416 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1417 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001418 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1419 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001420 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1421 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001422 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001423 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001424 "src/f16-relu/gen/neonfp16arith-x16.c",
1425 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001426 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1427 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1428 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1429 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1430 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1431 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1432 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1433 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1434 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1435 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1436 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1437 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1438 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1439 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1440 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1441 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1442 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1443 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1444 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1445 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1446 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1447 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1448 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1449 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1450 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1451 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1452 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1453 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001454 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1455 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001456 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1457 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1458 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1459 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1460 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1461 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1462 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1463 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001464 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1465 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1466 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1467 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001468 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1469 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1470 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1471 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1472 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1473 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1474 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1475 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1476 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1477 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1478 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1479 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001480 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1481 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1482 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1483 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1484 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1485 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1486 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1487 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001488]
1489
1490SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001491 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1492 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001493 "src/f32-clamp/gen/sse-x4.c",
1494 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001495 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1496 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001497 "src/f32-dwconv-chw/3x3p1-sse.c",
1498 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001499 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1500 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1501 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1502 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1503 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1504 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1505 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1506 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1507 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1508 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1509 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1510 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001511 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001512 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1513 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001514 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1515 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1516 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1517 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1518 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1519 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1520 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1521 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1522 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1523 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1524 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1525 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1526 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001527 "src/f32-hswish/gen/hswish-sse-x4.c",
1528 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001529 "src/f32-ibilinear/gen/sse-c4.c",
1530 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1532 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1533 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1534 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1535 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1536 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1537 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001538 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1539 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1540 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001541 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001542 "src/f32-prelu/gen/sse-2x4.c",
1543 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001544 "src/f32-relu/gen/sse-x4.c",
1545 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001546 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001547 "src/f32-spmm/gen/4x1-minmax-sse.c",
1548 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001549 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001550 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1551 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1552 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1553 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1554 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1555 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1556 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1557 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001558 "src/f32-vbinary/gen/vmax-sse-x4.c",
1559 "src/f32-vbinary/gen/vmax-sse-x8.c",
1560 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1561 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1562 "src/f32-vbinary/gen/vmin-sse-x4.c",
1563 "src/f32-vbinary/gen/vmin-sse-x8.c",
1564 "src/f32-vbinary/gen/vminc-sse-x4.c",
1565 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001566 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1567 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1568 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1569 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1570 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1571 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1572 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1573 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001574 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1575 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1576 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1577 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001578 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1579 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1580 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1581 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001582 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1583 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001584 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1585 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001586 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1587 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001588 "src/f32-vunary/gen/vabs-sse-x4.c",
1589 "src/f32-vunary/gen/vabs-sse-x8.c",
1590 "src/f32-vunary/gen/vneg-sse-x4.c",
1591 "src/f32-vunary/gen/vneg-sse-x8.c",
1592 "src/f32-vunary/gen/vsqr-sse-x4.c",
1593 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001594 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001596 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001597 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001598 "src/math/roundd-sse-addsub.c",
1599 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001600 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001601 "src/math/sqrt-sse-hh1mac.c",
1602 "src/math/sqrt-sse-nr1mac.c",
1603 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604]
1605
1606SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001607 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1608 "src/f32-argmaxpool/4x-sse2-c4.c",
1609 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001610 "src/f32-prelu/gen/sse2-2x4.c",
1611 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001612 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1613 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1614 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1615 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1616 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1617 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1618 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1619 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1620 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1621 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1622 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1623 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001624 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001625 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001626 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001627 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001628 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1629 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001630 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1631 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001632 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1633 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1634 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1635 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1636 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1637 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1638 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1639 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001640 "src/qs8-requantization/precise-sse2.c",
1641 "src/qs8-requantization/fp32-sse2.c",
1642 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001643 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1644 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001645 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1646 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001647 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1648 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001649 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1650 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001651 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1652 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001653 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1654 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001655 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1656 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
1657 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1658 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001659 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1660 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1661 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1662 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1663 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1664 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1665 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1666 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001667 "src/qu8-requantization/precise-sse2.c",
1668 "src/qu8-requantization/fp32-sse2.c",
1669 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001670 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001671 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001672 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001673 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001674 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001675 "src/x32-zip/x2-sse2.c",
1676 "src/x32-zip/x3-sse2.c",
1677 "src/x32-zip/x4-sse2.c",
1678 "src/x32-zip/xm-sse2.c",
1679 "src/x8-zip/x2-sse2.c",
1680 "src/x8-zip/x3-sse2.c",
1681 "src/x8-zip/x4-sse2.c",
1682 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001683 "src/math/exp-sse2-p5.c",
1684 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001685 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001686 "src/math/roundd-sse2-cvt.c",
1687 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001688 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001689 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001690]
1691
1692SSSE3_UKERNELS = [
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001693 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1694 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001695 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1696 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001697 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1698 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001699 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1700 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001701 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1702 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001703 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1704 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001705 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1706 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
1707 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1708 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001709 "src/qs8-requantization/precise-ssse3.c",
1710 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001711 "src/qu8-requantization/precise-ssse3.c",
1712 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001713]
1714
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001715SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001716 "src/f32-prelu/gen/sse41-2x4.c",
1717 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001718 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001719 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001720 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001721 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001722 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1723 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001724 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1725 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001726 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1727 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1728 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1729 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1730 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1731 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1732 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1733 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001734 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1735 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001736 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1737 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001738 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1739 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001740 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1741 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001742 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1743 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001744 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1745 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001746 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1747 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
1748 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1749 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001750 "src/qs8-requantization/fp32-sse4.c",
1751 "src/qs8-requantization/precise-sse4.c",
1752 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001753 "src/qu8-requantization/precise-sse4.c",
1754 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001755 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001756 "src/math/roundd-sse41.c",
1757 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001758 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001759]
1760
Marat Dukhan08c4a432019-10-03 09:29:21 -07001761AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001762 "src/f32-clamp/gen/avx-x8.c",
1763 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001764 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1765 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1766 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1767 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1768 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1769 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1770 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1771 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1772 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1773 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1774 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1775 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1776 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1777 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1778 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1779 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1780 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1781 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1782 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1783 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1784 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1785 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1786 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1787 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1788 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1789 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1790 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1791 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1792 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1793 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001794 "src/f32-hswish/gen/hswish-avx-x8.c",
1795 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001796 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1797 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1798 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1799 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1800 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1801 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1802 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1803 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1804 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001805 "src/f32-prelu/gen/avx-2x8.c",
1806 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001807 "src/f32-relu/gen/avx-x16.c",
1808 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001809 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001810 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1811 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1812 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1813 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1814 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1815 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1816 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1817 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001818 "src/f32-vbinary/gen/vmax-avx-x8.c",
1819 "src/f32-vbinary/gen/vmax-avx-x16.c",
1820 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1821 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1822 "src/f32-vbinary/gen/vmin-avx-x8.c",
1823 "src/f32-vbinary/gen/vmin-avx-x16.c",
1824 "src/f32-vbinary/gen/vminc-avx-x8.c",
1825 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001826 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1827 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1828 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1829 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1830 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1831 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1832 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1833 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001834 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1835 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1836 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1837 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001838 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1839 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1840 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1841 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001842 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1843 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001844 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001845 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1846 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1847 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1848 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1849 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1850 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1851 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1852 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001853 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1854 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001855 "src/f32-vunary/gen/vabs-avx-x8.c",
1856 "src/f32-vunary/gen/vabs-avx-x16.c",
1857 "src/f32-vunary/gen/vneg-avx-x8.c",
1858 "src/f32-vunary/gen/vneg-avx-x16.c",
1859 "src/f32-vunary/gen/vsqr-avx-x8.c",
1860 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001861]
1862
Marat Dukhan1566fee2020-08-02 21:55:41 -07001863XOP_UKERNELS = [
1864 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1865 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1866 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1867 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
1868 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1869 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
1870 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1871 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
1872 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1873 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1874 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1875 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1876 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1877 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
1878 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1879 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
1880]
1881
Marat Dukhanfda12b82019-11-21 12:27:59 -08001882FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001883 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1884 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1885 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1886 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1887 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1888 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1889 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1890 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1891 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1892 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1893 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1894 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1895 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1896 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1897 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1898 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1899 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1900 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1901 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1902 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1903 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1904 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1905 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1906 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1907 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1908 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1909 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1910 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1911 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1912 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1913 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1914 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1915 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1916 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1917 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1918 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1919 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1920 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1921 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1922 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001923 "src/f32-hswish/gen/hswish-fma3-x8.c",
1924 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001925 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1926 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1927 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1928 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1929 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1930 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1931 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1932 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1933 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1934 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1935 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1936 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1937 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1938 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001939 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1940 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1941 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1942 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1943 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1944 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1945 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1946 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001947 "src/math/sqrt-fma3-nr1fma.c",
1948 "src/math/sqrt-fma3-nr2fma.c",
1949 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001950]
1951
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001952AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001953 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1954 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1955 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1956 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1957 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1958 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1959 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1960 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1961 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1962 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1963 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1964 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1965 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1966 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1967 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1968 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1969 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1970 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1971 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1972 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1973 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1974 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1975 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1976 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1977 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1978 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1979 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1980 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1981 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1982 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1983 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1984 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1985 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1986 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1987 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1988 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001989 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1990 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1991 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1992 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1993 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1994 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1995 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1996 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1997 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1998 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1999 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2000 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2001 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2002 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2003 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2004 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2005 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2006 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2007 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2008 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2009 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2010 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2011 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2012 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2013 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2014 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2015 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2016 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2017 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2018 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002019 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2020 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2021 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2022 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2023 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2024 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2025 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2026 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2027 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2028 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2029 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2030 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2031 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2032 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2033 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2034 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2035 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2036 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2037 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2038 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2039 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2040 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2041 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2042 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002043 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2044 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2045 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
2046 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2047 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2048 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002049 "src/math/exp-avx2-p5.c",
2050 "src/math/exp-avx2-perm-p3.c",
2051 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002052 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002053 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002054 "src/math/sigmoid-avx2-rr2-p5-div.c",
2055 "src/math/sigmoid-avx2-rr1-p5-div.c",
2056 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2057 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2058 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2059 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002060]
2061
Marat Dukhan08c4a432019-10-03 09:29:21 -07002062AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002063 "src/f32-clamp/gen/avx512f-x16.c",
2064 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002065 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2066 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2067 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2068 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2069 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2070 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2071 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2072 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2073 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2074 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2075 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2076 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2077 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2078 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2079 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2080 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2081 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2082 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2083 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2084 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2085 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2086 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2087 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2088 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002089 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2090 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002091 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2092 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2093 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2094 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2095 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2096 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002097 "src/f32-prelu/gen/avx512f-2x16.c",
2098 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002099 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2100 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2101 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2102 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2103 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2104 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2105 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2106 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2107 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2108 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2109 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2110 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2111 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2112 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2113 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2114 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2115 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2116 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2117 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2118 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2119 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2120 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2121 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2122 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2123 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2124 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2125 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2126 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2127 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2128 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2129 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2130 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2131 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2132 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2133 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2134 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002135 "src/f32-relu/gen/avx512f-x16.c",
2136 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002137 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002138 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2139 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2140 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2141 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2142 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2143 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2144 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2145 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002146 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2147 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2148 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2149 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2150 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2151 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2152 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2153 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002154 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2155 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2156 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2157 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2158 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2159 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2160 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2161 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002162 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2163 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2164 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2165 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002166 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2167 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2168 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2169 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002170 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2171 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002172 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002173 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2174 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2175 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2176 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2177 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2178 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2179 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2180 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2181 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2182 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2183 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2184 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2185 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2186 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2187 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2188 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2189 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2190 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2191 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2192 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2193 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2194 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2195 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2196 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002197 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2198 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2199 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2200 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2201 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2202 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2203 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2204 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002205 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2206 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2207 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2208 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2209 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2210 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2211 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2212 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002213 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2214 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2215 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2216 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2217 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2218 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002219 "src/math/exp-avx512f-p5-scalef.c",
2220 "src/math/exp-avx512f-p5.c",
2221 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002222 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002223 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002224 "src/math/sqrt-avx512f-nr1fma.c",
2225 "src/math/sqrt-avx512f-nr2fma.c",
2226 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002227]
2228
2229AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002230 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002231 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002232 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2233 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002234 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002235 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002236 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002237 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002238 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2239 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002240 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2241 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2242 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2243 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244]
2245
2246AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002247 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002248 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2249 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002250 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002251 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2252 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002253 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2254 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2255 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002256 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002257 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2258 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2259 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002260 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002261 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2262 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002263 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002264 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2265 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2266 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2267 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2268 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2269 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2270 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2271 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2272 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2273 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2274 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2275 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2276 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2277 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2278 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2279 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2280 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2281 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002282 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2283 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002284 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002285 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2286 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2287 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2288 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2289 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2290 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2291 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2292 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2293 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2294 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2295 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2296 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2297 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2298 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2299 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2300 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2301 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2302 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002303 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2304 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2305 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2306 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2307 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2308 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2309 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2310 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2311 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2312 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2313 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2314 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2315 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2316 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2317 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2318 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2319 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2320 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002321]
2322
Marat Dukhan1b354632020-03-23 12:50:22 -07002323INTERNAL_MICROKERNEL_HDRS = [
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 Dukhan1566fee2020-08-02 21:55:41 -07002916 name = "xop_ukernels",
2917 hdrs = INTERNAL_HDRS,
2918 gcc_copts = xnnpack_gcc_std_copts(),
2919 gcc_x86_copts = ["-mxop"],
2920 msvc_copts = xnnpack_msvc_std_copts(),
2921 msvc_x86_32_copts = ["/arch:AVX"],
2922 msvc_x86_64_copts = ["/arch:AVX"],
2923 x86_srcs = XOP_UKERNELS,
2924 deps = [
2925 ":tables",
2926 "@FP16",
2927 "@pthreadpool",
2928 ],
2929)
2930
2931xnnpack_cc_library(
2932 name = "xop_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 = ["-mxop"],
2940 msvc_copts = xnnpack_msvc_std_copts(),
2941 msvc_x86_32_copts = ["/arch:AVX"],
2942 msvc_x86_64_copts = ["/arch:AVX"],
2943 x86_srcs = XOP_UKERNELS,
2944 deps = [
2945 ":tables",
2946 "@FP16",
2947 "@pthreadpool",
2948 ],
2949)
2950
2951xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002952 name = "fma3_ukernels",
2953 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002954 gcc_copts = xnnpack_gcc_std_copts(),
2955 gcc_x86_copts = ["-mfma"],
2956 msvc_copts = xnnpack_msvc_std_copts(),
2957 msvc_x86_32_copts = ["/arch:AVX"],
2958 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002959 x86_srcs = FMA3_UKERNELS,
2960 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002961 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002962 "@FP16",
2963 "@pthreadpool",
2964 ],
2965)
2966
2967xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002968 name = "fma3_ukernels_test_mode",
2969 hdrs = INTERNAL_HDRS,
2970 copts = [
2971 "-UNDEBUG",
2972 "-DXNN_TEST_MODE=1",
2973 ],
2974 gcc_copts = xnnpack_gcc_std_copts(),
2975 gcc_x86_copts = ["-mfma"],
2976 msvc_copts = xnnpack_msvc_std_copts(),
2977 msvc_x86_32_copts = ["/arch:AVX"],
2978 msvc_x86_64_copts = ["/arch:AVX"],
2979 x86_srcs = FMA3_UKERNELS,
2980 deps = [
2981 ":tables",
2982 "@FP16",
2983 "@pthreadpool",
2984 ],
2985)
2986
2987xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002988 name = "avx2_ukernels",
2989 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002990 gcc_copts = xnnpack_gcc_std_copts(),
2991 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002992 "-mfma",
2993 "-mavx2",
2994 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002995 msvc_copts = xnnpack_msvc_std_copts(),
2996 msvc_x86_32_copts = ["/arch:AVX2"],
2997 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002998 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002999 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003000 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003001 "@FP16",
3002 "@pthreadpool",
3003 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003004)
3005
3006xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003007 name = "avx2_ukernels_test_mode",
3008 hdrs = INTERNAL_HDRS,
3009 copts = [
3010 "-UNDEBUG",
3011 "-DXNN_TEST_MODE=1",
3012 ],
3013 gcc_copts = xnnpack_gcc_std_copts(),
3014 gcc_x86_copts = [
3015 "-mfma",
3016 "-mavx2",
3017 ],
3018 msvc_copts = xnnpack_msvc_std_copts(),
3019 msvc_x86_32_copts = ["/arch:AVX2"],
3020 msvc_x86_64_copts = ["/arch:AVX2"],
3021 x86_srcs = AVX2_UKERNELS,
3022 deps = [
3023 ":tables",
3024 "@FP16",
3025 "@pthreadpool",
3026 ],
3027)
3028
3029xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030 name = "avx512f_ukernels",
3031 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003032 gcc_copts = xnnpack_gcc_std_copts(),
3033 gcc_x86_copts = ["-mavx512f"],
3034 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3035 msvc_copts = xnnpack_msvc_std_copts(),
3036 msvc_x86_32_copts = ["/arch:AVX512"],
3037 msvc_x86_64_copts = ["/arch:AVX512"],
3038 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003040 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003041 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003042 "@FP16",
3043 "@pthreadpool",
3044 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003045)
3046
3047xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003048 name = "avx512f_ukernels_test_mode",
3049 hdrs = INTERNAL_HDRS,
3050 copts = [
3051 "-UNDEBUG",
3052 "-DXNN_TEST_MODE=1",
3053 ],
3054 gcc_copts = xnnpack_gcc_std_copts(),
3055 gcc_x86_copts = ["-mavx512f"],
3056 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3057 msvc_copts = xnnpack_msvc_std_copts(),
3058 msvc_x86_32_copts = ["/arch:AVX512"],
3059 msvc_x86_64_copts = ["/arch:AVX512"],
3060 msys_copts = ["-fno-asynchronous-unwind-tables"],
3061 x86_srcs = AVX512F_UKERNELS,
3062 deps = [
3063 ":tables",
3064 "@FP16",
3065 "@pthreadpool",
3066 ],
3067)
3068
3069xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003070 name = "asm_ukernels",
3071 hdrs = ["src/xnnpack/assembly.h"],
3072 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003073 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003074 aarch64_srcs = AARCH64_ASM_UKERNELS,
3075)
3076
Marat Dukhan3b59de22020-06-03 20:15:19 -07003077xnnpack_cc_library(
3078 name = "logging_utils",
3079 srcs = LOGGING_SRCS,
3080 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3081 copts = LOGGING_COPTS + [
3082 "-Isrc",
3083 "-Iinclude",
3084 ] + select({
3085 ":debug_build": [],
3086 "//conditions:default": xnnpack_min_size_copts(),
3087 }),
3088 gcc_copts = xnnpack_gcc_std_copts(),
3089 msvc_copts = xnnpack_msvc_std_copts(),
3090 visibility = xnnpack_visibility(),
3091 deps = [
3092 "@FP16",
3093 "@clog",
3094 "@pthreadpool",
3095 ],
3096)
3097
Marat Dukhan08c4a432019-10-03 09:29:21 -07003098xnnpack_aggregate_library(
3099 name = "ukernels",
3100 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003101 ":neon_ukernels",
3102 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003103 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003104 ":asm_ukernels",
3105 ],
3106 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003107 ":neon_ukernels",
3108 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003109 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003110 ":neonfp16arith_ukernels",
3111 ":asm_ukernels",
3112 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003113 generic_deps = [
3114 ":scalar_ukernels",
3115 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003116 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003117 ":psimd_fastmath_ukernels",
3118 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003120 wasm_deps = [
3121 ":wasm_ukernels",
3122 ],
3123 wasmsimd_deps = [
3124 ":wasm_ukernels",
3125 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003126 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003127 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003128 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003129 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003131 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003132 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003133 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003134 ":avx512f_ukernels",
3135 ],
3136)
3137
Marat Dukhan33fcf782020-05-24 14:27:15 -07003138xnnpack_aggregate_library(
3139 name = "ukernels_test_mode",
3140 aarch32_deps = [
3141 ":neon_ukernels_test_mode",
3142 ":neonfma_ukernels_test_mode",
3143 ":neonv8_ukernels_test_mode",
3144 ":asm_ukernels",
3145 ],
3146 aarch64_deps = [
3147 ":neon_ukernels_test_mode",
3148 ":neonfma_ukernels_test_mode",
3149 ":neonv8_ukernels_test_mode",
3150 ":neonfp16arith_ukernels_test_mode",
3151 ":asm_ukernels",
3152 ],
3153 generic_deps = [
3154 ":scalar_ukernels_test_mode",
3155 ],
3156 psimd_deps = [
3157 ":psimd_fastmath_ukernels_test_mode",
3158 ":psimd_accmath_ukernels_test_mode",
3159 ],
3160 wasm_deps = [
3161 ":wasm_ukernels_test_mode",
3162 ],
3163 wasmsimd_deps = [
3164 ":wasm_ukernels_test_mode",
3165 ],
3166 x86_deps = [
3167 ":sse2_ukernels_test_mode",
3168 ":ssse3_ukernels_test_mode",
3169 ":sse41_ukernels_test_mode",
3170 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003171 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003172 ":fma3_ukernels_test_mode",
3173 ":avx2_ukernels_test_mode",
3174 ":avx512f_ukernels_test_mode",
3175 ],
3176)
3177
Marat Dukhan86df3842020-06-17 14:03:31 -07003178# This target must be compatible with Android NDK r17 and compile with Clang 6.
3179# It is similar to :ukernels target, but excludes microkernels which require
3180# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003181xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003182 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003183 aarch32_deps = [
3184 ":neon_ukernels",
3185 ":neonfma_ukernels",
3186 ":neonv8_ukernels",
3187 ":asm_ukernels",
3188 ],
3189 aarch64_deps = [
3190 ":neon_ukernels",
3191 ":neonfma_ukernels",
3192 ":neonv8_ukernels",
3193 ":asm_ukernels",
3194 ],
3195 generic_deps = [
3196 ":scalar_ukernels",
3197 ],
3198 psimd_deps = [
3199 ":psimd_fastmath_ukernels",
3200 ":psimd_accmath_ukernels",
3201 ],
3202 wasm_deps = [
3203 ":wasm_ukernels",
3204 ],
3205 wasmsimd_deps = [
3206 ":wasm_ukernels",
3207 ],
3208 x86_deps = [
3209 ":sse2_ukernels",
3210 ":ssse3_ukernels",
3211 ":sse41_ukernels",
3212 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003213 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003214 ":fma3_ukernels",
3215 ":avx2_ukernels",
3216 ":avx512f_ukernels",
3217 ],
3218)
3219
Marat Dukhan08c4a432019-10-03 09:29:21 -07003220xnnpack_cc_library(
3221 name = "im2col",
3222 srcs = ["src/im2col.c"],
3223 hdrs = [
3224 "src/xnnpack/common.h",
3225 "src/xnnpack/im2col.h",
3226 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003227 gcc_copts = xnnpack_gcc_std_copts(),
3228 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003229)
3230
3231xnnpack_cc_library(
3232 name = "indirection",
3233 srcs = ["src/indirection.c"],
3234 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003235 gcc_copts = xnnpack_gcc_std_copts(),
3236 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237 deps = [
3238 "@FP16",
3239 "@FXdiv",
3240 "@pthreadpool",
3241 ],
3242)
3243
3244xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003245 name = "indirection_test_mode",
3246 srcs = ["src/indirection.c"],
3247 hdrs = INTERNAL_HDRS,
3248 copts = [
3249 "-UNDEBUG",
3250 "-DXNN_TEST_MODE=1",
3251 ],
3252 gcc_copts = xnnpack_gcc_std_copts(),
3253 msvc_copts = xnnpack_msvc_std_copts(),
3254 deps = [
3255 "@FP16",
3256 "@FXdiv",
3257 "@pthreadpool",
3258 ],
3259)
3260
3261xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003262 name = "packing",
3263 srcs = ["src/packing.c"],
3264 hdrs = INTERNAL_HDRS,
3265 gcc_copts = xnnpack_gcc_std_copts(),
3266 msvc_copts = xnnpack_msvc_std_copts(),
3267 deps = [
3268 "@FP16",
3269 "@FXdiv",
3270 "@pthreadpool",
3271 ],
3272)
3273
3274xnnpack_cc_library(
3275 name = "packing_test_mode",
3276 srcs = ["src/packing.c"],
3277 hdrs = INTERNAL_HDRS,
3278 copts = [
3279 "-UNDEBUG",
3280 "-DXNN_TEST_MODE=1",
3281 ],
3282 gcc_copts = xnnpack_gcc_std_copts(),
3283 msvc_copts = xnnpack_msvc_std_copts(),
3284 deps = [
3285 "@FP16",
3286 "@FXdiv",
3287 "@pthreadpool",
3288 ],
3289)
3290
3291xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003292 name = "operator_run",
3293 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003294 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003295 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003296 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3297 "//conditions:default": [],
3298 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003299 gcc_copts = xnnpack_gcc_std_copts(),
3300 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003301 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003302 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003303 "@FP16",
3304 "@FXdiv",
3305 "@clog",
3306 "@pthreadpool",
3307 ],
3308)
3309
Chao Mei6ddfc602020-05-13 22:29:36 -07003310xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003311 name = "operator_run_test_mode",
3312 srcs = ["src/operator-run.c"],
3313 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3314 copts = LOGGING_COPTS + [
3315 "-UNDEBUG",
3316 "-DXNN_TEST_MODE=1",
3317 ] + select({
3318 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3319 "//conditions:default": [],
3320 }),
3321 gcc_copts = xnnpack_gcc_std_copts(),
3322 msvc_copts = xnnpack_msvc_std_copts(),
3323 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003324 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003325 "@FP16",
3326 "@FXdiv",
3327 "@clog",
3328 "@pthreadpool",
3329 ],
3330)
3331
3332xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003333 name = "memory_planner",
3334 srcs = ["src/memory-planner.c"],
3335 hdrs = INTERNAL_HDRS,
3336 defines = select({
3337 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3338 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3339 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3340 }),
3341 gcc_copts = xnnpack_gcc_std_copts(),
3342 msvc_copts = xnnpack_msvc_std_copts(),
3343 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003344 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003345 "@pthreadpool",
3346 ],
3347)
3348
Marat Dukhan33fcf782020-05-24 14:27:15 -07003349xnnpack_cc_library(
3350 name = "memory_planner_test_mode",
3351 srcs = ["src/memory-planner.c"],
3352 hdrs = INTERNAL_HDRS,
3353 copts = [
3354 "-UNDEBUG",
3355 "-DXNN_TEST_MODE=1",
3356 ],
3357 defines = select({
3358 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3359 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3360 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3361 }),
3362 gcc_copts = xnnpack_gcc_std_copts(),
3363 msvc_copts = xnnpack_msvc_std_copts(),
3364 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003365 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003366 "@pthreadpool",
3367 ],
3368)
3369
Marat Dukhan08c4a432019-10-03 09:29:21 -07003370cc_library(
3371 name = "enable_assembly",
3372 defines = select({
3373 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3374 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003375 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003376 }),
3377)
3378
Marat Dukhan9de90e02020-06-18 16:04:12 -07003379cc_library(
3380 name = "enable_sparse",
3381 defines = select({
3382 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3383 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003384 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003385 }),
3386)
3387
Marat Dukhancf056b22019-10-07 10:26:29 -07003388xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003389 name = "operators",
3390 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003391 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003392 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003393 ],
3394 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003395 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003396 "-Isrc",
3397 "-Iinclude",
3398 ] + select({
3399 ":debug_build": [],
3400 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003401 }) + select({
3402 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3403 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003404 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003405 gcc_copts = xnnpack_gcc_std_copts(),
3406 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003407 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003408 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003409 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003410 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003411 "@FP16",
3412 "@FXdiv",
3413 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003414 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003415 ],
3416)
3417
Marat Dukhan10a38082020-04-17 03:58:35 -07003418xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003419 name = "operators_test_mode",
3420 srcs = OPERATOR_SRCS + [
3421 "src/memory.c",
3422 "src/operator-delete.c",
3423 ],
3424 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3425 copts = LOGGING_COPTS + [
3426 "-Isrc",
3427 "-Iinclude",
3428 "-UNDEBUG",
3429 "-DXNN_TEST_MODE=1",
3430 ] + select({
3431 ":debug_build": [],
3432 "//conditions:default": xnnpack_min_size_copts(),
3433 }) + select({
3434 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3435 "//conditions:default": [],
3436 }),
3437 gcc_copts = xnnpack_gcc_std_copts(),
3438 msvc_copts = xnnpack_msvc_std_copts(),
3439 deps = [
3440 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003441 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003442 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003443 "@FP16",
3444 "@FXdiv",
3445 "@clog",
3446 "@pthreadpool",
3447 ],
3448)
3449
3450xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003451 name = "XNNPACK",
3452 srcs = [
3453 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003454 "src/runtime.c",
3455 "src/subgraph.c",
3456 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003457 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003458 hdrs = ["include/xnnpack.h"],
3459 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003460 "-Isrc",
3461 "-Iinclude",
3462 ] + select({
3463 ":debug_build": [],
3464 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003465 }) + select({
3466 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3467 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003468 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003469 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003470 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003471 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003472 visibility = xnnpack_visibility(),
3473 deps = [
3474 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003475 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003476 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003477 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003478 ":operator_run",
3479 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003480 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003481 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003482 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003483 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003484 ] + select({
3485 ":emscripten": [],
3486 "//conditions:default": ["@cpuinfo"],
3487 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003488)
3489
Marat Dukhan10a38082020-04-17 03:58:35 -07003490xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003491 name = "XNNPACK_test_mode",
3492 srcs = [
3493 "src/init.c",
3494 "src/runtime.c",
3495 "src/subgraph.c",
3496 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003497 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003498 hdrs = ["include/xnnpack.h"],
3499 copts = LOGGING_COPTS + [
3500 "-Isrc",
3501 "-Iinclude",
3502 "-UNDEBUG",
3503 "-DXNN_TEST_MODE=1",
3504 ] + select({
3505 ":debug_build": [],
3506 "//conditions:default": xnnpack_min_size_copts(),
3507 }) + select({
3508 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3509 "//conditions:default": [],
3510 }),
3511 gcc_copts = xnnpack_gcc_std_copts(),
3512 includes = ["include"],
3513 msvc_copts = xnnpack_msvc_std_copts(),
3514 visibility = xnnpack_visibility(),
3515 deps = [
3516 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003517 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003518 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003519 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003520 ":operator_run_test_mode",
3521 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003522 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003523 "@clog",
3524 "@FP16",
3525 "@pthreadpool",
3526 ] + select({
3527 ":emscripten": [],
3528 "//conditions:default": ["@cpuinfo"],
3529 }),
3530)
3531
3532xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003533 name = "xnnpack_f32",
3534 srcs = [
3535 "src/init.c",
3536 "src/runtime.c",
3537 "src/subgraph.c",
3538 "src/tensor.c",
3539 ] + SUBGRAPH_SRCS,
3540 hdrs = ["include/xnnpack.h"],
3541 copts = LOGGING_COPTS + [
3542 "-Isrc",
3543 "-Iinclude",
3544 ] + select({
3545 ":debug_build": [],
3546 "//conditions:default": xnnpack_min_size_copts(),
3547 }) + select({
3548 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3549 "//conditions:default": [],
3550 }),
3551 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003552 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003553 "XNN_NO_U8_OPERATORS",
3554 "XNN_NO_X8_OPERATORS",
3555 "XNN_NO_F16_OPERATORS",
3556 "XNN_NO_X16_OPERATORS",
3557 ],
3558 gcc_copts = xnnpack_gcc_std_copts(),
3559 includes = ["include"],
3560 msvc_copts = xnnpack_msvc_std_copts(),
3561 visibility = xnnpack_visibility(),
3562 deps = [
3563 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003564 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003565 ":logging_utils",
3566 ":memory_planner",
3567 ":operator_run",
3568 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003569 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003570 "@clog",
3571 "@FP16",
3572 "@pthreadpool",
3573 ] + select({
3574 ":emscripten": [],
3575 "//conditions:default": ["@cpuinfo"],
3576 }),
3577)
3578
3579xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003580 name = "xnnpack_operators_nhwc_f32",
3581 srcs = [
3582 "src/init.c",
3583 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003584 hdrs = ["include/xnnpack.h"],
3585 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003586 "-Isrc",
3587 "-Iinclude",
3588 ] + select({
3589 ":debug_build": [],
3590 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003591 }) + select({
3592 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3593 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003594 }),
3595 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003596 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003597 "XNN_NO_U8_OPERATORS",
3598 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003599 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003600 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003601 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003603 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003604 visibility = xnnpack_visibility(),
3605 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003606 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003607 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003608 ":operator_run",
3609 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003610 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003611 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003612 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003613 ] + select({
3614 ":emscripten": [],
3615 "//conditions:default": ["@cpuinfo"],
3616 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617)
3618
Marat Dukhancf056b22019-10-07 10:26:29 -07003619xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003620 name = "bench_utils",
3621 srcs = ["bench/utils.cc"],
3622 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003623 deps = [
3624 "@com_google_benchmark//:benchmark",
3625 "@cpuinfo",
3626 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003627)
3628
Frank Barchard7e955972019-10-11 10:34:25 -07003629######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003630
3631xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003632 name = "qs8_gemm_bench",
3633 srcs = [
3634 "bench/gemm.h",
3635 "bench/qs8-gemm.cc",
3636 "src/xnnpack/AlignedAllocator.h",
3637 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3638 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3639)
3640
3641xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003642 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643 srcs = [
3644 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003645 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003646 "src/xnnpack/AlignedAllocator.h",
3647 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003648 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003649 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003650)
3651
3652xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003653 name = "qu8_requantization_bench",
3654 srcs = [
3655 "bench/qu8-requantization.cc",
3656 "src/xnnpack/requantization-stubs.h",
3657 "src/xnnpack/AlignedAllocator.h",
3658 ] + MICROKERNEL_BENCHMARK_HDRS,
3659 deps = MICROKERNEL_BENCHMARK_DEPS,
3660)
3661
3662xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003663 name = "f16_igemm_bench",
3664 srcs = [
3665 "bench/f16-igemm.cc",
3666 "bench/conv.h",
3667 "bench/google/conv.h",
3668 "src/xnnpack/AlignedAllocator.h",
3669 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003670 deps = MICROKERNEL_BENCHMARK_DEPS + [
3671 ":indirection",
3672 ":packing",
3673 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003674)
3675
3676xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003677 name = "f16_gemm_bench",
3678 srcs = [
3679 "bench/f16-gemm.cc",
3680 "bench/gemm.h",
3681 "src/xnnpack/AlignedAllocator.h",
3682 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003683 deps = MICROKERNEL_BENCHMARK_DEPS + [
3684 ":packing",
3685 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003686)
3687
3688xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003689 name = "f16_spmm_bench",
3690 srcs = [
3691 "bench/f16-spmm.cc",
3692 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003693 "src/xnnpack/AlignedAllocator.h",
3694 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003695 deps = MICROKERNEL_BENCHMARK_DEPS,
3696)
3697
3698xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003699 name = "f32_igemm_bench",
3700 srcs = [
3701 "bench/f32-igemm.cc",
3702 "bench/conv.h",
3703 "src/xnnpack/AlignedAllocator.h",
3704 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003705 deps = MICROKERNEL_BENCHMARK_DEPS + [
3706 ":indirection",
3707 ":packing",
3708 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709)
3710
3711xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003712 name = "f16_relu_bench",
3713 srcs = [
3714 "bench/f16-relu.cc",
3715 "src/xnnpack/AlignedAllocator.h",
3716 ] + MICROKERNEL_BENCHMARK_HDRS,
3717 deps = MICROKERNEL_BENCHMARK_DEPS,
3718)
3719
3720xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003721 name = "f32_conv_hwc_bench",
3722 srcs = [
3723 "bench/f32-conv-hwc.cc",
3724 "bench/dconv.h",
3725 "src/xnnpack/AlignedAllocator.h",
3726 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003727 deps = MICROKERNEL_BENCHMARK_DEPS + [
3728 ":packing",
3729 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003730)
3731
3732xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003733 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003734 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003735 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003736 "bench/dconv.h",
3737 "src/xnnpack/AlignedAllocator.h",
3738 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003739 deps = MICROKERNEL_BENCHMARK_DEPS + [
3740 ":packing",
3741 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003742)
3743
3744xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003745 name = "f16_dwconv_bench",
3746 srcs = [
3747 "bench/f16-dwconv.cc",
3748 "bench/dwconv.h",
3749 "bench/google/dwconv.h",
3750 "src/xnnpack/AlignedAllocator.h",
3751 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003752 deps = MICROKERNEL_BENCHMARK_DEPS + [
3753 ":indirection",
3754 ":packing",
3755 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003756)
3757
3758xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759 name = "f32_dwconv_bench",
3760 srcs = [
3761 "bench/f32-dwconv.cc",
3762 "bench/dwconv.h",
3763 "src/xnnpack/AlignedAllocator.h",
3764 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003765 deps = MICROKERNEL_BENCHMARK_DEPS + [
3766 ":indirection",
3767 ":packing",
3768 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769)
3770
3771xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003772 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003773 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003774 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003775 "bench/dwconv.h",
3776 "src/xnnpack/AlignedAllocator.h",
3777 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003778 deps = MICROKERNEL_BENCHMARK_DEPS + [
3779 ":indirection",
3780 ":packing",
3781 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003782)
3783
3784xnnpack_benchmark(
3785 name = "f32_gemm_bench",
3786 srcs = [
3787 "bench/f32-gemm.cc",
3788 "bench/gemm.h",
3789 "src/xnnpack/AlignedAllocator.h",
3790 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003791 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003792 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003793)
3794
3795xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003796 name = "f32_hswish_bench",
3797 srcs = [
3798 "bench/f32-hswish.cc",
3799 "src/xnnpack/AlignedAllocator.h",
3800 ] + MICROKERNEL_BENCHMARK_HDRS,
3801 deps = MICROKERNEL_BENCHMARK_DEPS,
3802)
3803
3804xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003805 name = "f32_raddexpminusmax_bench",
3806 srcs = [
3807 "bench/f32-raddexpminusmax.cc",
3808 "src/xnnpack/AlignedAllocator.h",
3809 ] + MICROKERNEL_BENCHMARK_HDRS,
3810 deps = MICROKERNEL_BENCHMARK_DEPS,
3811)
3812
3813xnnpack_benchmark(
3814 name = "f32_raddextexp_bench",
3815 srcs = [
3816 "bench/f32-raddextexp.cc",
3817 "src/xnnpack/AlignedAllocator.h",
3818 ] + MICROKERNEL_BENCHMARK_HDRS,
3819 deps = MICROKERNEL_BENCHMARK_DEPS,
3820)
3821
3822xnnpack_benchmark(
3823 name = "f32_raddstoreexpminusmax_bench",
3824 srcs = [
3825 "bench/f32-raddstoreexpminusmax.cc",
3826 "src/xnnpack/AlignedAllocator.h",
3827 ] + MICROKERNEL_BENCHMARK_HDRS,
3828 deps = MICROKERNEL_BENCHMARK_DEPS,
3829)
3830
3831xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003832 name = "f32_relu_bench",
3833 srcs = [
3834 "bench/f32-relu.cc",
3835 "src/xnnpack/AlignedAllocator.h",
3836 ] + MICROKERNEL_BENCHMARK_HDRS,
3837 deps = MICROKERNEL_BENCHMARK_DEPS,
3838)
3839
3840xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003841 name = "f32_rmax_bench",
3842 srcs = [
3843 "bench/f32-rmax.cc",
3844 "src/xnnpack/AlignedAllocator.h",
3845 ] + MICROKERNEL_BENCHMARK_HDRS,
3846 deps = MICROKERNEL_BENCHMARK_DEPS,
3847)
3848
3849xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003850 name = "f32_sigmoid_bench",
3851 srcs = [
3852 "bench/f32-sigmoid.cc",
3853 "src/xnnpack/AlignedAllocator.h",
3854 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003855 deps = MICROKERNEL_BENCHMARK_DEPS,
3856)
3857
3858xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003859 name = "f32_spmm_bench",
3860 srcs = [
3861 "bench/f32-spmm.cc",
3862 "bench/gemm.h",
3863 "src/xnnpack/AlignedAllocator.h",
3864 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003865 deps = MICROKERNEL_BENCHMARK_DEPS,
3866)
3867
3868xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003869 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003870 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003871 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003872 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003873 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003874 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003875)
3876
3877xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003878 name = "f32_vscaleexpminusmax_bench",
3879 srcs = [
3880 "bench/f32-vscaleexpminusmax.cc",
3881 "src/xnnpack/AlignedAllocator.h",
3882 ] + MICROKERNEL_BENCHMARK_HDRS,
3883 deps = MICROKERNEL_BENCHMARK_DEPS,
3884)
3885
3886xnnpack_benchmark(
3887 name = "f32_vscaleextexp_bench",
3888 srcs = [
3889 "bench/f32-vscaleextexp.cc",
3890 "src/xnnpack/AlignedAllocator.h",
3891 ] + MICROKERNEL_BENCHMARK_HDRS,
3892 deps = MICROKERNEL_BENCHMARK_DEPS,
3893)
3894
3895xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003896 name = "f32_vsqrt_bench",
3897 srcs = [
3898 "bench/f32-vsqrt.cc",
3899 "src/xnnpack/AlignedAllocator.h",
3900 ] + MICROKERNEL_BENCHMARK_HDRS,
3901 deps = MICROKERNEL_BENCHMARK_DEPS,
3902)
3903
3904xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003905 name = "f32_im2col_gemm_bench",
3906 srcs = [
3907 "bench/f32-im2col-gemm.cc",
3908 "bench/conv.h",
3909 "src/xnnpack/AlignedAllocator.h",
3910 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003911 deps = MICROKERNEL_BENCHMARK_DEPS + [
3912 ":im2col",
3913 ":packing",
3914 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003915)
3916
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003917xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003918 name = "rounding_bench",
3919 srcs = [
3920 "bench/rounding.cc",
3921 "src/xnnpack/math-stubs.h",
3922 "src/xnnpack/AlignedAllocator.h",
3923 ] + MICROKERNEL_BENCHMARK_HDRS,
3924 deps = MICROKERNEL_BENCHMARK_DEPS,
3925)
3926
Marat Dukhan08c4a432019-10-03 09:29:21 -07003927########################### Benchmarks for operators ###########################
3928
3929xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003930 name = "average_pooling_bench",
3931 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003932 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003933 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003934 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003935)
3936
3937xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003938 name = "bankers_rounding_bench",
3939 srcs = ["bench/bankers-rounding.cc"],
3940 copts = xnnpack_optional_tflite_copts(),
3941 tags = ["nowin32"],
3942 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3943)
3944
3945xnnpack_benchmark(
3946 name = "ceiling_bench",
3947 srcs = ["bench/ceiling.cc"],
3948 copts = xnnpack_optional_tflite_copts(),
3949 tags = ["nowin32"],
3950 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3951)
3952
3953xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 name = "channel_shuffle_bench",
3955 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003956 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003957)
3958
3959xnnpack_benchmark(
3960 name = "convolution_bench",
3961 srcs = ["bench/convolution.cc"],
3962 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003963 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003964 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003965)
3966
3967xnnpack_benchmark(
3968 name = "deconvolution_bench",
3969 srcs = ["bench/deconvolution.cc"],
3970 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003971 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003972 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003973)
3974
3975xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003976 name = "floor_bench",
3977 srcs = ["bench/floor.cc"],
3978 copts = xnnpack_optional_tflite_copts(),
3979 tags = ["nowin32"],
3980 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3981)
3982
3983xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003984 name = "global_average_pooling_bench",
3985 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003986 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003987)
3988
3989xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003990 name = "hardswish_bench",
3991 srcs = ["bench/hardswish.cc"],
3992 copts = xnnpack_optional_tflite_copts(),
3993 tags = ["nowin32"],
3994 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3995)
3996
3997xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003998 name = "max_pooling_bench",
3999 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004000 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004001)
4002
4003xnnpack_benchmark(
4004 name = "sigmoid_bench",
4005 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004006 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004007 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004008 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004009)
4010
4011xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004012 name = "prelu_bench",
4013 srcs = ["bench/prelu.cc"],
4014 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004015 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004016 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004017)
4018
4019xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004020 name = "softmax_bench",
4021 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004022 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004023 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004024 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004025)
4026
Marat Dukhan87727142020-06-24 15:24:10 -07004027xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004028 name = "square_root_bench",
4029 srcs = ["bench/square-root.cc"],
4030 copts = xnnpack_optional_tflite_copts(),
4031 tags = ["nowin32"],
4032 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4033)
4034
4035xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004036 name = "truncation_bench",
4037 srcs = ["bench/truncation.cc"],
4038 deps = OPERATOR_BENCHMARK_DEPS,
4039)
4040
Marat Dukhanc068bb62019-10-04 13:24:39 -07004041############################# End-to-end benchmarks ############################
4042
4043cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004044 name = "fp32_mobilenet_v1",
4045 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004046 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004047 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004048 linkstatic = True,
4049 deps = [
4050 ":XNNPACK",
4051 "@pthreadpool",
4052 ],
4053)
4054
4055cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004056 name = "fp16_mobilenet_v1",
4057 srcs = ["models/fp16-mobilenet-v1.cc"],
4058 hdrs = ["models/models.h"],
4059 copts = xnnpack_std_cxxopts(),
4060 linkstatic = True,
4061 deps = [
4062 ":XNNPACK",
4063 "@FP16",
4064 "@pthreadpool",
4065 ],
4066)
4067
4068cc_library(
4069 name = "fp32_mobilenet_v2",
4070 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004071 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004072 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004073 linkstatic = True,
4074 deps = [
4075 ":XNNPACK",
4076 "@pthreadpool",
4077 ],
4078)
4079
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004080cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004081 name = "fp16_mobilenet_v2",
4082 srcs = ["models/fp16-mobilenet-v2.cc"],
4083 hdrs = ["models/models.h"],
4084 copts = xnnpack_std_cxxopts(),
4085 linkstatic = True,
4086 deps = [
4087 ":XNNPACK",
4088 "@FP16",
4089 "@pthreadpool",
4090 ],
4091)
4092
4093cc_library(
4094 name = "fp32_mobilenet_v3_large",
4095 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004096 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004097 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004098 linkstatic = True,
4099 deps = [
4100 ":XNNPACK",
4101 "@pthreadpool",
4102 ],
4103)
4104
4105cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004106 name = "fp32_mobilenet_v3_small",
4107 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004108 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004109 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004110 linkstatic = True,
4111 deps = [
4112 ":XNNPACK",
4113 "@pthreadpool",
4114 ],
4115)
4116
Marat Dukhanc068bb62019-10-04 13:24:39 -07004117xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004118 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004119 srcs = [
4120 "bench/f32-dwconv-e2e.cc",
4121 "bench/end2end.h",
4122 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004123 deps = MICROKERNEL_BENCHMARK_DEPS + [
4124 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004125 ":fp32_mobilenet_v1",
4126 ":fp32_mobilenet_v2",
4127 ":fp32_mobilenet_v3_large",
4128 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004129 ],
4130)
4131
4132xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004133 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004134 srcs = [
4135 "bench/f32-gemm-e2e.cc",
4136 "bench/end2end.h",
4137 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004138 deps = MICROKERNEL_BENCHMARK_DEPS + [
4139 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004140 ":fp32_mobilenet_v1",
4141 ":fp32_mobilenet_v2",
4142 ":fp32_mobilenet_v3_large",
4143 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004144 ],
4145)
4146
4147xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004148 name = "end2end_bench",
4149 srcs = ["bench/end2end.cc"],
4150 deps = [
4151 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004152 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004153 ":fp16_mobilenet_v1",
4154 ":fp16_mobilenet_v2",
4155 ":fp32_mobilenet_v1",
4156 ":fp32_mobilenet_v2",
4157 ":fp32_mobilenet_v3_large",
4158 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004159 "@pthreadpool",
4160 ],
4161)
4162
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004163#################### Accuracy evaluation for math functions ####################
4164
4165xnnpack_benchmark(
4166 name = "f32_exp_eval",
4167 srcs = [
4168 "eval/f32-exp.cc",
4169 "src/xnnpack/AlignedAllocator.h",
4170 ] + ACCURACY_EVAL_HDRS,
4171 deps = ACCURACY_EVAL_DEPS,
4172)
4173
Marat Dukhan515c9772019-10-17 18:07:57 -07004174xnnpack_benchmark(
4175 name = "f32_expminus_eval",
4176 srcs = [
4177 "eval/f32-expminus.cc",
4178 "src/xnnpack/AlignedAllocator.h",
4179 ] + ACCURACY_EVAL_HDRS,
4180 deps = ACCURACY_EVAL_DEPS,
4181)
4182
Marat Dukhan98ba4412019-10-23 02:14:28 -07004183xnnpack_benchmark(
4184 name = "f32_extexp_eval",
4185 srcs = [
4186 "eval/f32-extexp.cc",
4187 "src/xnnpack/AlignedAllocator.h",
4188 ] + ACCURACY_EVAL_HDRS,
4189 deps = ACCURACY_EVAL_DEPS,
4190)
4191
Marat Dukhan8853b822020-05-07 12:19:01 -07004192xnnpack_unit_test(
4193 name = "f32_roundne_eval",
4194 srcs = [
4195 "eval/f32-roundne.cc",
4196 "src/xnnpack/AlignedAllocator.h",
4197 "src/xnnpack/math-stubs.h",
4198 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004199 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004200 deps = MICROKERNEL_TEST_DEPS,
4201)
4202
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004203xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004204 name = "f32_roundd_eval",
4205 srcs = [
4206 "eval/f32-roundd.cc",
4207 "src/xnnpack/AlignedAllocator.h",
4208 "src/xnnpack/math-stubs.h",
4209 ] + MICROKERNEL_TEST_HDRS,
4210 automatic = False,
4211 deps = MICROKERNEL_TEST_DEPS,
4212)
4213
4214xnnpack_unit_test(
4215 name = "f32_roundu_eval",
4216 srcs = [
4217 "eval/f32-roundu.cc",
4218 "src/xnnpack/AlignedAllocator.h",
4219 "src/xnnpack/math-stubs.h",
4220 ] + MICROKERNEL_TEST_HDRS,
4221 automatic = False,
4222 deps = MICROKERNEL_TEST_DEPS,
4223)
4224
4225xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004226 name = "f32_roundz_eval",
4227 srcs = [
4228 "eval/f32-roundz.cc",
4229 "src/xnnpack/AlignedAllocator.h",
4230 "src/xnnpack/math-stubs.h",
4231 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004232 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004233 deps = MICROKERNEL_TEST_DEPS,
4234)
4235
Marat Dukhan346a9e52019-11-15 09:06:30 -08004236xnnpack_benchmark(
4237 name = "f32_sigmoid_eval",
4238 srcs = [
4239 "eval/f32-sigmoid.cc",
4240 "src/xnnpack/AlignedAllocator.h",
4241 ] + ACCURACY_EVAL_HDRS,
4242 deps = ACCURACY_EVAL_DEPS,
4243)
4244
Marat Dukhan84000762020-06-29 18:38:43 -07004245xnnpack_benchmark(
4246 name = "f32_sqrt_eval",
4247 srcs = [
4248 "eval/f32-sqrt.cc",
4249 "src/xnnpack/AlignedAllocator.h",
4250 ] + ACCURACY_EVAL_HDRS,
4251 deps = ACCURACY_EVAL_DEPS + [
4252 ":bench_utils",
4253 ],
4254)
4255
Marat Dukhan08c4a432019-10-03 09:29:21 -07004256######################### Unit tests for micro-kernels #########################
4257
4258xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004259 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004260 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004261 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004262 "test/gemm-microkernel-tester.h",
4263 "src/xnnpack/AlignedAllocator.h",
4264 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004265 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004266)
4267
4268xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004269 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004270 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004271 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004272 "test/spmm-microkernel-tester.h",
4273 "src/xnnpack/AlignedAllocator.h",
4274 ] + MICROKERNEL_TEST_HDRS,
4275 deps = MICROKERNEL_TEST_DEPS,
4276)
4277
4278xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004279 name = "f32_argmaxpool_test",
4280 srcs = [
4281 "test/f32-argmaxpool.cc",
4282 "test/argmaxpool-microkernel-tester.h",
4283 "src/xnnpack/AlignedAllocator.h",
4284 ] + MICROKERNEL_TEST_HDRS,
4285 deps = MICROKERNEL_TEST_DEPS,
4286)
4287
4288xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004289 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004290 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004291 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004292 "test/avgpool-microkernel-tester.h",
4293 "src/xnnpack/AlignedAllocator.h",
4294 ] + MICROKERNEL_TEST_HDRS,
4295 deps = MICROKERNEL_TEST_DEPS,
4296)
4297
4298xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004299 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004300 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004301 "test/f32-ibilinear.cc",
4302 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004303 "src/xnnpack/AlignedAllocator.h",
4304 ] + MICROKERNEL_TEST_HDRS,
4305 deps = MICROKERNEL_TEST_DEPS,
4306)
4307
4308xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004309 name = "f16_clamp_test",
4310 srcs = [
4311 "test/f16-clamp.cc",
4312 "test/clamp-microkernel-tester.h",
4313 ] + MICROKERNEL_TEST_HDRS,
4314 deps = MICROKERNEL_TEST_DEPS,
4315)
4316
4317xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004318 name = "f32_clamp_test",
4319 srcs = [
4320 "test/f32-clamp.cc",
4321 "test/clamp-microkernel-tester.h",
4322 ] + MICROKERNEL_TEST_HDRS,
4323 deps = MICROKERNEL_TEST_DEPS,
4324)
4325
4326xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004327 name = "f32_igemm_test",
4328 srcs = [
4329 "test/f32-igemm.cc",
4330 "test/gemm-microkernel-tester.h",
4331 "src/xnnpack/AlignedAllocator.h",
4332 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004333 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004334)
4335
4336xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004337 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004338 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004339 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004340 "test/gemm-microkernel-tester.h",
4341 "src/xnnpack/AlignedAllocator.h",
4342 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004343 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004344)
4345
4346xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004347 name = "f32_igemm_minmax_test",
4348 srcs = [
4349 "test/f32-igemm-minmax.cc",
4350 "test/gemm-microkernel-tester.h",
4351 "src/xnnpack/AlignedAllocator.h",
4352 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004353 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004354)
4355
4356xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004357 name = "f16_igemm_minmax_test",
4358 srcs = [
4359 "test/f16-igemm-minmax.cc",
4360 "test/gemm-microkernel-tester.h",
4361 "src/xnnpack/AlignedAllocator.h",
4362 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004363 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004364)
4365
4366xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004367 name = "f32_conv_hwc_test",
4368 srcs = [
4369 "test/f32-conv-hwc.cc",
4370 "test/conv-hwc-microkernel-tester.h",
4371 "src/xnnpack/AlignedAllocator.h",
4372 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004373 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004374)
4375
4376xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004377 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004378 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004379 "test/f32-conv-hwc2chw.cc",
4380 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004381 "src/xnnpack/AlignedAllocator.h",
4382 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004383 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384)
4385
4386xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004387 name = "f32_dwconv_test",
4388 srcs = [
4389 "test/f32-dwconv.cc",
4390 "test/dwconv-microkernel-tester.h",
4391 "src/xnnpack/AlignedAllocator.h",
4392 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004393 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004394)
4395
4396xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004397 name = "f16_dwconv_minmax_test",
4398 srcs = [
4399 "test/f16-dwconv-minmax.cc",
4400 "test/dwconv-microkernel-tester.h",
4401 "src/xnnpack/AlignedAllocator.h",
4402 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004403 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004404)
4405
4406xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004407 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004408 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004409 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004410 "test/dwconv-microkernel-tester.h",
4411 "src/xnnpack/AlignedAllocator.h",
4412 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004413 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004414)
4415
4416xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004417 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004418 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004419 "test/f32-dwconv-chw.cc",
4420 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004421 "src/xnnpack/AlignedAllocator.h",
4422 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004423 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004424)
4425
4426xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004427 name = "f16_gavgpool_minmax_test",
4428 srcs = [
4429 "test/f16-gavgpool-minmax.cc",
4430 "test/gavgpool-microkernel-tester.h",
4431 "src/xnnpack/AlignedAllocator.h",
4432 ] + MICROKERNEL_TEST_HDRS,
4433 deps = MICROKERNEL_TEST_DEPS,
4434)
4435
4436xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004437 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004438 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004439 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004440 "test/gavgpool-microkernel-tester.h",
4441 "src/xnnpack/AlignedAllocator.h",
4442 ] + MICROKERNEL_TEST_HDRS,
4443 deps = MICROKERNEL_TEST_DEPS,
4444)
4445
4446xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004447 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004448 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004449 "test/f32-gavgpool-cw.cc",
4450 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 "src/xnnpack/AlignedAllocator.h",
4452 ] + MICROKERNEL_TEST_HDRS,
4453 deps = MICROKERNEL_TEST_DEPS,
4454)
4455
4456xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004457 name = "f32_gemm_test",
4458 srcs = [
4459 "test/f32-gemm.cc",
4460 "test/gemm-microkernel-tester.h",
4461 "src/xnnpack/AlignedAllocator.h",
4462 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004463 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004464)
4465
4466xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004467 name = "f32_gemm_relu_test",
4468 srcs = [
4469 "test/f32-gemm-relu.cc",
4470 "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 Dukhan467f6362020-05-22 23:21:55 -07004474)
4475
4476xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004477 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004478 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004479 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480 "test/gemm-microkernel-tester.h",
4481 "src/xnnpack/AlignedAllocator.h",
4482 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004483 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004484)
4485
4486xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004487 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004488 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004489 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490 "test/gemm-microkernel-tester.h",
4491 "src/xnnpack/AlignedAllocator.h",
4492 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004493 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004494)
4495
4496xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004497 name = "f16_hswish_test",
4498 srcs = [
4499 "test/f16-hswish.cc",
4500 "test/hswish-microkernel-tester.h",
4501 ] + MICROKERNEL_TEST_HDRS,
4502 deps = MICROKERNEL_TEST_DEPS,
4503)
4504
4505xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004506 name = "f32_hswish_test",
4507 srcs = [
4508 "test/f32-hswish.cc",
4509 "test/hswish-microkernel-tester.h",
4510 ] + MICROKERNEL_TEST_HDRS,
4511 deps = MICROKERNEL_TEST_DEPS,
4512)
4513
4514xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004515 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004516 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004517 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004518 "test/maxpool-microkernel-tester.h",
4519 ] + MICROKERNEL_TEST_HDRS,
4520 deps = MICROKERNEL_TEST_DEPS,
4521)
4522
4523xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004524 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004526 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004527 "test/avgpool-microkernel-tester.h",
4528 "src/xnnpack/AlignedAllocator.h",
4529 ] + MICROKERNEL_TEST_HDRS,
4530 deps = MICROKERNEL_TEST_DEPS,
4531)
4532
4533xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004534 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004535 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004536 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004537 "test/gemm-microkernel-tester.h",
4538 "src/xnnpack/AlignedAllocator.h",
4539 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004540 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004541)
4542
4543xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004544 name = "f16_prelu_test",
4545 srcs = [
4546 "test/f16-prelu.cc",
4547 "test/prelu-microkernel-tester.h",
4548 "src/xnnpack/AlignedAllocator.h",
4549 ] + MICROKERNEL_TEST_HDRS,
4550 deps = MICROKERNEL_TEST_DEPS,
4551)
4552
4553xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004554 name = "f32_prelu_test",
4555 srcs = [
4556 "test/f32-prelu.cc",
4557 "test/prelu-microkernel-tester.h",
4558 "src/xnnpack/AlignedAllocator.h",
4559 ] + MICROKERNEL_TEST_HDRS,
4560 deps = MICROKERNEL_TEST_DEPS,
4561)
4562
4563xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004564 name = "f32_raddexpminusmax_test",
4565 srcs = [
4566 "test/f32-raddexpminusmax.cc",
4567 "test/raddexpminusmax-microkernel-tester.h",
4568 ] + MICROKERNEL_TEST_HDRS,
4569 deps = MICROKERNEL_TEST_DEPS,
4570)
4571
4572xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004573 name = "f32_raddextexp_test",
4574 srcs = [
4575 "test/f32-raddextexp.cc",
4576 "test/raddextexp-microkernel-tester.h",
4577 ] + MICROKERNEL_TEST_HDRS,
4578 deps = MICROKERNEL_TEST_DEPS,
4579)
4580
4581xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004582 name = "f32_raddstoreexpminusmax_test",
4583 srcs = [
4584 "test/f32-raddstoreexpminusmax.cc",
4585 "test/raddstoreexpminusmax-microkernel-tester.h",
4586 ] + MICROKERNEL_TEST_HDRS,
4587 deps = MICROKERNEL_TEST_DEPS,
4588)
4589
4590xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004591 name = "f32_relu_test",
4592 srcs = [
4593 "test/f32-relu.cc",
4594 "test/vunary-microkernel-tester.h",
4595 ] + MICROKERNEL_TEST_HDRS,
4596 deps = MICROKERNEL_TEST_DEPS,
4597)
4598
4599xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004600 name = "f32_rmax_test",
4601 srcs = [
4602 "test/f32-rmax.cc",
4603 "test/rmax-microkernel-tester.h",
4604 ] + MICROKERNEL_TEST_HDRS,
4605 deps = MICROKERNEL_TEST_DEPS,
4606)
4607
4608xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004609 name = "f32_sigmoid_test",
4610 srcs = [
4611 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004612 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004613 ] + MICROKERNEL_TEST_HDRS,
4614 deps = MICROKERNEL_TEST_DEPS,
4615)
4616
4617xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004618 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004619 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004620 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004621 "test/spmm-microkernel-tester.h",
4622 "src/xnnpack/AlignedAllocator.h",
4623 ] + MICROKERNEL_TEST_HDRS,
4624 deps = MICROKERNEL_TEST_DEPS,
4625)
4626
4627xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004628 name = "f16_vadd_minmax_test",
4629 srcs = [
4630 "test/f16-vadd-minmax.cc",
4631 "test/vbinary-microkernel-tester.h",
4632 ] + MICROKERNEL_TEST_HDRS,
4633 deps = MICROKERNEL_TEST_DEPS,
4634)
4635
4636xnnpack_unit_test(
4637 name = "f16_vaddc_minmax_test",
4638 srcs = [
4639 "test/f16-vaddc-minmax.cc",
4640 "test/vbinaryc-microkernel-tester.h",
4641 ] + MICROKERNEL_TEST_HDRS,
4642 deps = MICROKERNEL_TEST_DEPS,
4643)
4644
4645xnnpack_unit_test(
4646 name = "f16_vdiv_minmax_test",
4647 srcs = [
4648 "test/f16-vdiv-minmax.cc",
4649 "test/vbinary-microkernel-tester.h",
4650 ] + MICROKERNEL_TEST_HDRS,
4651 deps = MICROKERNEL_TEST_DEPS,
4652)
4653
4654xnnpack_unit_test(
4655 name = "f16_vdivc_minmax_test",
4656 srcs = [
4657 "test/f16-vdivc-minmax.cc",
4658 "test/vbinaryc-microkernel-tester.h",
4659 ] + MICROKERNEL_TEST_HDRS,
4660 deps = MICROKERNEL_TEST_DEPS,
4661)
4662
4663xnnpack_unit_test(
4664 name = "f16_vrdivc_minmax_test",
4665 srcs = [
4666 "test/f16-vrdivc-minmax.cc",
4667 "test/vbinaryc-microkernel-tester.h",
4668 ] + MICROKERNEL_TEST_HDRS,
4669 deps = MICROKERNEL_TEST_DEPS,
4670)
4671
4672xnnpack_unit_test(
4673 name = "f16_vmax_test",
4674 srcs = [
4675 "test/f16-vmax.cc",
4676 "test/vbinary-microkernel-tester.h",
4677 ] + MICROKERNEL_TEST_HDRS,
4678 deps = MICROKERNEL_TEST_DEPS,
4679)
4680
4681xnnpack_unit_test(
4682 name = "f16_vmaxc_test",
4683 srcs = [
4684 "test/f16-vmaxc.cc",
4685 "test/vbinaryc-microkernel-tester.h",
4686 ] + MICROKERNEL_TEST_HDRS,
4687 deps = MICROKERNEL_TEST_DEPS,
4688)
4689
4690xnnpack_unit_test(
4691 name = "f16_vmin_test",
4692 srcs = [
4693 "test/f16-vmin.cc",
4694 "test/vbinary-microkernel-tester.h",
4695 ] + MICROKERNEL_TEST_HDRS,
4696 deps = MICROKERNEL_TEST_DEPS,
4697)
4698
4699xnnpack_unit_test(
4700 name = "f16_vminc_test",
4701 srcs = [
4702 "test/f16-vminc.cc",
4703 "test/vbinaryc-microkernel-tester.h",
4704 ] + MICROKERNEL_TEST_HDRS,
4705 deps = MICROKERNEL_TEST_DEPS,
4706)
4707
4708xnnpack_unit_test(
4709 name = "f16_vmul_minmax_test",
4710 srcs = [
4711 "test/f16-vmul-minmax.cc",
4712 "test/vbinary-microkernel-tester.h",
4713 ] + MICROKERNEL_TEST_HDRS,
4714 deps = MICROKERNEL_TEST_DEPS,
4715)
4716
4717xnnpack_unit_test(
4718 name = "f16_vmulc_minmax_test",
4719 srcs = [
4720 "test/f16-vmulc-minmax.cc",
4721 "test/vbinaryc-microkernel-tester.h",
4722 ] + MICROKERNEL_TEST_HDRS,
4723 deps = MICROKERNEL_TEST_DEPS,
4724)
4725
4726xnnpack_unit_test(
4727 name = "f16_vsub_minmax_test",
4728 srcs = [
4729 "test/f16-vsub-minmax.cc",
4730 "test/vbinary-microkernel-tester.h",
4731 ] + MICROKERNEL_TEST_HDRS,
4732 deps = MICROKERNEL_TEST_DEPS,
4733)
4734
4735xnnpack_unit_test(
4736 name = "f16_vsubc_minmax_test",
4737 srcs = [
4738 "test/f16-vsubc-minmax.cc",
4739 "test/vbinaryc-microkernel-tester.h",
4740 ] + MICROKERNEL_TEST_HDRS,
4741 deps = MICROKERNEL_TEST_DEPS,
4742)
4743
4744xnnpack_unit_test(
4745 name = "f16_vrsubc_minmax_test",
4746 srcs = [
4747 "test/f16-vrsubc-minmax.cc",
4748 "test/vbinaryc-microkernel-tester.h",
4749 ] + MICROKERNEL_TEST_HDRS,
4750 deps = MICROKERNEL_TEST_DEPS,
4751)
4752
4753xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004754 name = "f32_vabs_test",
4755 srcs = [
4756 "test/f32-vabs.cc",
4757 "test/vunary-microkernel-tester.h",
4758 ] + MICROKERNEL_TEST_HDRS,
4759 deps = MICROKERNEL_TEST_DEPS,
4760)
4761
4762xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004763 name = "f32_vadd_test",
4764 srcs = [
4765 "test/f32-vadd.cc",
4766 "test/vbinary-microkernel-tester.h",
4767 ] + MICROKERNEL_TEST_HDRS,
4768 deps = MICROKERNEL_TEST_DEPS,
4769)
4770
4771xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004772 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004773 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004774 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004775 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004776 ] + MICROKERNEL_TEST_HDRS,
4777 deps = MICROKERNEL_TEST_DEPS,
4778)
4779
4780xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004781 name = "f32_vaddc_test",
4782 srcs = [
4783 "test/f32-vaddc.cc",
4784 "test/vbinaryc-microkernel-tester.h",
4785 ] + MICROKERNEL_TEST_HDRS,
4786 deps = MICROKERNEL_TEST_DEPS,
4787)
4788
4789xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004790 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004791 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004792 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004793 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004794 ] + MICROKERNEL_TEST_HDRS,
4795 deps = MICROKERNEL_TEST_DEPS,
4796)
4797
4798xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004799 name = "f32_vdiv_test",
4800 srcs = [
4801 "test/f32-vdiv.cc",
4802 "test/vbinary-microkernel-tester.h",
4803 ] + MICROKERNEL_TEST_HDRS,
4804 deps = MICROKERNEL_TEST_DEPS,
4805)
4806
4807xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004808 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004809 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004810 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004811 "test/vbinary-microkernel-tester.h",
4812 ] + MICROKERNEL_TEST_HDRS,
4813 deps = MICROKERNEL_TEST_DEPS,
4814)
4815
4816xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004817 name = "f32_vdivc_test",
4818 srcs = [
4819 "test/f32-vdivc.cc",
4820 "test/vbinaryc-microkernel-tester.h",
4821 ] + MICROKERNEL_TEST_HDRS,
4822 deps = MICROKERNEL_TEST_DEPS,
4823)
4824
4825xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004826 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004827 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004828 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004829 "test/vbinaryc-microkernel-tester.h",
4830 ] + MICROKERNEL_TEST_HDRS,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004835 name = "f32_vrdivc_test",
4836 srcs = [
4837 "test/f32-vrdivc.cc",
4838 "test/vbinaryc-microkernel-tester.h",
4839 ] + MICROKERNEL_TEST_HDRS,
4840 deps = MICROKERNEL_TEST_DEPS,
4841)
4842
4843xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004844 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004845 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004846 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004847 "test/vbinaryc-microkernel-tester.h",
4848 ] + MICROKERNEL_TEST_HDRS,
4849 deps = MICROKERNEL_TEST_DEPS,
4850)
4851
4852xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004853 name = "f32_vmax_test",
4854 srcs = [
4855 "test/f32-vmax.cc",
4856 "test/vbinary-microkernel-tester.h",
4857 ] + MICROKERNEL_TEST_HDRS,
4858 deps = MICROKERNEL_TEST_DEPS,
4859)
4860
4861xnnpack_unit_test(
4862 name = "f32_vmaxc_test",
4863 srcs = [
4864 "test/f32-vmaxc.cc",
4865 "test/vbinaryc-microkernel-tester.h",
4866 ] + MICROKERNEL_TEST_HDRS,
4867 deps = MICROKERNEL_TEST_DEPS,
4868)
4869
4870xnnpack_unit_test(
4871 name = "f32_vmin_test",
4872 srcs = [
4873 "test/f32-vmin.cc",
4874 "test/vbinary-microkernel-tester.h",
4875 ] + MICROKERNEL_TEST_HDRS,
4876 deps = MICROKERNEL_TEST_DEPS,
4877)
4878
4879xnnpack_unit_test(
4880 name = "f32_vminc_test",
4881 srcs = [
4882 "test/f32-vminc.cc",
4883 "test/vbinaryc-microkernel-tester.h",
4884 ] + MICROKERNEL_TEST_HDRS,
4885 deps = MICROKERNEL_TEST_DEPS,
4886)
4887
4888xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004889 name = "f32_vmul_test",
4890 srcs = [
4891 "test/f32-vmul.cc",
4892 "test/vbinary-microkernel-tester.h",
4893 ] + MICROKERNEL_TEST_HDRS,
4894 deps = MICROKERNEL_TEST_DEPS,
4895)
4896
4897xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004898 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004899 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004900 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004901 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004902 ] + MICROKERNEL_TEST_HDRS,
4903 deps = MICROKERNEL_TEST_DEPS,
4904)
4905
4906xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004907 name = "f32_vmulc_test",
4908 srcs = [
4909 "test/f32-vmulc.cc",
4910 "test/vbinaryc-microkernel-tester.h",
4911 ] + MICROKERNEL_TEST_HDRS,
4912 deps = MICROKERNEL_TEST_DEPS,
4913)
4914
4915xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004916 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004917 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004918 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004919 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004920 ] + MICROKERNEL_TEST_HDRS,
4921 deps = MICROKERNEL_TEST_DEPS,
4922)
4923
4924xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004925 name = "f16_vmulcaddc_minmax_test",
4926 srcs = [
4927 "test/f16-vmulcaddc-minmax.cc",
4928 "test/vmulcaddc-microkernel-tester.h",
4929 "src/xnnpack/AlignedAllocator.h",
4930 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004931 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004932)
4933
4934xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004935 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004936 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004937 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004938 "test/vmulcaddc-microkernel-tester.h",
4939 "src/xnnpack/AlignedAllocator.h",
4940 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004941 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004942)
4943
4944xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004945 name = "f32_vlrelu_test",
4946 srcs = [
4947 "test/f32-vlrelu.cc",
4948 "test/vunary-microkernel-tester.h",
4949 ] + MICROKERNEL_TEST_HDRS,
4950 deps = MICROKERNEL_TEST_DEPS,
4951)
4952
4953xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004954 name = "f32_vneg_test",
4955 srcs = [
4956 "test/f32-vneg.cc",
4957 "test/vunary-microkernel-tester.h",
4958 ] + MICROKERNEL_TEST_HDRS,
4959 deps = MICROKERNEL_TEST_DEPS,
4960)
4961
4962xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004963 name = "f32_vrndne_test",
4964 srcs = [
4965 "test/f32-vrndne.cc",
4966 "test/vunary-microkernel-tester.h",
4967 ] + MICROKERNEL_TEST_HDRS,
4968 deps = MICROKERNEL_TEST_DEPS,
4969)
4970
4971xnnpack_unit_test(
4972 name = "f32_vrndz_test",
4973 srcs = [
4974 "test/f32-vrndz.cc",
4975 "test/vunary-microkernel-tester.h",
4976 ] + MICROKERNEL_TEST_HDRS,
4977 deps = MICROKERNEL_TEST_DEPS,
4978)
4979
4980xnnpack_unit_test(
4981 name = "f32_vrndu_test",
4982 srcs = [
4983 "test/f32-vrndu.cc",
4984 "test/vunary-microkernel-tester.h",
4985 ] + MICROKERNEL_TEST_HDRS,
4986 deps = MICROKERNEL_TEST_DEPS,
4987)
4988
4989xnnpack_unit_test(
4990 name = "f32_vrndd_test",
4991 srcs = [
4992 "test/f32-vrndd.cc",
4993 "test/vunary-microkernel-tester.h",
4994 ] + MICROKERNEL_TEST_HDRS,
4995 deps = MICROKERNEL_TEST_DEPS,
4996)
4997
4998xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004999 name = "f32_vscale_test",
5000 srcs = [
5001 "test/f32-vscale.cc",
5002 "test/vscale-microkernel-tester.h",
5003 ] + MICROKERNEL_TEST_HDRS,
5004 deps = MICROKERNEL_TEST_DEPS,
5005)
5006
5007xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005008 name = "f32_vscaleexpminusmax_test",
5009 srcs = [
5010 "test/f32-vscaleexpminusmax.cc",
5011 "test/vscaleexpminusmax-microkernel-tester.h",
5012 ] + MICROKERNEL_TEST_HDRS,
5013 deps = MICROKERNEL_TEST_DEPS,
5014)
5015
5016xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005017 name = "f32_vscaleextexp_test",
5018 srcs = [
5019 "test/f32-vscaleextexp.cc",
5020 "test/vscaleextexp-microkernel-tester.h",
5021 ] + MICROKERNEL_TEST_HDRS,
5022 deps = MICROKERNEL_TEST_DEPS,
5023)
5024
5025xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005026 name = "f32_vsqr_test",
5027 srcs = [
5028 "test/f32-vsqr.cc",
5029 "test/vunary-microkernel-tester.h",
5030 ] + MICROKERNEL_TEST_HDRS,
5031 deps = MICROKERNEL_TEST_DEPS,
5032)
5033
5034xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005035 name = "f32_vsqrdiff_test",
5036 srcs = [
5037 "test/f32-vsqrdiff.cc",
5038 "test/vbinary-microkernel-tester.h",
5039 ] + MICROKERNEL_TEST_HDRS,
5040 deps = MICROKERNEL_TEST_DEPS,
5041)
5042
5043xnnpack_unit_test(
5044 name = "f32_vsqrdiffc_test",
5045 srcs = [
5046 "test/f32-vsqrdiffc.cc",
5047 "test/vbinaryc-microkernel-tester.h",
5048 ] + MICROKERNEL_TEST_HDRS,
5049 deps = MICROKERNEL_TEST_DEPS,
5050)
5051
5052xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005053 name = "f32_vsqrt_test",
5054 srcs = [
5055 "test/f32-vsqrt.cc",
5056 "test/vunary-microkernel-tester.h",
5057 ] + MICROKERNEL_TEST_HDRS,
5058 deps = MICROKERNEL_TEST_DEPS,
5059)
5060
5061xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005062 name = "f32_vsub_test",
5063 srcs = [
5064 "test/f32-vsub.cc",
5065 "test/vbinary-microkernel-tester.h",
5066 ] + MICROKERNEL_TEST_HDRS,
5067 deps = MICROKERNEL_TEST_DEPS,
5068)
5069
5070xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005071 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005072 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005073 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005074 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005075 ] + MICROKERNEL_TEST_HDRS,
5076 deps = MICROKERNEL_TEST_DEPS,
5077)
5078
5079xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005080 name = "f32_vsubc_test",
5081 srcs = [
5082 "test/f32-vsubc.cc",
5083 "test/vbinaryc-microkernel-tester.h",
5084 ] + MICROKERNEL_TEST_HDRS,
5085 deps = MICROKERNEL_TEST_DEPS,
5086)
5087
5088xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005089 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005090 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005091 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005092 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005093 ] + MICROKERNEL_TEST_HDRS,
5094 deps = MICROKERNEL_TEST_DEPS,
5095)
5096
5097xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005098 name = "f32_vrsubc_test",
5099 srcs = [
5100 "test/f32-vrsubc.cc",
5101 "test/vbinaryc-microkernel-tester.h",
5102 ] + MICROKERNEL_TEST_HDRS,
5103 deps = MICROKERNEL_TEST_DEPS,
5104)
5105
5106xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005107 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005108 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005109 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005110 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005111 ] + MICROKERNEL_TEST_HDRS,
5112 deps = MICROKERNEL_TEST_DEPS,
5113)
5114
5115xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005116 name = "qs8_gemm_minmax_test",
5117 srcs = [
5118 "test/qs8-gemm-minmax.cc",
5119 "test/gemm-microkernel-tester.h",
5120 "src/xnnpack/AlignedAllocator.h",
5121 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5122 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5123)
5124
5125xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005126 name = "qs8_igemm_minmax_test",
5127 srcs = [
5128 "test/qs8-igemm-minmax.cc",
5129 "test/gemm-microkernel-tester.h",
5130 "src/xnnpack/AlignedAllocator.h",
5131 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5132 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5133)
5134
5135xnnpack_unit_test(
5136 name = "qs8_requantization_test",
5137 srcs = [
5138 "src/xnnpack/requantization-stubs.h",
5139 "test/qs8-requantization.cc",
5140 "test/requantization-tester.h",
5141 ] + MICROKERNEL_TEST_HDRS,
5142 deps = MICROKERNEL_TEST_DEPS,
5143)
5144
5145xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005146 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005147 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005148 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005149 "test/avgpool-microkernel-tester.h",
5150 "src/xnnpack/AlignedAllocator.h",
5151 ] + MICROKERNEL_TEST_HDRS,
5152 deps = MICROKERNEL_TEST_DEPS,
5153)
5154
5155xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005156 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005157 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005158 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159 "test/gemm-microkernel-tester.h",
5160 "src/xnnpack/AlignedAllocator.h",
5161 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005162 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005163)
5164
5165xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005166 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005167 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005168 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005169 "test/dwconv-microkernel-tester.h",
5170 "src/xnnpack/AlignedAllocator.h",
5171 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005172 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005173)
5174
5175xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005176 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005177 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005178 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005179 "test/gavgpool-microkernel-tester.h",
5180 "src/xnnpack/AlignedAllocator.h",
5181 ] + MICROKERNEL_TEST_HDRS,
5182 deps = MICROKERNEL_TEST_DEPS,
5183)
5184
5185xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005186 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005188 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005189 "test/gemm-microkernel-tester.h",
5190 "src/xnnpack/AlignedAllocator.h",
5191 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005192 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193)
5194
5195xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005196 name = "qu8_requantization_test",
5197 srcs = [
5198 "src/xnnpack/requantization-stubs.h",
5199 "test/qu8-requantization.cc",
5200 "test/requantization-tester.h",
5201 ] + MICROKERNEL_TEST_HDRS,
5202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
5205xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005206 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005207 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005208 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005209 "test/vadd-microkernel-tester.h",
5210 ] + MICROKERNEL_TEST_HDRS,
5211 deps = MICROKERNEL_TEST_DEPS,
5212)
5213
5214xnnpack_unit_test(
5215 name = "u8_clamp_test",
5216 srcs = [
5217 "test/u8-clamp.cc",
5218 "test/clamp-microkernel-tester.h",
5219 ] + MICROKERNEL_TEST_HDRS,
5220 deps = MICROKERNEL_TEST_DEPS,
5221)
5222
5223xnnpack_unit_test(
5224 name = "u8_lut32norm_test",
5225 srcs = [
5226 "test/u8-lut32norm.cc",
5227 "test/lut-norm-microkernel-tester.h",
5228 ] + MICROKERNEL_TEST_HDRS,
5229 deps = MICROKERNEL_TEST_DEPS,
5230)
5231
5232xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005233 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005234 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005235 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005236 "test/maxpool-microkernel-tester.h",
5237 ] + MICROKERNEL_TEST_HDRS,
5238 deps = MICROKERNEL_TEST_DEPS,
5239)
5240
5241xnnpack_unit_test(
5242 name = "u8_rmax_test",
5243 srcs = [
5244 "test/u8-rmax.cc",
5245 "test/rmax-microkernel-tester.h",
5246 ] + MICROKERNEL_TEST_HDRS,
5247 deps = MICROKERNEL_TEST_DEPS,
5248)
5249
5250xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005251 name = "x32_fill_test",
5252 srcs = [
5253 "test/x32-fill.cc",
5254 "test/fill-microkernel-tester.h",
5255 ] + MICROKERNEL_TEST_HDRS,
5256 deps = MICROKERNEL_TEST_DEPS,
5257)
5258
5259xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260 name = "x32_packx_test",
5261 srcs = [
5262 "test/x32-packx.cc",
5263 "test/pack-microkernel-tester.h",
5264 "src/xnnpack/AlignedAllocator.h",
5265 ] + MICROKERNEL_TEST_HDRS,
5266 deps = MICROKERNEL_TEST_DEPS,
5267)
5268
5269xnnpack_unit_test(
5270 name = "x32_pad_test",
5271 srcs = [
5272 "test/x32-pad.cc",
5273 "test/pad-microkernel-tester.h",
5274 ] + MICROKERNEL_TEST_HDRS,
5275 deps = MICROKERNEL_TEST_DEPS,
5276)
5277
5278xnnpack_unit_test(
5279 name = "x32_unpool_test",
5280 srcs = [
5281 "test/x32-unpool.cc",
5282 "test/unpool-microkernel-tester.h",
5283 ] + MICROKERNEL_TEST_HDRS,
5284 deps = MICROKERNEL_TEST_DEPS,
5285)
5286
5287xnnpack_unit_test(
5288 name = "x32_zip_test",
5289 srcs = [
5290 "test/x32-zip.cc",
5291 "test/zip-microkernel-tester.h",
5292 ] + MICROKERNEL_TEST_HDRS,
5293 deps = MICROKERNEL_TEST_DEPS,
5294)
5295
5296xnnpack_unit_test(
5297 name = "x8_lut_test",
5298 srcs = [
5299 "test/x8-lut.cc",
5300 "test/lut-microkernel-tester.h",
5301 ] + MICROKERNEL_TEST_HDRS,
5302 deps = MICROKERNEL_TEST_DEPS,
5303)
5304
5305xnnpack_unit_test(
5306 name = "x8_zip_test",
5307 srcs = [
5308 "test/x8-zip.cc",
5309 "test/zip-microkernel-tester.h",
5310 ] + MICROKERNEL_TEST_HDRS,
5311 deps = MICROKERNEL_TEST_DEPS,
5312)
5313
Marat Dukhan20c3b922020-03-10 03:45:06 -07005314########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005315
5316xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005317 name = "operator_size_test",
5318 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005319 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005320)
5321
Marat Dukhan20c3b922020-03-10 03:45:06 -07005322xnnpack_binary(
5323 name = "subgraph_size_test",
5324 srcs = ["test/subgraph-size.c"],
5325 deps = [":XNNPACK"],
5326)
5327
5328########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005329
5330xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005331 name = "abs_nc_test",
5332 srcs = [
5333 "test/abs-nc.cc",
5334 "test/abs-operator-tester.h",
5335 ],
5336 deps = OPERATOR_TEST_DEPS,
5337)
5338
5339xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005340 name = "add_nd_test",
5341 srcs = [
5342 "test/add-nd.cc",
5343 "test/binary-elementwise-operator-tester.h",
5344 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005345 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005346)
5347
5348xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005349 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005350 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005351 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005352 "test/argmax-pooling-operator-tester.h",
5353 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005354 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005355)
5356
5357xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005358 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005359 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005360 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005361 "test/average-pooling-operator-tester.h",
5362 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005363 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005364)
5365
5366xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005367 name = "bankers_rounding_nc_test",
5368 srcs = [
5369 "test/bankers-rounding-nc.cc",
5370 "test/bankers-rounding-operator-tester.h",
5371 ],
5372 deps = OPERATOR_TEST_DEPS,
5373)
5374
5375xnnpack_unit_test(
5376 name = "ceiling_nc_test",
5377 srcs = [
5378 "test/ceiling-nc.cc",
5379 "test/ceiling-operator-tester.h",
5380 ],
5381 deps = OPERATOR_TEST_DEPS,
5382)
5383
5384xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005385 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005386 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005387 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005388 "test/channel-shuffle-operator-tester.h",
5389 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005390 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005391)
5392
5393xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005394 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005395 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005396 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005397 "test/clamp-operator-tester.h",
5398 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005399 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005400)
5401
5402xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005403 name = "constant_pad_nd_test",
5404 srcs = [
5405 "test/constant-pad-nd.cc",
5406 "test/constant-pad-operator-tester.h",
5407 ],
5408 deps = OPERATOR_TEST_DEPS,
5409)
5410
5411xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005412 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005414 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415 "test/convolution-operator-tester.h",
5416 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005417 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005418)
5419
5420xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005421 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005422 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005423 "test/convolution-nchw.cc",
5424 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005425 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005426 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005427)
5428
5429xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005430 name = "copy_nc_test",
5431 srcs = [
5432 "test/copy-nc.cc",
5433 "test/copy-operator-tester.h",
5434 ],
5435 deps = OPERATOR_TEST_DEPS,
5436)
5437
5438xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005439 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005440 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005441 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005442 "test/deconvolution-operator-tester.h",
5443 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005444 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005445)
5446
5447xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005448 name = "divide_nd_test",
5449 srcs = [
5450 "test/binary-elementwise-operator-tester.h",
5451 "test/divide-nd.cc",
5452 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005453 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005454)
5455
5456xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005457 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005458 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005459 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005460 "test/fully-connected-operator-tester.h",
5461 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005462 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005463)
5464
5465xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005466 name = "floor_nc_test",
5467 srcs = [
5468 "test/floor-nc.cc",
5469 "test/floor-operator-tester.h",
5470 ],
5471 deps = OPERATOR_TEST_DEPS,
5472)
5473
5474xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005475 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005476 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005477 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005479 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005480 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005481)
5482
5483xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005484 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005485 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005486 "test/global-average-pooling-ncw.cc",
5487 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005488 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005489 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490)
5491
5492xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005493 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005494 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005495 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496 "test/hardswish-operator-tester.h",
5497 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005498 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005499)
5500
5501xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005502 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005503 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005504 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005505 "test/leaky-relu-operator-tester.h",
5506 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005507 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005508)
5509
5510xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005511 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005512 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005513 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005514 "test/max-pooling-operator-tester.h",
5515 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005516 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005517)
5518
5519xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005520 name = "maximum_nd_test",
5521 srcs = [
5522 "test/binary-elementwise-operator-tester.h",
5523 "test/maximum-nd.cc",
5524 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005525 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005526)
5527
5528xnnpack_unit_test(
5529 name = "minimum_nd_test",
5530 srcs = [
5531 "test/binary-elementwise-operator-tester.h",
5532 "test/minimum-nd.cc",
5533 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005534 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005535)
5536
5537xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005538 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005539 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005540 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005541 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005542 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005543 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005544)
5545
5546xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005547 name = "negate_nc_test",
5548 srcs = [
5549 "test/negate-nc.cc",
5550 "test/negate-operator-tester.h",
5551 ],
5552 deps = OPERATOR_TEST_DEPS,
5553)
5554
5555xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005556 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005557 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005558 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005559 "test/prelu-operator-tester.h",
5560 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005561 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005562)
5563
5564xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005565 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005566 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005567 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005568 "test/resize-bilinear-operator-tester.h",
5569 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005570 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005571)
5572
5573xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005574 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005575 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005576 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005577 "test/sigmoid-operator-tester.h",
5578 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005579 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005580)
5581
5582xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005583 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005584 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005585 "test/softmax-nc.cc",
5586 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005587 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005588 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005589)
5590
5591xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005592 name = "square_nc_test",
5593 srcs = [
5594 "test/square-nc.cc",
5595 "test/square-operator-tester.h",
5596 ],
5597 deps = OPERATOR_TEST_DEPS,
5598)
5599
5600xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005601 name = "square_root_nc_test",
5602 srcs = [
5603 "test/square-root-nc.cc",
5604 "test/square-root-operator-tester.h",
5605 ],
5606 deps = OPERATOR_TEST_DEPS,
5607)
5608
5609xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005610 name = "squared_difference_nd_test",
5611 srcs = [
5612 "test/binary-elementwise-operator-tester.h",
5613 "test/squared-difference-nd.cc",
5614 ],
5615 deps = OPERATOR_TEST_DEPS,
5616)
5617
5618xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005619 name = "subtract_nd_test",
5620 srcs = [
5621 "test/binary-elementwise-operator-tester.h",
5622 "test/subtract-nd.cc",
5623 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005624 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005625)
5626
5627xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005628 name = "truncation_nc_test",
5629 srcs = [
5630 "test/truncation-nc.cc",
5631 "test/truncation-operator-tester.h",
5632 ],
5633 deps = OPERATOR_TEST_DEPS,
5634)
5635
5636xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005637 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005638 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005639 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005640 "test/unpooling-operator-tester.h",
5641 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005642 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005643)
5644
Chao Mei6ddfc602020-05-13 22:29:36 -07005645############################### Misc unit tests ###############################
5646
5647xnnpack_unit_test(
5648 name = "memory_planner_test",
5649 srcs = [
5650 "test/memory-planner-test.cc",
5651 ],
5652 deps = [
5653 ":XNNPACK",
5654 ":memory_planner",
5655 ],
5656)
5657
Marat Dukhan08c4a432019-10-03 09:29:21 -07005658############################# Build configurations #############################
5659
Marat Dukhanb8642352019-10-30 15:43:02 -07005660# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005661config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005662 name = "xnn_enable_assembly_explicit_true",
5663 define_values = {"xnn_enable_assembly": "true"},
5664)
5665
5666# Disables usage of assembly kernels.
5667config_setting(
5668 name = "xnn_enable_assembly_explicit_false",
5669 define_values = {"xnn_enable_assembly": "false"},
5670)
5671
Marat Dukhan9de90e02020-06-18 16:04:12 -07005672# Enables usage of sparse inference.
5673config_setting(
5674 name = "xnn_enable_sparse_explicit_true",
5675 define_values = {"xnn_enable_sparse": "true"},
5676)
5677
5678# Disables usage of sparse inference.
5679config_setting(
5680 name = "xnn_enable_sparse_explicit_false",
5681 define_values = {"xnn_enable_sparse": "false"},
5682)
5683
Marat Dukhan05702cf2020-03-26 15:41:33 -07005684# Disables usage of HMP-aware optimizations.
5685config_setting(
5686 name = "xnn_enable_hmp_explicit_false",
5687 define_values = {"xnn_enable_hmp": "false"},
5688)
5689
Chao Mei6ddfc602020-05-13 22:29:36 -07005690# Enable usage of optimized memory allocation
5691config_setting(
5692 name = "xnn_enable_memopt_explicit_true",
5693 define_values = {"xnnpack_enable_memopt": "true"},
5694)
5695
5696# Disable usage of optimized memory allocation
5697config_setting(
5698 name = "xnn_enable_memopt_explicit_false",
5699 define_values = {"xnnpack_enable_memopt": "false"},
5700)
5701
Marat Dukhanb8642352019-10-30 15:43:02 -07005702# Builds with -c dbg
5703config_setting(
5704 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005705 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005706 "compilation_mode": "dbg",
5707 },
5708)
5709
5710# Builds with -c opt
5711config_setting(
5712 name = "optimized_build",
5713 values = {
5714 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005715 },
5716)
5717
5718config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005719 name = "linux_k8",
5720 values = {"cpu": "k8"},
5721)
5722
5723config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005724 name = "linux_arm",
5725 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005726)
5727
5728config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005729 name = "linux_armeabi",
5730 values = {"cpu": "armeabi"},
5731)
5732
5733config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005734 name = "linux_armhf",
5735 values = {"cpu": "armhf"},
5736)
5737
5738config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005739 name = "linux_armv7a",
5740 values = {"cpu": "armv7a"},
5741)
5742
5743config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005744 name = "linux_aarch64",
5745 values = {"cpu": "aarch64"},
5746)
5747
5748config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005749 name = "android",
5750 values = {"crosstool_top": "//external:android/crosstool"},
5751)
5752
5753config_setting(
5754 name = "android_armv7",
5755 values = {
5756 "crosstool_top": "//external:android/crosstool",
5757 "cpu": "armeabi-v7a",
5758 },
5759)
5760
5761config_setting(
5762 name = "android_arm64",
5763 values = {
5764 "crosstool_top": "//external:android/crosstool",
5765 "cpu": "arm64-v8a",
5766 },
5767)
5768
5769config_setting(
5770 name = "android_x86",
5771 values = {
5772 "crosstool_top": "//external:android/crosstool",
5773 "cpu": "x86",
5774 },
5775)
5776
5777config_setting(
5778 name = "android_x86_64",
5779 values = {
5780 "crosstool_top": "//external:android/crosstool",
5781 "cpu": "x86_64",
5782 },
5783)
5784
5785config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005786 name = "windows_x86_64",
5787 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005788)
5789
5790config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005791 name = "windows_x86_64_clang",
5792 values = {
5793 "compiler": "clang-cl",
5794 "cpu": "x64_windows",
5795 },
5796)
5797
5798config_setting(
5799 name = "windows_x86_64_mingw",
5800 values = {
5801 "compiler": "mingw-gcc",
5802 "cpu": "x64_windows",
5803 },
5804)
5805
5806config_setting(
5807 name = "windows_x86_64_msys",
5808 values = {
5809 "compiler": "msys-gcc",
5810 "cpu": "x64_windows",
5811 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005812)
5813
5814config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005815 name = "macos_x86_64",
5816 values = {
5817 "apple_platform_type": "macos",
5818 "cpu": "darwin",
5819 },
5820)
5821
5822config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005823 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005824 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005825)
5826
5827config_setting(
5828 name = "emscripten_wasm",
5829 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005830 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005831 "cpu": "wasm",
5832 },
5833)
5834
5835config_setting(
5836 name = "emscripten_wasmsimd",
5837 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005838 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005839 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005840 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005841 },
5842)
5843
5844config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005845 name = "ios_armv7",
5846 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005847 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005848 "cpu": "ios_armv7",
5849 },
5850)
5851
5852config_setting(
5853 name = "ios_arm64",
5854 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005855 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005856 "cpu": "ios_arm64",
5857 },
5858)
5859
5860config_setting(
5861 name = "ios_arm64e",
5862 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005863 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005864 "cpu": "ios_arm64e",
5865 },
5866)
5867
5868config_setting(
5869 name = "ios_x86",
5870 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005871 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005872 "cpu": "ios_i386",
5873 },
5874)
5875
5876config_setting(
5877 name = "ios_x86_64",
5878 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005879 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005880 "cpu": "ios_x86_64",
5881 },
5882)
5883
5884config_setting(
5885 name = "watchos_armv7k",
5886 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005887 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005888 "cpu": "watchos_armv7k",
5889 },
5890)
5891
5892config_setting(
5893 name = "watchos_arm64_32",
5894 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005895 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005896 "cpu": "watchos_arm64_32",
5897 },
5898)
5899
5900config_setting(
5901 name = "watchos_x86",
5902 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005903 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005904 "cpu": "watchos_i386",
5905 },
5906)
5907
5908config_setting(
5909 name = "watchos_x86_64",
5910 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005911 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005912 "cpu": "watchos_x86_64",
5913 },
5914)
5915
5916config_setting(
5917 name = "tvos_arm64",
5918 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005919 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005920 "cpu": "tvos_arm64",
5921 },
5922)
5923
5924config_setting(
5925 name = "tvos_x86_64",
5926 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005927 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005928 "cpu": "tvos_x86_64",
5929 },
5930)