blob: 1833cd7436ea6fbf44870b719935cec9b04fef29 [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 Dukhan138560c2020-08-03 18:57:34 -0700852 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700853 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700854 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700855 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700856 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700857 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700858 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700859 "src/x32-zip/x2-wasmsimd.c",
860 "src/x32-zip/x3-wasmsimd.c",
861 "src/x32-zip/x4-wasmsimd.c",
862 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700863 "src/math/roundne-wasmsimd-addsub.c",
864 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700865 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700866 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700867 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700868 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700869 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700870 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
871 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700872]
873
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800874PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700875 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700876 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700877 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700878 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700879 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700880 "src/f32-spmm/gen/4x1-minmax-psimd.c",
881 "src/f32-spmm/gen/8x1-minmax-psimd.c",
882 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700883]
884
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800885PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700886 "src/qs8-requantization/precise-psimd.c",
887 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700888 "src/qu8-requantization/precise-psimd.c",
889 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800890]
891
Marat Dukhan08c4a432019-10-03 09:29:21 -0700892# ISA-specific micro-kernels
893NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700894 "src/f32-argmaxpool/4x-neon-c4.c",
895 "src/f32-argmaxpool/9p8x-neon-c4.c",
896 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700897 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
898 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700899 "src/f32-clamp/gen/neon-x4.c",
900 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700903 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
904 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700905 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
906 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700907 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
908 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700909 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
910 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
911 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
912 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700913 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
914 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
915 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
916 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700917 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
918 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
919 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
920 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700921 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700922 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
923 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700924 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
925 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
926 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
927 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
928 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
929 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
930 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
931 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
932 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
933 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
934 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
935 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
936 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
937 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
938 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
939 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
940 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
941 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
942 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
944 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
945 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
946 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
947 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
948 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
949 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
950 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
951 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
952 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
953 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
954 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700955 "src/f32-hswish/gen/hswish-neon-x4.c",
956 "src/f32-hswish/gen/hswish-neon-x8.c",
957 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700958 "src/f32-ibilinear/gen/neon-c4.c",
959 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700960 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
961 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
962 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
963 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
964 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
965 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
966 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
967 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
968 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
969 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
970 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
971 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
972 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
973 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
974 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
975 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700976 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
977 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
978 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700979 "src/f32-ppmm/gen/4x8-minmax-neon.c",
980 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700981 "src/f32-prelu/gen/neon-1x4.c",
982 "src/f32-prelu/gen/neon-1x8.c",
983 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800984 "src/f32-prelu/gen/neon-2x4.c",
985 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700986 "src/f32-prelu/gen/neon-2x16.c",
987 "src/f32-prelu/gen/neon-4x4.c",
988 "src/f32-prelu/gen/neon-4x8.c",
989 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1005 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1013 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001014 "src/f32-relu/gen/neon-x4.c",
1015 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001016 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001017 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001018 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1019 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1020 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1021 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1022 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1023 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1026 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1027 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1028 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1029 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1030 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1031 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1032 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1033 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1034 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1035 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001036 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1037 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1038 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1039 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001040 "src/f32-vbinary/gen/vmax-neon-x4.c",
1041 "src/f32-vbinary/gen/vmax-neon-x8.c",
1042 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1043 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1044 "src/f32-vbinary/gen/vmin-neon-x4.c",
1045 "src/f32-vbinary/gen/vmin-neon-x8.c",
1046 "src/f32-vbinary/gen/vminc-neon-x4.c",
1047 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001048 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1049 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1050 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1051 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1052 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1053 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001054 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1055 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1056 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1057 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001058 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1059 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1060 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1061 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001062 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1063 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001064 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1065 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001066 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1067 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1068 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1069 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1070 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1071 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1072 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1073 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001074 "src/f32-vunary/gen/vabs-neon-x4.c",
1075 "src/f32-vunary/gen/vabs-neon-x8.c",
1076 "src/f32-vunary/gen/vneg-neon-x4.c",
1077 "src/f32-vunary/gen/vneg-neon-x8.c",
1078 "src/f32-vunary/gen/vsqr-neon-x4.c",
1079 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001080 "src/qs8-requantization/precise-neon.c",
1081 "src/qs8-requantization/fp32-neon.c",
1082 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001083 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1084 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1085 "src/qu8-dwconv/up8x9-minmax-neon.c",
1086 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1087 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1088 "src/qu8-gemm/4x8-minmax-neon.c",
1089 "src/qu8-gemm/8x8-minmax-neon.c",
1090 "src/qu8-igemm/4x8-minmax-neon.c",
1091 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001092 "src/qu8-requantization/precise-neon.c",
1093 "src/qu8-requantization/fp32-neon.c",
1094 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001095 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001096 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001097 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001098 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001099 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001100 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001101 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001102 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001103 "src/x32-zip/x2-neon.c",
1104 "src/x32-zip/x3-neon.c",
1105 "src/x32-zip/x4-neon.c",
1106 "src/x32-zip/xm-neon.c",
1107 "src/x8-zip/x2-neon.c",
1108 "src/x8-zip/x3-neon.c",
1109 "src/x8-zip/x4-neon.c",
1110 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001111 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001112 "src/math/roundd-neon-addsub.c",
1113 "src/math/roundd-neon-cvt.c",
1114 "src/math/roundu-neon-addsub.c",
1115 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001116 "src/math/roundz-neon-addsub.c",
1117 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001118 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001119 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1120 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1121 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1122 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1123 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1124 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001125 "src/math/sqrt-neon-nr1rsqrts.c",
1126 "src/math/sqrt-neon-nr2rsqrts.c",
1127 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001128]
1129
1130NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001131 "src/f32-ibilinear/gen/neonfma-c4.c",
1132 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001133 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1134 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1135 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1136 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1137 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1138 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1139 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1140 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1141 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001142 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1143 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1144 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1145 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001146 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1147 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1148 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1149 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001150 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1151 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1152 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1153 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001154 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1155 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1156 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1157 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1158 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1159 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1160 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1161 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1162 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1163 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1164 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1165 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1166 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1167 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1168 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1169 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1170 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1171 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001172 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1173 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001174 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1177 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1178 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1179 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1180 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1181 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1182 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1185 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1187 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1188 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1189 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1190 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1191 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1196 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1197 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001198 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001252 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1253 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001254 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1255 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1256 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1257 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1258 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1259 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1260 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1261 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1262 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1263 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1264 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1265 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1266 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1267 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1268 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1269 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1270 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1271 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1272 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1273 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001274 "src/math/exp-neonfma-lut64-p2.c",
1275 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001276 "src/math/expminus-neonfma-lut2048-p1.c",
1277 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001278 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001279 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1280 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1281 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1282 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1283 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1284 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1285 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1286 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1287 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1288 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1289 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1290 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1291 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1292 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1293 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1294 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1295 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1296 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001297 "src/math/sqrt-neonfma-nr1fma.c",
1298 "src/math/sqrt-neonfma-nr2fma.c",
1299 "src/math/sqrt-neonfma-nr3fma.c",
1300 "src/math/sqrt-neonfma-nr2fma1adj.c",
1301 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001302]
1303
1304AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001305 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1306 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1307 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1308 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1309 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1310 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001311 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1312 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1313 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1314 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1315 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1316 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1317 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1318 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1319 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1320 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1321 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1322 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1323 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1324 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1325 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1326 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1327 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1328 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1329 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1330 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001331 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1332 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001333 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1334 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001335 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1336 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001337 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1338 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001339 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1340 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1341 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1342 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1343 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001344 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1361 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001362 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1363 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1364 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1365 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1366 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1367 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1368 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1369 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1370 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1371 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1372 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1373 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1374 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1375 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1376 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1377 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1378 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1379 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001380 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1381 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001382 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1383 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1384 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1385 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1386 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1387 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001388]
1389
Marat Dukhan8853b822020-05-07 12:19:01 -07001390NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001391 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1392 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1393 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1394 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1395 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1396 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1397 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1398 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001399 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001400 "src/math/roundd-neonv8.c",
1401 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001402 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001403]
1404
Marat Dukhan08c4a432019-10-03 09:29:21 -07001405AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001406 "src/f16-clamp/gen/neonfp16arith-x16.c",
1407 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001408 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1409 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1410 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1411 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1412 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1413 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1414 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1415 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1416 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1417 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1418 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1419 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001420 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1421 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001422 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1423 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001424 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001425 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001426 "src/f16-relu/gen/neonfp16arith-x16.c",
1427 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001428 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1429 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1430 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1431 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1432 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1433 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1434 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1435 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1436 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1437 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1438 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1439 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1440 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1441 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1442 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1443 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1444 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1445 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1446 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1447 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1448 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1449 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1450 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1451 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1452 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1453 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1454 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1455 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001456 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1457 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001458 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1459 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1460 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1461 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1462 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1463 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1464 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1465 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001466 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1467 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1468 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1469 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001470 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1471 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1472 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1473 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1474 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1475 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1476 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1477 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1478 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1479 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1480 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1481 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001482 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1483 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1484 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1485 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1486 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1487 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1488 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1489 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001490]
1491
1492SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001493 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1494 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001495 "src/f32-clamp/gen/sse-x4.c",
1496 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001497 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1498 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001499 "src/f32-dwconv-chw/3x3p1-sse.c",
1500 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001501 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1502 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1503 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1504 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1505 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1506 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1507 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1508 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1509 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1510 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1511 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1512 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001513 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001514 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1515 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001516 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1517 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1518 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1519 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1520 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1521 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1522 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1523 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1524 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1525 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1526 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1527 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1528 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001529 "src/f32-hswish/gen/hswish-sse-x4.c",
1530 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001531 "src/f32-ibilinear/gen/sse-c4.c",
1532 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001533 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1534 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1535 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1536 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1537 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1538 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1539 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001540 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1541 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1542 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001543 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001544 "src/f32-prelu/gen/sse-2x4.c",
1545 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001546 "src/f32-relu/gen/sse-x4.c",
1547 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001548 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001549 "src/f32-spmm/gen/4x1-minmax-sse.c",
1550 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001551 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001552 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1553 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1554 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1555 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1556 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1557 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1558 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1559 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001560 "src/f32-vbinary/gen/vmax-sse-x4.c",
1561 "src/f32-vbinary/gen/vmax-sse-x8.c",
1562 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1563 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1564 "src/f32-vbinary/gen/vmin-sse-x4.c",
1565 "src/f32-vbinary/gen/vmin-sse-x8.c",
1566 "src/f32-vbinary/gen/vminc-sse-x4.c",
1567 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001568 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1569 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1570 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1571 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1572 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1573 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1574 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1575 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001576 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1577 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1578 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1579 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001580 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1581 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1582 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1583 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001584 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1585 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001586 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1587 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001588 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1589 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001590 "src/f32-vunary/gen/vabs-sse-x4.c",
1591 "src/f32-vunary/gen/vabs-sse-x8.c",
1592 "src/f32-vunary/gen/vneg-sse-x4.c",
1593 "src/f32-vunary/gen/vneg-sse-x8.c",
1594 "src/f32-vunary/gen/vsqr-sse-x4.c",
1595 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001596 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001597 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001598 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001599 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001600 "src/math/roundd-sse-addsub.c",
1601 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001602 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001603 "src/math/sqrt-sse-hh1mac.c",
1604 "src/math/sqrt-sse-nr1mac.c",
1605 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001606]
1607
1608SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001609 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1610 "src/f32-argmaxpool/4x-sse2-c4.c",
1611 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001612 "src/f32-prelu/gen/sse2-2x4.c",
1613 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001614 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1615 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1616 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1617 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1618 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1619 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1620 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1621 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1622 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1623 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1624 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1625 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001626 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001627 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001628 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001629 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001630 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1631 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001632 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1633 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001634 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1635 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1636 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1637 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1638 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1639 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1640 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1641 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001642 "src/qs8-requantization/precise-sse2.c",
1643 "src/qs8-requantization/fp32-sse2.c",
1644 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001645 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1646 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001647 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1648 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001649 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1650 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001651 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1652 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001653 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001654 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1655 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001656 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001657 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1658 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1659 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001660 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1661 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001662 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1663 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001664 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1665 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001666 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001667 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1668 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001669 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001670 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1671 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1672 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1673 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1674 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1675 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1676 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1677 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001678 "src/qu8-requantization/precise-sse2.c",
1679 "src/qu8-requantization/fp32-sse2.c",
1680 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001681 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001682 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001683 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001684 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001685 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001686 "src/x32-zip/x2-sse2.c",
1687 "src/x32-zip/x3-sse2.c",
1688 "src/x32-zip/x4-sse2.c",
1689 "src/x32-zip/xm-sse2.c",
1690 "src/x8-zip/x2-sse2.c",
1691 "src/x8-zip/x3-sse2.c",
1692 "src/x8-zip/x4-sse2.c",
1693 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001694 "src/math/exp-sse2-p5.c",
1695 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001696 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001697 "src/math/roundd-sse2-cvt.c",
1698 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001699 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001700 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001701]
1702
1703SSSE3_UKERNELS = [
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001704 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1705 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001706 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1707 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001708 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1709 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001710 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1711 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001712 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001713 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1714 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001715 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001716 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1717 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1718 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001719 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1720 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001721 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1722 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001723 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1724 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001725 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001726 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1727 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001728 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001729 "src/qs8-requantization/precise-ssse3.c",
1730 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001731 "src/qu8-requantization/precise-ssse3.c",
1732 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001733]
1734
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001735SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001736 "src/f32-prelu/gen/sse41-2x4.c",
1737 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001738 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001739 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001740 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001741 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001742 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1743 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001744 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1745 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001746 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1747 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1748 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1749 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1750 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1751 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1752 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1753 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001754 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1755 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001756 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1757 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001758 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1759 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001760 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1761 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001762 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001763 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1764 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001765 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001766 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1767 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1768 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001769 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1770 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001771 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1772 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001773 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1774 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001775 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001776 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1777 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001778 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001779 "src/qs8-requantization/fp32-sse4.c",
1780 "src/qs8-requantization/precise-sse4.c",
1781 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001782 "src/qu8-requantization/precise-sse4.c",
1783 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001784 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001785 "src/math/roundd-sse41.c",
1786 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001787 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001788]
1789
Marat Dukhan08c4a432019-10-03 09:29:21 -07001790AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001791 "src/f32-clamp/gen/avx-x8.c",
1792 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001793 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1794 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1795 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1796 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1797 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1798 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1799 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1800 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1801 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1802 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1803 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1804 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1805 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1806 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1807 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1808 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1809 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1810 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1811 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1812 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1813 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1814 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1815 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1816 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1817 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1818 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1819 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1820 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1821 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1822 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001823 "src/f32-hswish/gen/hswish-avx-x8.c",
1824 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001825 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1826 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1827 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1828 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1829 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1830 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1831 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1832 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1833 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001834 "src/f32-prelu/gen/avx-2x8.c",
1835 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001836 "src/f32-relu/gen/avx-x16.c",
1837 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001839 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1840 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1841 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1842 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1843 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1844 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1845 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1846 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001847 "src/f32-vbinary/gen/vmax-avx-x8.c",
1848 "src/f32-vbinary/gen/vmax-avx-x16.c",
1849 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1850 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1851 "src/f32-vbinary/gen/vmin-avx-x8.c",
1852 "src/f32-vbinary/gen/vmin-avx-x16.c",
1853 "src/f32-vbinary/gen/vminc-avx-x8.c",
1854 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001855 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1856 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1857 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1858 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1859 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1860 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1861 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1862 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001863 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1864 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1865 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1866 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001867 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1868 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1869 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1870 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001871 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1872 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001873 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001874 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1875 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1876 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1877 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1878 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1879 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1880 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1881 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001882 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1883 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001884 "src/f32-vunary/gen/vabs-avx-x8.c",
1885 "src/f32-vunary/gen/vabs-avx-x16.c",
1886 "src/f32-vunary/gen/vneg-avx-x8.c",
1887 "src/f32-vunary/gen/vneg-avx-x16.c",
1888 "src/f32-vunary/gen/vsqr-avx-x8.c",
1889 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001890]
1891
Marat Dukhan1566fee2020-08-02 21:55:41 -07001892XOP_UKERNELS = [
1893 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1894 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1895 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1896 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001897 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
1898 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001899 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1900 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001901 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001902 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1903 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001904 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001905 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
1906 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
1907 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001908 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1909 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1910 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1911 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1912 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1913 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001914 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001915 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1916 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001917 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001918]
1919
Marat Dukhanfda12b82019-11-21 12:27:59 -08001920FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001921 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1922 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1923 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1924 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1925 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1926 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1927 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1928 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1929 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1930 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1931 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1932 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1933 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1934 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1935 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1936 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1937 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1938 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1939 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1940 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1941 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1942 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1943 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1944 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1945 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1946 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1947 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1948 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1949 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1950 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1951 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1952 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1953 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1954 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1955 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1956 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1957 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1958 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1959 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1960 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001961 "src/f32-hswish/gen/hswish-fma3-x8.c",
1962 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001963 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1964 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1965 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1966 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1967 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1968 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1969 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1970 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1971 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1972 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1973 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1974 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1975 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1976 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001977 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1978 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1979 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1980 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1981 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1982 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1983 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1984 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001985 "src/math/sqrt-fma3-nr1fma.c",
1986 "src/math/sqrt-fma3-nr2fma.c",
1987 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001988]
1989
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001990AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001991 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1992 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1993 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1994 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1995 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1996 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1997 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1998 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1999 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2000 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2001 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2002 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2003 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2004 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2005 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2006 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2007 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2008 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2009 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2010 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2011 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2012 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2013 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2014 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2015 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2016 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2017 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2018 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2019 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2020 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2021 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2022 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2023 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2024 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2025 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2026 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002027 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2028 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2029 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2030 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2031 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2032 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2033 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2034 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2035 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2036 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2037 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2038 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2039 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2040 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2041 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2042 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2043 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2044 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2045 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2046 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2047 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2048 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2049 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2050 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2051 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2052 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2053 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2054 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2055 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2056 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002057 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2058 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2059 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2060 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2061 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2062 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2063 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2064 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2065 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2066 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2067 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2068 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2069 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2070 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2071 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2072 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2073 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2074 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2075 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2076 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2077 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2078 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2079 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2080 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002081 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2082 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2083 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002084 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2085 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2086 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002087 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2088 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2089 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002090 "src/math/exp-avx2-p5.c",
2091 "src/math/exp-avx2-perm-p3.c",
2092 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002093 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002094 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002095 "src/math/sigmoid-avx2-rr2-p5-div.c",
2096 "src/math/sigmoid-avx2-rr1-p5-div.c",
2097 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2098 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2099 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2100 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002101]
2102
Marat Dukhan08c4a432019-10-03 09:29:21 -07002103AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002104 "src/f32-clamp/gen/avx512f-x16.c",
2105 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002106 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2107 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2108 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2109 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2110 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2111 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2112 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2113 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2114 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2115 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2116 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2117 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2118 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2119 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2120 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2121 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2122 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2123 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2124 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2125 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2126 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2127 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2128 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2129 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002130 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2131 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002132 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2133 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2134 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2135 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2136 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2137 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002138 "src/f32-prelu/gen/avx512f-2x16.c",
2139 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002140 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2141 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2142 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2143 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2144 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2145 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2146 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2147 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2148 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2149 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2150 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2151 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2152 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2153 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2154 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2155 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2156 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2157 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2158 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2159 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2160 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2161 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2162 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2163 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2164 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2165 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2166 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2167 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2168 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2169 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2170 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2171 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2172 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2173 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2174 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2175 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002176 "src/f32-relu/gen/avx512f-x16.c",
2177 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002178 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002179 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2180 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2181 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2182 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2183 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2184 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2185 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2186 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002187 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2188 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2189 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2190 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2191 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2192 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2193 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2194 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002195 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2196 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2197 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2198 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2199 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2200 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2201 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2202 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002203 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2204 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2205 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2206 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002207 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2208 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2209 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2210 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002211 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2212 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002213 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002214 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2215 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2216 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2217 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2218 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2219 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2220 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2221 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2222 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2223 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2224 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2225 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2226 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2227 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2228 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2229 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2230 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2231 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2232 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2233 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2234 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2235 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2236 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2237 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002238 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2239 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2240 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2241 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2242 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2243 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2244 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2245 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002246 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2247 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2248 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2249 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2250 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2251 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2252 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2253 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002254 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2255 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2256 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2257 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2258 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2259 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002260 "src/math/exp-avx512f-p5-scalef.c",
2261 "src/math/exp-avx512f-p5.c",
2262 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002263 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002264 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002265 "src/math/sqrt-avx512f-nr1fma.c",
2266 "src/math/sqrt-avx512f-nr2fma.c",
2267 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268]
2269
2270AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002271 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002272 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002273 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2274 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002275 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002276 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002277 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002278 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002279 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2280 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002281 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2282 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2283 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2284 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002285]
2286
2287AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002288 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002289 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2290 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002291 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002292 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2293 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002294 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2295 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2296 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002297 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002298 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2299 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2300 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002301 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002302 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2303 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002304 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002305 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2306 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2307 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2308 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2309 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2310 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2311 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2312 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2313 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2314 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2315 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2316 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2317 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2318 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2319 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2320 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2321 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2322 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002323 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2324 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002325 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002326 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2327 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2328 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2329 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2330 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2331 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2332 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2333 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2334 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2335 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2336 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2337 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2338 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2339 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2340 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2341 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2342 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2343 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002344 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2345 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2346 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2347 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2348 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2349 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2350 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2351 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2352 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2353 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2354 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2355 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2356 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2357 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2358 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2359 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2360 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2361 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002362]
2363
Marat Dukhan1b354632020-03-23 12:50:22 -07002364INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002365 "src/xnnpack/argmaxpool.h",
2366 "src/xnnpack/avgpool.h",
2367 "src/xnnpack/clamp.h",
2368 "src/xnnpack/common.h",
2369 "src/xnnpack/conv.h",
2370 "src/xnnpack/dwconv.h",
2371 "src/xnnpack/gavgpool.h",
2372 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002373 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002374 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002375 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002376 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002377 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002378 "src/xnnpack/lut.h",
2379 "src/xnnpack/math.h",
2380 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002381 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002382 "src/xnnpack/packx.h",
2383 "src/xnnpack/pad.h",
2384 "src/xnnpack/params.h",
2385 "src/xnnpack/pavgpool.h",
2386 "src/xnnpack/ppmm.h",
2387 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002388 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002389 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002390 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391 "src/xnnpack/rmax.h",
2392 "src/xnnpack/scalar-utils.h",
2393 "src/xnnpack/spmm.h",
2394 "src/xnnpack/unpool.h",
2395 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002396 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002397 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002398 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002399 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002400 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002401 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002402 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002403]
2404
2405INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002406 "include/xnnpack.h",
2407 "src/xnnpack/allocator.h",
2408 "src/xnnpack/compute.h",
2409 "src/xnnpack/im2col.h",
2410 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002411 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002412 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002413 "src/xnnpack/operator.h",
2414 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002415 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002416 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002417 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002418 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002419]
2420
Marat Dukhan1b354632020-03-23 12:50:22 -07002421ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002422 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002423]
2424
Marat Dukhan1b354632020-03-23 12:50:22 -07002425MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002426 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002427 "include/xnnpack.h",
2428]
2429
Marat Dukhan1b354632020-03-23 12:50:22 -07002430MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002431 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002432 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002433 "src/xnnpack/requantization.h",
2434 "include/xnnpack.h",
2435]
2436
2437OPERATOR_TEST_PARAMS_HDRS = [
2438 "src/xnnpack/params.h",
2439 "src/xnnpack/common.h",
2440]
2441
2442WEIGHTS_PACK_HDRS = [
2443 "src/xnnpack/pack.h",
2444 "src/xnnpack/operator.h",
2445 "src/xnnpack/compute.h",
2446]
2447
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002448LOGGING_COPTS = select({
2449 # No logging in optimized mode
2450 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2451 # Full logging in debug mode
2452 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2453 # Error-only logging in default (fastbuild) mode
2454 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2455})
2456
Marat Dukhan3b59de22020-06-03 20:15:19 -07002457LOGGING_SRCS = select({
2458 # No logging in optimized mode
2459 ":optimized_build": [],
2460 "//conditions:default": [
2461 "src/operator-strings.c",
2462 "src/subgraph-strings.c",
2463 ],
2464})
2465
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002466LOGGING_HDRS = [
2467 "src/xnnpack/log.h",
2468]
2469
Marat Dukhan08c4a432019-10-03 09:29:21 -07002470xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002471 name = "tables",
2472 srcs = TABLE_SRCS,
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 Dukhan3a77ea72019-12-23 12:10:24 -08002476)
2477
2478xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002479 name = "scalar_ukernels",
2480 srcs = SCALAR_UKERNELS,
2481 hdrs = INTERNAL_HDRS,
2482 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002483 gcc_copts = xnnpack_gcc_std_copts(),
2484 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002485 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002486 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002487 "@FP16",
2488 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002489 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002490 ],
2491)
2492
2493xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002494 name = "scalar_ukernels_test_mode",
2495 srcs = SCALAR_UKERNELS,
2496 hdrs = INTERNAL_HDRS,
2497 aarch32_copts = ["-marm"],
2498 copts = [
2499 "-UNDEBUG",
2500 "-DXNN_TEST_MODE=1",
2501 ],
2502 gcc_copts = xnnpack_gcc_std_copts(),
2503 msvc_copts = xnnpack_msvc_std_copts(),
2504 deps = [
2505 ":tables",
2506 "@FP16",
2507 "@FXdiv",
2508 "@pthreadpool",
2509 ],
2510)
2511
2512xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002513 name = "wasm_ukernels",
2514 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002515 gcc_copts = xnnpack_gcc_std_copts(),
2516 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002517 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002518 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002519 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002520 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002521 "@FP16",
2522 "@FXdiv",
2523 "@pthreadpool",
2524 ],
2525)
2526
2527xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002528 name = "wasm_ukernels_test_mode",
2529 hdrs = INTERNAL_HDRS,
2530 copts = [
2531 "-UNDEBUG",
2532 "-DXNN_TEST_MODE=1",
2533 ],
2534 gcc_copts = xnnpack_gcc_std_copts(),
2535 msvc_copts = xnnpack_msvc_std_copts(),
2536 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002537 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002538 deps = [
2539 ":tables",
2540 "@FP16",
2541 "@FXdiv",
2542 "@pthreadpool",
2543 ],
2544)
2545
2546xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002547 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002548 hdrs = INTERNAL_HDRS,
2549 aarch32_copts = [
2550 "-marm",
2551 "-mfpu=neon",
2552 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002553 gcc_copts = xnnpack_gcc_std_copts(),
2554 gcc_x86_copts = ["-msse2"],
2555 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002556 optimized_copts = [
2557 "-O3",
2558 "-ffast-math",
2559 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002560 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002561 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002562 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002563 "@FP16",
2564 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002565 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002566 ],
2567)
2568
2569xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002570 name = "psimd_fastmath_ukernels_test_mode",
2571 hdrs = INTERNAL_HDRS,
2572 aarch32_copts = [
2573 "-marm",
2574 "-mfpu=neon",
2575 ],
2576 copts = [
2577 "-UNDEBUG",
2578 "-DXNN_TEST_MODE=1",
2579 ],
2580 gcc_copts = xnnpack_gcc_std_copts(),
2581 gcc_x86_copts = ["-msse2"],
2582 msvc_copts = xnnpack_msvc_std_copts(),
2583 optimized_copts = [
2584 "-O3",
2585 "-ffast-math",
2586 ],
2587 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2588 deps = [
2589 ":tables",
2590 "@FP16",
2591 "@psimd",
2592 "@pthreadpool",
2593 ],
2594)
2595
2596xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002597 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002598 hdrs = INTERNAL_HDRS,
2599 aarch32_copts = [
2600 "-marm",
2601 "-mfpu=neon",
2602 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002603 gcc_copts = xnnpack_gcc_std_copts(),
2604 gcc_x86_copts = ["-msse2"],
2605 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002606 optimized_copts = [
2607 "-O3",
2608 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002609 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002610 deps = [
2611 ":tables",
2612 "@FP16",
2613 "@psimd",
2614 "@pthreadpool",
2615 ],
2616)
2617
2618xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002619 name = "psimd_accmath_ukernels_test_mode",
2620 hdrs = INTERNAL_HDRS,
2621 aarch32_copts = [
2622 "-marm",
2623 "-mfpu=neon",
2624 ],
2625 copts = [
2626 "-UNDEBUG",
2627 "-DXNN_TEST_MODE=1",
2628 ],
2629 gcc_copts = xnnpack_gcc_std_copts(),
2630 gcc_x86_copts = ["-msse2"],
2631 msvc_copts = xnnpack_msvc_std_copts(),
2632 optimized_copts = [
2633 "-O3",
2634 ],
2635 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2636 deps = [
2637 ":tables",
2638 "@FP16",
2639 "@psimd",
2640 "@pthreadpool",
2641 ],
2642)
2643
2644xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002645 name = "neon_ukernels",
2646 hdrs = INTERNAL_HDRS,
2647 aarch32_copts = [
2648 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002649 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002650 "-mfpu=neon",
2651 ],
2652 aarch32_srcs = NEON_UKERNELS,
2653 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002654 gcc_copts = xnnpack_gcc_std_copts(),
2655 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002656 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002657 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002658 "@FP16",
2659 "@pthreadpool",
2660 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002661)
2662
2663xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002664 name = "neon_ukernels_test_mode",
2665 hdrs = INTERNAL_HDRS,
2666 aarch32_copts = [
2667 "-marm",
2668 "-march=armv7-a",
2669 "-mfpu=neon",
2670 ],
2671 aarch32_srcs = NEON_UKERNELS,
2672 aarch64_srcs = NEON_UKERNELS,
2673 copts = [
2674 "-UNDEBUG",
2675 "-DXNN_TEST_MODE=1",
2676 ],
2677 gcc_copts = xnnpack_gcc_std_copts(),
2678 msvc_copts = xnnpack_msvc_std_copts(),
2679 deps = [
2680 ":tables",
2681 "@FP16",
2682 "@pthreadpool",
2683 ],
2684)
2685
2686xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002687 name = "neonfma_ukernels",
2688 hdrs = INTERNAL_HDRS,
2689 aarch32_copts = [
2690 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002691 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002692 "-mfpu=neon-vfpv4",
2693 ],
2694 aarch32_srcs = NEONFMA_UKERNELS,
2695 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002696 apple_aarch32_copts = [
2697 "-mcpu=swift",
2698 "-mtune=generic",
2699 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002700 gcc_copts = xnnpack_gcc_std_copts(),
2701 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002702 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002703 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002704 "@FP16",
2705 "@pthreadpool",
2706 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002707)
2708
2709xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002710 name = "neonfma_ukernels_test_mode",
2711 hdrs = INTERNAL_HDRS,
2712 aarch32_copts = [
2713 "-marm",
2714 "-march=armv7-a",
2715 "-mfpu=neon-vfpv4",
2716 ],
2717 aarch32_srcs = NEONFMA_UKERNELS,
2718 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002719 apple_aarch32_copts = [
2720 "-mcpu=swift",
2721 "-mtune=generic",
2722 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002723 copts = [
2724 "-UNDEBUG",
2725 "-DXNN_TEST_MODE=1",
2726 ],
2727 gcc_copts = xnnpack_gcc_std_copts(),
2728 msvc_copts = xnnpack_msvc_std_copts(),
2729 deps = [
2730 ":tables",
2731 "@FP16",
2732 "@pthreadpool",
2733 ],
2734)
2735
2736xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002737 name = "neonv8_ukernels",
2738 hdrs = INTERNAL_HDRS,
2739 aarch32_copts = [
2740 "-marm",
2741 "-march=armv8-a",
2742 "-mfpu=neon-fp-armv8",
2743 ],
2744 aarch32_srcs = NEONV8_UKERNELS,
2745 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002746 apple_aarch32_copts = [
2747 "-mcpu=cyclone",
2748 "-mtune=generic",
2749 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002750 gcc_copts = xnnpack_gcc_std_copts(),
2751 msvc_copts = xnnpack_msvc_std_copts(),
2752 deps = [
2753 ":tables",
2754 "@FP16",
2755 "@pthreadpool",
2756 ],
2757)
2758
2759xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002760 name = "neonv8_ukernels_test_mode",
2761 hdrs = INTERNAL_HDRS,
2762 aarch32_copts = [
2763 "-marm",
2764 "-march=armv8-a",
2765 "-mfpu=neon-fp-armv8",
2766 ],
2767 aarch32_srcs = NEONV8_UKERNELS,
2768 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002769 apple_aarch32_copts = [
2770 "-mcpu=cyclone",
2771 "-mtune=generic",
2772 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002773 copts = [
2774 "-UNDEBUG",
2775 "-DXNN_TEST_MODE=1",
2776 ],
2777 gcc_copts = xnnpack_gcc_std_copts(),
2778 msvc_copts = xnnpack_msvc_std_copts(),
2779 deps = [
2780 ":tables",
2781 "@FP16",
2782 "@pthreadpool",
2783 ],
2784)
2785
2786xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002787 name = "neonfp16arith_ukernels",
2788 hdrs = INTERNAL_HDRS,
2789 aarch64_copts = ["-march=armv8.2-a+fp16"],
2790 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002791 gcc_copts = xnnpack_gcc_std_copts(),
2792 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002793 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002794 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002795 "@FP16",
2796 "@pthreadpool",
2797 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002798)
2799
2800xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002801 name = "neonfp16arith_ukernels_test_mode",
2802 hdrs = INTERNAL_HDRS,
2803 aarch64_copts = ["-march=armv8.2-a+fp16"],
2804 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2805 copts = [
2806 "-UNDEBUG",
2807 "-DXNN_TEST_MODE=1",
2808 ],
2809 gcc_copts = xnnpack_gcc_std_copts(),
2810 msvc_copts = xnnpack_msvc_std_copts(),
2811 deps = [
2812 ":tables",
2813 "@FP16",
2814 "@pthreadpool",
2815 ],
2816)
2817
2818xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002819 name = "sse2_ukernels",
2820 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002821 gcc_copts = xnnpack_gcc_std_copts(),
2822 gcc_x86_copts = ["-msse2"],
2823 msvc_copts = xnnpack_msvc_std_copts(),
2824 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002825 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002826 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002827 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002828 "@FP16",
2829 "@pthreadpool",
2830 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002831)
2832
2833xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002834 name = "sse2_ukernels_test_mode",
2835 hdrs = INTERNAL_HDRS,
2836 copts = [
2837 "-UNDEBUG",
2838 "-DXNN_TEST_MODE=1",
2839 ],
2840 gcc_copts = xnnpack_gcc_std_copts(),
2841 gcc_x86_copts = ["-msse2"],
2842 msvc_copts = xnnpack_msvc_std_copts(),
2843 msvc_x86_32_copts = ["/arch:SSE2"],
2844 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2845 deps = [
2846 ":tables",
2847 "@FP16",
2848 "@pthreadpool",
2849 ],
2850)
2851
2852xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002853 name = "ssse3_ukernels",
2854 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002855 gcc_copts = xnnpack_gcc_std_copts(),
2856 gcc_x86_copts = ["-mssse3"],
2857 msvc_copts = xnnpack_msvc_std_copts(),
2858 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002859 x86_srcs = SSSE3_UKERNELS,
2860 deps = [
2861 ":tables",
2862 "@FP16",
2863 "@pthreadpool",
2864 ],
2865)
2866
2867xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002868 name = "ssse3_ukernels_test_mode",
2869 hdrs = INTERNAL_HDRS,
2870 copts = [
2871 "-UNDEBUG",
2872 "-DXNN_TEST_MODE=1",
2873 ],
2874 gcc_copts = xnnpack_gcc_std_copts(),
2875 gcc_x86_copts = ["-mssse3"],
2876 msvc_copts = xnnpack_msvc_std_copts(),
2877 msvc_x86_32_copts = ["/arch:SSE2"],
2878 x86_srcs = SSSE3_UKERNELS,
2879 deps = [
2880 ":tables",
2881 "@FP16",
2882 "@pthreadpool",
2883 ],
2884)
2885
2886xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002887 name = "sse41_ukernels",
2888 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002889 gcc_copts = xnnpack_gcc_std_copts(),
2890 gcc_x86_copts = ["-msse4.1"],
2891 msvc_copts = xnnpack_msvc_std_copts(),
2892 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002893 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002894 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002895 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002896 "@FP16",
2897 "@pthreadpool",
2898 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002899)
2900
2901xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002902 name = "sse41_ukernels_test_mode",
2903 hdrs = INTERNAL_HDRS,
2904 copts = [
2905 "-UNDEBUG",
2906 "-DXNN_TEST_MODE=1",
2907 ],
2908 gcc_copts = xnnpack_gcc_std_copts(),
2909 gcc_x86_copts = ["-msse4.1"],
2910 msvc_copts = xnnpack_msvc_std_copts(),
2911 msvc_x86_32_copts = ["/arch:SSE2"],
2912 x86_srcs = SSE41_UKERNELS,
2913 deps = [
2914 ":tables",
2915 "@FP16",
2916 "@pthreadpool",
2917 ],
2918)
2919
2920xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002921 name = "avx_ukernels",
2922 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002923 gcc_copts = xnnpack_gcc_std_copts(),
2924 gcc_x86_copts = ["-mavx"],
2925 msvc_copts = xnnpack_msvc_std_copts(),
2926 msvc_x86_32_copts = ["/arch:AVX"],
2927 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002928 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002929 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002930 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002931 "@FP16",
2932 "@pthreadpool",
2933 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934)
2935
2936xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002937 name = "avx_ukernels_test_mode",
2938 hdrs = INTERNAL_HDRS,
2939 copts = [
2940 "-UNDEBUG",
2941 "-DXNN_TEST_MODE=1",
2942 ],
2943 gcc_copts = xnnpack_gcc_std_copts(),
2944 gcc_x86_copts = ["-mavx"],
2945 msvc_copts = xnnpack_msvc_std_copts(),
2946 msvc_x86_32_copts = ["/arch:AVX"],
2947 msvc_x86_64_copts = ["/arch:AVX"],
2948 x86_srcs = AVX_UKERNELS,
2949 deps = [
2950 ":tables",
2951 "@FP16",
2952 "@pthreadpool",
2953 ],
2954)
2955
2956xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07002957 name = "xop_ukernels",
2958 hdrs = INTERNAL_HDRS,
2959 gcc_copts = xnnpack_gcc_std_copts(),
2960 gcc_x86_copts = ["-mxop"],
2961 msvc_copts = xnnpack_msvc_std_copts(),
2962 msvc_x86_32_copts = ["/arch:AVX"],
2963 msvc_x86_64_copts = ["/arch:AVX"],
2964 x86_srcs = XOP_UKERNELS,
2965 deps = [
2966 ":tables",
2967 "@FP16",
2968 "@pthreadpool",
2969 ],
2970)
2971
2972xnnpack_cc_library(
2973 name = "xop_ukernels_test_mode",
2974 hdrs = INTERNAL_HDRS,
2975 copts = [
2976 "-UNDEBUG",
2977 "-DXNN_TEST_MODE=1",
2978 ],
2979 gcc_copts = xnnpack_gcc_std_copts(),
2980 gcc_x86_copts = ["-mxop"],
2981 msvc_copts = xnnpack_msvc_std_copts(),
2982 msvc_x86_32_copts = ["/arch:AVX"],
2983 msvc_x86_64_copts = ["/arch:AVX"],
2984 x86_srcs = XOP_UKERNELS,
2985 deps = [
2986 ":tables",
2987 "@FP16",
2988 "@pthreadpool",
2989 ],
2990)
2991
2992xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002993 name = "fma3_ukernels",
2994 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002995 gcc_copts = xnnpack_gcc_std_copts(),
2996 gcc_x86_copts = ["-mfma"],
2997 msvc_copts = xnnpack_msvc_std_copts(),
2998 msvc_x86_32_copts = ["/arch:AVX"],
2999 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003000 x86_srcs = FMA3_UKERNELS,
3001 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003002 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003003 "@FP16",
3004 "@pthreadpool",
3005 ],
3006)
3007
3008xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003009 name = "fma3_ukernels_test_mode",
3010 hdrs = INTERNAL_HDRS,
3011 copts = [
3012 "-UNDEBUG",
3013 "-DXNN_TEST_MODE=1",
3014 ],
3015 gcc_copts = xnnpack_gcc_std_copts(),
3016 gcc_x86_copts = ["-mfma"],
3017 msvc_copts = xnnpack_msvc_std_copts(),
3018 msvc_x86_32_copts = ["/arch:AVX"],
3019 msvc_x86_64_copts = ["/arch:AVX"],
3020 x86_srcs = FMA3_UKERNELS,
3021 deps = [
3022 ":tables",
3023 "@FP16",
3024 "@pthreadpool",
3025 ],
3026)
3027
3028xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003029 name = "avx2_ukernels",
3030 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003031 gcc_copts = xnnpack_gcc_std_copts(),
3032 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003033 "-mfma",
3034 "-mavx2",
3035 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003036 msvc_copts = xnnpack_msvc_std_copts(),
3037 msvc_x86_32_copts = ["/arch:AVX2"],
3038 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003039 x86_srcs = AVX2_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 Dukhan6adff4e2019-10-14 18:32:07 -07003045)
3046
3047xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003048 name = "avx2_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 = [
3056 "-mfma",
3057 "-mavx2",
3058 ],
3059 msvc_copts = xnnpack_msvc_std_copts(),
3060 msvc_x86_32_copts = ["/arch:AVX2"],
3061 msvc_x86_64_copts = ["/arch:AVX2"],
3062 x86_srcs = AVX2_UKERNELS,
3063 deps = [
3064 ":tables",
3065 "@FP16",
3066 "@pthreadpool",
3067 ],
3068)
3069
3070xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003071 name = "avx512f_ukernels",
3072 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003073 gcc_copts = xnnpack_gcc_std_copts(),
3074 gcc_x86_copts = ["-mavx512f"],
3075 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3076 msvc_copts = xnnpack_msvc_std_copts(),
3077 msvc_x86_32_copts = ["/arch:AVX512"],
3078 msvc_x86_64_copts = ["/arch:AVX512"],
3079 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003080 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003081 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003082 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003083 "@FP16",
3084 "@pthreadpool",
3085 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086)
3087
3088xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003089 name = "avx512f_ukernels_test_mode",
3090 hdrs = INTERNAL_HDRS,
3091 copts = [
3092 "-UNDEBUG",
3093 "-DXNN_TEST_MODE=1",
3094 ],
3095 gcc_copts = xnnpack_gcc_std_copts(),
3096 gcc_x86_copts = ["-mavx512f"],
3097 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3098 msvc_copts = xnnpack_msvc_std_copts(),
3099 msvc_x86_32_copts = ["/arch:AVX512"],
3100 msvc_x86_64_copts = ["/arch:AVX512"],
3101 msys_copts = ["-fno-asynchronous-unwind-tables"],
3102 x86_srcs = AVX512F_UKERNELS,
3103 deps = [
3104 ":tables",
3105 "@FP16",
3106 "@pthreadpool",
3107 ],
3108)
3109
3110xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003111 name = "asm_ukernels",
3112 hdrs = ["src/xnnpack/assembly.h"],
3113 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003114 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003115 aarch64_srcs = AARCH64_ASM_UKERNELS,
3116)
3117
Marat Dukhan3b59de22020-06-03 20:15:19 -07003118xnnpack_cc_library(
3119 name = "logging_utils",
3120 srcs = LOGGING_SRCS,
3121 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3122 copts = LOGGING_COPTS + [
3123 "-Isrc",
3124 "-Iinclude",
3125 ] + select({
3126 ":debug_build": [],
3127 "//conditions:default": xnnpack_min_size_copts(),
3128 }),
3129 gcc_copts = xnnpack_gcc_std_copts(),
3130 msvc_copts = xnnpack_msvc_std_copts(),
3131 visibility = xnnpack_visibility(),
3132 deps = [
3133 "@FP16",
3134 "@clog",
3135 "@pthreadpool",
3136 ],
3137)
3138
Marat Dukhan08c4a432019-10-03 09:29:21 -07003139xnnpack_aggregate_library(
3140 name = "ukernels",
3141 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142 ":neon_ukernels",
3143 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003144 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003145 ":asm_ukernels",
3146 ],
3147 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003148 ":neon_ukernels",
3149 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003150 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003151 ":neonfp16arith_ukernels",
3152 ":asm_ukernels",
3153 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003154 generic_deps = [
3155 ":scalar_ukernels",
3156 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003157 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003158 ":psimd_fastmath_ukernels",
3159 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003160 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003161 wasm_deps = [
3162 ":wasm_ukernels",
3163 ],
3164 wasmsimd_deps = [
3165 ":wasm_ukernels",
3166 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003168 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003169 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003170 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003171 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003172 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003173 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003174 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003175 ":avx512f_ukernels",
3176 ],
3177)
3178
Marat Dukhan33fcf782020-05-24 14:27:15 -07003179xnnpack_aggregate_library(
3180 name = "ukernels_test_mode",
3181 aarch32_deps = [
3182 ":neon_ukernels_test_mode",
3183 ":neonfma_ukernels_test_mode",
3184 ":neonv8_ukernels_test_mode",
3185 ":asm_ukernels",
3186 ],
3187 aarch64_deps = [
3188 ":neon_ukernels_test_mode",
3189 ":neonfma_ukernels_test_mode",
3190 ":neonv8_ukernels_test_mode",
3191 ":neonfp16arith_ukernels_test_mode",
3192 ":asm_ukernels",
3193 ],
3194 generic_deps = [
3195 ":scalar_ukernels_test_mode",
3196 ],
3197 psimd_deps = [
3198 ":psimd_fastmath_ukernels_test_mode",
3199 ":psimd_accmath_ukernels_test_mode",
3200 ],
3201 wasm_deps = [
3202 ":wasm_ukernels_test_mode",
3203 ],
3204 wasmsimd_deps = [
3205 ":wasm_ukernels_test_mode",
3206 ],
3207 x86_deps = [
3208 ":sse2_ukernels_test_mode",
3209 ":ssse3_ukernels_test_mode",
3210 ":sse41_ukernels_test_mode",
3211 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003212 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003213 ":fma3_ukernels_test_mode",
3214 ":avx2_ukernels_test_mode",
3215 ":avx512f_ukernels_test_mode",
3216 ],
3217)
3218
Marat Dukhan86df3842020-06-17 14:03:31 -07003219# This target must be compatible with Android NDK r17 and compile with Clang 6.
3220# It is similar to :ukernels target, but excludes microkernels which require
3221# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003222xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003223 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003224 aarch32_deps = [
3225 ":neon_ukernels",
3226 ":neonfma_ukernels",
3227 ":neonv8_ukernels",
3228 ":asm_ukernels",
3229 ],
3230 aarch64_deps = [
3231 ":neon_ukernels",
3232 ":neonfma_ukernels",
3233 ":neonv8_ukernels",
3234 ":asm_ukernels",
3235 ],
3236 generic_deps = [
3237 ":scalar_ukernels",
3238 ],
3239 psimd_deps = [
3240 ":psimd_fastmath_ukernels",
3241 ":psimd_accmath_ukernels",
3242 ],
3243 wasm_deps = [
3244 ":wasm_ukernels",
3245 ],
3246 wasmsimd_deps = [
3247 ":wasm_ukernels",
3248 ],
3249 x86_deps = [
3250 ":sse2_ukernels",
3251 ":ssse3_ukernels",
3252 ":sse41_ukernels",
3253 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003254 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003255 ":fma3_ukernels",
3256 ":avx2_ukernels",
3257 ":avx512f_ukernels",
3258 ],
3259)
3260
Marat Dukhan08c4a432019-10-03 09:29:21 -07003261xnnpack_cc_library(
3262 name = "im2col",
3263 srcs = ["src/im2col.c"],
3264 hdrs = [
3265 "src/xnnpack/common.h",
3266 "src/xnnpack/im2col.h",
3267 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003268 gcc_copts = xnnpack_gcc_std_copts(),
3269 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003270)
3271
3272xnnpack_cc_library(
3273 name = "indirection",
3274 srcs = ["src/indirection.c"],
3275 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003276 gcc_copts = xnnpack_gcc_std_copts(),
3277 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278 deps = [
3279 "@FP16",
3280 "@FXdiv",
3281 "@pthreadpool",
3282 ],
3283)
3284
3285xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003286 name = "indirection_test_mode",
3287 srcs = ["src/indirection.c"],
3288 hdrs = INTERNAL_HDRS,
3289 copts = [
3290 "-UNDEBUG",
3291 "-DXNN_TEST_MODE=1",
3292 ],
3293 gcc_copts = xnnpack_gcc_std_copts(),
3294 msvc_copts = xnnpack_msvc_std_copts(),
3295 deps = [
3296 "@FP16",
3297 "@FXdiv",
3298 "@pthreadpool",
3299 ],
3300)
3301
3302xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003303 name = "packing",
3304 srcs = ["src/packing.c"],
3305 hdrs = INTERNAL_HDRS,
3306 gcc_copts = xnnpack_gcc_std_copts(),
3307 msvc_copts = xnnpack_msvc_std_copts(),
3308 deps = [
3309 "@FP16",
3310 "@FXdiv",
3311 "@pthreadpool",
3312 ],
3313)
3314
3315xnnpack_cc_library(
3316 name = "packing_test_mode",
3317 srcs = ["src/packing.c"],
3318 hdrs = INTERNAL_HDRS,
3319 copts = [
3320 "-UNDEBUG",
3321 "-DXNN_TEST_MODE=1",
3322 ],
3323 gcc_copts = xnnpack_gcc_std_copts(),
3324 msvc_copts = xnnpack_msvc_std_copts(),
3325 deps = [
3326 "@FP16",
3327 "@FXdiv",
3328 "@pthreadpool",
3329 ],
3330)
3331
3332xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003333 name = "operator_run",
3334 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003335 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003336 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003337 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3338 "//conditions:default": [],
3339 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003340 gcc_copts = xnnpack_gcc_std_copts(),
3341 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003342 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003343 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003344 "@FP16",
3345 "@FXdiv",
3346 "@clog",
3347 "@pthreadpool",
3348 ],
3349)
3350
Chao Mei6ddfc602020-05-13 22:29:36 -07003351xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003352 name = "operator_run_test_mode",
3353 srcs = ["src/operator-run.c"],
3354 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3355 copts = LOGGING_COPTS + [
3356 "-UNDEBUG",
3357 "-DXNN_TEST_MODE=1",
3358 ] + select({
3359 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3360 "//conditions:default": [],
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 "@FP16",
3367 "@FXdiv",
3368 "@clog",
3369 "@pthreadpool",
3370 ],
3371)
3372
3373xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003374 name = "memory_planner",
3375 srcs = ["src/memory-planner.c"],
3376 hdrs = INTERNAL_HDRS,
3377 defines = select({
3378 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3379 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3380 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3381 }),
3382 gcc_copts = xnnpack_gcc_std_copts(),
3383 msvc_copts = xnnpack_msvc_std_copts(),
3384 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003385 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003386 "@pthreadpool",
3387 ],
3388)
3389
Marat Dukhan33fcf782020-05-24 14:27:15 -07003390xnnpack_cc_library(
3391 name = "memory_planner_test_mode",
3392 srcs = ["src/memory-planner.c"],
3393 hdrs = INTERNAL_HDRS,
3394 copts = [
3395 "-UNDEBUG",
3396 "-DXNN_TEST_MODE=1",
3397 ],
3398 defines = select({
3399 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3400 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3401 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3402 }),
3403 gcc_copts = xnnpack_gcc_std_copts(),
3404 msvc_copts = xnnpack_msvc_std_copts(),
3405 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003406 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003407 "@pthreadpool",
3408 ],
3409)
3410
Marat Dukhan08c4a432019-10-03 09:29:21 -07003411cc_library(
3412 name = "enable_assembly",
3413 defines = select({
3414 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3415 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003416 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003417 }),
3418)
3419
Marat Dukhan9de90e02020-06-18 16:04:12 -07003420cc_library(
3421 name = "enable_sparse",
3422 defines = select({
3423 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3424 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003425 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003426 }),
3427)
3428
Marat Dukhancf056b22019-10-07 10:26:29 -07003429xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430 name = "operators",
3431 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003432 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003433 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003434 ],
3435 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003436 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003437 "-Isrc",
3438 "-Iinclude",
3439 ] + select({
3440 ":debug_build": [],
3441 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003442 }) + select({
3443 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3444 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003445 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003446 gcc_copts = xnnpack_gcc_std_copts(),
3447 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003449 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003450 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003451 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003452 "@FP16",
3453 "@FXdiv",
3454 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003455 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003456 ],
3457)
3458
Marat Dukhan10a38082020-04-17 03:58:35 -07003459xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003460 name = "operators_test_mode",
3461 srcs = OPERATOR_SRCS + [
3462 "src/memory.c",
3463 "src/operator-delete.c",
3464 ],
3465 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3466 copts = LOGGING_COPTS + [
3467 "-Isrc",
3468 "-Iinclude",
3469 "-UNDEBUG",
3470 "-DXNN_TEST_MODE=1",
3471 ] + select({
3472 ":debug_build": [],
3473 "//conditions:default": xnnpack_min_size_copts(),
3474 }) + select({
3475 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3476 "//conditions:default": [],
3477 }),
3478 gcc_copts = xnnpack_gcc_std_copts(),
3479 msvc_copts = xnnpack_msvc_std_copts(),
3480 deps = [
3481 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003482 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003483 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003484 "@FP16",
3485 "@FXdiv",
3486 "@clog",
3487 "@pthreadpool",
3488 ],
3489)
3490
3491xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003492 name = "XNNPACK",
3493 srcs = [
3494 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003495 "src/runtime.c",
3496 "src/subgraph.c",
3497 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003498 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003499 hdrs = ["include/xnnpack.h"],
3500 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003501 "-Isrc",
3502 "-Iinclude",
3503 ] + select({
3504 ":debug_build": [],
3505 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003506 }) + select({
3507 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3508 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003509 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003510 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003511 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003512 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003513 visibility = xnnpack_visibility(),
3514 deps = [
3515 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003516 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003517 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003518 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003519 ":operator_run",
3520 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003521 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003522 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003523 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003524 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003525 ] + select({
3526 ":emscripten": [],
3527 "//conditions:default": ["@cpuinfo"],
3528 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003529)
3530
Marat Dukhan10a38082020-04-17 03:58:35 -07003531xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003532 name = "XNNPACK_test_mode",
3533 srcs = [
3534 "src/init.c",
3535 "src/runtime.c",
3536 "src/subgraph.c",
3537 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003538 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003539 hdrs = ["include/xnnpack.h"],
3540 copts = LOGGING_COPTS + [
3541 "-Isrc",
3542 "-Iinclude",
3543 "-UNDEBUG",
3544 "-DXNN_TEST_MODE=1",
3545 ] + select({
3546 ":debug_build": [],
3547 "//conditions:default": xnnpack_min_size_copts(),
3548 }) + select({
3549 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3550 "//conditions:default": [],
3551 }),
3552 gcc_copts = xnnpack_gcc_std_copts(),
3553 includes = ["include"],
3554 msvc_copts = xnnpack_msvc_std_copts(),
3555 visibility = xnnpack_visibility(),
3556 deps = [
3557 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003558 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003559 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003560 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003561 ":operator_run_test_mode",
3562 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003563 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003564 "@clog",
3565 "@FP16",
3566 "@pthreadpool",
3567 ] + select({
3568 ":emscripten": [],
3569 "//conditions:default": ["@cpuinfo"],
3570 }),
3571)
3572
3573xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003574 name = "xnnpack_f32",
3575 srcs = [
3576 "src/init.c",
3577 "src/runtime.c",
3578 "src/subgraph.c",
3579 "src/tensor.c",
3580 ] + SUBGRAPH_SRCS,
3581 hdrs = ["include/xnnpack.h"],
3582 copts = LOGGING_COPTS + [
3583 "-Isrc",
3584 "-Iinclude",
3585 ] + select({
3586 ":debug_build": [],
3587 "//conditions:default": xnnpack_min_size_copts(),
3588 }) + select({
3589 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3590 "//conditions:default": [],
3591 }),
3592 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003593 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003594 "XNN_NO_U8_OPERATORS",
3595 "XNN_NO_X8_OPERATORS",
3596 "XNN_NO_F16_OPERATORS",
3597 "XNN_NO_X16_OPERATORS",
3598 ],
3599 gcc_copts = xnnpack_gcc_std_copts(),
3600 includes = ["include"],
3601 msvc_copts = xnnpack_msvc_std_copts(),
3602 visibility = xnnpack_visibility(),
3603 deps = [
3604 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003605 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003606 ":logging_utils",
3607 ":memory_planner",
3608 ":operator_run",
3609 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003610 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003611 "@clog",
3612 "@FP16",
3613 "@pthreadpool",
3614 ] + select({
3615 ":emscripten": [],
3616 "//conditions:default": ["@cpuinfo"],
3617 }),
3618)
3619
3620xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003621 name = "xnnpack_operators_nhwc_f32",
3622 srcs = [
3623 "src/init.c",
3624 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003625 hdrs = ["include/xnnpack.h"],
3626 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003627 "-Isrc",
3628 "-Iinclude",
3629 ] + select({
3630 ":debug_build": [],
3631 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003632 }) + select({
3633 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3634 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003635 }),
3636 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003637 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003638 "XNN_NO_U8_OPERATORS",
3639 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003640 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003641 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003642 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003644 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 visibility = xnnpack_visibility(),
3646 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003647 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003648 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003649 ":operator_run",
3650 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003651 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003652 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003654 ] + select({
3655 ":emscripten": [],
3656 "//conditions:default": ["@cpuinfo"],
3657 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003658)
3659
Marat Dukhancf056b22019-10-07 10:26:29 -07003660xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003661 name = "bench_utils",
3662 srcs = ["bench/utils.cc"],
3663 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003664 deps = [
3665 "@com_google_benchmark//:benchmark",
3666 "@cpuinfo",
3667 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003668)
3669
Frank Barchard7e955972019-10-11 10:34:25 -07003670######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003671
3672xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003673 name = "qs8_gemm_bench",
3674 srcs = [
3675 "bench/gemm.h",
3676 "bench/qs8-gemm.cc",
3677 "src/xnnpack/AlignedAllocator.h",
3678 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3679 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3680)
3681
3682xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003683 name = "qs8_requantization_bench",
3684 srcs = [
3685 "bench/qs8-requantization.cc",
3686 "src/xnnpack/requantization-stubs.h",
3687 "src/xnnpack/AlignedAllocator.h",
3688 ] + MICROKERNEL_BENCHMARK_HDRS,
3689 deps = MICROKERNEL_BENCHMARK_DEPS,
3690)
3691
3692xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003693 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003694 srcs = [
3695 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003696 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003697 "src/xnnpack/AlignedAllocator.h",
3698 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003699 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003700 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003701)
3702
3703xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003704 name = "qu8_requantization_bench",
3705 srcs = [
3706 "bench/qu8-requantization.cc",
3707 "src/xnnpack/requantization-stubs.h",
3708 "src/xnnpack/AlignedAllocator.h",
3709 ] + MICROKERNEL_BENCHMARK_HDRS,
3710 deps = MICROKERNEL_BENCHMARK_DEPS,
3711)
3712
3713xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003714 name = "f16_igemm_bench",
3715 srcs = [
3716 "bench/f16-igemm.cc",
3717 "bench/conv.h",
3718 "bench/google/conv.h",
3719 "src/xnnpack/AlignedAllocator.h",
3720 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003721 deps = MICROKERNEL_BENCHMARK_DEPS + [
3722 ":indirection",
3723 ":packing",
3724 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003725)
3726
3727xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728 name = "f16_gemm_bench",
3729 srcs = [
3730 "bench/f16-gemm.cc",
3731 "bench/gemm.h",
3732 "src/xnnpack/AlignedAllocator.h",
3733 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003734 deps = MICROKERNEL_BENCHMARK_DEPS + [
3735 ":packing",
3736 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737)
3738
3739xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003740 name = "f16_spmm_bench",
3741 srcs = [
3742 "bench/f16-spmm.cc",
3743 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003744 "src/xnnpack/AlignedAllocator.h",
3745 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003746 deps = MICROKERNEL_BENCHMARK_DEPS,
3747)
3748
3749xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 name = "f32_igemm_bench",
3751 srcs = [
3752 "bench/f32-igemm.cc",
3753 "bench/conv.h",
3754 "src/xnnpack/AlignedAllocator.h",
3755 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003756 deps = MICROKERNEL_BENCHMARK_DEPS + [
3757 ":indirection",
3758 ":packing",
3759 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003760)
3761
3762xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003763 name = "f16_relu_bench",
3764 srcs = [
3765 "bench/f16-relu.cc",
3766 "src/xnnpack/AlignedAllocator.h",
3767 ] + MICROKERNEL_BENCHMARK_HDRS,
3768 deps = MICROKERNEL_BENCHMARK_DEPS,
3769)
3770
3771xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772 name = "f32_conv_hwc_bench",
3773 srcs = [
3774 "bench/f32-conv-hwc.cc",
3775 "bench/dconv.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 ":packing",
3780 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003781)
3782
3783xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003784 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003785 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003786 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003787 "bench/dconv.h",
3788 "src/xnnpack/AlignedAllocator.h",
3789 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003790 deps = MICROKERNEL_BENCHMARK_DEPS + [
3791 ":packing",
3792 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003793)
3794
3795xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003796 name = "f16_dwconv_bench",
3797 srcs = [
3798 "bench/f16-dwconv.cc",
3799 "bench/dwconv.h",
3800 "bench/google/dwconv.h",
3801 "src/xnnpack/AlignedAllocator.h",
3802 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003803 deps = MICROKERNEL_BENCHMARK_DEPS + [
3804 ":indirection",
3805 ":packing",
3806 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003807)
3808
3809xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003810 name = "f32_dwconv_bench",
3811 srcs = [
3812 "bench/f32-dwconv.cc",
3813 "bench/dwconv.h",
3814 "src/xnnpack/AlignedAllocator.h",
3815 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003816 deps = MICROKERNEL_BENCHMARK_DEPS + [
3817 ":indirection",
3818 ":packing",
3819 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003820)
3821
3822xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003823 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003824 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003825 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003826 "bench/dwconv.h",
3827 "src/xnnpack/AlignedAllocator.h",
3828 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003829 deps = MICROKERNEL_BENCHMARK_DEPS + [
3830 ":indirection",
3831 ":packing",
3832 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003833)
3834
3835xnnpack_benchmark(
3836 name = "f32_gemm_bench",
3837 srcs = [
3838 "bench/f32-gemm.cc",
3839 "bench/gemm.h",
3840 "src/xnnpack/AlignedAllocator.h",
3841 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003842 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003843 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003844)
3845
3846xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003847 name = "f32_hswish_bench",
3848 srcs = [
3849 "bench/f32-hswish.cc",
3850 "src/xnnpack/AlignedAllocator.h",
3851 ] + MICROKERNEL_BENCHMARK_HDRS,
3852 deps = MICROKERNEL_BENCHMARK_DEPS,
3853)
3854
3855xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003856 name = "f32_raddexpminusmax_bench",
3857 srcs = [
3858 "bench/f32-raddexpminusmax.cc",
3859 "src/xnnpack/AlignedAllocator.h",
3860 ] + MICROKERNEL_BENCHMARK_HDRS,
3861 deps = MICROKERNEL_BENCHMARK_DEPS,
3862)
3863
3864xnnpack_benchmark(
3865 name = "f32_raddextexp_bench",
3866 srcs = [
3867 "bench/f32-raddextexp.cc",
3868 "src/xnnpack/AlignedAllocator.h",
3869 ] + MICROKERNEL_BENCHMARK_HDRS,
3870 deps = MICROKERNEL_BENCHMARK_DEPS,
3871)
3872
3873xnnpack_benchmark(
3874 name = "f32_raddstoreexpminusmax_bench",
3875 srcs = [
3876 "bench/f32-raddstoreexpminusmax.cc",
3877 "src/xnnpack/AlignedAllocator.h",
3878 ] + MICROKERNEL_BENCHMARK_HDRS,
3879 deps = MICROKERNEL_BENCHMARK_DEPS,
3880)
3881
3882xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003883 name = "f32_relu_bench",
3884 srcs = [
3885 "bench/f32-relu.cc",
3886 "src/xnnpack/AlignedAllocator.h",
3887 ] + MICROKERNEL_BENCHMARK_HDRS,
3888 deps = MICROKERNEL_BENCHMARK_DEPS,
3889)
3890
3891xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003892 name = "f32_rmax_bench",
3893 srcs = [
3894 "bench/f32-rmax.cc",
3895 "src/xnnpack/AlignedAllocator.h",
3896 ] + MICROKERNEL_BENCHMARK_HDRS,
3897 deps = MICROKERNEL_BENCHMARK_DEPS,
3898)
3899
3900xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003901 name = "f32_sigmoid_bench",
3902 srcs = [
3903 "bench/f32-sigmoid.cc",
3904 "src/xnnpack/AlignedAllocator.h",
3905 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003906 deps = MICROKERNEL_BENCHMARK_DEPS,
3907)
3908
3909xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003910 name = "f32_spmm_bench",
3911 srcs = [
3912 "bench/f32-spmm.cc",
3913 "bench/gemm.h",
3914 "src/xnnpack/AlignedAllocator.h",
3915 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003916 deps = MICROKERNEL_BENCHMARK_DEPS,
3917)
3918
3919xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003920 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003921 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003922 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003923 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003924 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003925 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003926)
3927
3928xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003929 name = "f32_vscaleexpminusmax_bench",
3930 srcs = [
3931 "bench/f32-vscaleexpminusmax.cc",
3932 "src/xnnpack/AlignedAllocator.h",
3933 ] + MICROKERNEL_BENCHMARK_HDRS,
3934 deps = MICROKERNEL_BENCHMARK_DEPS,
3935)
3936
3937xnnpack_benchmark(
3938 name = "f32_vscaleextexp_bench",
3939 srcs = [
3940 "bench/f32-vscaleextexp.cc",
3941 "src/xnnpack/AlignedAllocator.h",
3942 ] + MICROKERNEL_BENCHMARK_HDRS,
3943 deps = MICROKERNEL_BENCHMARK_DEPS,
3944)
3945
3946xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003947 name = "f32_vsqrt_bench",
3948 srcs = [
3949 "bench/f32-vsqrt.cc",
3950 "src/xnnpack/AlignedAllocator.h",
3951 ] + MICROKERNEL_BENCHMARK_HDRS,
3952 deps = MICROKERNEL_BENCHMARK_DEPS,
3953)
3954
3955xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003956 name = "f32_im2col_gemm_bench",
3957 srcs = [
3958 "bench/f32-im2col-gemm.cc",
3959 "bench/conv.h",
3960 "src/xnnpack/AlignedAllocator.h",
3961 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003962 deps = MICROKERNEL_BENCHMARK_DEPS + [
3963 ":im2col",
3964 ":packing",
3965 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966)
3967
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003968xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003969 name = "rounding_bench",
3970 srcs = [
3971 "bench/rounding.cc",
3972 "src/xnnpack/math-stubs.h",
3973 "src/xnnpack/AlignedAllocator.h",
3974 ] + MICROKERNEL_BENCHMARK_HDRS,
3975 deps = MICROKERNEL_BENCHMARK_DEPS,
3976)
3977
Marat Dukhan08c4a432019-10-03 09:29:21 -07003978########################### Benchmarks for operators ###########################
3979
3980xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003981 name = "average_pooling_bench",
3982 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003983 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003984 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003985 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003986)
3987
3988xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003989 name = "bankers_rounding_bench",
3990 srcs = ["bench/bankers-rounding.cc"],
3991 copts = xnnpack_optional_tflite_copts(),
3992 tags = ["nowin32"],
3993 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3994)
3995
3996xnnpack_benchmark(
3997 name = "ceiling_bench",
3998 srcs = ["bench/ceiling.cc"],
3999 copts = xnnpack_optional_tflite_copts(),
4000 tags = ["nowin32"],
4001 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4002)
4003
4004xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004005 name = "channel_shuffle_bench",
4006 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004007 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004008)
4009
4010xnnpack_benchmark(
4011 name = "convolution_bench",
4012 srcs = ["bench/convolution.cc"],
4013 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004014 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004015 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004016)
4017
4018xnnpack_benchmark(
4019 name = "deconvolution_bench",
4020 srcs = ["bench/deconvolution.cc"],
4021 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004022 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004023 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004024)
4025
4026xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004027 name = "floor_bench",
4028 srcs = ["bench/floor.cc"],
4029 copts = xnnpack_optional_tflite_copts(),
4030 tags = ["nowin32"],
4031 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4032)
4033
4034xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004035 name = "global_average_pooling_bench",
4036 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004037 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004038)
4039
4040xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004041 name = "hardswish_bench",
4042 srcs = ["bench/hardswish.cc"],
4043 copts = xnnpack_optional_tflite_copts(),
4044 tags = ["nowin32"],
4045 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4046)
4047
4048xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004049 name = "max_pooling_bench",
4050 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004051 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004052)
4053
4054xnnpack_benchmark(
4055 name = "sigmoid_bench",
4056 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004057 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004058 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004059 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004060)
4061
4062xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004063 name = "prelu_bench",
4064 srcs = ["bench/prelu.cc"],
4065 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004066 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004067 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004068)
4069
4070xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004071 name = "softmax_bench",
4072 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004073 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004074 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004075 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004076)
4077
Marat Dukhan87727142020-06-24 15:24:10 -07004078xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004079 name = "square_root_bench",
4080 srcs = ["bench/square-root.cc"],
4081 copts = xnnpack_optional_tflite_copts(),
4082 tags = ["nowin32"],
4083 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4084)
4085
4086xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004087 name = "truncation_bench",
4088 srcs = ["bench/truncation.cc"],
4089 deps = OPERATOR_BENCHMARK_DEPS,
4090)
4091
Marat Dukhanc068bb62019-10-04 13:24:39 -07004092############################# End-to-end benchmarks ############################
4093
4094cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004095 name = "fp32_mobilenet_v1",
4096 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004097 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004098 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004099 linkstatic = True,
4100 deps = [
4101 ":XNNPACK",
4102 "@pthreadpool",
4103 ],
4104)
4105
4106cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004107 name = "fp16_mobilenet_v1",
4108 srcs = ["models/fp16-mobilenet-v1.cc"],
4109 hdrs = ["models/models.h"],
4110 copts = xnnpack_std_cxxopts(),
4111 linkstatic = True,
4112 deps = [
4113 ":XNNPACK",
4114 "@FP16",
4115 "@pthreadpool",
4116 ],
4117)
4118
4119cc_library(
4120 name = "fp32_mobilenet_v2",
4121 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004122 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004123 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004124 linkstatic = True,
4125 deps = [
4126 ":XNNPACK",
4127 "@pthreadpool",
4128 ],
4129)
4130
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004131cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004132 name = "fp16_mobilenet_v2",
4133 srcs = ["models/fp16-mobilenet-v2.cc"],
4134 hdrs = ["models/models.h"],
4135 copts = xnnpack_std_cxxopts(),
4136 linkstatic = True,
4137 deps = [
4138 ":XNNPACK",
4139 "@FP16",
4140 "@pthreadpool",
4141 ],
4142)
4143
4144cc_library(
4145 name = "fp32_mobilenet_v3_large",
4146 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004147 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004148 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004149 linkstatic = True,
4150 deps = [
4151 ":XNNPACK",
4152 "@pthreadpool",
4153 ],
4154)
4155
4156cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004157 name = "fp32_mobilenet_v3_small",
4158 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004159 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004160 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004161 linkstatic = True,
4162 deps = [
4163 ":XNNPACK",
4164 "@pthreadpool",
4165 ],
4166)
4167
Marat Dukhanc068bb62019-10-04 13:24:39 -07004168xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004169 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004170 srcs = [
4171 "bench/f32-dwconv-e2e.cc",
4172 "bench/end2end.h",
4173 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004174 deps = MICROKERNEL_BENCHMARK_DEPS + [
4175 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004176 ":fp32_mobilenet_v1",
4177 ":fp32_mobilenet_v2",
4178 ":fp32_mobilenet_v3_large",
4179 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004180 ],
4181)
4182
4183xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004184 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004185 srcs = [
4186 "bench/f32-gemm-e2e.cc",
4187 "bench/end2end.h",
4188 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004189 deps = MICROKERNEL_BENCHMARK_DEPS + [
4190 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004191 ":fp32_mobilenet_v1",
4192 ":fp32_mobilenet_v2",
4193 ":fp32_mobilenet_v3_large",
4194 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004195 ],
4196)
4197
4198xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004199 name = "end2end_bench",
4200 srcs = ["bench/end2end.cc"],
4201 deps = [
4202 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004203 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004204 ":fp16_mobilenet_v1",
4205 ":fp16_mobilenet_v2",
4206 ":fp32_mobilenet_v1",
4207 ":fp32_mobilenet_v2",
4208 ":fp32_mobilenet_v3_large",
4209 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004210 "@pthreadpool",
4211 ],
4212)
4213
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004214#################### Accuracy evaluation for math functions ####################
4215
4216xnnpack_benchmark(
4217 name = "f32_exp_eval",
4218 srcs = [
4219 "eval/f32-exp.cc",
4220 "src/xnnpack/AlignedAllocator.h",
4221 ] + ACCURACY_EVAL_HDRS,
4222 deps = ACCURACY_EVAL_DEPS,
4223)
4224
Marat Dukhan515c9772019-10-17 18:07:57 -07004225xnnpack_benchmark(
4226 name = "f32_expminus_eval",
4227 srcs = [
4228 "eval/f32-expminus.cc",
4229 "src/xnnpack/AlignedAllocator.h",
4230 ] + ACCURACY_EVAL_HDRS,
4231 deps = ACCURACY_EVAL_DEPS,
4232)
4233
Marat Dukhan98ba4412019-10-23 02:14:28 -07004234xnnpack_benchmark(
4235 name = "f32_extexp_eval",
4236 srcs = [
4237 "eval/f32-extexp.cc",
4238 "src/xnnpack/AlignedAllocator.h",
4239 ] + ACCURACY_EVAL_HDRS,
4240 deps = ACCURACY_EVAL_DEPS,
4241)
4242
Marat Dukhan8853b822020-05-07 12:19:01 -07004243xnnpack_unit_test(
4244 name = "f32_roundne_eval",
4245 srcs = [
4246 "eval/f32-roundne.cc",
4247 "src/xnnpack/AlignedAllocator.h",
4248 "src/xnnpack/math-stubs.h",
4249 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004250 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004251 deps = MICROKERNEL_TEST_DEPS,
4252)
4253
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004254xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004255 name = "f32_roundd_eval",
4256 srcs = [
4257 "eval/f32-roundd.cc",
4258 "src/xnnpack/AlignedAllocator.h",
4259 "src/xnnpack/math-stubs.h",
4260 ] + MICROKERNEL_TEST_HDRS,
4261 automatic = False,
4262 deps = MICROKERNEL_TEST_DEPS,
4263)
4264
4265xnnpack_unit_test(
4266 name = "f32_roundu_eval",
4267 srcs = [
4268 "eval/f32-roundu.cc",
4269 "src/xnnpack/AlignedAllocator.h",
4270 "src/xnnpack/math-stubs.h",
4271 ] + MICROKERNEL_TEST_HDRS,
4272 automatic = False,
4273 deps = MICROKERNEL_TEST_DEPS,
4274)
4275
4276xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004277 name = "f32_roundz_eval",
4278 srcs = [
4279 "eval/f32-roundz.cc",
4280 "src/xnnpack/AlignedAllocator.h",
4281 "src/xnnpack/math-stubs.h",
4282 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004283 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004284 deps = MICROKERNEL_TEST_DEPS,
4285)
4286
Marat Dukhan346a9e52019-11-15 09:06:30 -08004287xnnpack_benchmark(
4288 name = "f32_sigmoid_eval",
4289 srcs = [
4290 "eval/f32-sigmoid.cc",
4291 "src/xnnpack/AlignedAllocator.h",
4292 ] + ACCURACY_EVAL_HDRS,
4293 deps = ACCURACY_EVAL_DEPS,
4294)
4295
Marat Dukhan84000762020-06-29 18:38:43 -07004296xnnpack_benchmark(
4297 name = "f32_sqrt_eval",
4298 srcs = [
4299 "eval/f32-sqrt.cc",
4300 "src/xnnpack/AlignedAllocator.h",
4301 ] + ACCURACY_EVAL_HDRS,
4302 deps = ACCURACY_EVAL_DEPS + [
4303 ":bench_utils",
4304 ],
4305)
4306
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307######################### Unit tests for micro-kernels #########################
4308
4309xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004310 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004311 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004312 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004313 "test/gemm-microkernel-tester.h",
4314 "src/xnnpack/AlignedAllocator.h",
4315 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004316 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004317)
4318
4319xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004320 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004321 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004322 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004323 "test/spmm-microkernel-tester.h",
4324 "src/xnnpack/AlignedAllocator.h",
4325 ] + MICROKERNEL_TEST_HDRS,
4326 deps = MICROKERNEL_TEST_DEPS,
4327)
4328
4329xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004330 name = "f32_argmaxpool_test",
4331 srcs = [
4332 "test/f32-argmaxpool.cc",
4333 "test/argmaxpool-microkernel-tester.h",
4334 "src/xnnpack/AlignedAllocator.h",
4335 ] + MICROKERNEL_TEST_HDRS,
4336 deps = MICROKERNEL_TEST_DEPS,
4337)
4338
4339xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004340 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004341 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004342 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004343 "test/avgpool-microkernel-tester.h",
4344 "src/xnnpack/AlignedAllocator.h",
4345 ] + MICROKERNEL_TEST_HDRS,
4346 deps = MICROKERNEL_TEST_DEPS,
4347)
4348
4349xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004350 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004351 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004352 "test/f32-ibilinear.cc",
4353 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004354 "src/xnnpack/AlignedAllocator.h",
4355 ] + MICROKERNEL_TEST_HDRS,
4356 deps = MICROKERNEL_TEST_DEPS,
4357)
4358
4359xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004360 name = "f16_clamp_test",
4361 srcs = [
4362 "test/f16-clamp.cc",
4363 "test/clamp-microkernel-tester.h",
4364 ] + MICROKERNEL_TEST_HDRS,
4365 deps = MICROKERNEL_TEST_DEPS,
4366)
4367
4368xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004369 name = "f32_clamp_test",
4370 srcs = [
4371 "test/f32-clamp.cc",
4372 "test/clamp-microkernel-tester.h",
4373 ] + MICROKERNEL_TEST_HDRS,
4374 deps = MICROKERNEL_TEST_DEPS,
4375)
4376
4377xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004378 name = "f32_igemm_test",
4379 srcs = [
4380 "test/f32-igemm.cc",
4381 "test/gemm-microkernel-tester.h",
4382 "src/xnnpack/AlignedAllocator.h",
4383 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004384 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004385)
4386
4387xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004388 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004389 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004390 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391 "test/gemm-microkernel-tester.h",
4392 "src/xnnpack/AlignedAllocator.h",
4393 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004394 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004395)
4396
4397xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004398 name = "f32_igemm_minmax_test",
4399 srcs = [
4400 "test/f32-igemm-minmax.cc",
4401 "test/gemm-microkernel-tester.h",
4402 "src/xnnpack/AlignedAllocator.h",
4403 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004404 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004405)
4406
4407xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004408 name = "f16_igemm_minmax_test",
4409 srcs = [
4410 "test/f16-igemm-minmax.cc",
4411 "test/gemm-microkernel-tester.h",
4412 "src/xnnpack/AlignedAllocator.h",
4413 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004414 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004415)
4416
4417xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004418 name = "f32_conv_hwc_test",
4419 srcs = [
4420 "test/f32-conv-hwc.cc",
4421 "test/conv-hwc-microkernel-tester.h",
4422 "src/xnnpack/AlignedAllocator.h",
4423 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004424 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004425)
4426
4427xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004428 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004429 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004430 "test/f32-conv-hwc2chw.cc",
4431 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004432 "src/xnnpack/AlignedAllocator.h",
4433 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004434 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004435)
4436
4437xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004438 name = "f32_dwconv_test",
4439 srcs = [
4440 "test/f32-dwconv.cc",
4441 "test/dwconv-microkernel-tester.h",
4442 "src/xnnpack/AlignedAllocator.h",
4443 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004444 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004445)
4446
4447xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004448 name = "f16_dwconv_minmax_test",
4449 srcs = [
4450 "test/f16-dwconv-minmax.cc",
4451 "test/dwconv-microkernel-tester.h",
4452 "src/xnnpack/AlignedAllocator.h",
4453 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004454 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004455)
4456
4457xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004458 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004460 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461 "test/dwconv-microkernel-tester.h",
4462 "src/xnnpack/AlignedAllocator.h",
4463 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004464 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465)
4466
4467xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004468 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004469 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004470 "test/f32-dwconv-chw.cc",
4471 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004472 "src/xnnpack/AlignedAllocator.h",
4473 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004474 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004475)
4476
4477xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004478 name = "f16_gavgpool_minmax_test",
4479 srcs = [
4480 "test/f16-gavgpool-minmax.cc",
4481 "test/gavgpool-microkernel-tester.h",
4482 "src/xnnpack/AlignedAllocator.h",
4483 ] + MICROKERNEL_TEST_HDRS,
4484 deps = MICROKERNEL_TEST_DEPS,
4485)
4486
4487xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004488 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004489 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004490 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004491 "test/gavgpool-microkernel-tester.h",
4492 "src/xnnpack/AlignedAllocator.h",
4493 ] + MICROKERNEL_TEST_HDRS,
4494 deps = MICROKERNEL_TEST_DEPS,
4495)
4496
4497xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004498 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004500 "test/f32-gavgpool-cw.cc",
4501 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004502 "src/xnnpack/AlignedAllocator.h",
4503 ] + MICROKERNEL_TEST_HDRS,
4504 deps = MICROKERNEL_TEST_DEPS,
4505)
4506
4507xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004508 name = "f32_gemm_test",
4509 srcs = [
4510 "test/f32-gemm.cc",
4511 "test/gemm-microkernel-tester.h",
4512 "src/xnnpack/AlignedAllocator.h",
4513 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004514 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004515)
4516
4517xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004518 name = "f32_gemm_relu_test",
4519 srcs = [
4520 "test/f32-gemm-relu.cc",
4521 "test/gemm-microkernel-tester.h",
4522 "src/xnnpack/AlignedAllocator.h",
4523 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004524 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004525)
4526
4527xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004528 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004529 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004530 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004531 "test/gemm-microkernel-tester.h",
4532 "src/xnnpack/AlignedAllocator.h",
4533 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004534 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004535)
4536
4537xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004538 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004539 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004540 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004541 "test/gemm-microkernel-tester.h",
4542 "src/xnnpack/AlignedAllocator.h",
4543 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004544 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004545)
4546
4547xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004548 name = "f16_hswish_test",
4549 srcs = [
4550 "test/f16-hswish.cc",
4551 "test/hswish-microkernel-tester.h",
4552 ] + MICROKERNEL_TEST_HDRS,
4553 deps = MICROKERNEL_TEST_DEPS,
4554)
4555
4556xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004557 name = "f32_hswish_test",
4558 srcs = [
4559 "test/f32-hswish.cc",
4560 "test/hswish-microkernel-tester.h",
4561 ] + MICROKERNEL_TEST_HDRS,
4562 deps = MICROKERNEL_TEST_DEPS,
4563)
4564
4565xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004566 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004567 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004568 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004569 "test/maxpool-microkernel-tester.h",
4570 ] + MICROKERNEL_TEST_HDRS,
4571 deps = MICROKERNEL_TEST_DEPS,
4572)
4573
4574xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004575 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004576 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004577 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004578 "test/avgpool-microkernel-tester.h",
4579 "src/xnnpack/AlignedAllocator.h",
4580 ] + MICROKERNEL_TEST_HDRS,
4581 deps = MICROKERNEL_TEST_DEPS,
4582)
4583
4584xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004585 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004586 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004587 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004588 "test/gemm-microkernel-tester.h",
4589 "src/xnnpack/AlignedAllocator.h",
4590 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004591 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004592)
4593
4594xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004595 name = "f16_prelu_test",
4596 srcs = [
4597 "test/f16-prelu.cc",
4598 "test/prelu-microkernel-tester.h",
4599 "src/xnnpack/AlignedAllocator.h",
4600 ] + MICROKERNEL_TEST_HDRS,
4601 deps = MICROKERNEL_TEST_DEPS,
4602)
4603
4604xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004605 name = "f32_prelu_test",
4606 srcs = [
4607 "test/f32-prelu.cc",
4608 "test/prelu-microkernel-tester.h",
4609 "src/xnnpack/AlignedAllocator.h",
4610 ] + MICROKERNEL_TEST_HDRS,
4611 deps = MICROKERNEL_TEST_DEPS,
4612)
4613
4614xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004615 name = "f32_raddexpminusmax_test",
4616 srcs = [
4617 "test/f32-raddexpminusmax.cc",
4618 "test/raddexpminusmax-microkernel-tester.h",
4619 ] + MICROKERNEL_TEST_HDRS,
4620 deps = MICROKERNEL_TEST_DEPS,
4621)
4622
4623xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004624 name = "f32_raddextexp_test",
4625 srcs = [
4626 "test/f32-raddextexp.cc",
4627 "test/raddextexp-microkernel-tester.h",
4628 ] + MICROKERNEL_TEST_HDRS,
4629 deps = MICROKERNEL_TEST_DEPS,
4630)
4631
4632xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004633 name = "f32_raddstoreexpminusmax_test",
4634 srcs = [
4635 "test/f32-raddstoreexpminusmax.cc",
4636 "test/raddstoreexpminusmax-microkernel-tester.h",
4637 ] + MICROKERNEL_TEST_HDRS,
4638 deps = MICROKERNEL_TEST_DEPS,
4639)
4640
4641xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004642 name = "f32_relu_test",
4643 srcs = [
4644 "test/f32-relu.cc",
4645 "test/vunary-microkernel-tester.h",
4646 ] + MICROKERNEL_TEST_HDRS,
4647 deps = MICROKERNEL_TEST_DEPS,
4648)
4649
4650xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004651 name = "f32_rmax_test",
4652 srcs = [
4653 "test/f32-rmax.cc",
4654 "test/rmax-microkernel-tester.h",
4655 ] + MICROKERNEL_TEST_HDRS,
4656 deps = MICROKERNEL_TEST_DEPS,
4657)
4658
4659xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004660 name = "f32_sigmoid_test",
4661 srcs = [
4662 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004663 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004664 ] + MICROKERNEL_TEST_HDRS,
4665 deps = MICROKERNEL_TEST_DEPS,
4666)
4667
4668xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004669 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004670 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004671 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004672 "test/spmm-microkernel-tester.h",
4673 "src/xnnpack/AlignedAllocator.h",
4674 ] + MICROKERNEL_TEST_HDRS,
4675 deps = MICROKERNEL_TEST_DEPS,
4676)
4677
4678xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004679 name = "f16_vadd_minmax_test",
4680 srcs = [
4681 "test/f16-vadd-minmax.cc",
4682 "test/vbinary-microkernel-tester.h",
4683 ] + MICROKERNEL_TEST_HDRS,
4684 deps = MICROKERNEL_TEST_DEPS,
4685)
4686
4687xnnpack_unit_test(
4688 name = "f16_vaddc_minmax_test",
4689 srcs = [
4690 "test/f16-vaddc-minmax.cc",
4691 "test/vbinaryc-microkernel-tester.h",
4692 ] + MICROKERNEL_TEST_HDRS,
4693 deps = MICROKERNEL_TEST_DEPS,
4694)
4695
4696xnnpack_unit_test(
4697 name = "f16_vdiv_minmax_test",
4698 srcs = [
4699 "test/f16-vdiv-minmax.cc",
4700 "test/vbinary-microkernel-tester.h",
4701 ] + MICROKERNEL_TEST_HDRS,
4702 deps = MICROKERNEL_TEST_DEPS,
4703)
4704
4705xnnpack_unit_test(
4706 name = "f16_vdivc_minmax_test",
4707 srcs = [
4708 "test/f16-vdivc-minmax.cc",
4709 "test/vbinaryc-microkernel-tester.h",
4710 ] + MICROKERNEL_TEST_HDRS,
4711 deps = MICROKERNEL_TEST_DEPS,
4712)
4713
4714xnnpack_unit_test(
4715 name = "f16_vrdivc_minmax_test",
4716 srcs = [
4717 "test/f16-vrdivc-minmax.cc",
4718 "test/vbinaryc-microkernel-tester.h",
4719 ] + MICROKERNEL_TEST_HDRS,
4720 deps = MICROKERNEL_TEST_DEPS,
4721)
4722
4723xnnpack_unit_test(
4724 name = "f16_vmax_test",
4725 srcs = [
4726 "test/f16-vmax.cc",
4727 "test/vbinary-microkernel-tester.h",
4728 ] + MICROKERNEL_TEST_HDRS,
4729 deps = MICROKERNEL_TEST_DEPS,
4730)
4731
4732xnnpack_unit_test(
4733 name = "f16_vmaxc_test",
4734 srcs = [
4735 "test/f16-vmaxc.cc",
4736 "test/vbinaryc-microkernel-tester.h",
4737 ] + MICROKERNEL_TEST_HDRS,
4738 deps = MICROKERNEL_TEST_DEPS,
4739)
4740
4741xnnpack_unit_test(
4742 name = "f16_vmin_test",
4743 srcs = [
4744 "test/f16-vmin.cc",
4745 "test/vbinary-microkernel-tester.h",
4746 ] + MICROKERNEL_TEST_HDRS,
4747 deps = MICROKERNEL_TEST_DEPS,
4748)
4749
4750xnnpack_unit_test(
4751 name = "f16_vminc_test",
4752 srcs = [
4753 "test/f16-vminc.cc",
4754 "test/vbinaryc-microkernel-tester.h",
4755 ] + MICROKERNEL_TEST_HDRS,
4756 deps = MICROKERNEL_TEST_DEPS,
4757)
4758
4759xnnpack_unit_test(
4760 name = "f16_vmul_minmax_test",
4761 srcs = [
4762 "test/f16-vmul-minmax.cc",
4763 "test/vbinary-microkernel-tester.h",
4764 ] + MICROKERNEL_TEST_HDRS,
4765 deps = MICROKERNEL_TEST_DEPS,
4766)
4767
4768xnnpack_unit_test(
4769 name = "f16_vmulc_minmax_test",
4770 srcs = [
4771 "test/f16-vmulc-minmax.cc",
4772 "test/vbinaryc-microkernel-tester.h",
4773 ] + MICROKERNEL_TEST_HDRS,
4774 deps = MICROKERNEL_TEST_DEPS,
4775)
4776
4777xnnpack_unit_test(
4778 name = "f16_vsub_minmax_test",
4779 srcs = [
4780 "test/f16-vsub-minmax.cc",
4781 "test/vbinary-microkernel-tester.h",
4782 ] + MICROKERNEL_TEST_HDRS,
4783 deps = MICROKERNEL_TEST_DEPS,
4784)
4785
4786xnnpack_unit_test(
4787 name = "f16_vsubc_minmax_test",
4788 srcs = [
4789 "test/f16-vsubc-minmax.cc",
4790 "test/vbinaryc-microkernel-tester.h",
4791 ] + MICROKERNEL_TEST_HDRS,
4792 deps = MICROKERNEL_TEST_DEPS,
4793)
4794
4795xnnpack_unit_test(
4796 name = "f16_vrsubc_minmax_test",
4797 srcs = [
4798 "test/f16-vrsubc-minmax.cc",
4799 "test/vbinaryc-microkernel-tester.h",
4800 ] + MICROKERNEL_TEST_HDRS,
4801 deps = MICROKERNEL_TEST_DEPS,
4802)
4803
4804xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004805 name = "f32_vabs_test",
4806 srcs = [
4807 "test/f32-vabs.cc",
4808 "test/vunary-microkernel-tester.h",
4809 ] + MICROKERNEL_TEST_HDRS,
4810 deps = MICROKERNEL_TEST_DEPS,
4811)
4812
4813xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004814 name = "f32_vadd_test",
4815 srcs = [
4816 "test/f32-vadd.cc",
4817 "test/vbinary-microkernel-tester.h",
4818 ] + MICROKERNEL_TEST_HDRS,
4819 deps = MICROKERNEL_TEST_DEPS,
4820)
4821
4822xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004823 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004824 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004825 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004826 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004827 ] + MICROKERNEL_TEST_HDRS,
4828 deps = MICROKERNEL_TEST_DEPS,
4829)
4830
4831xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004832 name = "f32_vaddc_test",
4833 srcs = [
4834 "test/f32-vaddc.cc",
4835 "test/vbinaryc-microkernel-tester.h",
4836 ] + MICROKERNEL_TEST_HDRS,
4837 deps = MICROKERNEL_TEST_DEPS,
4838)
4839
4840xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004841 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004842 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004843 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004844 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004845 ] + MICROKERNEL_TEST_HDRS,
4846 deps = MICROKERNEL_TEST_DEPS,
4847)
4848
4849xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004850 name = "f32_vdiv_test",
4851 srcs = [
4852 "test/f32-vdiv.cc",
4853 "test/vbinary-microkernel-tester.h",
4854 ] + MICROKERNEL_TEST_HDRS,
4855 deps = MICROKERNEL_TEST_DEPS,
4856)
4857
4858xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004859 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004860 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004861 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004862 "test/vbinary-microkernel-tester.h",
4863 ] + MICROKERNEL_TEST_HDRS,
4864 deps = MICROKERNEL_TEST_DEPS,
4865)
4866
4867xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004868 name = "f32_vdivc_test",
4869 srcs = [
4870 "test/f32-vdivc.cc",
4871 "test/vbinaryc-microkernel-tester.h",
4872 ] + MICROKERNEL_TEST_HDRS,
4873 deps = MICROKERNEL_TEST_DEPS,
4874)
4875
4876xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004877 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004878 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004879 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004880 "test/vbinaryc-microkernel-tester.h",
4881 ] + MICROKERNEL_TEST_HDRS,
4882 deps = MICROKERNEL_TEST_DEPS,
4883)
4884
4885xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004886 name = "f32_vrdivc_test",
4887 srcs = [
4888 "test/f32-vrdivc.cc",
4889 "test/vbinaryc-microkernel-tester.h",
4890 ] + MICROKERNEL_TEST_HDRS,
4891 deps = MICROKERNEL_TEST_DEPS,
4892)
4893
4894xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004895 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004896 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004897 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004898 "test/vbinaryc-microkernel-tester.h",
4899 ] + MICROKERNEL_TEST_HDRS,
4900 deps = MICROKERNEL_TEST_DEPS,
4901)
4902
4903xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004904 name = "f32_vmax_test",
4905 srcs = [
4906 "test/f32-vmax.cc",
4907 "test/vbinary-microkernel-tester.h",
4908 ] + MICROKERNEL_TEST_HDRS,
4909 deps = MICROKERNEL_TEST_DEPS,
4910)
4911
4912xnnpack_unit_test(
4913 name = "f32_vmaxc_test",
4914 srcs = [
4915 "test/f32-vmaxc.cc",
4916 "test/vbinaryc-microkernel-tester.h",
4917 ] + MICROKERNEL_TEST_HDRS,
4918 deps = MICROKERNEL_TEST_DEPS,
4919)
4920
4921xnnpack_unit_test(
4922 name = "f32_vmin_test",
4923 srcs = [
4924 "test/f32-vmin.cc",
4925 "test/vbinary-microkernel-tester.h",
4926 ] + MICROKERNEL_TEST_HDRS,
4927 deps = MICROKERNEL_TEST_DEPS,
4928)
4929
4930xnnpack_unit_test(
4931 name = "f32_vminc_test",
4932 srcs = [
4933 "test/f32-vminc.cc",
4934 "test/vbinaryc-microkernel-tester.h",
4935 ] + MICROKERNEL_TEST_HDRS,
4936 deps = MICROKERNEL_TEST_DEPS,
4937)
4938
4939xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004940 name = "f32_vmul_test",
4941 srcs = [
4942 "test/f32-vmul.cc",
4943 "test/vbinary-microkernel-tester.h",
4944 ] + MICROKERNEL_TEST_HDRS,
4945 deps = MICROKERNEL_TEST_DEPS,
4946)
4947
4948xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004949 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004950 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004951 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004952 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004953 ] + MICROKERNEL_TEST_HDRS,
4954 deps = MICROKERNEL_TEST_DEPS,
4955)
4956
4957xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004958 name = "f32_vmulc_test",
4959 srcs = [
4960 "test/f32-vmulc.cc",
4961 "test/vbinaryc-microkernel-tester.h",
4962 ] + MICROKERNEL_TEST_HDRS,
4963 deps = MICROKERNEL_TEST_DEPS,
4964)
4965
4966xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004967 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004968 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004969 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004970 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004971 ] + MICROKERNEL_TEST_HDRS,
4972 deps = MICROKERNEL_TEST_DEPS,
4973)
4974
4975xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004976 name = "f16_vmulcaddc_minmax_test",
4977 srcs = [
4978 "test/f16-vmulcaddc-minmax.cc",
4979 "test/vmulcaddc-microkernel-tester.h",
4980 "src/xnnpack/AlignedAllocator.h",
4981 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004982 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004983)
4984
4985xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004986 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004988 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 "test/vmulcaddc-microkernel-tester.h",
4990 "src/xnnpack/AlignedAllocator.h",
4991 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004992 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993)
4994
4995xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004996 name = "f32_vlrelu_test",
4997 srcs = [
4998 "test/f32-vlrelu.cc",
4999 "test/vunary-microkernel-tester.h",
5000 ] + MICROKERNEL_TEST_HDRS,
5001 deps = MICROKERNEL_TEST_DEPS,
5002)
5003
5004xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005005 name = "f32_vneg_test",
5006 srcs = [
5007 "test/f32-vneg.cc",
5008 "test/vunary-microkernel-tester.h",
5009 ] + MICROKERNEL_TEST_HDRS,
5010 deps = MICROKERNEL_TEST_DEPS,
5011)
5012
5013xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005014 name = "f32_vrndne_test",
5015 srcs = [
5016 "test/f32-vrndne.cc",
5017 "test/vunary-microkernel-tester.h",
5018 ] + MICROKERNEL_TEST_HDRS,
5019 deps = MICROKERNEL_TEST_DEPS,
5020)
5021
5022xnnpack_unit_test(
5023 name = "f32_vrndz_test",
5024 srcs = [
5025 "test/f32-vrndz.cc",
5026 "test/vunary-microkernel-tester.h",
5027 ] + MICROKERNEL_TEST_HDRS,
5028 deps = MICROKERNEL_TEST_DEPS,
5029)
5030
5031xnnpack_unit_test(
5032 name = "f32_vrndu_test",
5033 srcs = [
5034 "test/f32-vrndu.cc",
5035 "test/vunary-microkernel-tester.h",
5036 ] + MICROKERNEL_TEST_HDRS,
5037 deps = MICROKERNEL_TEST_DEPS,
5038)
5039
5040xnnpack_unit_test(
5041 name = "f32_vrndd_test",
5042 srcs = [
5043 "test/f32-vrndd.cc",
5044 "test/vunary-microkernel-tester.h",
5045 ] + MICROKERNEL_TEST_HDRS,
5046 deps = MICROKERNEL_TEST_DEPS,
5047)
5048
5049xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005050 name = "f32_vscale_test",
5051 srcs = [
5052 "test/f32-vscale.cc",
5053 "test/vscale-microkernel-tester.h",
5054 ] + MICROKERNEL_TEST_HDRS,
5055 deps = MICROKERNEL_TEST_DEPS,
5056)
5057
5058xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005059 name = "f32_vscaleexpminusmax_test",
5060 srcs = [
5061 "test/f32-vscaleexpminusmax.cc",
5062 "test/vscaleexpminusmax-microkernel-tester.h",
5063 ] + MICROKERNEL_TEST_HDRS,
5064 deps = MICROKERNEL_TEST_DEPS,
5065)
5066
5067xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005068 name = "f32_vscaleextexp_test",
5069 srcs = [
5070 "test/f32-vscaleextexp.cc",
5071 "test/vscaleextexp-microkernel-tester.h",
5072 ] + MICROKERNEL_TEST_HDRS,
5073 deps = MICROKERNEL_TEST_DEPS,
5074)
5075
5076xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005077 name = "f32_vsqr_test",
5078 srcs = [
5079 "test/f32-vsqr.cc",
5080 "test/vunary-microkernel-tester.h",
5081 ] + MICROKERNEL_TEST_HDRS,
5082 deps = MICROKERNEL_TEST_DEPS,
5083)
5084
5085xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005086 name = "f32_vsqrdiff_test",
5087 srcs = [
5088 "test/f32-vsqrdiff.cc",
5089 "test/vbinary-microkernel-tester.h",
5090 ] + MICROKERNEL_TEST_HDRS,
5091 deps = MICROKERNEL_TEST_DEPS,
5092)
5093
5094xnnpack_unit_test(
5095 name = "f32_vsqrdiffc_test",
5096 srcs = [
5097 "test/f32-vsqrdiffc.cc",
5098 "test/vbinaryc-microkernel-tester.h",
5099 ] + MICROKERNEL_TEST_HDRS,
5100 deps = MICROKERNEL_TEST_DEPS,
5101)
5102
5103xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005104 name = "f32_vsqrt_test",
5105 srcs = [
5106 "test/f32-vsqrt.cc",
5107 "test/vunary-microkernel-tester.h",
5108 ] + MICROKERNEL_TEST_HDRS,
5109 deps = MICROKERNEL_TEST_DEPS,
5110)
5111
5112xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005113 name = "f32_vsub_test",
5114 srcs = [
5115 "test/f32-vsub.cc",
5116 "test/vbinary-microkernel-tester.h",
5117 ] + MICROKERNEL_TEST_HDRS,
5118 deps = MICROKERNEL_TEST_DEPS,
5119)
5120
5121xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005122 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005123 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005124 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005125 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005126 ] + MICROKERNEL_TEST_HDRS,
5127 deps = MICROKERNEL_TEST_DEPS,
5128)
5129
5130xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005131 name = "f32_vsubc_test",
5132 srcs = [
5133 "test/f32-vsubc.cc",
5134 "test/vbinaryc-microkernel-tester.h",
5135 ] + MICROKERNEL_TEST_HDRS,
5136 deps = MICROKERNEL_TEST_DEPS,
5137)
5138
5139xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005140 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005141 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005142 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005143 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005144 ] + MICROKERNEL_TEST_HDRS,
5145 deps = MICROKERNEL_TEST_DEPS,
5146)
5147
5148xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005149 name = "f32_vrsubc_test",
5150 srcs = [
5151 "test/f32-vrsubc.cc",
5152 "test/vbinaryc-microkernel-tester.h",
5153 ] + MICROKERNEL_TEST_HDRS,
5154 deps = MICROKERNEL_TEST_DEPS,
5155)
5156
5157xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005158 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005159 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005160 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005161 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005162 ] + MICROKERNEL_TEST_HDRS,
5163 deps = MICROKERNEL_TEST_DEPS,
5164)
5165
5166xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005167 name = "qs8_gemm_minmax_test",
5168 srcs = [
5169 "test/qs8-gemm-minmax.cc",
5170 "test/gemm-microkernel-tester.h",
5171 "src/xnnpack/AlignedAllocator.h",
5172 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5173 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5174)
5175
5176xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005177 name = "qs8_igemm_minmax_test",
5178 srcs = [
5179 "test/qs8-igemm-minmax.cc",
5180 "test/gemm-microkernel-tester.h",
5181 "src/xnnpack/AlignedAllocator.h",
5182 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5183 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5184)
5185
5186xnnpack_unit_test(
5187 name = "qs8_requantization_test",
5188 srcs = [
5189 "src/xnnpack/requantization-stubs.h",
5190 "test/qs8-requantization.cc",
5191 "test/requantization-tester.h",
5192 ] + MICROKERNEL_TEST_HDRS,
5193 deps = MICROKERNEL_TEST_DEPS,
5194)
5195
5196xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005197 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005198 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005199 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200 "test/avgpool-microkernel-tester.h",
5201 "src/xnnpack/AlignedAllocator.h",
5202 ] + MICROKERNEL_TEST_HDRS,
5203 deps = MICROKERNEL_TEST_DEPS,
5204)
5205
5206xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005207 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005208 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005209 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210 "test/gemm-microkernel-tester.h",
5211 "src/xnnpack/AlignedAllocator.h",
5212 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005213 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005214)
5215
5216xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005217 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005218 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005219 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005220 "test/dwconv-microkernel-tester.h",
5221 "src/xnnpack/AlignedAllocator.h",
5222 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005223 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224)
5225
5226xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005227 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005228 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005229 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230 "test/gavgpool-microkernel-tester.h",
5231 "src/xnnpack/AlignedAllocator.h",
5232 ] + MICROKERNEL_TEST_HDRS,
5233 deps = MICROKERNEL_TEST_DEPS,
5234)
5235
5236xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005237 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005238 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005239 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005240 "test/gemm-microkernel-tester.h",
5241 "src/xnnpack/AlignedAllocator.h",
5242 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005243 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244)
5245
5246xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005247 name = "qu8_requantization_test",
5248 srcs = [
5249 "src/xnnpack/requantization-stubs.h",
5250 "test/qu8-requantization.cc",
5251 "test/requantization-tester.h",
5252 ] + MICROKERNEL_TEST_HDRS,
5253 deps = MICROKERNEL_TEST_DEPS,
5254)
5255
5256xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005257 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005259 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260 "test/vadd-microkernel-tester.h",
5261 ] + MICROKERNEL_TEST_HDRS,
5262 deps = MICROKERNEL_TEST_DEPS,
5263)
5264
5265xnnpack_unit_test(
5266 name = "u8_clamp_test",
5267 srcs = [
5268 "test/u8-clamp.cc",
5269 "test/clamp-microkernel-tester.h",
5270 ] + MICROKERNEL_TEST_HDRS,
5271 deps = MICROKERNEL_TEST_DEPS,
5272)
5273
5274xnnpack_unit_test(
5275 name = "u8_lut32norm_test",
5276 srcs = [
5277 "test/u8-lut32norm.cc",
5278 "test/lut-norm-microkernel-tester.h",
5279 ] + MICROKERNEL_TEST_HDRS,
5280 deps = MICROKERNEL_TEST_DEPS,
5281)
5282
5283xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005284 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005286 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287 "test/maxpool-microkernel-tester.h",
5288 ] + MICROKERNEL_TEST_HDRS,
5289 deps = MICROKERNEL_TEST_DEPS,
5290)
5291
5292xnnpack_unit_test(
5293 name = "u8_rmax_test",
5294 srcs = [
5295 "test/u8-rmax.cc",
5296 "test/rmax-microkernel-tester.h",
5297 ] + MICROKERNEL_TEST_HDRS,
5298 deps = MICROKERNEL_TEST_DEPS,
5299)
5300
5301xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005302 name = "x32_fill_test",
5303 srcs = [
5304 "test/x32-fill.cc",
5305 "test/fill-microkernel-tester.h",
5306 ] + MICROKERNEL_TEST_HDRS,
5307 deps = MICROKERNEL_TEST_DEPS,
5308)
5309
5310xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005311 name = "x32_packx_test",
5312 srcs = [
5313 "test/x32-packx.cc",
5314 "test/pack-microkernel-tester.h",
5315 "src/xnnpack/AlignedAllocator.h",
5316 ] + MICROKERNEL_TEST_HDRS,
5317 deps = MICROKERNEL_TEST_DEPS,
5318)
5319
5320xnnpack_unit_test(
5321 name = "x32_pad_test",
5322 srcs = [
5323 "test/x32-pad.cc",
5324 "test/pad-microkernel-tester.h",
5325 ] + MICROKERNEL_TEST_HDRS,
5326 deps = MICROKERNEL_TEST_DEPS,
5327)
5328
5329xnnpack_unit_test(
5330 name = "x32_unpool_test",
5331 srcs = [
5332 "test/x32-unpool.cc",
5333 "test/unpool-microkernel-tester.h",
5334 ] + MICROKERNEL_TEST_HDRS,
5335 deps = MICROKERNEL_TEST_DEPS,
5336)
5337
5338xnnpack_unit_test(
5339 name = "x32_zip_test",
5340 srcs = [
5341 "test/x32-zip.cc",
5342 "test/zip-microkernel-tester.h",
5343 ] + MICROKERNEL_TEST_HDRS,
5344 deps = MICROKERNEL_TEST_DEPS,
5345)
5346
5347xnnpack_unit_test(
5348 name = "x8_lut_test",
5349 srcs = [
5350 "test/x8-lut.cc",
5351 "test/lut-microkernel-tester.h",
5352 ] + MICROKERNEL_TEST_HDRS,
5353 deps = MICROKERNEL_TEST_DEPS,
5354)
5355
5356xnnpack_unit_test(
5357 name = "x8_zip_test",
5358 srcs = [
5359 "test/x8-zip.cc",
5360 "test/zip-microkernel-tester.h",
5361 ] + MICROKERNEL_TEST_HDRS,
5362 deps = MICROKERNEL_TEST_DEPS,
5363)
5364
Marat Dukhan20c3b922020-03-10 03:45:06 -07005365########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366
5367xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005368 name = "operator_size_test",
5369 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005370 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005371)
5372
Marat Dukhan20c3b922020-03-10 03:45:06 -07005373xnnpack_binary(
5374 name = "subgraph_size_test",
5375 srcs = ["test/subgraph-size.c"],
5376 deps = [":XNNPACK"],
5377)
5378
5379########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005380
5381xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005382 name = "abs_nc_test",
5383 srcs = [
5384 "test/abs-nc.cc",
5385 "test/abs-operator-tester.h",
5386 ],
5387 deps = OPERATOR_TEST_DEPS,
5388)
5389
5390xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005391 name = "add_nd_test",
5392 srcs = [
5393 "test/add-nd.cc",
5394 "test/binary-elementwise-operator-tester.h",
5395 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005396 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005397)
5398
5399xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005400 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005401 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005402 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403 "test/argmax-pooling-operator-tester.h",
5404 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005405 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406)
5407
5408xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005409 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005410 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005411 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412 "test/average-pooling-operator-tester.h",
5413 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005414 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415)
5416
5417xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005418 name = "bankers_rounding_nc_test",
5419 srcs = [
5420 "test/bankers-rounding-nc.cc",
5421 "test/bankers-rounding-operator-tester.h",
5422 ],
5423 deps = OPERATOR_TEST_DEPS,
5424)
5425
5426xnnpack_unit_test(
5427 name = "ceiling_nc_test",
5428 srcs = [
5429 "test/ceiling-nc.cc",
5430 "test/ceiling-operator-tester.h",
5431 ],
5432 deps = OPERATOR_TEST_DEPS,
5433)
5434
5435xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005436 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005437 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005438 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439 "test/channel-shuffle-operator-tester.h",
5440 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005441 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005442)
5443
5444xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005445 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005446 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005447 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005448 "test/clamp-operator-tester.h",
5449 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005450 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005451)
5452
5453xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005454 name = "constant_pad_nd_test",
5455 srcs = [
5456 "test/constant-pad-nd.cc",
5457 "test/constant-pad-operator-tester.h",
5458 ],
5459 deps = OPERATOR_TEST_DEPS,
5460)
5461
5462xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005463 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005464 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005465 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005466 "test/convolution-operator-tester.h",
5467 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005468 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005469)
5470
5471xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005472 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005474 "test/convolution-nchw.cc",
5475 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005476 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005477 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478)
5479
5480xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005481 name = "copy_nc_test",
5482 srcs = [
5483 "test/copy-nc.cc",
5484 "test/copy-operator-tester.h",
5485 ],
5486 deps = OPERATOR_TEST_DEPS,
5487)
5488
5489xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005490 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005491 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005492 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005493 "test/deconvolution-operator-tester.h",
5494 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005495 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496)
5497
5498xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005499 name = "divide_nd_test",
5500 srcs = [
5501 "test/binary-elementwise-operator-tester.h",
5502 "test/divide-nd.cc",
5503 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005504 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005505)
5506
5507xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005508 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005509 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005510 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005511 "test/fully-connected-operator-tester.h",
5512 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005513 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005514)
5515
5516xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005517 name = "floor_nc_test",
5518 srcs = [
5519 "test/floor-nc.cc",
5520 "test/floor-operator-tester.h",
5521 ],
5522 deps = OPERATOR_TEST_DEPS,
5523)
5524
5525xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005526 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005527 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005528 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005529 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005530 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005531 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005532)
5533
5534xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005535 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005536 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005537 "test/global-average-pooling-ncw.cc",
5538 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005539 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005540 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005541)
5542
5543xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005544 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005545 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005546 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005547 "test/hardswish-operator-tester.h",
5548 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005549 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005550)
5551
5552xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005553 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005554 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005555 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005556 "test/leaky-relu-operator-tester.h",
5557 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005558 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005559)
5560
5561xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005562 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005563 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005564 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005565 "test/max-pooling-operator-tester.h",
5566 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005567 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005568)
5569
5570xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005571 name = "maximum_nd_test",
5572 srcs = [
5573 "test/binary-elementwise-operator-tester.h",
5574 "test/maximum-nd.cc",
5575 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005576 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005577)
5578
5579xnnpack_unit_test(
5580 name = "minimum_nd_test",
5581 srcs = [
5582 "test/binary-elementwise-operator-tester.h",
5583 "test/minimum-nd.cc",
5584 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005585 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005586)
5587
5588xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005589 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005590 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005591 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005592 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005593 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005594 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005595)
5596
5597xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005598 name = "negate_nc_test",
5599 srcs = [
5600 "test/negate-nc.cc",
5601 "test/negate-operator-tester.h",
5602 ],
5603 deps = OPERATOR_TEST_DEPS,
5604)
5605
5606xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005607 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005608 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005609 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005610 "test/prelu-operator-tester.h",
5611 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005612 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005613)
5614
5615xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005616 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005617 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005618 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005619 "test/resize-bilinear-operator-tester.h",
5620 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005621 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005622)
5623
5624xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005625 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005626 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005627 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005628 "test/sigmoid-operator-tester.h",
5629 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005630 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005631)
5632
5633xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005634 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005635 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005636 "test/softmax-nc.cc",
5637 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005638 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005639 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005640)
5641
5642xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005643 name = "square_nc_test",
5644 srcs = [
5645 "test/square-nc.cc",
5646 "test/square-operator-tester.h",
5647 ],
5648 deps = OPERATOR_TEST_DEPS,
5649)
5650
5651xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005652 name = "square_root_nc_test",
5653 srcs = [
5654 "test/square-root-nc.cc",
5655 "test/square-root-operator-tester.h",
5656 ],
5657 deps = OPERATOR_TEST_DEPS,
5658)
5659
5660xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005661 name = "squared_difference_nd_test",
5662 srcs = [
5663 "test/binary-elementwise-operator-tester.h",
5664 "test/squared-difference-nd.cc",
5665 ],
5666 deps = OPERATOR_TEST_DEPS,
5667)
5668
5669xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005670 name = "subtract_nd_test",
5671 srcs = [
5672 "test/binary-elementwise-operator-tester.h",
5673 "test/subtract-nd.cc",
5674 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005675 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005676)
5677
5678xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005679 name = "truncation_nc_test",
5680 srcs = [
5681 "test/truncation-nc.cc",
5682 "test/truncation-operator-tester.h",
5683 ],
5684 deps = OPERATOR_TEST_DEPS,
5685)
5686
5687xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005688 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005689 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005690 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005691 "test/unpooling-operator-tester.h",
5692 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005693 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005694)
5695
Chao Mei6ddfc602020-05-13 22:29:36 -07005696############################### Misc unit tests ###############################
5697
5698xnnpack_unit_test(
5699 name = "memory_planner_test",
5700 srcs = [
5701 "test/memory-planner-test.cc",
5702 ],
5703 deps = [
5704 ":XNNPACK",
5705 ":memory_planner",
5706 ],
5707)
5708
Marat Dukhan08c4a432019-10-03 09:29:21 -07005709############################# Build configurations #############################
5710
Marat Dukhanb8642352019-10-30 15:43:02 -07005711# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005712config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005713 name = "xnn_enable_assembly_explicit_true",
5714 define_values = {"xnn_enable_assembly": "true"},
5715)
5716
5717# Disables usage of assembly kernels.
5718config_setting(
5719 name = "xnn_enable_assembly_explicit_false",
5720 define_values = {"xnn_enable_assembly": "false"},
5721)
5722
Marat Dukhan9de90e02020-06-18 16:04:12 -07005723# Enables usage of sparse inference.
5724config_setting(
5725 name = "xnn_enable_sparse_explicit_true",
5726 define_values = {"xnn_enable_sparse": "true"},
5727)
5728
5729# Disables usage of sparse inference.
5730config_setting(
5731 name = "xnn_enable_sparse_explicit_false",
5732 define_values = {"xnn_enable_sparse": "false"},
5733)
5734
Marat Dukhan05702cf2020-03-26 15:41:33 -07005735# Disables usage of HMP-aware optimizations.
5736config_setting(
5737 name = "xnn_enable_hmp_explicit_false",
5738 define_values = {"xnn_enable_hmp": "false"},
5739)
5740
Chao Mei6ddfc602020-05-13 22:29:36 -07005741# Enable usage of optimized memory allocation
5742config_setting(
5743 name = "xnn_enable_memopt_explicit_true",
5744 define_values = {"xnnpack_enable_memopt": "true"},
5745)
5746
5747# Disable usage of optimized memory allocation
5748config_setting(
5749 name = "xnn_enable_memopt_explicit_false",
5750 define_values = {"xnnpack_enable_memopt": "false"},
5751)
5752
Marat Dukhanb8642352019-10-30 15:43:02 -07005753# Builds with -c dbg
5754config_setting(
5755 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005756 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005757 "compilation_mode": "dbg",
5758 },
5759)
5760
5761# Builds with -c opt
5762config_setting(
5763 name = "optimized_build",
5764 values = {
5765 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005766 },
5767)
5768
5769config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005770 name = "linux_k8",
5771 values = {"cpu": "k8"},
5772)
5773
5774config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005775 name = "linux_arm",
5776 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005777)
5778
5779config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005780 name = "linux_armeabi",
5781 values = {"cpu": "armeabi"},
5782)
5783
5784config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005785 name = "linux_armhf",
5786 values = {"cpu": "armhf"},
5787)
5788
5789config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005790 name = "linux_armv7a",
5791 values = {"cpu": "armv7a"},
5792)
5793
5794config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005795 name = "linux_aarch64",
5796 values = {"cpu": "aarch64"},
5797)
5798
5799config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005800 name = "android",
5801 values = {"crosstool_top": "//external:android/crosstool"},
5802)
5803
5804config_setting(
5805 name = "android_armv7",
5806 values = {
5807 "crosstool_top": "//external:android/crosstool",
5808 "cpu": "armeabi-v7a",
5809 },
5810)
5811
5812config_setting(
5813 name = "android_arm64",
5814 values = {
5815 "crosstool_top": "//external:android/crosstool",
5816 "cpu": "arm64-v8a",
5817 },
5818)
5819
5820config_setting(
5821 name = "android_x86",
5822 values = {
5823 "crosstool_top": "//external:android/crosstool",
5824 "cpu": "x86",
5825 },
5826)
5827
5828config_setting(
5829 name = "android_x86_64",
5830 values = {
5831 "crosstool_top": "//external:android/crosstool",
5832 "cpu": "x86_64",
5833 },
5834)
5835
5836config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005837 name = "windows_x86_64",
5838 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005839)
5840
5841config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005842 name = "windows_x86_64_clang",
5843 values = {
5844 "compiler": "clang-cl",
5845 "cpu": "x64_windows",
5846 },
5847)
5848
5849config_setting(
5850 name = "windows_x86_64_mingw",
5851 values = {
5852 "compiler": "mingw-gcc",
5853 "cpu": "x64_windows",
5854 },
5855)
5856
5857config_setting(
5858 name = "windows_x86_64_msys",
5859 values = {
5860 "compiler": "msys-gcc",
5861 "cpu": "x64_windows",
5862 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005863)
5864
5865config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005866 name = "macos_x86_64",
5867 values = {
5868 "apple_platform_type": "macos",
5869 "cpu": "darwin",
5870 },
5871)
5872
5873config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005874 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005875 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005876)
5877
5878config_setting(
5879 name = "emscripten_wasm",
5880 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005881 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005882 "cpu": "wasm",
5883 },
5884)
5885
5886config_setting(
5887 name = "emscripten_wasmsimd",
5888 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005889 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005890 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005891 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005892 },
5893)
5894
5895config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005896 name = "ios_armv7",
5897 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005898 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005899 "cpu": "ios_armv7",
5900 },
5901)
5902
5903config_setting(
5904 name = "ios_arm64",
5905 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005906 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005907 "cpu": "ios_arm64",
5908 },
5909)
5910
5911config_setting(
5912 name = "ios_arm64e",
5913 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005914 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005915 "cpu": "ios_arm64e",
5916 },
5917)
5918
5919config_setting(
5920 name = "ios_x86",
5921 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005922 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005923 "cpu": "ios_i386",
5924 },
5925)
5926
5927config_setting(
5928 name = "ios_x86_64",
5929 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005930 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005931 "cpu": "ios_x86_64",
5932 },
5933)
5934
5935config_setting(
5936 name = "watchos_armv7k",
5937 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005938 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005939 "cpu": "watchos_armv7k",
5940 },
5941)
5942
5943config_setting(
5944 name = "watchos_arm64_32",
5945 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005946 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005947 "cpu": "watchos_arm64_32",
5948 },
5949)
5950
5951config_setting(
5952 name = "watchos_x86",
5953 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005954 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005955 "cpu": "watchos_i386",
5956 },
5957)
5958
5959config_setting(
5960 name = "watchos_x86_64",
5961 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005962 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005963 "cpu": "watchos_x86_64",
5964 },
5965)
5966
5967config_setting(
5968 name = "tvos_arm64",
5969 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005970 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005971 "cpu": "tvos_arm64",
5972 },
5973)
5974
5975config_setting(
5976 name = "tvos_x86_64",
5977 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005978 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005979 "cpu": "tvos_x86_64",
5980 },
5981)