blob: a6ebd3ab79ae399636d5b0d489cf0e62b022dab3 [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 Dukhan8ee37012020-07-16 13:17:13 -0700854 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700855 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700856 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700857 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700858 "src/x32-zip/x2-wasmsimd.c",
859 "src/x32-zip/x3-wasmsimd.c",
860 "src/x32-zip/x4-wasmsimd.c",
861 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700862 "src/math/roundne-wasmsimd-addsub.c",
863 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700864 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700865 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700866 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700867 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700868 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700869 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
870 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700871]
872
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800873PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700874 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700875 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700876 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700877 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700878 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700879 "src/f32-spmm/gen/4x1-minmax-psimd.c",
880 "src/f32-spmm/gen/8x1-minmax-psimd.c",
881 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700882]
883
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800884PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700885 "src/qs8-requantization/precise-psimd.c",
886 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700887 "src/qu8-requantization/precise-psimd.c",
888 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800889]
890
Marat Dukhan08c4a432019-10-03 09:29:21 -0700891# ISA-specific micro-kernels
892NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700893 "src/f32-argmaxpool/4x-neon-c4.c",
894 "src/f32-argmaxpool/9p8x-neon-c4.c",
895 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700896 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
897 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700898 "src/f32-clamp/gen/neon-x4.c",
899 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700900 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
903 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700904 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
905 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700906 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
907 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700908 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
909 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
910 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
911 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700912 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
913 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
914 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
915 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700916 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
917 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
918 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
919 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700920 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700921 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
922 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700923 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
924 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
925 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
926 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
927 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
928 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
929 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
930 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
931 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
932 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
933 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
934 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
935 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
936 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
937 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
938 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
939 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
940 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
941 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
942 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
944 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
945 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
946 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
947 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
948 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
949 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
950 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
951 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
952 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
953 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700954 "src/f32-hswish/gen/hswish-neon-x4.c",
955 "src/f32-hswish/gen/hswish-neon-x8.c",
956 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700957 "src/f32-ibilinear/gen/neon-c4.c",
958 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700959 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
960 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
961 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
962 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
963 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
964 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
965 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
966 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
967 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
968 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
969 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
970 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
971 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
972 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
973 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
974 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700975 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
976 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
977 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700978 "src/f32-ppmm/gen/4x8-minmax-neon.c",
979 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700980 "src/f32-prelu/gen/neon-1x4.c",
981 "src/f32-prelu/gen/neon-1x8.c",
982 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800983 "src/f32-prelu/gen/neon-2x4.c",
984 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700985 "src/f32-prelu/gen/neon-2x16.c",
986 "src/f32-prelu/gen/neon-4x4.c",
987 "src/f32-prelu/gen/neon-4x8.c",
988 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800989 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1005 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001013 "src/f32-relu/gen/neon-x4.c",
1014 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001015 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001016 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001017 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1018 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1019 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1020 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1021 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1022 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1023 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1026 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1027 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1028 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1029 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1030 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1031 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1032 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1033 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1034 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001035 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1036 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1037 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1038 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001039 "src/f32-vbinary/gen/vmax-neon-x4.c",
1040 "src/f32-vbinary/gen/vmax-neon-x8.c",
1041 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1042 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1043 "src/f32-vbinary/gen/vmin-neon-x4.c",
1044 "src/f32-vbinary/gen/vmin-neon-x8.c",
1045 "src/f32-vbinary/gen/vminc-neon-x4.c",
1046 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001047 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1048 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1049 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1050 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1051 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1052 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001053 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1054 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1055 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1056 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001057 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1058 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1059 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1060 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001061 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1062 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001063 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1064 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001065 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1066 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1067 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1068 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1069 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1070 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1071 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1072 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001073 "src/f32-vunary/gen/vabs-neon-x4.c",
1074 "src/f32-vunary/gen/vabs-neon-x8.c",
1075 "src/f32-vunary/gen/vneg-neon-x4.c",
1076 "src/f32-vunary/gen/vneg-neon-x8.c",
1077 "src/f32-vunary/gen/vsqr-neon-x4.c",
1078 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001079 "src/qs8-requantization/precise-neon.c",
1080 "src/qs8-requantization/fp32-neon.c",
1081 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001082 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1083 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1084 "src/qu8-dwconv/up8x9-minmax-neon.c",
1085 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1086 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1087 "src/qu8-gemm/4x8-minmax-neon.c",
1088 "src/qu8-gemm/8x8-minmax-neon.c",
1089 "src/qu8-igemm/4x8-minmax-neon.c",
1090 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001091 "src/qu8-requantization/precise-neon.c",
1092 "src/qu8-requantization/fp32-neon.c",
1093 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001094 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001095 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001096 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001097 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001098 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001099 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001100 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001101 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001102 "src/x32-zip/x2-neon.c",
1103 "src/x32-zip/x3-neon.c",
1104 "src/x32-zip/x4-neon.c",
1105 "src/x32-zip/xm-neon.c",
1106 "src/x8-zip/x2-neon.c",
1107 "src/x8-zip/x3-neon.c",
1108 "src/x8-zip/x4-neon.c",
1109 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001110 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001111 "src/math/roundd-neon-addsub.c",
1112 "src/math/roundd-neon-cvt.c",
1113 "src/math/roundu-neon-addsub.c",
1114 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001115 "src/math/roundz-neon-addsub.c",
1116 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001117 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001118 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1119 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1120 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1121 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1122 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1123 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001124 "src/math/sqrt-neon-nr1rsqrts.c",
1125 "src/math/sqrt-neon-nr2rsqrts.c",
1126 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001127]
1128
1129NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001130 "src/f32-ibilinear/gen/neonfma-c4.c",
1131 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001132 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1133 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1134 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1135 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1136 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1137 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1138 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1139 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1140 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001141 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1142 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1143 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1144 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001145 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1146 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1147 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1148 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001149 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1150 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1151 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1152 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1154 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1155 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1156 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1157 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1158 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1159 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1160 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1161 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1162 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1164 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1165 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1166 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1167 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1168 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1169 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1170 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001171 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1172 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001173 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1177 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1178 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1179 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1180 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1181 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1182 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1185 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1187 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1188 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1189 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1190 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1191 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1196 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001197 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001251 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1252 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001253 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1254 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1255 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1256 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1257 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1258 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1259 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1260 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1261 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1262 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1263 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1264 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1265 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1266 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1267 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1268 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1269 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1270 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1271 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1272 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001273 "src/math/exp-neonfma-lut64-p2.c",
1274 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001275 "src/math/expminus-neonfma-lut2048-p1.c",
1276 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001277 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001278 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1279 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1280 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1281 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1282 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1283 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1284 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1285 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1286 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1287 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1288 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1289 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1290 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1291 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1292 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1293 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1294 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1295 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001296 "src/math/sqrt-neonfma-nr1fma.c",
1297 "src/math/sqrt-neonfma-nr2fma.c",
1298 "src/math/sqrt-neonfma-nr3fma.c",
1299 "src/math/sqrt-neonfma-nr2fma1adj.c",
1300 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001301]
1302
1303AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001304 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1305 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1306 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1307 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1308 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1309 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001310 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1311 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1312 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1313 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1314 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1315 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1316 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1317 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1318 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1319 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1320 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1321 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1322 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1323 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1324 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1325 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1326 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1327 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1328 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1329 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001330 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1331 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001332 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1333 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001334 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1335 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001336 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1337 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001338 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1339 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1340 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1341 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1342 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001343 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001361 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1362 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1363 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1364 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1365 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1366 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1367 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1368 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1369 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1370 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1371 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1372 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1373 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1374 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1375 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1376 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1377 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1378 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001379 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1380 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001381 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1382 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1383 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1384 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1385 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1386 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001387]
1388
Marat Dukhan8853b822020-05-07 12:19:01 -07001389NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001390 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1391 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1392 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1393 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1394 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1395 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1396 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1397 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001398 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001399 "src/math/roundd-neonv8.c",
1400 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001401 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001402]
1403
Marat Dukhan08c4a432019-10-03 09:29:21 -07001404AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001405 "src/f16-clamp/gen/neonfp16arith-x16.c",
1406 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001407 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1408 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1409 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1410 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1411 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1412 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1413 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1414 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1415 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1416 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1417 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1418 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001419 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1420 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001421 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1422 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001423 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001424 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001425 "src/f16-relu/gen/neonfp16arith-x16.c",
1426 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001427 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1428 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1429 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1430 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1431 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1432 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1433 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1434 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1435 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1436 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1437 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1438 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1439 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1440 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1441 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1442 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1443 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1444 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1445 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1446 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1447 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1448 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1449 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1450 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1451 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1452 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1453 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1454 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001455 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1456 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001457 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1458 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1459 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1460 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1461 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1462 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1463 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1464 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001465 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1466 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1467 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1468 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001469 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1470 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1471 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1472 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1473 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1474 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1475 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1476 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1477 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1478 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1479 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1480 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001481 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1482 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1483 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1484 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1485 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1486 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1487 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1488 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001489]
1490
1491SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001492 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1493 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001494 "src/f32-clamp/gen/sse-x4.c",
1495 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001496 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1497 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001498 "src/f32-dwconv-chw/3x3p1-sse.c",
1499 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001500 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1501 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1502 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1503 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1504 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1505 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1506 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1507 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1508 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1509 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1510 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1511 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001512 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001513 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1514 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001515 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1516 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1517 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1518 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1519 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1520 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1521 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1522 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1523 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1524 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1525 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1526 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1527 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001528 "src/f32-hswish/gen/hswish-sse-x4.c",
1529 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001530 "src/f32-ibilinear/gen/sse-c4.c",
1531 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001532 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1533 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1534 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1535 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1536 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1537 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1538 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001539 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1540 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1541 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001542 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001543 "src/f32-prelu/gen/sse-2x4.c",
1544 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001545 "src/f32-relu/gen/sse-x4.c",
1546 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001547 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001548 "src/f32-spmm/gen/4x1-minmax-sse.c",
1549 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001550 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001551 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1552 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1553 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1554 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1555 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1556 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1557 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1558 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001559 "src/f32-vbinary/gen/vmax-sse-x4.c",
1560 "src/f32-vbinary/gen/vmax-sse-x8.c",
1561 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1562 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1563 "src/f32-vbinary/gen/vmin-sse-x4.c",
1564 "src/f32-vbinary/gen/vmin-sse-x8.c",
1565 "src/f32-vbinary/gen/vminc-sse-x4.c",
1566 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001567 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1568 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1569 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1570 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1571 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1572 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1573 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1574 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001575 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1576 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1577 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1578 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001579 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1580 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1581 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1582 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001583 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1584 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001585 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1586 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001587 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1588 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001589 "src/f32-vunary/gen/vabs-sse-x4.c",
1590 "src/f32-vunary/gen/vabs-sse-x8.c",
1591 "src/f32-vunary/gen/vneg-sse-x4.c",
1592 "src/f32-vunary/gen/vneg-sse-x8.c",
1593 "src/f32-vunary/gen/vsqr-sse-x4.c",
1594 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001595 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001596 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001597 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001598 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001599 "src/math/roundd-sse-addsub.c",
1600 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001601 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001602 "src/math/sqrt-sse-hh1mac.c",
1603 "src/math/sqrt-sse-nr1mac.c",
1604 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001605]
1606
1607SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001608 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1609 "src/f32-argmaxpool/4x-sse2-c4.c",
1610 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001611 "src/f32-prelu/gen/sse2-2x4.c",
1612 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001613 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1614 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1615 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1616 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1617 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1618 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1619 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1620 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1621 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1622 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1623 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1624 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001625 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001626 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001627 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001628 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001629 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1630 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001631 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1632 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001633 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1634 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1635 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1636 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1637 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1638 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1639 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1640 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001641 "src/qs8-requantization/precise-sse2.c",
1642 "src/qs8-requantization/fp32-sse2.c",
1643 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001644 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1645 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001646 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1647 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001648 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1649 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001650 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1651 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001652 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001653 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1654 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001655 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001656 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1657 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1658 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001659 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1660 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001661 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1662 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001663 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1664 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001665 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001666 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1667 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001668 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001669 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1670 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1671 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1672 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1673 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1674 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1675 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1676 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001677 "src/qu8-requantization/precise-sse2.c",
1678 "src/qu8-requantization/fp32-sse2.c",
1679 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001680 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001681 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001682 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001683 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001684 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001685 "src/x32-zip/x2-sse2.c",
1686 "src/x32-zip/x3-sse2.c",
1687 "src/x32-zip/x4-sse2.c",
1688 "src/x32-zip/xm-sse2.c",
1689 "src/x8-zip/x2-sse2.c",
1690 "src/x8-zip/x3-sse2.c",
1691 "src/x8-zip/x4-sse2.c",
1692 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001693 "src/math/exp-sse2-p5.c",
1694 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001695 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001696 "src/math/roundd-sse2-cvt.c",
1697 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001698 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001699 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001700]
1701
1702SSSE3_UKERNELS = [
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001703 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1704 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001705 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1706 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001707 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1708 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001709 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1710 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001711 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001712 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1713 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001714 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001715 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1716 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1717 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001718 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1719 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001720 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1721 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001722 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1723 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001724 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001725 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1726 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001727 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001728 "src/qs8-requantization/precise-ssse3.c",
1729 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001730 "src/qu8-requantization/precise-ssse3.c",
1731 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001732]
1733
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001734SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001735 "src/f32-prelu/gen/sse41-2x4.c",
1736 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001737 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001738 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001739 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001740 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001741 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1742 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001743 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1744 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001745 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1746 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1747 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1748 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1749 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1750 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1751 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1752 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001753 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1754 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001755 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1756 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001757 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1758 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001759 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1760 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001761 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001762 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1763 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001764 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001765 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1766 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1767 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001768 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1769 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001770 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1771 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001772 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1773 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001774 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001775 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1776 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001777 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001778 "src/qs8-requantization/fp32-sse4.c",
1779 "src/qs8-requantization/precise-sse4.c",
1780 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001781 "src/qu8-requantization/precise-sse4.c",
1782 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001783 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001784 "src/math/roundd-sse41.c",
1785 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001786 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001787]
1788
Marat Dukhan08c4a432019-10-03 09:29:21 -07001789AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001790 "src/f32-clamp/gen/avx-x8.c",
1791 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001792 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1793 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1794 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1795 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1796 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1797 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1798 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1799 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1800 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1801 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1802 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1803 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1804 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1805 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1806 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1807 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1808 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1809 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1810 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1811 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1812 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1813 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1814 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1815 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1816 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1817 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1818 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1819 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1820 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1821 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001822 "src/f32-hswish/gen/hswish-avx-x8.c",
1823 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001824 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1825 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1826 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1827 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1828 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1829 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1830 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1831 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1832 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001833 "src/f32-prelu/gen/avx-2x8.c",
1834 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001835 "src/f32-relu/gen/avx-x16.c",
1836 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001837 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001838 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1839 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1840 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1841 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1842 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1843 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1844 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1845 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001846 "src/f32-vbinary/gen/vmax-avx-x8.c",
1847 "src/f32-vbinary/gen/vmax-avx-x16.c",
1848 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1849 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1850 "src/f32-vbinary/gen/vmin-avx-x8.c",
1851 "src/f32-vbinary/gen/vmin-avx-x16.c",
1852 "src/f32-vbinary/gen/vminc-avx-x8.c",
1853 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001854 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1855 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1856 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1857 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1858 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1859 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1860 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1861 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001862 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1863 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1864 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1865 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001866 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1867 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1868 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1869 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001870 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1871 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001872 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001873 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1874 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1875 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1876 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1877 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1878 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1879 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1880 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001881 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1882 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001883 "src/f32-vunary/gen/vabs-avx-x8.c",
1884 "src/f32-vunary/gen/vabs-avx-x16.c",
1885 "src/f32-vunary/gen/vneg-avx-x8.c",
1886 "src/f32-vunary/gen/vneg-avx-x16.c",
1887 "src/f32-vunary/gen/vsqr-avx-x8.c",
1888 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001889]
1890
Marat Dukhan1566fee2020-08-02 21:55:41 -07001891XOP_UKERNELS = [
1892 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1893 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1894 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1895 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001896 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
1897 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001898 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1899 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001900 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001901 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1902 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001903 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001904 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
1905 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
1906 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001907 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1908 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1909 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1910 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1911 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1912 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001913 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001914 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1915 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001916 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001917]
1918
Marat Dukhanfda12b82019-11-21 12:27:59 -08001919FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001920 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1921 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1922 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1923 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1924 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1925 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1926 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1927 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1928 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1929 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1930 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1931 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1932 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1933 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1934 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1935 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1936 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1937 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1938 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1939 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1940 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1941 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1942 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1943 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1944 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1945 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1946 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1947 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1948 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1949 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1950 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1951 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1952 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1953 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1954 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1955 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1956 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1957 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1958 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1959 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001960 "src/f32-hswish/gen/hswish-fma3-x8.c",
1961 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001962 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1963 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1964 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1965 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1966 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1967 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1968 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1969 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1970 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1971 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1972 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1973 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1974 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1975 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001976 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1977 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1978 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1979 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1980 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1981 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1982 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1983 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001984 "src/math/sqrt-fma3-nr1fma.c",
1985 "src/math/sqrt-fma3-nr2fma.c",
1986 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001987]
1988
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001989AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001990 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1991 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1992 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1993 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1994 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1995 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1996 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1997 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1998 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1999 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2000 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2001 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2002 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2003 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2004 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2005 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2006 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2007 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2008 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2009 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2010 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2011 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2012 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2013 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2014 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2015 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2016 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2017 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2018 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2019 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2020 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2021 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2022 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2023 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2024 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2025 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002026 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2027 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2028 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2029 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2030 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2031 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2032 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2033 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2034 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2035 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2036 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2037 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2038 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2039 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2040 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2041 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2042 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2043 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2044 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2045 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2046 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2047 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2048 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2049 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2050 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2051 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2052 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2053 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2054 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2055 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002056 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2057 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2058 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2059 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2060 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2061 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2062 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2063 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2064 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2065 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2066 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2067 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2068 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2069 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2070 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2071 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2072 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2073 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2074 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2075 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2076 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2077 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2078 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2079 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002080 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2081 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2082 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002083 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2084 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2085 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002086 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2087 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2088 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002089 "src/math/exp-avx2-p5.c",
2090 "src/math/exp-avx2-perm-p3.c",
2091 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002092 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002093 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002094 "src/math/sigmoid-avx2-rr2-p5-div.c",
2095 "src/math/sigmoid-avx2-rr1-p5-div.c",
2096 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2097 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2098 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2099 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002100]
2101
Marat Dukhan08c4a432019-10-03 09:29:21 -07002102AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002103 "src/f32-clamp/gen/avx512f-x16.c",
2104 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002105 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2106 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2107 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2108 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2109 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2110 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2111 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2112 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2113 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2114 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2115 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2116 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2117 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2118 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2119 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2120 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2121 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2122 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2123 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2124 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2125 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2126 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2127 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2128 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002129 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2130 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002131 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2132 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2133 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2134 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2135 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2136 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002137 "src/f32-prelu/gen/avx512f-2x16.c",
2138 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002139 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2140 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2141 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2142 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2143 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2144 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2145 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2146 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2147 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2148 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2149 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2150 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2151 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2152 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2153 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2154 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2155 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2156 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2157 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2158 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2159 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2160 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2161 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2162 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2163 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2164 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2165 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2166 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2167 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2168 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2169 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2170 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2171 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2172 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2173 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2174 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002175 "src/f32-relu/gen/avx512f-x16.c",
2176 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002177 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002178 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2179 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2180 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2181 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2182 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2183 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2184 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2185 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002186 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2187 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2188 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2189 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2190 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2191 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2192 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2193 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002194 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2195 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2196 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2197 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2198 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2199 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2200 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2201 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002202 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2203 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2204 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2205 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002206 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2207 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2208 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2209 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002210 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2211 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002212 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002213 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2214 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2215 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2216 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2217 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2218 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2219 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2220 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2221 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2222 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2223 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2224 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2225 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2226 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2227 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2228 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2229 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2230 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2231 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2232 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2233 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2234 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2235 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2236 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002237 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2238 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2239 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2240 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2241 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2242 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2243 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2244 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002245 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2246 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2247 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2248 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2249 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2250 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2251 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2252 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002253 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2254 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2255 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2256 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2257 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2258 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002259 "src/math/exp-avx512f-p5-scalef.c",
2260 "src/math/exp-avx512f-p5.c",
2261 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002262 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002263 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002264 "src/math/sqrt-avx512f-nr1fma.c",
2265 "src/math/sqrt-avx512f-nr2fma.c",
2266 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002267]
2268
2269AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002270 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002271 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002272 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2273 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002274 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002275 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002276 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002277 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002278 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2279 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002280 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2281 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2282 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2283 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002284]
2285
2286AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002287 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002288 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2289 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002290 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002291 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2292 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002293 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2294 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2295 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002296 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002297 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2298 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2299 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002300 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002301 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2302 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002303 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002304 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2305 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2306 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2307 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2308 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2309 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2310 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2311 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2312 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2313 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2314 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2315 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2316 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2317 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2318 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2319 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2320 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2321 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002322 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2323 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002324 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002325 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2326 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2327 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2328 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2329 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2330 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2331 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2332 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2333 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2334 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2335 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2336 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2337 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2338 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2339 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2340 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2341 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2342 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002343 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2344 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2345 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2346 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2347 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2348 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2349 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2350 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2351 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2352 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2353 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2354 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2355 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2356 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2357 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2358 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2359 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2360 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002361]
2362
Marat Dukhan1b354632020-03-23 12:50:22 -07002363INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002364 "src/xnnpack/argmaxpool.h",
2365 "src/xnnpack/avgpool.h",
2366 "src/xnnpack/clamp.h",
2367 "src/xnnpack/common.h",
2368 "src/xnnpack/conv.h",
2369 "src/xnnpack/dwconv.h",
2370 "src/xnnpack/gavgpool.h",
2371 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002372 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002373 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002374 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002375 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002376 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002377 "src/xnnpack/lut.h",
2378 "src/xnnpack/math.h",
2379 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002380 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002381 "src/xnnpack/packx.h",
2382 "src/xnnpack/pad.h",
2383 "src/xnnpack/params.h",
2384 "src/xnnpack/pavgpool.h",
2385 "src/xnnpack/ppmm.h",
2386 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002387 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002388 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002389 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002390 "src/xnnpack/rmax.h",
2391 "src/xnnpack/scalar-utils.h",
2392 "src/xnnpack/spmm.h",
2393 "src/xnnpack/unpool.h",
2394 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002395 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002396 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002397 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002398 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002399 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002400 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002401 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002402]
2403
2404INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002405 "include/xnnpack.h",
2406 "src/xnnpack/allocator.h",
2407 "src/xnnpack/compute.h",
2408 "src/xnnpack/im2col.h",
2409 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002410 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002411 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002412 "src/xnnpack/operator.h",
2413 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002414 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002415 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002416 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002417 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002418]
2419
Marat Dukhan1b354632020-03-23 12:50:22 -07002420ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002421 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002422]
2423
Marat Dukhan1b354632020-03-23 12:50:22 -07002424MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002425 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002426 "include/xnnpack.h",
2427]
2428
Marat Dukhan1b354632020-03-23 12:50:22 -07002429MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002430 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002431 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002432 "src/xnnpack/requantization.h",
2433 "include/xnnpack.h",
2434]
2435
2436OPERATOR_TEST_PARAMS_HDRS = [
2437 "src/xnnpack/params.h",
2438 "src/xnnpack/common.h",
2439]
2440
2441WEIGHTS_PACK_HDRS = [
2442 "src/xnnpack/pack.h",
2443 "src/xnnpack/operator.h",
2444 "src/xnnpack/compute.h",
2445]
2446
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002447LOGGING_COPTS = select({
2448 # No logging in optimized mode
2449 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2450 # Full logging in debug mode
2451 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2452 # Error-only logging in default (fastbuild) mode
2453 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2454})
2455
Marat Dukhan3b59de22020-06-03 20:15:19 -07002456LOGGING_SRCS = select({
2457 # No logging in optimized mode
2458 ":optimized_build": [],
2459 "//conditions:default": [
2460 "src/operator-strings.c",
2461 "src/subgraph-strings.c",
2462 ],
2463})
2464
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002465LOGGING_HDRS = [
2466 "src/xnnpack/log.h",
2467]
2468
Marat Dukhan08c4a432019-10-03 09:29:21 -07002469xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002470 name = "tables",
2471 srcs = TABLE_SRCS,
2472 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002473 gcc_copts = xnnpack_gcc_std_copts(),
2474 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002475)
2476
2477xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002478 name = "scalar_ukernels",
2479 srcs = SCALAR_UKERNELS,
2480 hdrs = INTERNAL_HDRS,
2481 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002482 gcc_copts = xnnpack_gcc_std_copts(),
2483 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002484 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002485 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002486 "@FP16",
2487 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002488 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002489 ],
2490)
2491
2492xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002493 name = "scalar_ukernels_test_mode",
2494 srcs = SCALAR_UKERNELS,
2495 hdrs = INTERNAL_HDRS,
2496 aarch32_copts = ["-marm"],
2497 copts = [
2498 "-UNDEBUG",
2499 "-DXNN_TEST_MODE=1",
2500 ],
2501 gcc_copts = xnnpack_gcc_std_copts(),
2502 msvc_copts = xnnpack_msvc_std_copts(),
2503 deps = [
2504 ":tables",
2505 "@FP16",
2506 "@FXdiv",
2507 "@pthreadpool",
2508 ],
2509)
2510
2511xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002512 name = "wasm_ukernels",
2513 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002514 gcc_copts = xnnpack_gcc_std_copts(),
2515 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002516 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002517 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002518 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002519 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002520 "@FP16",
2521 "@FXdiv",
2522 "@pthreadpool",
2523 ],
2524)
2525
2526xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002527 name = "wasm_ukernels_test_mode",
2528 hdrs = INTERNAL_HDRS,
2529 copts = [
2530 "-UNDEBUG",
2531 "-DXNN_TEST_MODE=1",
2532 ],
2533 gcc_copts = xnnpack_gcc_std_copts(),
2534 msvc_copts = xnnpack_msvc_std_copts(),
2535 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002536 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002537 deps = [
2538 ":tables",
2539 "@FP16",
2540 "@FXdiv",
2541 "@pthreadpool",
2542 ],
2543)
2544
2545xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002546 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002547 hdrs = INTERNAL_HDRS,
2548 aarch32_copts = [
2549 "-marm",
2550 "-mfpu=neon",
2551 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002552 gcc_copts = xnnpack_gcc_std_copts(),
2553 gcc_x86_copts = ["-msse2"],
2554 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002555 optimized_copts = [
2556 "-O3",
2557 "-ffast-math",
2558 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002559 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002560 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002561 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562 "@FP16",
2563 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002564 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002565 ],
2566)
2567
2568xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002569 name = "psimd_fastmath_ukernels_test_mode",
2570 hdrs = INTERNAL_HDRS,
2571 aarch32_copts = [
2572 "-marm",
2573 "-mfpu=neon",
2574 ],
2575 copts = [
2576 "-UNDEBUG",
2577 "-DXNN_TEST_MODE=1",
2578 ],
2579 gcc_copts = xnnpack_gcc_std_copts(),
2580 gcc_x86_copts = ["-msse2"],
2581 msvc_copts = xnnpack_msvc_std_copts(),
2582 optimized_copts = [
2583 "-O3",
2584 "-ffast-math",
2585 ],
2586 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2587 deps = [
2588 ":tables",
2589 "@FP16",
2590 "@psimd",
2591 "@pthreadpool",
2592 ],
2593)
2594
2595xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002596 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002597 hdrs = INTERNAL_HDRS,
2598 aarch32_copts = [
2599 "-marm",
2600 "-mfpu=neon",
2601 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002602 gcc_copts = xnnpack_gcc_std_copts(),
2603 gcc_x86_copts = ["-msse2"],
2604 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002605 optimized_copts = [
2606 "-O3",
2607 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002608 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002609 deps = [
2610 ":tables",
2611 "@FP16",
2612 "@psimd",
2613 "@pthreadpool",
2614 ],
2615)
2616
2617xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002618 name = "psimd_accmath_ukernels_test_mode",
2619 hdrs = INTERNAL_HDRS,
2620 aarch32_copts = [
2621 "-marm",
2622 "-mfpu=neon",
2623 ],
2624 copts = [
2625 "-UNDEBUG",
2626 "-DXNN_TEST_MODE=1",
2627 ],
2628 gcc_copts = xnnpack_gcc_std_copts(),
2629 gcc_x86_copts = ["-msse2"],
2630 msvc_copts = xnnpack_msvc_std_copts(),
2631 optimized_copts = [
2632 "-O3",
2633 ],
2634 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2635 deps = [
2636 ":tables",
2637 "@FP16",
2638 "@psimd",
2639 "@pthreadpool",
2640 ],
2641)
2642
2643xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002644 name = "neon_ukernels",
2645 hdrs = INTERNAL_HDRS,
2646 aarch32_copts = [
2647 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002648 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002649 "-mfpu=neon",
2650 ],
2651 aarch32_srcs = NEON_UKERNELS,
2652 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002653 gcc_copts = xnnpack_gcc_std_copts(),
2654 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002655 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002656 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002657 "@FP16",
2658 "@pthreadpool",
2659 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002660)
2661
2662xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002663 name = "neon_ukernels_test_mode",
2664 hdrs = INTERNAL_HDRS,
2665 aarch32_copts = [
2666 "-marm",
2667 "-march=armv7-a",
2668 "-mfpu=neon",
2669 ],
2670 aarch32_srcs = NEON_UKERNELS,
2671 aarch64_srcs = NEON_UKERNELS,
2672 copts = [
2673 "-UNDEBUG",
2674 "-DXNN_TEST_MODE=1",
2675 ],
2676 gcc_copts = xnnpack_gcc_std_copts(),
2677 msvc_copts = xnnpack_msvc_std_copts(),
2678 deps = [
2679 ":tables",
2680 "@FP16",
2681 "@pthreadpool",
2682 ],
2683)
2684
2685xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686 name = "neonfma_ukernels",
2687 hdrs = INTERNAL_HDRS,
2688 aarch32_copts = [
2689 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002690 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002691 "-mfpu=neon-vfpv4",
2692 ],
2693 aarch32_srcs = NEONFMA_UKERNELS,
2694 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002695 apple_aarch32_copts = [
2696 "-mcpu=swift",
2697 "-mtune=generic",
2698 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002699 gcc_copts = xnnpack_gcc_std_copts(),
2700 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002701 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002702 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002703 "@FP16",
2704 "@pthreadpool",
2705 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002706)
2707
2708xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002709 name = "neonfma_ukernels_test_mode",
2710 hdrs = INTERNAL_HDRS,
2711 aarch32_copts = [
2712 "-marm",
2713 "-march=armv7-a",
2714 "-mfpu=neon-vfpv4",
2715 ],
2716 aarch32_srcs = NEONFMA_UKERNELS,
2717 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002718 apple_aarch32_copts = [
2719 "-mcpu=swift",
2720 "-mtune=generic",
2721 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002722 copts = [
2723 "-UNDEBUG",
2724 "-DXNN_TEST_MODE=1",
2725 ],
2726 gcc_copts = xnnpack_gcc_std_copts(),
2727 msvc_copts = xnnpack_msvc_std_copts(),
2728 deps = [
2729 ":tables",
2730 "@FP16",
2731 "@pthreadpool",
2732 ],
2733)
2734
2735xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002736 name = "neonv8_ukernels",
2737 hdrs = INTERNAL_HDRS,
2738 aarch32_copts = [
2739 "-marm",
2740 "-march=armv8-a",
2741 "-mfpu=neon-fp-armv8",
2742 ],
2743 aarch32_srcs = NEONV8_UKERNELS,
2744 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002745 apple_aarch32_copts = [
2746 "-mcpu=cyclone",
2747 "-mtune=generic",
2748 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002749 gcc_copts = xnnpack_gcc_std_copts(),
2750 msvc_copts = xnnpack_msvc_std_copts(),
2751 deps = [
2752 ":tables",
2753 "@FP16",
2754 "@pthreadpool",
2755 ],
2756)
2757
2758xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002759 name = "neonv8_ukernels_test_mode",
2760 hdrs = INTERNAL_HDRS,
2761 aarch32_copts = [
2762 "-marm",
2763 "-march=armv8-a",
2764 "-mfpu=neon-fp-armv8",
2765 ],
2766 aarch32_srcs = NEONV8_UKERNELS,
2767 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002768 apple_aarch32_copts = [
2769 "-mcpu=cyclone",
2770 "-mtune=generic",
2771 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002772 copts = [
2773 "-UNDEBUG",
2774 "-DXNN_TEST_MODE=1",
2775 ],
2776 gcc_copts = xnnpack_gcc_std_copts(),
2777 msvc_copts = xnnpack_msvc_std_copts(),
2778 deps = [
2779 ":tables",
2780 "@FP16",
2781 "@pthreadpool",
2782 ],
2783)
2784
2785xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002786 name = "neonfp16arith_ukernels",
2787 hdrs = INTERNAL_HDRS,
2788 aarch64_copts = ["-march=armv8.2-a+fp16"],
2789 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002790 gcc_copts = xnnpack_gcc_std_copts(),
2791 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002792 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002793 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002794 "@FP16",
2795 "@pthreadpool",
2796 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002797)
2798
2799xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002800 name = "neonfp16arith_ukernels_test_mode",
2801 hdrs = INTERNAL_HDRS,
2802 aarch64_copts = ["-march=armv8.2-a+fp16"],
2803 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2804 copts = [
2805 "-UNDEBUG",
2806 "-DXNN_TEST_MODE=1",
2807 ],
2808 gcc_copts = xnnpack_gcc_std_copts(),
2809 msvc_copts = xnnpack_msvc_std_copts(),
2810 deps = [
2811 ":tables",
2812 "@FP16",
2813 "@pthreadpool",
2814 ],
2815)
2816
2817xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002818 name = "sse2_ukernels",
2819 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002820 gcc_copts = xnnpack_gcc_std_copts(),
2821 gcc_x86_copts = ["-msse2"],
2822 msvc_copts = xnnpack_msvc_std_copts(),
2823 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002824 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002825 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002826 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002827 "@FP16",
2828 "@pthreadpool",
2829 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002830)
2831
2832xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002833 name = "sse2_ukernels_test_mode",
2834 hdrs = INTERNAL_HDRS,
2835 copts = [
2836 "-UNDEBUG",
2837 "-DXNN_TEST_MODE=1",
2838 ],
2839 gcc_copts = xnnpack_gcc_std_copts(),
2840 gcc_x86_copts = ["-msse2"],
2841 msvc_copts = xnnpack_msvc_std_copts(),
2842 msvc_x86_32_copts = ["/arch:SSE2"],
2843 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2844 deps = [
2845 ":tables",
2846 "@FP16",
2847 "@pthreadpool",
2848 ],
2849)
2850
2851xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002852 name = "ssse3_ukernels",
2853 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002854 gcc_copts = xnnpack_gcc_std_copts(),
2855 gcc_x86_copts = ["-mssse3"],
2856 msvc_copts = xnnpack_msvc_std_copts(),
2857 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002858 x86_srcs = SSSE3_UKERNELS,
2859 deps = [
2860 ":tables",
2861 "@FP16",
2862 "@pthreadpool",
2863 ],
2864)
2865
2866xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002867 name = "ssse3_ukernels_test_mode",
2868 hdrs = INTERNAL_HDRS,
2869 copts = [
2870 "-UNDEBUG",
2871 "-DXNN_TEST_MODE=1",
2872 ],
2873 gcc_copts = xnnpack_gcc_std_copts(),
2874 gcc_x86_copts = ["-mssse3"],
2875 msvc_copts = xnnpack_msvc_std_copts(),
2876 msvc_x86_32_copts = ["/arch:SSE2"],
2877 x86_srcs = SSSE3_UKERNELS,
2878 deps = [
2879 ":tables",
2880 "@FP16",
2881 "@pthreadpool",
2882 ],
2883)
2884
2885xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002886 name = "sse41_ukernels",
2887 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002888 gcc_copts = xnnpack_gcc_std_copts(),
2889 gcc_x86_copts = ["-msse4.1"],
2890 msvc_copts = xnnpack_msvc_std_copts(),
2891 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002892 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002893 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002894 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002895 "@FP16",
2896 "@pthreadpool",
2897 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002898)
2899
2900xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002901 name = "sse41_ukernels_test_mode",
2902 hdrs = INTERNAL_HDRS,
2903 copts = [
2904 "-UNDEBUG",
2905 "-DXNN_TEST_MODE=1",
2906 ],
2907 gcc_copts = xnnpack_gcc_std_copts(),
2908 gcc_x86_copts = ["-msse4.1"],
2909 msvc_copts = xnnpack_msvc_std_copts(),
2910 msvc_x86_32_copts = ["/arch:SSE2"],
2911 x86_srcs = SSE41_UKERNELS,
2912 deps = [
2913 ":tables",
2914 "@FP16",
2915 "@pthreadpool",
2916 ],
2917)
2918
2919xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002920 name = "avx_ukernels",
2921 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002922 gcc_copts = xnnpack_gcc_std_copts(),
2923 gcc_x86_copts = ["-mavx"],
2924 msvc_copts = xnnpack_msvc_std_copts(),
2925 msvc_x86_32_copts = ["/arch:AVX"],
2926 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002927 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002928 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002929 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002930 "@FP16",
2931 "@pthreadpool",
2932 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002933)
2934
2935xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002936 name = "avx_ukernels_test_mode",
2937 hdrs = INTERNAL_HDRS,
2938 copts = [
2939 "-UNDEBUG",
2940 "-DXNN_TEST_MODE=1",
2941 ],
2942 gcc_copts = xnnpack_gcc_std_copts(),
2943 gcc_x86_copts = ["-mavx"],
2944 msvc_copts = xnnpack_msvc_std_copts(),
2945 msvc_x86_32_copts = ["/arch:AVX"],
2946 msvc_x86_64_copts = ["/arch:AVX"],
2947 x86_srcs = AVX_UKERNELS,
2948 deps = [
2949 ":tables",
2950 "@FP16",
2951 "@pthreadpool",
2952 ],
2953)
2954
2955xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07002956 name = "xop_ukernels",
2957 hdrs = INTERNAL_HDRS,
2958 gcc_copts = xnnpack_gcc_std_copts(),
2959 gcc_x86_copts = ["-mxop"],
2960 msvc_copts = xnnpack_msvc_std_copts(),
2961 msvc_x86_32_copts = ["/arch:AVX"],
2962 msvc_x86_64_copts = ["/arch:AVX"],
2963 x86_srcs = XOP_UKERNELS,
2964 deps = [
2965 ":tables",
2966 "@FP16",
2967 "@pthreadpool",
2968 ],
2969)
2970
2971xnnpack_cc_library(
2972 name = "xop_ukernels_test_mode",
2973 hdrs = INTERNAL_HDRS,
2974 copts = [
2975 "-UNDEBUG",
2976 "-DXNN_TEST_MODE=1",
2977 ],
2978 gcc_copts = xnnpack_gcc_std_copts(),
2979 gcc_x86_copts = ["-mxop"],
2980 msvc_copts = xnnpack_msvc_std_copts(),
2981 msvc_x86_32_copts = ["/arch:AVX"],
2982 msvc_x86_64_copts = ["/arch:AVX"],
2983 x86_srcs = XOP_UKERNELS,
2984 deps = [
2985 ":tables",
2986 "@FP16",
2987 "@pthreadpool",
2988 ],
2989)
2990
2991xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002992 name = "fma3_ukernels",
2993 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002994 gcc_copts = xnnpack_gcc_std_copts(),
2995 gcc_x86_copts = ["-mfma"],
2996 msvc_copts = xnnpack_msvc_std_copts(),
2997 msvc_x86_32_copts = ["/arch:AVX"],
2998 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002999 x86_srcs = FMA3_UKERNELS,
3000 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003001 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003002 "@FP16",
3003 "@pthreadpool",
3004 ],
3005)
3006
3007xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003008 name = "fma3_ukernels_test_mode",
3009 hdrs = INTERNAL_HDRS,
3010 copts = [
3011 "-UNDEBUG",
3012 "-DXNN_TEST_MODE=1",
3013 ],
3014 gcc_copts = xnnpack_gcc_std_copts(),
3015 gcc_x86_copts = ["-mfma"],
3016 msvc_copts = xnnpack_msvc_std_copts(),
3017 msvc_x86_32_copts = ["/arch:AVX"],
3018 msvc_x86_64_copts = ["/arch:AVX"],
3019 x86_srcs = FMA3_UKERNELS,
3020 deps = [
3021 ":tables",
3022 "@FP16",
3023 "@pthreadpool",
3024 ],
3025)
3026
3027xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003028 name = "avx2_ukernels",
3029 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003030 gcc_copts = xnnpack_gcc_std_copts(),
3031 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003032 "-mfma",
3033 "-mavx2",
3034 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003035 msvc_copts = xnnpack_msvc_std_copts(),
3036 msvc_x86_32_copts = ["/arch:AVX2"],
3037 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003038 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003039 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003040 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003041 "@FP16",
3042 "@pthreadpool",
3043 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003044)
3045
3046xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003047 name = "avx2_ukernels_test_mode",
3048 hdrs = INTERNAL_HDRS,
3049 copts = [
3050 "-UNDEBUG",
3051 "-DXNN_TEST_MODE=1",
3052 ],
3053 gcc_copts = xnnpack_gcc_std_copts(),
3054 gcc_x86_copts = [
3055 "-mfma",
3056 "-mavx2",
3057 ],
3058 msvc_copts = xnnpack_msvc_std_copts(),
3059 msvc_x86_32_copts = ["/arch:AVX2"],
3060 msvc_x86_64_copts = ["/arch:AVX2"],
3061 x86_srcs = AVX2_UKERNELS,
3062 deps = [
3063 ":tables",
3064 "@FP16",
3065 "@pthreadpool",
3066 ],
3067)
3068
3069xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003070 name = "avx512f_ukernels",
3071 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003072 gcc_copts = xnnpack_gcc_std_copts(),
3073 gcc_x86_copts = ["-mavx512f"],
3074 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3075 msvc_copts = xnnpack_msvc_std_copts(),
3076 msvc_x86_32_copts = ["/arch:AVX512"],
3077 msvc_x86_64_copts = ["/arch:AVX512"],
3078 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003079 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003080 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003081 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003082 "@FP16",
3083 "@pthreadpool",
3084 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003085)
3086
3087xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003088 name = "avx512f_ukernels_test_mode",
3089 hdrs = INTERNAL_HDRS,
3090 copts = [
3091 "-UNDEBUG",
3092 "-DXNN_TEST_MODE=1",
3093 ],
3094 gcc_copts = xnnpack_gcc_std_copts(),
3095 gcc_x86_copts = ["-mavx512f"],
3096 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3097 msvc_copts = xnnpack_msvc_std_copts(),
3098 msvc_x86_32_copts = ["/arch:AVX512"],
3099 msvc_x86_64_copts = ["/arch:AVX512"],
3100 msys_copts = ["-fno-asynchronous-unwind-tables"],
3101 x86_srcs = AVX512F_UKERNELS,
3102 deps = [
3103 ":tables",
3104 "@FP16",
3105 "@pthreadpool",
3106 ],
3107)
3108
3109xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003110 name = "asm_ukernels",
3111 hdrs = ["src/xnnpack/assembly.h"],
3112 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003113 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114 aarch64_srcs = AARCH64_ASM_UKERNELS,
3115)
3116
Marat Dukhan3b59de22020-06-03 20:15:19 -07003117xnnpack_cc_library(
3118 name = "logging_utils",
3119 srcs = LOGGING_SRCS,
3120 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3121 copts = LOGGING_COPTS + [
3122 "-Isrc",
3123 "-Iinclude",
3124 ] + select({
3125 ":debug_build": [],
3126 "//conditions:default": xnnpack_min_size_copts(),
3127 }),
3128 gcc_copts = xnnpack_gcc_std_copts(),
3129 msvc_copts = xnnpack_msvc_std_copts(),
3130 visibility = xnnpack_visibility(),
3131 deps = [
3132 "@FP16",
3133 "@clog",
3134 "@pthreadpool",
3135 ],
3136)
3137
Marat Dukhan08c4a432019-10-03 09:29:21 -07003138xnnpack_aggregate_library(
3139 name = "ukernels",
3140 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003141 ":neon_ukernels",
3142 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003143 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003144 ":asm_ukernels",
3145 ],
3146 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003147 ":neon_ukernels",
3148 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003149 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 ":neonfp16arith_ukernels",
3151 ":asm_ukernels",
3152 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003153 generic_deps = [
3154 ":scalar_ukernels",
3155 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003156 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003157 ":psimd_fastmath_ukernels",
3158 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003159 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003160 wasm_deps = [
3161 ":wasm_ukernels",
3162 ],
3163 wasmsimd_deps = [
3164 ":wasm_ukernels",
3165 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003168 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003169 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003171 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003172 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003173 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003174 ":avx512f_ukernels",
3175 ],
3176)
3177
Marat Dukhan33fcf782020-05-24 14:27:15 -07003178xnnpack_aggregate_library(
3179 name = "ukernels_test_mode",
3180 aarch32_deps = [
3181 ":neon_ukernels_test_mode",
3182 ":neonfma_ukernels_test_mode",
3183 ":neonv8_ukernels_test_mode",
3184 ":asm_ukernels",
3185 ],
3186 aarch64_deps = [
3187 ":neon_ukernels_test_mode",
3188 ":neonfma_ukernels_test_mode",
3189 ":neonv8_ukernels_test_mode",
3190 ":neonfp16arith_ukernels_test_mode",
3191 ":asm_ukernels",
3192 ],
3193 generic_deps = [
3194 ":scalar_ukernels_test_mode",
3195 ],
3196 psimd_deps = [
3197 ":psimd_fastmath_ukernels_test_mode",
3198 ":psimd_accmath_ukernels_test_mode",
3199 ],
3200 wasm_deps = [
3201 ":wasm_ukernels_test_mode",
3202 ],
3203 wasmsimd_deps = [
3204 ":wasm_ukernels_test_mode",
3205 ],
3206 x86_deps = [
3207 ":sse2_ukernels_test_mode",
3208 ":ssse3_ukernels_test_mode",
3209 ":sse41_ukernels_test_mode",
3210 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003211 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003212 ":fma3_ukernels_test_mode",
3213 ":avx2_ukernels_test_mode",
3214 ":avx512f_ukernels_test_mode",
3215 ],
3216)
3217
Marat Dukhan86df3842020-06-17 14:03:31 -07003218# This target must be compatible with Android NDK r17 and compile with Clang 6.
3219# It is similar to :ukernels target, but excludes microkernels which require
3220# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003221xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003222 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003223 aarch32_deps = [
3224 ":neon_ukernels",
3225 ":neonfma_ukernels",
3226 ":neonv8_ukernels",
3227 ":asm_ukernels",
3228 ],
3229 aarch64_deps = [
3230 ":neon_ukernels",
3231 ":neonfma_ukernels",
3232 ":neonv8_ukernels",
3233 ":asm_ukernels",
3234 ],
3235 generic_deps = [
3236 ":scalar_ukernels",
3237 ],
3238 psimd_deps = [
3239 ":psimd_fastmath_ukernels",
3240 ":psimd_accmath_ukernels",
3241 ],
3242 wasm_deps = [
3243 ":wasm_ukernels",
3244 ],
3245 wasmsimd_deps = [
3246 ":wasm_ukernels",
3247 ],
3248 x86_deps = [
3249 ":sse2_ukernels",
3250 ":ssse3_ukernels",
3251 ":sse41_ukernels",
3252 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003253 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003254 ":fma3_ukernels",
3255 ":avx2_ukernels",
3256 ":avx512f_ukernels",
3257 ],
3258)
3259
Marat Dukhan08c4a432019-10-03 09:29:21 -07003260xnnpack_cc_library(
3261 name = "im2col",
3262 srcs = ["src/im2col.c"],
3263 hdrs = [
3264 "src/xnnpack/common.h",
3265 "src/xnnpack/im2col.h",
3266 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003267 gcc_copts = xnnpack_gcc_std_copts(),
3268 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003269)
3270
3271xnnpack_cc_library(
3272 name = "indirection",
3273 srcs = ["src/indirection.c"],
3274 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003275 gcc_copts = xnnpack_gcc_std_copts(),
3276 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003277 deps = [
3278 "@FP16",
3279 "@FXdiv",
3280 "@pthreadpool",
3281 ],
3282)
3283
3284xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003285 name = "indirection_test_mode",
3286 srcs = ["src/indirection.c"],
3287 hdrs = INTERNAL_HDRS,
3288 copts = [
3289 "-UNDEBUG",
3290 "-DXNN_TEST_MODE=1",
3291 ],
3292 gcc_copts = xnnpack_gcc_std_copts(),
3293 msvc_copts = xnnpack_msvc_std_copts(),
3294 deps = [
3295 "@FP16",
3296 "@FXdiv",
3297 "@pthreadpool",
3298 ],
3299)
3300
3301xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003302 name = "packing",
3303 srcs = ["src/packing.c"],
3304 hdrs = INTERNAL_HDRS,
3305 gcc_copts = xnnpack_gcc_std_copts(),
3306 msvc_copts = xnnpack_msvc_std_copts(),
3307 deps = [
3308 "@FP16",
3309 "@FXdiv",
3310 "@pthreadpool",
3311 ],
3312)
3313
3314xnnpack_cc_library(
3315 name = "packing_test_mode",
3316 srcs = ["src/packing.c"],
3317 hdrs = INTERNAL_HDRS,
3318 copts = [
3319 "-UNDEBUG",
3320 "-DXNN_TEST_MODE=1",
3321 ],
3322 gcc_copts = xnnpack_gcc_std_copts(),
3323 msvc_copts = xnnpack_msvc_std_copts(),
3324 deps = [
3325 "@FP16",
3326 "@FXdiv",
3327 "@pthreadpool",
3328 ],
3329)
3330
3331xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003332 name = "operator_run",
3333 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003334 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003335 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003336 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3337 "//conditions:default": [],
3338 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003339 gcc_copts = xnnpack_gcc_std_copts(),
3340 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003341 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003342 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003343 "@FP16",
3344 "@FXdiv",
3345 "@clog",
3346 "@pthreadpool",
3347 ],
3348)
3349
Chao Mei6ddfc602020-05-13 22:29:36 -07003350xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003351 name = "operator_run_test_mode",
3352 srcs = ["src/operator-run.c"],
3353 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3354 copts = LOGGING_COPTS + [
3355 "-UNDEBUG",
3356 "-DXNN_TEST_MODE=1",
3357 ] + select({
3358 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3359 "//conditions:default": [],
3360 }),
3361 gcc_copts = xnnpack_gcc_std_copts(),
3362 msvc_copts = xnnpack_msvc_std_copts(),
3363 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003364 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003365 "@FP16",
3366 "@FXdiv",
3367 "@clog",
3368 "@pthreadpool",
3369 ],
3370)
3371
3372xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003373 name = "memory_planner",
3374 srcs = ["src/memory-planner.c"],
3375 hdrs = INTERNAL_HDRS,
3376 defines = select({
3377 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3378 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3379 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3380 }),
3381 gcc_copts = xnnpack_gcc_std_copts(),
3382 msvc_copts = xnnpack_msvc_std_copts(),
3383 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003384 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003385 "@pthreadpool",
3386 ],
3387)
3388
Marat Dukhan33fcf782020-05-24 14:27:15 -07003389xnnpack_cc_library(
3390 name = "memory_planner_test_mode",
3391 srcs = ["src/memory-planner.c"],
3392 hdrs = INTERNAL_HDRS,
3393 copts = [
3394 "-UNDEBUG",
3395 "-DXNN_TEST_MODE=1",
3396 ],
3397 defines = select({
3398 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3399 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3400 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3401 }),
3402 gcc_copts = xnnpack_gcc_std_copts(),
3403 msvc_copts = xnnpack_msvc_std_copts(),
3404 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003405 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003406 "@pthreadpool",
3407 ],
3408)
3409
Marat Dukhan08c4a432019-10-03 09:29:21 -07003410cc_library(
3411 name = "enable_assembly",
3412 defines = select({
3413 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3414 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003415 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003416 }),
3417)
3418
Marat Dukhan9de90e02020-06-18 16:04:12 -07003419cc_library(
3420 name = "enable_sparse",
3421 defines = select({
3422 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3423 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003424 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003425 }),
3426)
3427
Marat Dukhancf056b22019-10-07 10:26:29 -07003428xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003429 name = "operators",
3430 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003431 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003432 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003433 ],
3434 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003435 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436 "-Isrc",
3437 "-Iinclude",
3438 ] + select({
3439 ":debug_build": [],
3440 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003441 }) + select({
3442 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3443 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003444 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003445 gcc_copts = xnnpack_gcc_std_copts(),
3446 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003447 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003449 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003450 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003451 "@FP16",
3452 "@FXdiv",
3453 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003454 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003455 ],
3456)
3457
Marat Dukhan10a38082020-04-17 03:58:35 -07003458xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003459 name = "operators_test_mode",
3460 srcs = OPERATOR_SRCS + [
3461 "src/memory.c",
3462 "src/operator-delete.c",
3463 ],
3464 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3465 copts = LOGGING_COPTS + [
3466 "-Isrc",
3467 "-Iinclude",
3468 "-UNDEBUG",
3469 "-DXNN_TEST_MODE=1",
3470 ] + select({
3471 ":debug_build": [],
3472 "//conditions:default": xnnpack_min_size_copts(),
3473 }) + select({
3474 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3475 "//conditions:default": [],
3476 }),
3477 gcc_copts = xnnpack_gcc_std_copts(),
3478 msvc_copts = xnnpack_msvc_std_copts(),
3479 deps = [
3480 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003481 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003482 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003483 "@FP16",
3484 "@FXdiv",
3485 "@clog",
3486 "@pthreadpool",
3487 ],
3488)
3489
3490xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003491 name = "XNNPACK",
3492 srcs = [
3493 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003494 "src/runtime.c",
3495 "src/subgraph.c",
3496 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003497 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003498 hdrs = ["include/xnnpack.h"],
3499 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003500 "-Isrc",
3501 "-Iinclude",
3502 ] + select({
3503 ":debug_build": [],
3504 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003505 }) + select({
3506 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3507 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003508 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003509 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003510 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003511 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003512 visibility = xnnpack_visibility(),
3513 deps = [
3514 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003515 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003516 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003517 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003518 ":operator_run",
3519 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003520 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003521 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003522 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003523 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003524 ] + select({
3525 ":emscripten": [],
3526 "//conditions:default": ["@cpuinfo"],
3527 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003528)
3529
Marat Dukhan10a38082020-04-17 03:58:35 -07003530xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003531 name = "XNNPACK_test_mode",
3532 srcs = [
3533 "src/init.c",
3534 "src/runtime.c",
3535 "src/subgraph.c",
3536 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003537 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003538 hdrs = ["include/xnnpack.h"],
3539 copts = LOGGING_COPTS + [
3540 "-Isrc",
3541 "-Iinclude",
3542 "-UNDEBUG",
3543 "-DXNN_TEST_MODE=1",
3544 ] + select({
3545 ":debug_build": [],
3546 "//conditions:default": xnnpack_min_size_copts(),
3547 }) + select({
3548 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3549 "//conditions:default": [],
3550 }),
3551 gcc_copts = xnnpack_gcc_std_copts(),
3552 includes = ["include"],
3553 msvc_copts = xnnpack_msvc_std_copts(),
3554 visibility = xnnpack_visibility(),
3555 deps = [
3556 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003557 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003558 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003559 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003560 ":operator_run_test_mode",
3561 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003562 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003563 "@clog",
3564 "@FP16",
3565 "@pthreadpool",
3566 ] + select({
3567 ":emscripten": [],
3568 "//conditions:default": ["@cpuinfo"],
3569 }),
3570)
3571
3572xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003573 name = "xnnpack_f32",
3574 srcs = [
3575 "src/init.c",
3576 "src/runtime.c",
3577 "src/subgraph.c",
3578 "src/tensor.c",
3579 ] + SUBGRAPH_SRCS,
3580 hdrs = ["include/xnnpack.h"],
3581 copts = LOGGING_COPTS + [
3582 "-Isrc",
3583 "-Iinclude",
3584 ] + select({
3585 ":debug_build": [],
3586 "//conditions:default": xnnpack_min_size_copts(),
3587 }) + select({
3588 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3589 "//conditions:default": [],
3590 }),
3591 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003592 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003593 "XNN_NO_U8_OPERATORS",
3594 "XNN_NO_X8_OPERATORS",
3595 "XNN_NO_F16_OPERATORS",
3596 "XNN_NO_X16_OPERATORS",
3597 ],
3598 gcc_copts = xnnpack_gcc_std_copts(),
3599 includes = ["include"],
3600 msvc_copts = xnnpack_msvc_std_copts(),
3601 visibility = xnnpack_visibility(),
3602 deps = [
3603 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003604 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003605 ":logging_utils",
3606 ":memory_planner",
3607 ":operator_run",
3608 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003609 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003610 "@clog",
3611 "@FP16",
3612 "@pthreadpool",
3613 ] + select({
3614 ":emscripten": [],
3615 "//conditions:default": ["@cpuinfo"],
3616 }),
3617)
3618
3619xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003620 name = "xnnpack_operators_nhwc_f32",
3621 srcs = [
3622 "src/init.c",
3623 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003624 hdrs = ["include/xnnpack.h"],
3625 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003626 "-Isrc",
3627 "-Iinclude",
3628 ] + select({
3629 ":debug_build": [],
3630 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003631 }) + select({
3632 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3633 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003634 }),
3635 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003636 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003637 "XNN_NO_U8_OPERATORS",
3638 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003639 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003640 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003641 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003643 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644 visibility = xnnpack_visibility(),
3645 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003646 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003647 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648 ":operator_run",
3649 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003650 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003651 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003653 ] + select({
3654 ":emscripten": [],
3655 "//conditions:default": ["@cpuinfo"],
3656 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003657)
3658
Marat Dukhancf056b22019-10-07 10:26:29 -07003659xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660 name = "bench_utils",
3661 srcs = ["bench/utils.cc"],
3662 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003663 deps = [
3664 "@com_google_benchmark//:benchmark",
3665 "@cpuinfo",
3666 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003667)
3668
Frank Barchard7e955972019-10-11 10:34:25 -07003669######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003670
3671xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003672 name = "qs8_gemm_bench",
3673 srcs = [
3674 "bench/gemm.h",
3675 "bench/qs8-gemm.cc",
3676 "src/xnnpack/AlignedAllocator.h",
3677 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3678 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3679)
3680
3681xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003682 name = "qs8_requantization_bench",
3683 srcs = [
3684 "bench/qs8-requantization.cc",
3685 "src/xnnpack/requantization-stubs.h",
3686 "src/xnnpack/AlignedAllocator.h",
3687 ] + MICROKERNEL_BENCHMARK_HDRS,
3688 deps = MICROKERNEL_BENCHMARK_DEPS,
3689)
3690
3691xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003692 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003693 srcs = [
3694 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003695 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 "src/xnnpack/AlignedAllocator.h",
3697 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003698 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003699 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003700)
3701
3702xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003703 name = "qu8_requantization_bench",
3704 srcs = [
3705 "bench/qu8-requantization.cc",
3706 "src/xnnpack/requantization-stubs.h",
3707 "src/xnnpack/AlignedAllocator.h",
3708 ] + MICROKERNEL_BENCHMARK_HDRS,
3709 deps = MICROKERNEL_BENCHMARK_DEPS,
3710)
3711
3712xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003713 name = "f16_igemm_bench",
3714 srcs = [
3715 "bench/f16-igemm.cc",
3716 "bench/conv.h",
3717 "bench/google/conv.h",
3718 "src/xnnpack/AlignedAllocator.h",
3719 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003720 deps = MICROKERNEL_BENCHMARK_DEPS + [
3721 ":indirection",
3722 ":packing",
3723 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003724)
3725
3726xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003727 name = "f16_gemm_bench",
3728 srcs = [
3729 "bench/f16-gemm.cc",
3730 "bench/gemm.h",
3731 "src/xnnpack/AlignedAllocator.h",
3732 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003733 deps = MICROKERNEL_BENCHMARK_DEPS + [
3734 ":packing",
3735 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003736)
3737
3738xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003739 name = "f16_spmm_bench",
3740 srcs = [
3741 "bench/f16-spmm.cc",
3742 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003743 "src/xnnpack/AlignedAllocator.h",
3744 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003745 deps = MICROKERNEL_BENCHMARK_DEPS,
3746)
3747
3748xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003749 name = "f32_igemm_bench",
3750 srcs = [
3751 "bench/f32-igemm.cc",
3752 "bench/conv.h",
3753 "src/xnnpack/AlignedAllocator.h",
3754 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003755 deps = MICROKERNEL_BENCHMARK_DEPS + [
3756 ":indirection",
3757 ":packing",
3758 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759)
3760
3761xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003762 name = "f16_relu_bench",
3763 srcs = [
3764 "bench/f16-relu.cc",
3765 "src/xnnpack/AlignedAllocator.h",
3766 ] + MICROKERNEL_BENCHMARK_HDRS,
3767 deps = MICROKERNEL_BENCHMARK_DEPS,
3768)
3769
3770xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003771 name = "f32_conv_hwc_bench",
3772 srcs = [
3773 "bench/f32-conv-hwc.cc",
3774 "bench/dconv.h",
3775 "src/xnnpack/AlignedAllocator.h",
3776 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003777 deps = MICROKERNEL_BENCHMARK_DEPS + [
3778 ":packing",
3779 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003780)
3781
3782xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003783 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003784 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003785 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003786 "bench/dconv.h",
3787 "src/xnnpack/AlignedAllocator.h",
3788 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003789 deps = MICROKERNEL_BENCHMARK_DEPS + [
3790 ":packing",
3791 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003792)
3793
3794xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003795 name = "f16_dwconv_bench",
3796 srcs = [
3797 "bench/f16-dwconv.cc",
3798 "bench/dwconv.h",
3799 "bench/google/dwconv.h",
3800 "src/xnnpack/AlignedAllocator.h",
3801 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003802 deps = MICROKERNEL_BENCHMARK_DEPS + [
3803 ":indirection",
3804 ":packing",
3805 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003806)
3807
3808xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003809 name = "f32_dwconv_bench",
3810 srcs = [
3811 "bench/f32-dwconv.cc",
3812 "bench/dwconv.h",
3813 "src/xnnpack/AlignedAllocator.h",
3814 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003815 deps = MICROKERNEL_BENCHMARK_DEPS + [
3816 ":indirection",
3817 ":packing",
3818 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819)
3820
3821xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003822 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003823 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003824 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003825 "bench/dwconv.h",
3826 "src/xnnpack/AlignedAllocator.h",
3827 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003828 deps = MICROKERNEL_BENCHMARK_DEPS + [
3829 ":indirection",
3830 ":packing",
3831 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003832)
3833
3834xnnpack_benchmark(
3835 name = "f32_gemm_bench",
3836 srcs = [
3837 "bench/f32-gemm.cc",
3838 "bench/gemm.h",
3839 "src/xnnpack/AlignedAllocator.h",
3840 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003841 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003842 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003843)
3844
3845xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003846 name = "f32_hswish_bench",
3847 srcs = [
3848 "bench/f32-hswish.cc",
3849 "src/xnnpack/AlignedAllocator.h",
3850 ] + MICROKERNEL_BENCHMARK_HDRS,
3851 deps = MICROKERNEL_BENCHMARK_DEPS,
3852)
3853
3854xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003855 name = "f32_raddexpminusmax_bench",
3856 srcs = [
3857 "bench/f32-raddexpminusmax.cc",
3858 "src/xnnpack/AlignedAllocator.h",
3859 ] + MICROKERNEL_BENCHMARK_HDRS,
3860 deps = MICROKERNEL_BENCHMARK_DEPS,
3861)
3862
3863xnnpack_benchmark(
3864 name = "f32_raddextexp_bench",
3865 srcs = [
3866 "bench/f32-raddextexp.cc",
3867 "src/xnnpack/AlignedAllocator.h",
3868 ] + MICROKERNEL_BENCHMARK_HDRS,
3869 deps = MICROKERNEL_BENCHMARK_DEPS,
3870)
3871
3872xnnpack_benchmark(
3873 name = "f32_raddstoreexpminusmax_bench",
3874 srcs = [
3875 "bench/f32-raddstoreexpminusmax.cc",
3876 "src/xnnpack/AlignedAllocator.h",
3877 ] + MICROKERNEL_BENCHMARK_HDRS,
3878 deps = MICROKERNEL_BENCHMARK_DEPS,
3879)
3880
3881xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003882 name = "f32_relu_bench",
3883 srcs = [
3884 "bench/f32-relu.cc",
3885 "src/xnnpack/AlignedAllocator.h",
3886 ] + MICROKERNEL_BENCHMARK_HDRS,
3887 deps = MICROKERNEL_BENCHMARK_DEPS,
3888)
3889
3890xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003891 name = "f32_rmax_bench",
3892 srcs = [
3893 "bench/f32-rmax.cc",
3894 "src/xnnpack/AlignedAllocator.h",
3895 ] + MICROKERNEL_BENCHMARK_HDRS,
3896 deps = MICROKERNEL_BENCHMARK_DEPS,
3897)
3898
3899xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003900 name = "f32_sigmoid_bench",
3901 srcs = [
3902 "bench/f32-sigmoid.cc",
3903 "src/xnnpack/AlignedAllocator.h",
3904 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003905 deps = MICROKERNEL_BENCHMARK_DEPS,
3906)
3907
3908xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909 name = "f32_spmm_bench",
3910 srcs = [
3911 "bench/f32-spmm.cc",
3912 "bench/gemm.h",
3913 "src/xnnpack/AlignedAllocator.h",
3914 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003915 deps = MICROKERNEL_BENCHMARK_DEPS,
3916)
3917
3918xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003919 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003920 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003921 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003922 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003923 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003924 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003925)
3926
3927xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003928 name = "f32_vscaleexpminusmax_bench",
3929 srcs = [
3930 "bench/f32-vscaleexpminusmax.cc",
3931 "src/xnnpack/AlignedAllocator.h",
3932 ] + MICROKERNEL_BENCHMARK_HDRS,
3933 deps = MICROKERNEL_BENCHMARK_DEPS,
3934)
3935
3936xnnpack_benchmark(
3937 name = "f32_vscaleextexp_bench",
3938 srcs = [
3939 "bench/f32-vscaleextexp.cc",
3940 "src/xnnpack/AlignedAllocator.h",
3941 ] + MICROKERNEL_BENCHMARK_HDRS,
3942 deps = MICROKERNEL_BENCHMARK_DEPS,
3943)
3944
3945xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003946 name = "f32_vsqrt_bench",
3947 srcs = [
3948 "bench/f32-vsqrt.cc",
3949 "src/xnnpack/AlignedAllocator.h",
3950 ] + MICROKERNEL_BENCHMARK_HDRS,
3951 deps = MICROKERNEL_BENCHMARK_DEPS,
3952)
3953
3954xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003955 name = "f32_im2col_gemm_bench",
3956 srcs = [
3957 "bench/f32-im2col-gemm.cc",
3958 "bench/conv.h",
3959 "src/xnnpack/AlignedAllocator.h",
3960 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003961 deps = MICROKERNEL_BENCHMARK_DEPS + [
3962 ":im2col",
3963 ":packing",
3964 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003965)
3966
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003967xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003968 name = "rounding_bench",
3969 srcs = [
3970 "bench/rounding.cc",
3971 "src/xnnpack/math-stubs.h",
3972 "src/xnnpack/AlignedAllocator.h",
3973 ] + MICROKERNEL_BENCHMARK_HDRS,
3974 deps = MICROKERNEL_BENCHMARK_DEPS,
3975)
3976
Marat Dukhan08c4a432019-10-03 09:29:21 -07003977########################### Benchmarks for operators ###########################
3978
3979xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003980 name = "average_pooling_bench",
3981 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003982 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003983 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003984 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003985)
3986
3987xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003988 name = "bankers_rounding_bench",
3989 srcs = ["bench/bankers-rounding.cc"],
3990 copts = xnnpack_optional_tflite_copts(),
3991 tags = ["nowin32"],
3992 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3993)
3994
3995xnnpack_benchmark(
3996 name = "ceiling_bench",
3997 srcs = ["bench/ceiling.cc"],
3998 copts = xnnpack_optional_tflite_copts(),
3999 tags = ["nowin32"],
4000 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4001)
4002
4003xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004004 name = "channel_shuffle_bench",
4005 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004006 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004007)
4008
4009xnnpack_benchmark(
4010 name = "convolution_bench",
4011 srcs = ["bench/convolution.cc"],
4012 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004013 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004014 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004015)
4016
4017xnnpack_benchmark(
4018 name = "deconvolution_bench",
4019 srcs = ["bench/deconvolution.cc"],
4020 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004021 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004022 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004023)
4024
4025xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004026 name = "floor_bench",
4027 srcs = ["bench/floor.cc"],
4028 copts = xnnpack_optional_tflite_copts(),
4029 tags = ["nowin32"],
4030 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4031)
4032
4033xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004034 name = "global_average_pooling_bench",
4035 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004036 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004037)
4038
4039xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004040 name = "hardswish_bench",
4041 srcs = ["bench/hardswish.cc"],
4042 copts = xnnpack_optional_tflite_copts(),
4043 tags = ["nowin32"],
4044 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4045)
4046
4047xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004048 name = "max_pooling_bench",
4049 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004050 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004051)
4052
4053xnnpack_benchmark(
4054 name = "sigmoid_bench",
4055 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004056 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004057 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004058 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004059)
4060
4061xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004062 name = "prelu_bench",
4063 srcs = ["bench/prelu.cc"],
4064 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004065 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004066 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004067)
4068
4069xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004070 name = "softmax_bench",
4071 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004072 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004073 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004074 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004075)
4076
Marat Dukhan87727142020-06-24 15:24:10 -07004077xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004078 name = "square_root_bench",
4079 srcs = ["bench/square-root.cc"],
4080 copts = xnnpack_optional_tflite_copts(),
4081 tags = ["nowin32"],
4082 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4083)
4084
4085xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004086 name = "truncation_bench",
4087 srcs = ["bench/truncation.cc"],
4088 deps = OPERATOR_BENCHMARK_DEPS,
4089)
4090
Marat Dukhanc068bb62019-10-04 13:24:39 -07004091############################# End-to-end benchmarks ############################
4092
4093cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004094 name = "fp32_mobilenet_v1",
4095 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004096 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004097 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004098 linkstatic = True,
4099 deps = [
4100 ":XNNPACK",
4101 "@pthreadpool",
4102 ],
4103)
4104
4105cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004106 name = "fp16_mobilenet_v1",
4107 srcs = ["models/fp16-mobilenet-v1.cc"],
4108 hdrs = ["models/models.h"],
4109 copts = xnnpack_std_cxxopts(),
4110 linkstatic = True,
4111 deps = [
4112 ":XNNPACK",
4113 "@FP16",
4114 "@pthreadpool",
4115 ],
4116)
4117
4118cc_library(
4119 name = "fp32_mobilenet_v2",
4120 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004121 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004122 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004123 linkstatic = True,
4124 deps = [
4125 ":XNNPACK",
4126 "@pthreadpool",
4127 ],
4128)
4129
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004130cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004131 name = "fp16_mobilenet_v2",
4132 srcs = ["models/fp16-mobilenet-v2.cc"],
4133 hdrs = ["models/models.h"],
4134 copts = xnnpack_std_cxxopts(),
4135 linkstatic = True,
4136 deps = [
4137 ":XNNPACK",
4138 "@FP16",
4139 "@pthreadpool",
4140 ],
4141)
4142
4143cc_library(
4144 name = "fp32_mobilenet_v3_large",
4145 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004146 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004147 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004148 linkstatic = True,
4149 deps = [
4150 ":XNNPACK",
4151 "@pthreadpool",
4152 ],
4153)
4154
4155cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004156 name = "fp32_mobilenet_v3_small",
4157 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004158 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004159 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004160 linkstatic = True,
4161 deps = [
4162 ":XNNPACK",
4163 "@pthreadpool",
4164 ],
4165)
4166
Marat Dukhanc068bb62019-10-04 13:24:39 -07004167xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004168 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004169 srcs = [
4170 "bench/f32-dwconv-e2e.cc",
4171 "bench/end2end.h",
4172 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004173 deps = MICROKERNEL_BENCHMARK_DEPS + [
4174 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004175 ":fp32_mobilenet_v1",
4176 ":fp32_mobilenet_v2",
4177 ":fp32_mobilenet_v3_large",
4178 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004179 ],
4180)
4181
4182xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004183 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004184 srcs = [
4185 "bench/f32-gemm-e2e.cc",
4186 "bench/end2end.h",
4187 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004188 deps = MICROKERNEL_BENCHMARK_DEPS + [
4189 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004190 ":fp32_mobilenet_v1",
4191 ":fp32_mobilenet_v2",
4192 ":fp32_mobilenet_v3_large",
4193 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004194 ],
4195)
4196
4197xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004198 name = "end2end_bench",
4199 srcs = ["bench/end2end.cc"],
4200 deps = [
4201 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004202 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004203 ":fp16_mobilenet_v1",
4204 ":fp16_mobilenet_v2",
4205 ":fp32_mobilenet_v1",
4206 ":fp32_mobilenet_v2",
4207 ":fp32_mobilenet_v3_large",
4208 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004209 "@pthreadpool",
4210 ],
4211)
4212
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004213#################### Accuracy evaluation for math functions ####################
4214
4215xnnpack_benchmark(
4216 name = "f32_exp_eval",
4217 srcs = [
4218 "eval/f32-exp.cc",
4219 "src/xnnpack/AlignedAllocator.h",
4220 ] + ACCURACY_EVAL_HDRS,
4221 deps = ACCURACY_EVAL_DEPS,
4222)
4223
Marat Dukhan515c9772019-10-17 18:07:57 -07004224xnnpack_benchmark(
4225 name = "f32_expminus_eval",
4226 srcs = [
4227 "eval/f32-expminus.cc",
4228 "src/xnnpack/AlignedAllocator.h",
4229 ] + ACCURACY_EVAL_HDRS,
4230 deps = ACCURACY_EVAL_DEPS,
4231)
4232
Marat Dukhan98ba4412019-10-23 02:14:28 -07004233xnnpack_benchmark(
4234 name = "f32_extexp_eval",
4235 srcs = [
4236 "eval/f32-extexp.cc",
4237 "src/xnnpack/AlignedAllocator.h",
4238 ] + ACCURACY_EVAL_HDRS,
4239 deps = ACCURACY_EVAL_DEPS,
4240)
4241
Marat Dukhan8853b822020-05-07 12:19:01 -07004242xnnpack_unit_test(
4243 name = "f32_roundne_eval",
4244 srcs = [
4245 "eval/f32-roundne.cc",
4246 "src/xnnpack/AlignedAllocator.h",
4247 "src/xnnpack/math-stubs.h",
4248 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004249 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004250 deps = MICROKERNEL_TEST_DEPS,
4251)
4252
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004253xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004254 name = "f32_roundd_eval",
4255 srcs = [
4256 "eval/f32-roundd.cc",
4257 "src/xnnpack/AlignedAllocator.h",
4258 "src/xnnpack/math-stubs.h",
4259 ] + MICROKERNEL_TEST_HDRS,
4260 automatic = False,
4261 deps = MICROKERNEL_TEST_DEPS,
4262)
4263
4264xnnpack_unit_test(
4265 name = "f32_roundu_eval",
4266 srcs = [
4267 "eval/f32-roundu.cc",
4268 "src/xnnpack/AlignedAllocator.h",
4269 "src/xnnpack/math-stubs.h",
4270 ] + MICROKERNEL_TEST_HDRS,
4271 automatic = False,
4272 deps = MICROKERNEL_TEST_DEPS,
4273)
4274
4275xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004276 name = "f32_roundz_eval",
4277 srcs = [
4278 "eval/f32-roundz.cc",
4279 "src/xnnpack/AlignedAllocator.h",
4280 "src/xnnpack/math-stubs.h",
4281 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004282 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004283 deps = MICROKERNEL_TEST_DEPS,
4284)
4285
Marat Dukhan346a9e52019-11-15 09:06:30 -08004286xnnpack_benchmark(
4287 name = "f32_sigmoid_eval",
4288 srcs = [
4289 "eval/f32-sigmoid.cc",
4290 "src/xnnpack/AlignedAllocator.h",
4291 ] + ACCURACY_EVAL_HDRS,
4292 deps = ACCURACY_EVAL_DEPS,
4293)
4294
Marat Dukhan84000762020-06-29 18:38:43 -07004295xnnpack_benchmark(
4296 name = "f32_sqrt_eval",
4297 srcs = [
4298 "eval/f32-sqrt.cc",
4299 "src/xnnpack/AlignedAllocator.h",
4300 ] + ACCURACY_EVAL_HDRS,
4301 deps = ACCURACY_EVAL_DEPS + [
4302 ":bench_utils",
4303 ],
4304)
4305
Marat Dukhan08c4a432019-10-03 09:29:21 -07004306######################### Unit tests for micro-kernels #########################
4307
4308xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004309 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004311 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004312 "test/gemm-microkernel-tester.h",
4313 "src/xnnpack/AlignedAllocator.h",
4314 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004315 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004316)
4317
4318xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004319 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004320 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004321 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004322 "test/spmm-microkernel-tester.h",
4323 "src/xnnpack/AlignedAllocator.h",
4324 ] + MICROKERNEL_TEST_HDRS,
4325 deps = MICROKERNEL_TEST_DEPS,
4326)
4327
4328xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329 name = "f32_argmaxpool_test",
4330 srcs = [
4331 "test/f32-argmaxpool.cc",
4332 "test/argmaxpool-microkernel-tester.h",
4333 "src/xnnpack/AlignedAllocator.h",
4334 ] + MICROKERNEL_TEST_HDRS,
4335 deps = MICROKERNEL_TEST_DEPS,
4336)
4337
4338xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004339 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004340 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004341 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004342 "test/avgpool-microkernel-tester.h",
4343 "src/xnnpack/AlignedAllocator.h",
4344 ] + MICROKERNEL_TEST_HDRS,
4345 deps = MICROKERNEL_TEST_DEPS,
4346)
4347
4348xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004349 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004350 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004351 "test/f32-ibilinear.cc",
4352 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004353 "src/xnnpack/AlignedAllocator.h",
4354 ] + MICROKERNEL_TEST_HDRS,
4355 deps = MICROKERNEL_TEST_DEPS,
4356)
4357
4358xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004359 name = "f16_clamp_test",
4360 srcs = [
4361 "test/f16-clamp.cc",
4362 "test/clamp-microkernel-tester.h",
4363 ] + MICROKERNEL_TEST_HDRS,
4364 deps = MICROKERNEL_TEST_DEPS,
4365)
4366
4367xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004368 name = "f32_clamp_test",
4369 srcs = [
4370 "test/f32-clamp.cc",
4371 "test/clamp-microkernel-tester.h",
4372 ] + MICROKERNEL_TEST_HDRS,
4373 deps = MICROKERNEL_TEST_DEPS,
4374)
4375
4376xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004377 name = "f32_igemm_test",
4378 srcs = [
4379 "test/f32-igemm.cc",
4380 "test/gemm-microkernel-tester.h",
4381 "src/xnnpack/AlignedAllocator.h",
4382 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004383 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004384)
4385
4386xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004387 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004388 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004389 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390 "test/gemm-microkernel-tester.h",
4391 "src/xnnpack/AlignedAllocator.h",
4392 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004393 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394)
4395
4396xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004397 name = "f32_igemm_minmax_test",
4398 srcs = [
4399 "test/f32-igemm-minmax.cc",
4400 "test/gemm-microkernel-tester.h",
4401 "src/xnnpack/AlignedAllocator.h",
4402 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004403 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004404)
4405
4406xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004407 name = "f16_igemm_minmax_test",
4408 srcs = [
4409 "test/f16-igemm-minmax.cc",
4410 "test/gemm-microkernel-tester.h",
4411 "src/xnnpack/AlignedAllocator.h",
4412 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004413 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004414)
4415
4416xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004417 name = "f32_conv_hwc_test",
4418 srcs = [
4419 "test/f32-conv-hwc.cc",
4420 "test/conv-hwc-microkernel-tester.h",
4421 "src/xnnpack/AlignedAllocator.h",
4422 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004423 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004424)
4425
4426xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004427 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004428 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004429 "test/f32-conv-hwc2chw.cc",
4430 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004431 "src/xnnpack/AlignedAllocator.h",
4432 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004433 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004434)
4435
4436xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004437 name = "f32_dwconv_test",
4438 srcs = [
4439 "test/f32-dwconv.cc",
4440 "test/dwconv-microkernel-tester.h",
4441 "src/xnnpack/AlignedAllocator.h",
4442 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004443 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004444)
4445
4446xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004447 name = "f16_dwconv_minmax_test",
4448 srcs = [
4449 "test/f16-dwconv-minmax.cc",
4450 "test/dwconv-microkernel-tester.h",
4451 "src/xnnpack/AlignedAllocator.h",
4452 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004453 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004454)
4455
4456xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004457 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004459 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460 "test/dwconv-microkernel-tester.h",
4461 "src/xnnpack/AlignedAllocator.h",
4462 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004463 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004464)
4465
4466xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004467 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004468 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004469 "test/f32-dwconv-chw.cc",
4470 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004471 "src/xnnpack/AlignedAllocator.h",
4472 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004473 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004474)
4475
4476xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004477 name = "f16_gavgpool_minmax_test",
4478 srcs = [
4479 "test/f16-gavgpool-minmax.cc",
4480 "test/gavgpool-microkernel-tester.h",
4481 "src/xnnpack/AlignedAllocator.h",
4482 ] + MICROKERNEL_TEST_HDRS,
4483 deps = MICROKERNEL_TEST_DEPS,
4484)
4485
4486xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004487 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004488 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004489 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490 "test/gavgpool-microkernel-tester.h",
4491 "src/xnnpack/AlignedAllocator.h",
4492 ] + MICROKERNEL_TEST_HDRS,
4493 deps = MICROKERNEL_TEST_DEPS,
4494)
4495
4496xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004497 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004498 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004499 "test/f32-gavgpool-cw.cc",
4500 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004501 "src/xnnpack/AlignedAllocator.h",
4502 ] + MICROKERNEL_TEST_HDRS,
4503 deps = MICROKERNEL_TEST_DEPS,
4504)
4505
4506xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004507 name = "f32_gemm_test",
4508 srcs = [
4509 "test/f32-gemm.cc",
4510 "test/gemm-microkernel-tester.h",
4511 "src/xnnpack/AlignedAllocator.h",
4512 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004513 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004514)
4515
4516xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004517 name = "f32_gemm_relu_test",
4518 srcs = [
4519 "test/f32-gemm-relu.cc",
4520 "test/gemm-microkernel-tester.h",
4521 "src/xnnpack/AlignedAllocator.h",
4522 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004523 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004524)
4525
4526xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004527 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004528 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004529 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004530 "test/gemm-microkernel-tester.h",
4531 "src/xnnpack/AlignedAllocator.h",
4532 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004533 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534)
4535
4536xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004537 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004538 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004539 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004540 "test/gemm-microkernel-tester.h",
4541 "src/xnnpack/AlignedAllocator.h",
4542 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004543 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004544)
4545
4546xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004547 name = "f16_hswish_test",
4548 srcs = [
4549 "test/f16-hswish.cc",
4550 "test/hswish-microkernel-tester.h",
4551 ] + MICROKERNEL_TEST_HDRS,
4552 deps = MICROKERNEL_TEST_DEPS,
4553)
4554
4555xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556 name = "f32_hswish_test",
4557 srcs = [
4558 "test/f32-hswish.cc",
4559 "test/hswish-microkernel-tester.h",
4560 ] + MICROKERNEL_TEST_HDRS,
4561 deps = MICROKERNEL_TEST_DEPS,
4562)
4563
4564xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004565 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004566 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004567 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004568 "test/maxpool-microkernel-tester.h",
4569 ] + MICROKERNEL_TEST_HDRS,
4570 deps = MICROKERNEL_TEST_DEPS,
4571)
4572
4573xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004574 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004575 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004576 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004577 "test/avgpool-microkernel-tester.h",
4578 "src/xnnpack/AlignedAllocator.h",
4579 ] + MICROKERNEL_TEST_HDRS,
4580 deps = MICROKERNEL_TEST_DEPS,
4581)
4582
4583xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004584 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004585 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004586 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004587 "test/gemm-microkernel-tester.h",
4588 "src/xnnpack/AlignedAllocator.h",
4589 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004590 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004591)
4592
4593xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004594 name = "f16_prelu_test",
4595 srcs = [
4596 "test/f16-prelu.cc",
4597 "test/prelu-microkernel-tester.h",
4598 "src/xnnpack/AlignedAllocator.h",
4599 ] + MICROKERNEL_TEST_HDRS,
4600 deps = MICROKERNEL_TEST_DEPS,
4601)
4602
4603xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004604 name = "f32_prelu_test",
4605 srcs = [
4606 "test/f32-prelu.cc",
4607 "test/prelu-microkernel-tester.h",
4608 "src/xnnpack/AlignedAllocator.h",
4609 ] + MICROKERNEL_TEST_HDRS,
4610 deps = MICROKERNEL_TEST_DEPS,
4611)
4612
4613xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004614 name = "f32_raddexpminusmax_test",
4615 srcs = [
4616 "test/f32-raddexpminusmax.cc",
4617 "test/raddexpminusmax-microkernel-tester.h",
4618 ] + MICROKERNEL_TEST_HDRS,
4619 deps = MICROKERNEL_TEST_DEPS,
4620)
4621
4622xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004623 name = "f32_raddextexp_test",
4624 srcs = [
4625 "test/f32-raddextexp.cc",
4626 "test/raddextexp-microkernel-tester.h",
4627 ] + MICROKERNEL_TEST_HDRS,
4628 deps = MICROKERNEL_TEST_DEPS,
4629)
4630
4631xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004632 name = "f32_raddstoreexpminusmax_test",
4633 srcs = [
4634 "test/f32-raddstoreexpminusmax.cc",
4635 "test/raddstoreexpminusmax-microkernel-tester.h",
4636 ] + MICROKERNEL_TEST_HDRS,
4637 deps = MICROKERNEL_TEST_DEPS,
4638)
4639
4640xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004641 name = "f32_relu_test",
4642 srcs = [
4643 "test/f32-relu.cc",
4644 "test/vunary-microkernel-tester.h",
4645 ] + MICROKERNEL_TEST_HDRS,
4646 deps = MICROKERNEL_TEST_DEPS,
4647)
4648
4649xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004650 name = "f32_rmax_test",
4651 srcs = [
4652 "test/f32-rmax.cc",
4653 "test/rmax-microkernel-tester.h",
4654 ] + MICROKERNEL_TEST_HDRS,
4655 deps = MICROKERNEL_TEST_DEPS,
4656)
4657
4658xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004659 name = "f32_sigmoid_test",
4660 srcs = [
4661 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004662 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004663 ] + MICROKERNEL_TEST_HDRS,
4664 deps = MICROKERNEL_TEST_DEPS,
4665)
4666
4667xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004668 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004669 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004670 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004671 "test/spmm-microkernel-tester.h",
4672 "src/xnnpack/AlignedAllocator.h",
4673 ] + MICROKERNEL_TEST_HDRS,
4674 deps = MICROKERNEL_TEST_DEPS,
4675)
4676
4677xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004678 name = "f16_vadd_minmax_test",
4679 srcs = [
4680 "test/f16-vadd-minmax.cc",
4681 "test/vbinary-microkernel-tester.h",
4682 ] + MICROKERNEL_TEST_HDRS,
4683 deps = MICROKERNEL_TEST_DEPS,
4684)
4685
4686xnnpack_unit_test(
4687 name = "f16_vaddc_minmax_test",
4688 srcs = [
4689 "test/f16-vaddc-minmax.cc",
4690 "test/vbinaryc-microkernel-tester.h",
4691 ] + MICROKERNEL_TEST_HDRS,
4692 deps = MICROKERNEL_TEST_DEPS,
4693)
4694
4695xnnpack_unit_test(
4696 name = "f16_vdiv_minmax_test",
4697 srcs = [
4698 "test/f16-vdiv-minmax.cc",
4699 "test/vbinary-microkernel-tester.h",
4700 ] + MICROKERNEL_TEST_HDRS,
4701 deps = MICROKERNEL_TEST_DEPS,
4702)
4703
4704xnnpack_unit_test(
4705 name = "f16_vdivc_minmax_test",
4706 srcs = [
4707 "test/f16-vdivc-minmax.cc",
4708 "test/vbinaryc-microkernel-tester.h",
4709 ] + MICROKERNEL_TEST_HDRS,
4710 deps = MICROKERNEL_TEST_DEPS,
4711)
4712
4713xnnpack_unit_test(
4714 name = "f16_vrdivc_minmax_test",
4715 srcs = [
4716 "test/f16-vrdivc-minmax.cc",
4717 "test/vbinaryc-microkernel-tester.h",
4718 ] + MICROKERNEL_TEST_HDRS,
4719 deps = MICROKERNEL_TEST_DEPS,
4720)
4721
4722xnnpack_unit_test(
4723 name = "f16_vmax_test",
4724 srcs = [
4725 "test/f16-vmax.cc",
4726 "test/vbinary-microkernel-tester.h",
4727 ] + MICROKERNEL_TEST_HDRS,
4728 deps = MICROKERNEL_TEST_DEPS,
4729)
4730
4731xnnpack_unit_test(
4732 name = "f16_vmaxc_test",
4733 srcs = [
4734 "test/f16-vmaxc.cc",
4735 "test/vbinaryc-microkernel-tester.h",
4736 ] + MICROKERNEL_TEST_HDRS,
4737 deps = MICROKERNEL_TEST_DEPS,
4738)
4739
4740xnnpack_unit_test(
4741 name = "f16_vmin_test",
4742 srcs = [
4743 "test/f16-vmin.cc",
4744 "test/vbinary-microkernel-tester.h",
4745 ] + MICROKERNEL_TEST_HDRS,
4746 deps = MICROKERNEL_TEST_DEPS,
4747)
4748
4749xnnpack_unit_test(
4750 name = "f16_vminc_test",
4751 srcs = [
4752 "test/f16-vminc.cc",
4753 "test/vbinaryc-microkernel-tester.h",
4754 ] + MICROKERNEL_TEST_HDRS,
4755 deps = MICROKERNEL_TEST_DEPS,
4756)
4757
4758xnnpack_unit_test(
4759 name = "f16_vmul_minmax_test",
4760 srcs = [
4761 "test/f16-vmul-minmax.cc",
4762 "test/vbinary-microkernel-tester.h",
4763 ] + MICROKERNEL_TEST_HDRS,
4764 deps = MICROKERNEL_TEST_DEPS,
4765)
4766
4767xnnpack_unit_test(
4768 name = "f16_vmulc_minmax_test",
4769 srcs = [
4770 "test/f16-vmulc-minmax.cc",
4771 "test/vbinaryc-microkernel-tester.h",
4772 ] + MICROKERNEL_TEST_HDRS,
4773 deps = MICROKERNEL_TEST_DEPS,
4774)
4775
4776xnnpack_unit_test(
4777 name = "f16_vsub_minmax_test",
4778 srcs = [
4779 "test/f16-vsub-minmax.cc",
4780 "test/vbinary-microkernel-tester.h",
4781 ] + MICROKERNEL_TEST_HDRS,
4782 deps = MICROKERNEL_TEST_DEPS,
4783)
4784
4785xnnpack_unit_test(
4786 name = "f16_vsubc_minmax_test",
4787 srcs = [
4788 "test/f16-vsubc-minmax.cc",
4789 "test/vbinaryc-microkernel-tester.h",
4790 ] + MICROKERNEL_TEST_HDRS,
4791 deps = MICROKERNEL_TEST_DEPS,
4792)
4793
4794xnnpack_unit_test(
4795 name = "f16_vrsubc_minmax_test",
4796 srcs = [
4797 "test/f16-vrsubc-minmax.cc",
4798 "test/vbinaryc-microkernel-tester.h",
4799 ] + MICROKERNEL_TEST_HDRS,
4800 deps = MICROKERNEL_TEST_DEPS,
4801)
4802
4803xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004804 name = "f32_vabs_test",
4805 srcs = [
4806 "test/f32-vabs.cc",
4807 "test/vunary-microkernel-tester.h",
4808 ] + MICROKERNEL_TEST_HDRS,
4809 deps = MICROKERNEL_TEST_DEPS,
4810)
4811
4812xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004813 name = "f32_vadd_test",
4814 srcs = [
4815 "test/f32-vadd.cc",
4816 "test/vbinary-microkernel-tester.h",
4817 ] + MICROKERNEL_TEST_HDRS,
4818 deps = MICROKERNEL_TEST_DEPS,
4819)
4820
4821xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004822 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004823 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004824 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004825 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004826 ] + MICROKERNEL_TEST_HDRS,
4827 deps = MICROKERNEL_TEST_DEPS,
4828)
4829
4830xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004831 name = "f32_vaddc_test",
4832 srcs = [
4833 "test/f32-vaddc.cc",
4834 "test/vbinaryc-microkernel-tester.h",
4835 ] + MICROKERNEL_TEST_HDRS,
4836 deps = MICROKERNEL_TEST_DEPS,
4837)
4838
4839xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004840 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004841 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004842 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004843 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004844 ] + MICROKERNEL_TEST_HDRS,
4845 deps = MICROKERNEL_TEST_DEPS,
4846)
4847
4848xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004849 name = "f32_vdiv_test",
4850 srcs = [
4851 "test/f32-vdiv.cc",
4852 "test/vbinary-microkernel-tester.h",
4853 ] + MICROKERNEL_TEST_HDRS,
4854 deps = MICROKERNEL_TEST_DEPS,
4855)
4856
4857xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004858 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004859 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004860 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004861 "test/vbinary-microkernel-tester.h",
4862 ] + MICROKERNEL_TEST_HDRS,
4863 deps = MICROKERNEL_TEST_DEPS,
4864)
4865
4866xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004867 name = "f32_vdivc_test",
4868 srcs = [
4869 "test/f32-vdivc.cc",
4870 "test/vbinaryc-microkernel-tester.h",
4871 ] + MICROKERNEL_TEST_HDRS,
4872 deps = MICROKERNEL_TEST_DEPS,
4873)
4874
4875xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004876 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004877 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004878 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004879 "test/vbinaryc-microkernel-tester.h",
4880 ] + MICROKERNEL_TEST_HDRS,
4881 deps = MICROKERNEL_TEST_DEPS,
4882)
4883
4884xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004885 name = "f32_vrdivc_test",
4886 srcs = [
4887 "test/f32-vrdivc.cc",
4888 "test/vbinaryc-microkernel-tester.h",
4889 ] + MICROKERNEL_TEST_HDRS,
4890 deps = MICROKERNEL_TEST_DEPS,
4891)
4892
4893xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004894 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004895 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004896 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004897 "test/vbinaryc-microkernel-tester.h",
4898 ] + MICROKERNEL_TEST_HDRS,
4899 deps = MICROKERNEL_TEST_DEPS,
4900)
4901
4902xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004903 name = "f32_vmax_test",
4904 srcs = [
4905 "test/f32-vmax.cc",
4906 "test/vbinary-microkernel-tester.h",
4907 ] + MICROKERNEL_TEST_HDRS,
4908 deps = MICROKERNEL_TEST_DEPS,
4909)
4910
4911xnnpack_unit_test(
4912 name = "f32_vmaxc_test",
4913 srcs = [
4914 "test/f32-vmaxc.cc",
4915 "test/vbinaryc-microkernel-tester.h",
4916 ] + MICROKERNEL_TEST_HDRS,
4917 deps = MICROKERNEL_TEST_DEPS,
4918)
4919
4920xnnpack_unit_test(
4921 name = "f32_vmin_test",
4922 srcs = [
4923 "test/f32-vmin.cc",
4924 "test/vbinary-microkernel-tester.h",
4925 ] + MICROKERNEL_TEST_HDRS,
4926 deps = MICROKERNEL_TEST_DEPS,
4927)
4928
4929xnnpack_unit_test(
4930 name = "f32_vminc_test",
4931 srcs = [
4932 "test/f32-vminc.cc",
4933 "test/vbinaryc-microkernel-tester.h",
4934 ] + MICROKERNEL_TEST_HDRS,
4935 deps = MICROKERNEL_TEST_DEPS,
4936)
4937
4938xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004939 name = "f32_vmul_test",
4940 srcs = [
4941 "test/f32-vmul.cc",
4942 "test/vbinary-microkernel-tester.h",
4943 ] + MICROKERNEL_TEST_HDRS,
4944 deps = MICROKERNEL_TEST_DEPS,
4945)
4946
4947xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004948 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004949 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004950 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004951 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004952 ] + MICROKERNEL_TEST_HDRS,
4953 deps = MICROKERNEL_TEST_DEPS,
4954)
4955
4956xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004957 name = "f32_vmulc_test",
4958 srcs = [
4959 "test/f32-vmulc.cc",
4960 "test/vbinaryc-microkernel-tester.h",
4961 ] + MICROKERNEL_TEST_HDRS,
4962 deps = MICROKERNEL_TEST_DEPS,
4963)
4964
4965xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004966 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004967 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004968 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004969 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004970 ] + MICROKERNEL_TEST_HDRS,
4971 deps = MICROKERNEL_TEST_DEPS,
4972)
4973
4974xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004975 name = "f16_vmulcaddc_minmax_test",
4976 srcs = [
4977 "test/f16-vmulcaddc-minmax.cc",
4978 "test/vmulcaddc-microkernel-tester.h",
4979 "src/xnnpack/AlignedAllocator.h",
4980 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004981 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004982)
4983
4984xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004985 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004986 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004987 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004988 "test/vmulcaddc-microkernel-tester.h",
4989 "src/xnnpack/AlignedAllocator.h",
4990 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004991 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004992)
4993
4994xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004995 name = "f32_vlrelu_test",
4996 srcs = [
4997 "test/f32-vlrelu.cc",
4998 "test/vunary-microkernel-tester.h",
4999 ] + MICROKERNEL_TEST_HDRS,
5000 deps = MICROKERNEL_TEST_DEPS,
5001)
5002
5003xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005004 name = "f32_vneg_test",
5005 srcs = [
5006 "test/f32-vneg.cc",
5007 "test/vunary-microkernel-tester.h",
5008 ] + MICROKERNEL_TEST_HDRS,
5009 deps = MICROKERNEL_TEST_DEPS,
5010)
5011
5012xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005013 name = "f32_vrndne_test",
5014 srcs = [
5015 "test/f32-vrndne.cc",
5016 "test/vunary-microkernel-tester.h",
5017 ] + MICROKERNEL_TEST_HDRS,
5018 deps = MICROKERNEL_TEST_DEPS,
5019)
5020
5021xnnpack_unit_test(
5022 name = "f32_vrndz_test",
5023 srcs = [
5024 "test/f32-vrndz.cc",
5025 "test/vunary-microkernel-tester.h",
5026 ] + MICROKERNEL_TEST_HDRS,
5027 deps = MICROKERNEL_TEST_DEPS,
5028)
5029
5030xnnpack_unit_test(
5031 name = "f32_vrndu_test",
5032 srcs = [
5033 "test/f32-vrndu.cc",
5034 "test/vunary-microkernel-tester.h",
5035 ] + MICROKERNEL_TEST_HDRS,
5036 deps = MICROKERNEL_TEST_DEPS,
5037)
5038
5039xnnpack_unit_test(
5040 name = "f32_vrndd_test",
5041 srcs = [
5042 "test/f32-vrndd.cc",
5043 "test/vunary-microkernel-tester.h",
5044 ] + MICROKERNEL_TEST_HDRS,
5045 deps = MICROKERNEL_TEST_DEPS,
5046)
5047
5048xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005049 name = "f32_vscale_test",
5050 srcs = [
5051 "test/f32-vscale.cc",
5052 "test/vscale-microkernel-tester.h",
5053 ] + MICROKERNEL_TEST_HDRS,
5054 deps = MICROKERNEL_TEST_DEPS,
5055)
5056
5057xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005058 name = "f32_vscaleexpminusmax_test",
5059 srcs = [
5060 "test/f32-vscaleexpminusmax.cc",
5061 "test/vscaleexpminusmax-microkernel-tester.h",
5062 ] + MICROKERNEL_TEST_HDRS,
5063 deps = MICROKERNEL_TEST_DEPS,
5064)
5065
5066xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005067 name = "f32_vscaleextexp_test",
5068 srcs = [
5069 "test/f32-vscaleextexp.cc",
5070 "test/vscaleextexp-microkernel-tester.h",
5071 ] + MICROKERNEL_TEST_HDRS,
5072 deps = MICROKERNEL_TEST_DEPS,
5073)
5074
5075xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005076 name = "f32_vsqr_test",
5077 srcs = [
5078 "test/f32-vsqr.cc",
5079 "test/vunary-microkernel-tester.h",
5080 ] + MICROKERNEL_TEST_HDRS,
5081 deps = MICROKERNEL_TEST_DEPS,
5082)
5083
5084xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005085 name = "f32_vsqrdiff_test",
5086 srcs = [
5087 "test/f32-vsqrdiff.cc",
5088 "test/vbinary-microkernel-tester.h",
5089 ] + MICROKERNEL_TEST_HDRS,
5090 deps = MICROKERNEL_TEST_DEPS,
5091)
5092
5093xnnpack_unit_test(
5094 name = "f32_vsqrdiffc_test",
5095 srcs = [
5096 "test/f32-vsqrdiffc.cc",
5097 "test/vbinaryc-microkernel-tester.h",
5098 ] + MICROKERNEL_TEST_HDRS,
5099 deps = MICROKERNEL_TEST_DEPS,
5100)
5101
5102xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005103 name = "f32_vsqrt_test",
5104 srcs = [
5105 "test/f32-vsqrt.cc",
5106 "test/vunary-microkernel-tester.h",
5107 ] + MICROKERNEL_TEST_HDRS,
5108 deps = MICROKERNEL_TEST_DEPS,
5109)
5110
5111xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005112 name = "f32_vsub_test",
5113 srcs = [
5114 "test/f32-vsub.cc",
5115 "test/vbinary-microkernel-tester.h",
5116 ] + MICROKERNEL_TEST_HDRS,
5117 deps = MICROKERNEL_TEST_DEPS,
5118)
5119
5120xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005121 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005122 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005123 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005124 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005125 ] + MICROKERNEL_TEST_HDRS,
5126 deps = MICROKERNEL_TEST_DEPS,
5127)
5128
5129xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005130 name = "f32_vsubc_test",
5131 srcs = [
5132 "test/f32-vsubc.cc",
5133 "test/vbinaryc-microkernel-tester.h",
5134 ] + MICROKERNEL_TEST_HDRS,
5135 deps = MICROKERNEL_TEST_DEPS,
5136)
5137
5138xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005139 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005140 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005141 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005142 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005143 ] + MICROKERNEL_TEST_HDRS,
5144 deps = MICROKERNEL_TEST_DEPS,
5145)
5146
5147xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005148 name = "f32_vrsubc_test",
5149 srcs = [
5150 "test/f32-vrsubc.cc",
5151 "test/vbinaryc-microkernel-tester.h",
5152 ] + MICROKERNEL_TEST_HDRS,
5153 deps = MICROKERNEL_TEST_DEPS,
5154)
5155
5156xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005157 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005158 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005159 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005160 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005161 ] + MICROKERNEL_TEST_HDRS,
5162 deps = MICROKERNEL_TEST_DEPS,
5163)
5164
5165xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005166 name = "qs8_gemm_minmax_test",
5167 srcs = [
5168 "test/qs8-gemm-minmax.cc",
5169 "test/gemm-microkernel-tester.h",
5170 "src/xnnpack/AlignedAllocator.h",
5171 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5172 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5173)
5174
5175xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005176 name = "qs8_igemm_minmax_test",
5177 srcs = [
5178 "test/qs8-igemm-minmax.cc",
5179 "test/gemm-microkernel-tester.h",
5180 "src/xnnpack/AlignedAllocator.h",
5181 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5182 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5183)
5184
5185xnnpack_unit_test(
5186 name = "qs8_requantization_test",
5187 srcs = [
5188 "src/xnnpack/requantization-stubs.h",
5189 "test/qs8-requantization.cc",
5190 "test/requantization-tester.h",
5191 ] + MICROKERNEL_TEST_HDRS,
5192 deps = MICROKERNEL_TEST_DEPS,
5193)
5194
5195xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005196 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005198 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005199 "test/avgpool-microkernel-tester.h",
5200 "src/xnnpack/AlignedAllocator.h",
5201 ] + MICROKERNEL_TEST_HDRS,
5202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
5205xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005206 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005207 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005208 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005209 "test/gemm-microkernel-tester.h",
5210 "src/xnnpack/AlignedAllocator.h",
5211 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005212 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005213)
5214
5215xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005216 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005217 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005218 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219 "test/dwconv-microkernel-tester.h",
5220 "src/xnnpack/AlignedAllocator.h",
5221 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005222 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223)
5224
5225xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005226 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005227 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005228 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005229 "test/gavgpool-microkernel-tester.h",
5230 "src/xnnpack/AlignedAllocator.h",
5231 ] + MICROKERNEL_TEST_HDRS,
5232 deps = MICROKERNEL_TEST_DEPS,
5233)
5234
5235xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005236 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005238 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005239 "test/gemm-microkernel-tester.h",
5240 "src/xnnpack/AlignedAllocator.h",
5241 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005242 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005243)
5244
5245xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005246 name = "qu8_requantization_test",
5247 srcs = [
5248 "src/xnnpack/requantization-stubs.h",
5249 "test/qu8-requantization.cc",
5250 "test/requantization-tester.h",
5251 ] + MICROKERNEL_TEST_HDRS,
5252 deps = MICROKERNEL_TEST_DEPS,
5253)
5254
5255xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005256 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005257 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005258 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005259 "test/vadd-microkernel-tester.h",
5260 ] + MICROKERNEL_TEST_HDRS,
5261 deps = MICROKERNEL_TEST_DEPS,
5262)
5263
5264xnnpack_unit_test(
5265 name = "u8_clamp_test",
5266 srcs = [
5267 "test/u8-clamp.cc",
5268 "test/clamp-microkernel-tester.h",
5269 ] + MICROKERNEL_TEST_HDRS,
5270 deps = MICROKERNEL_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
5274 name = "u8_lut32norm_test",
5275 srcs = [
5276 "test/u8-lut32norm.cc",
5277 "test/lut-norm-microkernel-tester.h",
5278 ] + MICROKERNEL_TEST_HDRS,
5279 deps = MICROKERNEL_TEST_DEPS,
5280)
5281
5282xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005283 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005285 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005286 "test/maxpool-microkernel-tester.h",
5287 ] + MICROKERNEL_TEST_HDRS,
5288 deps = MICROKERNEL_TEST_DEPS,
5289)
5290
5291xnnpack_unit_test(
5292 name = "u8_rmax_test",
5293 srcs = [
5294 "test/u8-rmax.cc",
5295 "test/rmax-microkernel-tester.h",
5296 ] + MICROKERNEL_TEST_HDRS,
5297 deps = MICROKERNEL_TEST_DEPS,
5298)
5299
5300xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005301 name = "x32_fill_test",
5302 srcs = [
5303 "test/x32-fill.cc",
5304 "test/fill-microkernel-tester.h",
5305 ] + MICROKERNEL_TEST_HDRS,
5306 deps = MICROKERNEL_TEST_DEPS,
5307)
5308
5309xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310 name = "x32_packx_test",
5311 srcs = [
5312 "test/x32-packx.cc",
5313 "test/pack-microkernel-tester.h",
5314 "src/xnnpack/AlignedAllocator.h",
5315 ] + MICROKERNEL_TEST_HDRS,
5316 deps = MICROKERNEL_TEST_DEPS,
5317)
5318
5319xnnpack_unit_test(
5320 name = "x32_pad_test",
5321 srcs = [
5322 "test/x32-pad.cc",
5323 "test/pad-microkernel-tester.h",
5324 ] + MICROKERNEL_TEST_HDRS,
5325 deps = MICROKERNEL_TEST_DEPS,
5326)
5327
5328xnnpack_unit_test(
5329 name = "x32_unpool_test",
5330 srcs = [
5331 "test/x32-unpool.cc",
5332 "test/unpool-microkernel-tester.h",
5333 ] + MICROKERNEL_TEST_HDRS,
5334 deps = MICROKERNEL_TEST_DEPS,
5335)
5336
5337xnnpack_unit_test(
5338 name = "x32_zip_test",
5339 srcs = [
5340 "test/x32-zip.cc",
5341 "test/zip-microkernel-tester.h",
5342 ] + MICROKERNEL_TEST_HDRS,
5343 deps = MICROKERNEL_TEST_DEPS,
5344)
5345
5346xnnpack_unit_test(
5347 name = "x8_lut_test",
5348 srcs = [
5349 "test/x8-lut.cc",
5350 "test/lut-microkernel-tester.h",
5351 ] + MICROKERNEL_TEST_HDRS,
5352 deps = MICROKERNEL_TEST_DEPS,
5353)
5354
5355xnnpack_unit_test(
5356 name = "x8_zip_test",
5357 srcs = [
5358 "test/x8-zip.cc",
5359 "test/zip-microkernel-tester.h",
5360 ] + MICROKERNEL_TEST_HDRS,
5361 deps = MICROKERNEL_TEST_DEPS,
5362)
5363
Marat Dukhan20c3b922020-03-10 03:45:06 -07005364########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005365
5366xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005367 name = "operator_size_test",
5368 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005369 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005370)
5371
Marat Dukhan20c3b922020-03-10 03:45:06 -07005372xnnpack_binary(
5373 name = "subgraph_size_test",
5374 srcs = ["test/subgraph-size.c"],
5375 deps = [":XNNPACK"],
5376)
5377
5378########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005379
5380xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005381 name = "abs_nc_test",
5382 srcs = [
5383 "test/abs-nc.cc",
5384 "test/abs-operator-tester.h",
5385 ],
5386 deps = OPERATOR_TEST_DEPS,
5387)
5388
5389xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005390 name = "add_nd_test",
5391 srcs = [
5392 "test/add-nd.cc",
5393 "test/binary-elementwise-operator-tester.h",
5394 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005395 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005396)
5397
5398xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005399 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005400 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005401 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005402 "test/argmax-pooling-operator-tester.h",
5403 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005404 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005405)
5406
5407xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005408 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005409 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005410 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 "test/average-pooling-operator-tester.h",
5412 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005413 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005414)
5415
5416xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005417 name = "bankers_rounding_nc_test",
5418 srcs = [
5419 "test/bankers-rounding-nc.cc",
5420 "test/bankers-rounding-operator-tester.h",
5421 ],
5422 deps = OPERATOR_TEST_DEPS,
5423)
5424
5425xnnpack_unit_test(
5426 name = "ceiling_nc_test",
5427 srcs = [
5428 "test/ceiling-nc.cc",
5429 "test/ceiling-operator-tester.h",
5430 ],
5431 deps = OPERATOR_TEST_DEPS,
5432)
5433
5434xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005435 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005436 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005437 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005438 "test/channel-shuffle-operator-tester.h",
5439 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005440 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005441)
5442
5443xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005444 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005445 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005446 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005447 "test/clamp-operator-tester.h",
5448 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005449 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005450)
5451
5452xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005453 name = "constant_pad_nd_test",
5454 srcs = [
5455 "test/constant-pad-nd.cc",
5456 "test/constant-pad-operator-tester.h",
5457 ],
5458 deps = OPERATOR_TEST_DEPS,
5459)
5460
5461xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005462 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005463 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005464 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465 "test/convolution-operator-tester.h",
5466 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005467 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005468)
5469
5470xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005471 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005472 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005473 "test/convolution-nchw.cc",
5474 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005476 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005477)
5478
5479xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005480 name = "copy_nc_test",
5481 srcs = [
5482 "test/copy-nc.cc",
5483 "test/copy-operator-tester.h",
5484 ],
5485 deps = OPERATOR_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005489 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005491 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005492 "test/deconvolution-operator-tester.h",
5493 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005494 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005495)
5496
5497xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005498 name = "divide_nd_test",
5499 srcs = [
5500 "test/binary-elementwise-operator-tester.h",
5501 "test/divide-nd.cc",
5502 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005503 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005504)
5505
5506xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005507 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005508 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005509 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005510 "test/fully-connected-operator-tester.h",
5511 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005512 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005513)
5514
5515xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005516 name = "floor_nc_test",
5517 srcs = [
5518 "test/floor-nc.cc",
5519 "test/floor-operator-tester.h",
5520 ],
5521 deps = OPERATOR_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005525 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005526 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005527 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005528 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005529 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005530 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005531)
5532
5533xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005534 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005535 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005536 "test/global-average-pooling-ncw.cc",
5537 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005538 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005539 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005540)
5541
5542xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005543 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005544 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005545 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005546 "test/hardswish-operator-tester.h",
5547 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005548 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005549)
5550
5551xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005552 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005553 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005554 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005555 "test/leaky-relu-operator-tester.h",
5556 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005557 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005558)
5559
5560xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005561 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005562 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005563 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005564 "test/max-pooling-operator-tester.h",
5565 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005566 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005567)
5568
5569xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005570 name = "maximum_nd_test",
5571 srcs = [
5572 "test/binary-elementwise-operator-tester.h",
5573 "test/maximum-nd.cc",
5574 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005575 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005576)
5577
5578xnnpack_unit_test(
5579 name = "minimum_nd_test",
5580 srcs = [
5581 "test/binary-elementwise-operator-tester.h",
5582 "test/minimum-nd.cc",
5583 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005584 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005585)
5586
5587xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005588 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005589 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005590 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005591 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005592 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005593 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005594)
5595
5596xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005597 name = "negate_nc_test",
5598 srcs = [
5599 "test/negate-nc.cc",
5600 "test/negate-operator-tester.h",
5601 ],
5602 deps = OPERATOR_TEST_DEPS,
5603)
5604
5605xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005606 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005607 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005608 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005609 "test/prelu-operator-tester.h",
5610 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005611 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005612)
5613
5614xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005615 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005616 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005617 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005618 "test/resize-bilinear-operator-tester.h",
5619 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005620 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005621)
5622
5623xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005624 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005625 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005626 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005627 "test/sigmoid-operator-tester.h",
5628 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005629 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005630)
5631
5632xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005633 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005634 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005635 "test/softmax-nc.cc",
5636 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005637 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005638 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005639)
5640
5641xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005642 name = "square_nc_test",
5643 srcs = [
5644 "test/square-nc.cc",
5645 "test/square-operator-tester.h",
5646 ],
5647 deps = OPERATOR_TEST_DEPS,
5648)
5649
5650xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005651 name = "square_root_nc_test",
5652 srcs = [
5653 "test/square-root-nc.cc",
5654 "test/square-root-operator-tester.h",
5655 ],
5656 deps = OPERATOR_TEST_DEPS,
5657)
5658
5659xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005660 name = "squared_difference_nd_test",
5661 srcs = [
5662 "test/binary-elementwise-operator-tester.h",
5663 "test/squared-difference-nd.cc",
5664 ],
5665 deps = OPERATOR_TEST_DEPS,
5666)
5667
5668xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005669 name = "subtract_nd_test",
5670 srcs = [
5671 "test/binary-elementwise-operator-tester.h",
5672 "test/subtract-nd.cc",
5673 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005674 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005675)
5676
5677xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005678 name = "truncation_nc_test",
5679 srcs = [
5680 "test/truncation-nc.cc",
5681 "test/truncation-operator-tester.h",
5682 ],
5683 deps = OPERATOR_TEST_DEPS,
5684)
5685
5686xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005687 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005688 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005689 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005690 "test/unpooling-operator-tester.h",
5691 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005692 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005693)
5694
Chao Mei6ddfc602020-05-13 22:29:36 -07005695############################### Misc unit tests ###############################
5696
5697xnnpack_unit_test(
5698 name = "memory_planner_test",
5699 srcs = [
5700 "test/memory-planner-test.cc",
5701 ],
5702 deps = [
5703 ":XNNPACK",
5704 ":memory_planner",
5705 ],
5706)
5707
Marat Dukhan08c4a432019-10-03 09:29:21 -07005708############################# Build configurations #############################
5709
Marat Dukhanb8642352019-10-30 15:43:02 -07005710# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005711config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005712 name = "xnn_enable_assembly_explicit_true",
5713 define_values = {"xnn_enable_assembly": "true"},
5714)
5715
5716# Disables usage of assembly kernels.
5717config_setting(
5718 name = "xnn_enable_assembly_explicit_false",
5719 define_values = {"xnn_enable_assembly": "false"},
5720)
5721
Marat Dukhan9de90e02020-06-18 16:04:12 -07005722# Enables usage of sparse inference.
5723config_setting(
5724 name = "xnn_enable_sparse_explicit_true",
5725 define_values = {"xnn_enable_sparse": "true"},
5726)
5727
5728# Disables usage of sparse inference.
5729config_setting(
5730 name = "xnn_enable_sparse_explicit_false",
5731 define_values = {"xnn_enable_sparse": "false"},
5732)
5733
Marat Dukhan05702cf2020-03-26 15:41:33 -07005734# Disables usage of HMP-aware optimizations.
5735config_setting(
5736 name = "xnn_enable_hmp_explicit_false",
5737 define_values = {"xnn_enable_hmp": "false"},
5738)
5739
Chao Mei6ddfc602020-05-13 22:29:36 -07005740# Enable usage of optimized memory allocation
5741config_setting(
5742 name = "xnn_enable_memopt_explicit_true",
5743 define_values = {"xnnpack_enable_memopt": "true"},
5744)
5745
5746# Disable usage of optimized memory allocation
5747config_setting(
5748 name = "xnn_enable_memopt_explicit_false",
5749 define_values = {"xnnpack_enable_memopt": "false"},
5750)
5751
Marat Dukhanb8642352019-10-30 15:43:02 -07005752# Builds with -c dbg
5753config_setting(
5754 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005755 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005756 "compilation_mode": "dbg",
5757 },
5758)
5759
5760# Builds with -c opt
5761config_setting(
5762 name = "optimized_build",
5763 values = {
5764 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005765 },
5766)
5767
5768config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005769 name = "linux_k8",
5770 values = {"cpu": "k8"},
5771)
5772
5773config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005774 name = "linux_arm",
5775 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005776)
5777
5778config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005779 name = "linux_armeabi",
5780 values = {"cpu": "armeabi"},
5781)
5782
5783config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005784 name = "linux_armhf",
5785 values = {"cpu": "armhf"},
5786)
5787
5788config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005789 name = "linux_armv7a",
5790 values = {"cpu": "armv7a"},
5791)
5792
5793config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005794 name = "linux_aarch64",
5795 values = {"cpu": "aarch64"},
5796)
5797
5798config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005799 name = "android",
5800 values = {"crosstool_top": "//external:android/crosstool"},
5801)
5802
5803config_setting(
5804 name = "android_armv7",
5805 values = {
5806 "crosstool_top": "//external:android/crosstool",
5807 "cpu": "armeabi-v7a",
5808 },
5809)
5810
5811config_setting(
5812 name = "android_arm64",
5813 values = {
5814 "crosstool_top": "//external:android/crosstool",
5815 "cpu": "arm64-v8a",
5816 },
5817)
5818
5819config_setting(
5820 name = "android_x86",
5821 values = {
5822 "crosstool_top": "//external:android/crosstool",
5823 "cpu": "x86",
5824 },
5825)
5826
5827config_setting(
5828 name = "android_x86_64",
5829 values = {
5830 "crosstool_top": "//external:android/crosstool",
5831 "cpu": "x86_64",
5832 },
5833)
5834
5835config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005836 name = "windows_x86_64",
5837 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005838)
5839
5840config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005841 name = "windows_x86_64_clang",
5842 values = {
5843 "compiler": "clang-cl",
5844 "cpu": "x64_windows",
5845 },
5846)
5847
5848config_setting(
5849 name = "windows_x86_64_mingw",
5850 values = {
5851 "compiler": "mingw-gcc",
5852 "cpu": "x64_windows",
5853 },
5854)
5855
5856config_setting(
5857 name = "windows_x86_64_msys",
5858 values = {
5859 "compiler": "msys-gcc",
5860 "cpu": "x64_windows",
5861 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005862)
5863
5864config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005865 name = "macos_x86_64",
5866 values = {
5867 "apple_platform_type": "macos",
5868 "cpu": "darwin",
5869 },
5870)
5871
5872config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005873 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005874 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005875)
5876
5877config_setting(
5878 name = "emscripten_wasm",
5879 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005880 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005881 "cpu": "wasm",
5882 },
5883)
5884
5885config_setting(
5886 name = "emscripten_wasmsimd",
5887 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005888 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005889 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005890 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005891 },
5892)
5893
5894config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005895 name = "ios_armv7",
5896 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005897 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005898 "cpu": "ios_armv7",
5899 },
5900)
5901
5902config_setting(
5903 name = "ios_arm64",
5904 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005905 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005906 "cpu": "ios_arm64",
5907 },
5908)
5909
5910config_setting(
5911 name = "ios_arm64e",
5912 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005913 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005914 "cpu": "ios_arm64e",
5915 },
5916)
5917
5918config_setting(
5919 name = "ios_x86",
5920 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005921 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005922 "cpu": "ios_i386",
5923 },
5924)
5925
5926config_setting(
5927 name = "ios_x86_64",
5928 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005929 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005930 "cpu": "ios_x86_64",
5931 },
5932)
5933
5934config_setting(
5935 name = "watchos_armv7k",
5936 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005937 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005938 "cpu": "watchos_armv7k",
5939 },
5940)
5941
5942config_setting(
5943 name = "watchos_arm64_32",
5944 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005945 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005946 "cpu": "watchos_arm64_32",
5947 },
5948)
5949
5950config_setting(
5951 name = "watchos_x86",
5952 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005953 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005954 "cpu": "watchos_i386",
5955 },
5956)
5957
5958config_setting(
5959 name = "watchos_x86_64",
5960 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005961 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005962 "cpu": "watchos_x86_64",
5963 },
5964)
5965
5966config_setting(
5967 name = "tvos_arm64",
5968 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005969 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005970 "cpu": "tvos_arm64",
5971 },
5972)
5973
5974config_setting(
5975 name = "tvos_x86_64",
5976 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005977 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005978 "cpu": "tvos_x86_64",
5979 },
5980)