blob: f284547713b92c4976f563424f6e73ad5a01e781 [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 Dukhan08b7a972020-07-14 18:17:29 -0700361 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
362 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
363 "src/qu8-dwconv/up1x9-minmax-scalar.c",
364 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
365 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
366 "src/qu8-gemm/2x2-minmax-scalar.c",
367 "src/qu8-igemm/2x2-minmax-scalar.c",
368 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700369 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700370 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700371 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700372 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700373 "src/x32-fill/scalar-float.c",
374 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700375 "src/x32-packx/x2-scalar.c",
376 "src/x32-packx/x3-scalar.c",
377 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700378 "src/x32-pad/scalar-int.c",
379 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700380 "src/x32-unpool/scalar.c",
381 "src/x32-zip/x2-scalar.c",
382 "src/x32-zip/x3-scalar.c",
383 "src/x32-zip/x4-scalar.c",
384 "src/x32-zip/xm-scalar.c",
385 "src/x8-lut/scalar.c",
386 "src/x8-zip/x2-scalar.c",
387 "src/x8-zip/x3-scalar.c",
388 "src/x8-zip/x4-scalar.c",
389 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700390 "src/requantization/precise-scalar.c",
391 "src/requantization/fp32-scalar.c",
392 "src/requantization/q31-scalar.c",
393 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700394]
395
Marat Dukhan436ebe62019-12-04 15:10:12 -0800396WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700397 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
398 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700399 "src/f32-clamp/gen/wasm-x1.c",
400 "src/f32-clamp/gen/wasm-x2.c",
401 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700402 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
403 "src/f32-dwconv/gen/up1x4-wasm.c",
404 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
405 "src/f32-dwconv/gen/up1x9-wasm.c",
406 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
407 "src/f32-dwconv/gen/up1x25-wasm.c",
408 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
409 "src/f32-dwconv/gen/up2x4-wasm.c",
410 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
411 "src/f32-dwconv/gen/up2x9-wasm.c",
412 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
413 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700414 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
416 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
417 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700418 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
419 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700420 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
422 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700424 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700426 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
427 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700428 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
429 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
430 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700431 "src/f32-gemm/gen/1x4-wasm.c",
432 "src/f32-gemm/gen/2x4-wasm.c",
433 "src/f32-gemm/gen/4x2-wasm.c",
434 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700435 "src/f32-gemm/gen/1x4-relu-wasm.c",
436 "src/f32-gemm/gen/2x4-relu-wasm.c",
437 "src/f32-gemm/gen/4x2-relu-wasm.c",
438 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700439 "src/f32-gemm/gen/1x4-minmax-wasm.c",
440 "src/f32-gemm/gen/2x4-minmax-wasm.c",
441 "src/f32-gemm/gen/4x2-minmax-wasm.c",
442 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700443 "src/f32-hswish/gen/hswish-wasm-x1.c",
444 "src/f32-hswish/gen/hswish-wasm-x2.c",
445 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700446 "src/f32-igemm/gen/1x4-wasm.c",
447 "src/f32-igemm/gen/2x4-wasm.c",
448 "src/f32-igemm/gen/4x2-wasm.c",
449 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700450 "src/f32-igemm/gen/1x4-relu-wasm.c",
451 "src/f32-igemm/gen/2x4-relu-wasm.c",
452 "src/f32-igemm/gen/4x2-relu-wasm.c",
453 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700454 "src/f32-igemm/gen/1x4-minmax-wasm.c",
455 "src/f32-igemm/gen/2x4-minmax-wasm.c",
456 "src/f32-igemm/gen/4x2-minmax-wasm.c",
457 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700458 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
459 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
460 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700461 "src/f32-prelu/gen/wasm-2x1.c",
462 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700463 "src/f32-relu/gen/wasm-x1.c",
464 "src/f32-relu/gen/wasm-x2.c",
465 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700466 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
467 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
468 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
469 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
470 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
471 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
472 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
473 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
474 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
475 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
476 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
477 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800478 "src/f32-vbinary/gen/vmax-wasm-x1.c",
479 "src/f32-vbinary/gen/vmax-wasm-x2.c",
480 "src/f32-vbinary/gen/vmax-wasm-x4.c",
481 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
482 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
483 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
484 "src/f32-vbinary/gen/vmin-wasm-x1.c",
485 "src/f32-vbinary/gen/vmin-wasm-x2.c",
486 "src/f32-vbinary/gen/vmin-wasm-x4.c",
487 "src/f32-vbinary/gen/vminc-wasm-x1.c",
488 "src/f32-vbinary/gen/vminc-wasm-x2.c",
489 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700490 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
491 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
492 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
493 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
494 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
495 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
496 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
497 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
498 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
499 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
500 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
501 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
502 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
503 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
504 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
505 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
506 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
507 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700508 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
509 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
510 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700511 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
512 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
513 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800514]
515
Marat Dukhan290055c2020-06-09 12:24:29 -0700516WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700517 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
518 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
519 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700520 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
521 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
522 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
523 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700524 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
525 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
526 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
527 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700528 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
529 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
530 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
531 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
532 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
533 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700534 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
535 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
536 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
537 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
538 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
539 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
540 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
541 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
542 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
547 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
548 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
549 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
550 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
551 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
552 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
553 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
554 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700558 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
559 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
560 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700561 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700562 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
563 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
564 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
565 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700566 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
567 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
568 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
569 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
570 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
571 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
572 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
573 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
574 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
575 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
576 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
577 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
578 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
579 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
580 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
581 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
582 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
583 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
584 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
585 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
586 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
587 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
588 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
589 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
590 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
591 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
592 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
593 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
594 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
595 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700596 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
597 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700598 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
599 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
600 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
601 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
602 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
603 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
604 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
605 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
606 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
607 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
608 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
609 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
610 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
611 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
612 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
613 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
614 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
615 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
616 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
617 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
618 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
619 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
620 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
621 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
622 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
623 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
624 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
625 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
626 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
627 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700628 "src/f32-ibilinear/gen/wasmsimd-c4.c",
629 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700630 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
631 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
632 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700633 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700634 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
635 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
636 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
637 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700638 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
639 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
640 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
641 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
642 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
643 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
644 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
645 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
646 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
647 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
648 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
649 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
650 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
651 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
652 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
653 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
654 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
655 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
656 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
657 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
658 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
659 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
660 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
661 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
662 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
663 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
664 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
665 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
666 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
667 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700668 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
669 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700670 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
671 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
672 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
673 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700674 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
675 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
676 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700677 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
678 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700679 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
680 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
681 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
682 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700683 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
684 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700685 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
686 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
687 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700688 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
689 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700690 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
691 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
692 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
693 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
694 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
695 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
696 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700697 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
698 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700699 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
700 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
701 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
702 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700703 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
704 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
705 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
706 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
707 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
708 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
709 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
710 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
711 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
712 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
713 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
714 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700715 "src/f32-relu/gen/wasmsimd-x4.c",
716 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700717 "src/f32-rmax/wasmsimd-arm.c",
718 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700719 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
720 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
721 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
722 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
723 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
724 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700725 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
726 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
727 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
728 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
729 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
730 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700731 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
732 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
733 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
734 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700735 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
736 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700737 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
738 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
739 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
740 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700741 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
742 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700743 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
744 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
745 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
746 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700747 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
748 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700749 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
750 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
751 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
752 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700753 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
754 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700755 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
756 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
757 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
758 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700763 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
764 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
765 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
766 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700767 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
768 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
769 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
770 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700771 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
772 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
773 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
774 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700775 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
776 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700777 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
778 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
779 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
780 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700781 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
782 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700783 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
784 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
785 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
786 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700787 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
788 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700789 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
790 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
791 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
792 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700793 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
794 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700795 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
796 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700797 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
798 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700799 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
800 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
801 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
802 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700803 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
804 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700805 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
806 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
807 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
808 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700809 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
810 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700811 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
812 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
813 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
814 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700815 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
816 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
817 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
818 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700819 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
820 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
821 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
822 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
823 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
824 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
825 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
826 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
827 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
828 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
829 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
830 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
831 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
832 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700833 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
834 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700835 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
836 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
837 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
838 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
839 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
840 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700841 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700842 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700843 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700844 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700845 "src/x32-zip/x2-wasmsimd.c",
846 "src/x32-zip/x3-wasmsimd.c",
847 "src/x32-zip/x4-wasmsimd.c",
848 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700849 "src/math/roundne-wasmsimd-addsub.c",
850 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700851 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700852 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700853 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700854 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700855 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700856 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
857 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700858 "src/requantization/fp32-wasmsimd.c",
859]
860
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800861PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800862 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800863 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800864 "src/f32-argmaxpool/9x-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700865 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700866 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700867 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700868 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700869 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700870 "src/f32-gavgpool-cw/psimd-x4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700871 "src/f32-spmm/gen/4x1-minmax-psimd.c",
872 "src/f32-spmm/gen/8x1-minmax-psimd.c",
873 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700874 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700875 "src/x32-packx/x4-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700876]
877
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800878PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan249cd6a2020-07-17 14:05:30 -0700879 "src/requantization/precise-psimd.c",
880 "src/requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800881]
882
Marat Dukhan08c4a432019-10-03 09:29:21 -0700883# ISA-specific micro-kernels
884NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700885 "src/f32-argmaxpool/4x-neon-c4.c",
886 "src/f32-argmaxpool/9p8x-neon-c4.c",
887 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700888 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
889 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700890 "src/f32-clamp/gen/neon-x4.c",
891 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700892 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
893 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700894 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
895 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700896 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
897 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700898 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
899 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700900 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
901 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
902 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
903 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700904 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
905 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
906 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
907 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700908 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
909 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
910 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
911 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700912 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700913 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
914 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700915 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
916 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
917 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
918 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
919 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
920 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
921 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
922 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
923 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
924 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
925 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
926 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
927 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
928 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
929 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
930 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
931 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
932 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
933 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
934 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
935 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
936 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
937 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
938 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
939 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
940 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
941 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
942 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
943 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
944 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
945 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700946 "src/f32-hswish/gen/hswish-neon-x4.c",
947 "src/f32-hswish/gen/hswish-neon-x8.c",
948 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700949 "src/f32-ibilinear/gen/neon-c4.c",
950 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700951 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
952 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
953 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
954 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
955 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
956 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
957 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
958 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
959 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
960 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
961 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
962 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
963 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
964 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
965 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
966 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700967 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
968 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
969 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700970 "src/f32-ppmm/gen/4x8-minmax-neon.c",
971 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700972 "src/f32-prelu/gen/neon-1x4.c",
973 "src/f32-prelu/gen/neon-1x8.c",
974 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800975 "src/f32-prelu/gen/neon-2x4.c",
976 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700977 "src/f32-prelu/gen/neon-2x16.c",
978 "src/f32-prelu/gen/neon-4x4.c",
979 "src/f32-prelu/gen/neon-4x8.c",
980 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800981 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
982 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
983 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
984 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
985 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
986 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
987 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
988 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
989 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001005 "src/f32-relu/gen/neon-x4.c",
1006 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001007 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001008 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001009 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1010 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1011 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1012 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1013 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1014 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1015 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1016 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1017 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1018 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1019 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1020 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1021 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1022 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1023 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1026 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001027 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1028 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1029 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1030 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001031 "src/f32-vbinary/gen/vmax-neon-x4.c",
1032 "src/f32-vbinary/gen/vmax-neon-x8.c",
1033 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1034 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1035 "src/f32-vbinary/gen/vmin-neon-x4.c",
1036 "src/f32-vbinary/gen/vmin-neon-x8.c",
1037 "src/f32-vbinary/gen/vminc-neon-x4.c",
1038 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001039 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1040 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1041 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1042 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1043 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1044 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001045 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1046 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1047 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1048 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001049 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1050 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1051 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1052 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001053 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1054 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001055 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1056 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001057 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1058 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1059 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1060 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1061 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1062 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1063 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1064 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001065 "src/f32-vunary/gen/vabs-neon-x4.c",
1066 "src/f32-vunary/gen/vabs-neon-x8.c",
1067 "src/f32-vunary/gen/vneg-neon-x4.c",
1068 "src/f32-vunary/gen/vneg-neon-x8.c",
1069 "src/f32-vunary/gen/vsqr-neon-x4.c",
1070 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001071 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1072 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1073 "src/qu8-dwconv/up8x9-minmax-neon.c",
1074 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1075 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1076 "src/qu8-gemm/4x8-minmax-neon.c",
1077 "src/qu8-gemm/8x8-minmax-neon.c",
1078 "src/qu8-igemm/4x8-minmax-neon.c",
1079 "src/qu8-igemm/8x8-minmax-neon.c",
1080 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001081 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001082 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001083 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001084 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001085 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001086 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001087 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001088 "src/x32-zip/x2-neon.c",
1089 "src/x32-zip/x3-neon.c",
1090 "src/x32-zip/x4-neon.c",
1091 "src/x32-zip/xm-neon.c",
1092 "src/x8-zip/x2-neon.c",
1093 "src/x8-zip/x3-neon.c",
1094 "src/x8-zip/x4-neon.c",
1095 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001096 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001097 "src/math/roundd-neon-addsub.c",
1098 "src/math/roundd-neon-cvt.c",
1099 "src/math/roundu-neon-addsub.c",
1100 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001101 "src/math/roundz-neon-addsub.c",
1102 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001103 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001104 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1105 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1106 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1107 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1108 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1109 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001110 "src/math/sqrt-neon-nr1rsqrts.c",
1111 "src/math/sqrt-neon-nr2rsqrts.c",
1112 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001113 "src/requantization/precise-neon.c",
1114 "src/requantization/fp32-neon.c",
1115 "src/requantization/q31-neon.c",
1116 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001117]
1118
1119NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001120 "src/f32-ibilinear/gen/neonfma-c4.c",
1121 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001122 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1123 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1124 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1125 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1126 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1127 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1128 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1129 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1130 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001131 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1132 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1133 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1134 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001135 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1136 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1137 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1138 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001139 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1140 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1141 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1142 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001143 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1144 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1145 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1146 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1147 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1148 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1149 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1150 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1151 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1152 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1153 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1154 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1155 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1156 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1157 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1158 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1159 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1160 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001161 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1162 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001163 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1164 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1165 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1166 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1167 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1168 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1169 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1170 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1171 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1172 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1173 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1177 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1178 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1179 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1180 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1181 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1182 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1185 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001187 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001241 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1242 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001243 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1244 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1245 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1246 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1247 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1248 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1249 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1250 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1251 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1252 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1253 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1254 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1255 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1256 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1257 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1258 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1259 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1260 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1261 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1262 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001263 "src/math/exp-neonfma-lut64-p2.c",
1264 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001265 "src/math/expminus-neonfma-lut2048-p1.c",
1266 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001267 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001268 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1269 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1270 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1271 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1272 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1273 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1274 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1275 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1276 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1277 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1278 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1279 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1280 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1281 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1282 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1283 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1284 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1285 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001286 "src/math/sqrt-neonfma-nr1fma.c",
1287 "src/math/sqrt-neonfma-nr2fma.c",
1288 "src/math/sqrt-neonfma-nr3fma.c",
1289 "src/math/sqrt-neonfma-nr2fma1adj.c",
1290 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001291]
1292
1293AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001294 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1295 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1296 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1297 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1298 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1299 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001300 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1301 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1302 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1303 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1304 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1305 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1306 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1307 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1308 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1309 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1310 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1311 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1312 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1313 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1314 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1315 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1316 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1317 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1318 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1319 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001320 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1321 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001322 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1323 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001324 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1325 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001326 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1327 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001328 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1329 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1330 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1331 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1332 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001333 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1340 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1341 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1342 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001351 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1352 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1353 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1354 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1355 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1356 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1357 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1358 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1359 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1360 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1361 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1362 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1363 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1364 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1365 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1366 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1367 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1368 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001369 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1370 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001371 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1372 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1373 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1374 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1375 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1376 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001377]
1378
Marat Dukhan8853b822020-05-07 12:19:01 -07001379NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001380 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1381 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1382 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1383 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1384 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1385 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1386 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1387 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001388 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001389 "src/math/roundd-neonv8.c",
1390 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001391 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001392]
1393
Marat Dukhan08c4a432019-10-03 09:29:21 -07001394AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001395 "src/f16-clamp/gen/neonfp16arith-x16.c",
1396 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001397 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1398 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1399 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1400 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1401 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1402 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1403 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1404 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1405 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1406 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1407 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1408 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001409 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1410 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001411 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1412 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001413 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001414 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001415 "src/f16-relu/gen/neonfp16arith-x16.c",
1416 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001417 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1418 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1419 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1420 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1421 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1422 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1423 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1424 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1425 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1426 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1427 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1428 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1429 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1430 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1431 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1432 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1433 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1434 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1435 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1436 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1437 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1438 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1439 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1440 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1441 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1442 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1443 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1444 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001445 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1446 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001447 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1448 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1449 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1450 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1451 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1452 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1453 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1454 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001455 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1456 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1457 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1458 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001459 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1460 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1461 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1462 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1463 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1464 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1465 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1466 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1467 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1468 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1469 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1470 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001471 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1472 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1473 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1474 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1475 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1476 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1477 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1478 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001479]
1480
1481SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001482 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1483 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001484 "src/f32-clamp/gen/sse-x4.c",
1485 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001486 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1487 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001488 "src/f32-dwconv-chw/3x3p1-sse.c",
1489 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001490 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1491 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1492 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1493 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1494 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1495 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1496 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1497 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1498 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1499 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1500 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1501 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001502 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001503 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1504 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001505 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1506 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1507 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1508 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1509 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1510 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1511 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1512 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1513 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1514 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1515 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1516 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1517 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001518 "src/f32-hswish/gen/hswish-sse-x4.c",
1519 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001520 "src/f32-ibilinear/gen/sse-c4.c",
1521 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001522 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1523 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1524 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1525 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1526 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1527 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1528 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001529 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1530 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1531 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001532 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001533 "src/f32-prelu/gen/sse-2x4.c",
1534 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001535 "src/f32-relu/gen/sse-x4.c",
1536 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001537 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001538 "src/f32-spmm/gen/4x1-minmax-sse.c",
1539 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001540 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001541 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1542 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1543 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1544 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1545 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1546 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1547 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1548 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001549 "src/f32-vbinary/gen/vmax-sse-x4.c",
1550 "src/f32-vbinary/gen/vmax-sse-x8.c",
1551 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1552 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1553 "src/f32-vbinary/gen/vmin-sse-x4.c",
1554 "src/f32-vbinary/gen/vmin-sse-x8.c",
1555 "src/f32-vbinary/gen/vminc-sse-x4.c",
1556 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001557 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1558 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1559 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1560 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1561 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1562 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1563 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1564 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001565 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1566 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1567 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1568 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001569 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1570 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1571 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1572 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001573 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1574 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001575 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1576 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001577 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1578 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001579 "src/f32-vunary/gen/vabs-sse-x4.c",
1580 "src/f32-vunary/gen/vabs-sse-x8.c",
1581 "src/f32-vunary/gen/vneg-sse-x4.c",
1582 "src/f32-vunary/gen/vneg-sse-x8.c",
1583 "src/f32-vunary/gen/vsqr-sse-x4.c",
1584 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001585 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001586 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001587 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001588 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001589 "src/math/roundd-sse-addsub.c",
1590 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001591 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001592 "src/math/sqrt-sse-hh1mac.c",
1593 "src/math/sqrt-sse-nr1mac.c",
1594 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595]
1596
1597SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001598 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1599 "src/f32-argmaxpool/4x-sse2-c4.c",
1600 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001601 "src/f32-prelu/gen/sse2-2x4.c",
1602 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001603 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1604 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1605 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1606 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1607 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1608 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1609 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1610 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1611 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1612 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1613 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1614 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001615 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001616 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001617 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001618 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001619 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1620 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001621 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1622 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001623 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1624 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1625 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1626 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1627 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1628 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1629 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1630 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001631 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1632 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1633 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1634 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1635 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1636 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1637 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1638 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1639 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001640 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001641 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001642 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001643 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001644 "src/x32-zip/x2-sse2.c",
1645 "src/x32-zip/x3-sse2.c",
1646 "src/x32-zip/x4-sse2.c",
1647 "src/x32-zip/xm-sse2.c",
1648 "src/x8-zip/x2-sse2.c",
1649 "src/x8-zip/x3-sse2.c",
1650 "src/x8-zip/x4-sse2.c",
1651 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001652 "src/math/exp-sse2-p5.c",
1653 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001654 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001655 "src/math/roundd-sse2-cvt.c",
1656 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001657 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001658 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001659 "src/requantization/precise-sse2.c",
1660 "src/requantization/fp32-sse2.c",
1661 "src/requantization/q31-sse2.c",
1662 "src/requantization/gemmlowp-sse2.c",
1663]
1664
1665SSSE3_UKERNELS = [
1666 "src/requantization/precise-ssse3.c",
1667 "src/requantization/q31-ssse3.c",
1668 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001669]
1670
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001671SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001672 "src/f32-prelu/gen/sse41-2x4.c",
1673 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001674 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001675 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001676 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001677 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001678 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1679 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001680 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1681 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001682 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1683 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1684 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1685 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1686 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1687 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1688 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1689 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001690 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001691 "src/math/roundd-sse41.c",
1692 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001693 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001694 "src/requantization/precise-sse4.c",
1695 "src/requantization/q31-sse4.c",
1696 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001697]
1698
Marat Dukhan08c4a432019-10-03 09:29:21 -07001699AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001700 "src/f32-clamp/gen/avx-x8.c",
1701 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001702 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1703 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1704 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1705 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1706 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1707 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1708 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1709 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1710 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1711 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1712 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1713 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1714 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1715 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1716 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1717 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1718 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1719 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1720 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1721 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1722 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1723 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1724 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1725 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1726 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1727 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1728 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1729 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1730 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1731 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001732 "src/f32-hswish/gen/hswish-avx-x8.c",
1733 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001734 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1735 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1736 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1737 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1738 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1739 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1740 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1741 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1742 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001743 "src/f32-prelu/gen/avx-2x8.c",
1744 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001745 "src/f32-relu/gen/avx-x16.c",
1746 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001748 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1749 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1750 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1751 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1752 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1753 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1754 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1755 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001756 "src/f32-vbinary/gen/vmax-avx-x8.c",
1757 "src/f32-vbinary/gen/vmax-avx-x16.c",
1758 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1759 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1760 "src/f32-vbinary/gen/vmin-avx-x8.c",
1761 "src/f32-vbinary/gen/vmin-avx-x16.c",
1762 "src/f32-vbinary/gen/vminc-avx-x8.c",
1763 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001764 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1765 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1766 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1767 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1768 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1769 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1770 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1771 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001772 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1773 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1774 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1775 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001776 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1777 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1778 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1779 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001780 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1781 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001782 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001783 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1784 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1785 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1786 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1787 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1788 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1789 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1790 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001791 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1792 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001793 "src/f32-vunary/gen/vabs-avx-x8.c",
1794 "src/f32-vunary/gen/vabs-avx-x16.c",
1795 "src/f32-vunary/gen/vneg-avx-x8.c",
1796 "src/f32-vunary/gen/vneg-avx-x16.c",
1797 "src/f32-vunary/gen/vsqr-avx-x8.c",
1798 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001799]
1800
Marat Dukhanfda12b82019-11-21 12:27:59 -08001801FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001802 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1803 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1804 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1805 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1806 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1807 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1808 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1809 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1810 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1811 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1812 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1813 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1814 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1815 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1816 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1817 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1818 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1819 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1820 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1821 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1822 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1823 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1824 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1825 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1826 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1827 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1828 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1829 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1830 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1831 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1832 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1833 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1834 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1835 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1836 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1837 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1838 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1839 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1840 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1841 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001842 "src/f32-hswish/gen/hswish-fma3-x8.c",
1843 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001844 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1845 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1846 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1847 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1848 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1849 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1850 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1851 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1852 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1853 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1854 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1855 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1856 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1857 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001858 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1859 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1860 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1861 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1862 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1863 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1864 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1865 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001866 "src/math/sqrt-fma3-nr1fma.c",
1867 "src/math/sqrt-fma3-nr2fma.c",
1868 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001869]
1870
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001871AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001872 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1873 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1874 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1875 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1876 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1877 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1878 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1879 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1880 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1881 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1882 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1883 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1884 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1885 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1886 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1887 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1888 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1889 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1890 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1891 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1892 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1893 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1894 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1895 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1896 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1897 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1898 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1899 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1900 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1901 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1902 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1903 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1904 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1905 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1906 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1907 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001908 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1909 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1910 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1911 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1912 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1913 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1914 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1915 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1916 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1917 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1918 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1919 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1920 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1921 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1922 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1923 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1924 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1925 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1926 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1927 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1928 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1929 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1930 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1931 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1932 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1933 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1934 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1935 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1936 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1937 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001938 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1939 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1940 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1941 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1942 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1943 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1944 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1945 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1946 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1947 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1948 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1949 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1950 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1951 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1952 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1953 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1954 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1955 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1956 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1957 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1958 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1959 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1960 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1961 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001962 "src/math/exp-avx2-p5.c",
1963 "src/math/exp-avx2-perm-p3.c",
1964 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001965 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001966 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001967 "src/math/sigmoid-avx2-rr2-p5-div.c",
1968 "src/math/sigmoid-avx2-rr1-p5-div.c",
1969 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1970 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1971 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1972 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001973]
1974
Marat Dukhan08c4a432019-10-03 09:29:21 -07001975AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001976 "src/f32-clamp/gen/avx512f-x16.c",
1977 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001978 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1979 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1980 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1981 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1982 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1983 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1984 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1985 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1986 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1987 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1988 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1989 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1990 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1991 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1992 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1993 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1994 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1995 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1996 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1997 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1998 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1999 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2000 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2001 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002002 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2003 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002004 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2005 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2006 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2007 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2008 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2009 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002010 "src/f32-prelu/gen/avx512f-2x16.c",
2011 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002012 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2013 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2014 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2015 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2016 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2017 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2018 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2019 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2020 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2021 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2022 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2023 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2024 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2025 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2026 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2027 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2028 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2029 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2030 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2031 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2032 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2033 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2034 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2035 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2036 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2037 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2038 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2039 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2040 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2041 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2042 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2043 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2044 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2045 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2046 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2047 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002048 "src/f32-relu/gen/avx512f-x16.c",
2049 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002050 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002051 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2052 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2053 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2054 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2055 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2056 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2057 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2058 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002059 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2060 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2061 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2062 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2063 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2064 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2065 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2066 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002067 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2068 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2069 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2070 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2071 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2072 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2073 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2074 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002075 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2076 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2077 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2078 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002079 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2080 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2081 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2082 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002083 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2084 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002085 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002086 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2087 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2088 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2089 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2090 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2091 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2092 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2093 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2094 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2095 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2096 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2097 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2098 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2099 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2100 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2101 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2102 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2103 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2104 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2105 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2106 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2107 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2108 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2109 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002110 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2111 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2112 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2113 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2114 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2115 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2116 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2117 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002118 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2119 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2120 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2121 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2122 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2123 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2124 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2125 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002126 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2127 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2128 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2129 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2130 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2131 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002132 "src/math/exp-avx512f-p5-scalef.c",
2133 "src/math/exp-avx512f-p5.c",
2134 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002135 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002136 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002137 "src/math/sqrt-avx512f-nr1fma.c",
2138 "src/math/sqrt-avx512f-nr2fma.c",
2139 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002140]
2141
2142AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002143 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002144 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002145 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2146 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002147 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002148 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002149 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002150 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002151 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2152 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002153 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2154 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2155 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2156 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002157]
2158
2159AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002160 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002161 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2162 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002163 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002164 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2165 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002166 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2167 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2168 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002169 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002170 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2171 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2172 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002173 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002174 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2175 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002176 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002177 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2178 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2179 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2180 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2181 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2182 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2183 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2184 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2185 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2186 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2187 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2188 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2189 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2190 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2191 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2192 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2193 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2194 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002195 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2196 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002197 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002198 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2199 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2200 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2201 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2202 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2203 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2204 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2205 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2206 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2207 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2208 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2209 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2210 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2211 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2212 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2213 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2214 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2215 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002216 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2217 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2218 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2219 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2220 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2221 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2222 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2223 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2224 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2225 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2226 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2227 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2228 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2229 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2230 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2231 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2232 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2233 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002234]
2235
Marat Dukhan1b354632020-03-23 12:50:22 -07002236INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002237 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238 "src/xnnpack/argmaxpool.h",
2239 "src/xnnpack/avgpool.h",
2240 "src/xnnpack/clamp.h",
2241 "src/xnnpack/common.h",
2242 "src/xnnpack/conv.h",
2243 "src/xnnpack/dwconv.h",
2244 "src/xnnpack/gavgpool.h",
2245 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002246 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002247 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002248 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002249 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002250 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002251 "src/xnnpack/lut.h",
2252 "src/xnnpack/math.h",
2253 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002254 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002255 "src/xnnpack/packx.h",
2256 "src/xnnpack/pad.h",
2257 "src/xnnpack/params.h",
2258 "src/xnnpack/pavgpool.h",
2259 "src/xnnpack/ppmm.h",
2260 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002261 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002262 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002263 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002264 "src/xnnpack/rmax.h",
2265 "src/xnnpack/scalar-utils.h",
2266 "src/xnnpack/spmm.h",
2267 "src/xnnpack/unpool.h",
2268 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002269 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002270 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002271 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002272 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002273 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002274 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002275 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002276]
2277
2278INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002279 "include/xnnpack.h",
2280 "src/xnnpack/allocator.h",
2281 "src/xnnpack/compute.h",
2282 "src/xnnpack/im2col.h",
2283 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002284 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002285 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002286 "src/xnnpack/operator.h",
2287 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002288 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002290 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002291 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002292]
2293
Marat Dukhan1b354632020-03-23 12:50:22 -07002294ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002295 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002296]
2297
Marat Dukhan1b354632020-03-23 12:50:22 -07002298MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002299 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002300 "include/xnnpack.h",
2301]
2302
Marat Dukhan1b354632020-03-23 12:50:22 -07002303MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002304 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002305 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002306 "src/xnnpack/requantization.h",
2307 "include/xnnpack.h",
2308]
2309
2310OPERATOR_TEST_PARAMS_HDRS = [
2311 "src/xnnpack/params.h",
2312 "src/xnnpack/common.h",
2313]
2314
2315WEIGHTS_PACK_HDRS = [
2316 "src/xnnpack/pack.h",
2317 "src/xnnpack/operator.h",
2318 "src/xnnpack/compute.h",
2319]
2320
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002321LOGGING_COPTS = select({
2322 # No logging in optimized mode
2323 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2324 # Full logging in debug mode
2325 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2326 # Error-only logging in default (fastbuild) mode
2327 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2328})
2329
Marat Dukhan3b59de22020-06-03 20:15:19 -07002330LOGGING_SRCS = select({
2331 # No logging in optimized mode
2332 ":optimized_build": [],
2333 "//conditions:default": [
2334 "src/operator-strings.c",
2335 "src/subgraph-strings.c",
2336 ],
2337})
2338
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002339LOGGING_HDRS = [
2340 "src/xnnpack/log.h",
2341]
2342
Marat Dukhan08c4a432019-10-03 09:29:21 -07002343xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002344 name = "tables",
2345 srcs = TABLE_SRCS,
2346 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002347 gcc_copts = xnnpack_gcc_std_copts(),
2348 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002349)
2350
2351xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352 name = "scalar_ukernels",
2353 srcs = SCALAR_UKERNELS,
2354 hdrs = INTERNAL_HDRS,
2355 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002356 gcc_copts = xnnpack_gcc_std_copts(),
2357 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002358 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002359 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002360 "@FP16",
2361 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002362 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002363 ],
2364)
2365
2366xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002367 name = "scalar_ukernels_test_mode",
2368 srcs = SCALAR_UKERNELS,
2369 hdrs = INTERNAL_HDRS,
2370 aarch32_copts = ["-marm"],
2371 copts = [
2372 "-UNDEBUG",
2373 "-DXNN_TEST_MODE=1",
2374 ],
2375 gcc_copts = xnnpack_gcc_std_copts(),
2376 msvc_copts = xnnpack_msvc_std_copts(),
2377 deps = [
2378 ":tables",
2379 "@FP16",
2380 "@FXdiv",
2381 "@pthreadpool",
2382 ],
2383)
2384
2385xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002386 name = "wasm_ukernels",
2387 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002388 gcc_copts = xnnpack_gcc_std_copts(),
2389 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002390 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002391 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002392 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002393 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002394 "@FP16",
2395 "@FXdiv",
2396 "@pthreadpool",
2397 ],
2398)
2399
2400xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002401 name = "wasm_ukernels_test_mode",
2402 hdrs = INTERNAL_HDRS,
2403 copts = [
2404 "-UNDEBUG",
2405 "-DXNN_TEST_MODE=1",
2406 ],
2407 gcc_copts = xnnpack_gcc_std_copts(),
2408 msvc_copts = xnnpack_msvc_std_copts(),
2409 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002410 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002411 deps = [
2412 ":tables",
2413 "@FP16",
2414 "@FXdiv",
2415 "@pthreadpool",
2416 ],
2417)
2418
2419xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002420 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002421 hdrs = INTERNAL_HDRS,
2422 aarch32_copts = [
2423 "-marm",
2424 "-mfpu=neon",
2425 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002426 gcc_copts = xnnpack_gcc_std_copts(),
2427 gcc_x86_copts = ["-msse2"],
2428 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002429 optimized_copts = [
2430 "-O3",
2431 "-ffast-math",
2432 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002433 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002434 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002435 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002436 "@FP16",
2437 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002438 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002439 ],
2440)
2441
2442xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002443 name = "psimd_fastmath_ukernels_test_mode",
2444 hdrs = INTERNAL_HDRS,
2445 aarch32_copts = [
2446 "-marm",
2447 "-mfpu=neon",
2448 ],
2449 copts = [
2450 "-UNDEBUG",
2451 "-DXNN_TEST_MODE=1",
2452 ],
2453 gcc_copts = xnnpack_gcc_std_copts(),
2454 gcc_x86_copts = ["-msse2"],
2455 msvc_copts = xnnpack_msvc_std_copts(),
2456 optimized_copts = [
2457 "-O3",
2458 "-ffast-math",
2459 ],
2460 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2461 deps = [
2462 ":tables",
2463 "@FP16",
2464 "@psimd",
2465 "@pthreadpool",
2466 ],
2467)
2468
2469xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002470 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002471 hdrs = INTERNAL_HDRS,
2472 aarch32_copts = [
2473 "-marm",
2474 "-mfpu=neon",
2475 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002476 gcc_copts = xnnpack_gcc_std_copts(),
2477 gcc_x86_copts = ["-msse2"],
2478 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002479 optimized_copts = [
2480 "-O3",
2481 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002482 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002483 deps = [
2484 ":tables",
2485 "@FP16",
2486 "@psimd",
2487 "@pthreadpool",
2488 ],
2489)
2490
2491xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002492 name = "psimd_accmath_ukernels_test_mode",
2493 hdrs = INTERNAL_HDRS,
2494 aarch32_copts = [
2495 "-marm",
2496 "-mfpu=neon",
2497 ],
2498 copts = [
2499 "-UNDEBUG",
2500 "-DXNN_TEST_MODE=1",
2501 ],
2502 gcc_copts = xnnpack_gcc_std_copts(),
2503 gcc_x86_copts = ["-msse2"],
2504 msvc_copts = xnnpack_msvc_std_copts(),
2505 optimized_copts = [
2506 "-O3",
2507 ],
2508 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2509 deps = [
2510 ":tables",
2511 "@FP16",
2512 "@psimd",
2513 "@pthreadpool",
2514 ],
2515)
2516
2517xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002518 name = "neon_ukernels",
2519 hdrs = INTERNAL_HDRS,
2520 aarch32_copts = [
2521 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002522 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002523 "-mfpu=neon",
2524 ],
2525 aarch32_srcs = NEON_UKERNELS,
2526 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002527 gcc_copts = xnnpack_gcc_std_copts(),
2528 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002529 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002530 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002531 "@FP16",
2532 "@pthreadpool",
2533 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002534)
2535
2536xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002537 name = "neon_ukernels_test_mode",
2538 hdrs = INTERNAL_HDRS,
2539 aarch32_copts = [
2540 "-marm",
2541 "-march=armv7-a",
2542 "-mfpu=neon",
2543 ],
2544 aarch32_srcs = NEON_UKERNELS,
2545 aarch64_srcs = NEON_UKERNELS,
2546 copts = [
2547 "-UNDEBUG",
2548 "-DXNN_TEST_MODE=1",
2549 ],
2550 gcc_copts = xnnpack_gcc_std_copts(),
2551 msvc_copts = xnnpack_msvc_std_copts(),
2552 deps = [
2553 ":tables",
2554 "@FP16",
2555 "@pthreadpool",
2556 ],
2557)
2558
2559xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002560 name = "neonfma_ukernels",
2561 hdrs = INTERNAL_HDRS,
2562 aarch32_copts = [
2563 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002564 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002565 "-mfpu=neon-vfpv4",
2566 ],
2567 aarch32_srcs = NEONFMA_UKERNELS,
2568 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002569 apple_aarch32_copts = [
2570 "-mcpu=swift",
2571 "-mtune=generic",
2572 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002573 gcc_copts = xnnpack_gcc_std_copts(),
2574 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002575 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002576 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002577 "@FP16",
2578 "@pthreadpool",
2579 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002580)
2581
2582xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002583 name = "neonfma_ukernels_test_mode",
2584 hdrs = INTERNAL_HDRS,
2585 aarch32_copts = [
2586 "-marm",
2587 "-march=armv7-a",
2588 "-mfpu=neon-vfpv4",
2589 ],
2590 aarch32_srcs = NEONFMA_UKERNELS,
2591 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002592 apple_aarch32_copts = [
2593 "-mcpu=swift",
2594 "-mtune=generic",
2595 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002596 copts = [
2597 "-UNDEBUG",
2598 "-DXNN_TEST_MODE=1",
2599 ],
2600 gcc_copts = xnnpack_gcc_std_copts(),
2601 msvc_copts = xnnpack_msvc_std_copts(),
2602 deps = [
2603 ":tables",
2604 "@FP16",
2605 "@pthreadpool",
2606 ],
2607)
2608
2609xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002610 name = "neonv8_ukernels",
2611 hdrs = INTERNAL_HDRS,
2612 aarch32_copts = [
2613 "-marm",
2614 "-march=armv8-a",
2615 "-mfpu=neon-fp-armv8",
2616 ],
2617 aarch32_srcs = NEONV8_UKERNELS,
2618 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002619 apple_aarch32_copts = [
2620 "-mcpu=cyclone",
2621 "-mtune=generic",
2622 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002623 gcc_copts = xnnpack_gcc_std_copts(),
2624 msvc_copts = xnnpack_msvc_std_copts(),
2625 deps = [
2626 ":tables",
2627 "@FP16",
2628 "@pthreadpool",
2629 ],
2630)
2631
2632xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002633 name = "neonv8_ukernels_test_mode",
2634 hdrs = INTERNAL_HDRS,
2635 aarch32_copts = [
2636 "-marm",
2637 "-march=armv8-a",
2638 "-mfpu=neon-fp-armv8",
2639 ],
2640 aarch32_srcs = NEONV8_UKERNELS,
2641 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002642 apple_aarch32_copts = [
2643 "-mcpu=cyclone",
2644 "-mtune=generic",
2645 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002646 copts = [
2647 "-UNDEBUG",
2648 "-DXNN_TEST_MODE=1",
2649 ],
2650 gcc_copts = xnnpack_gcc_std_copts(),
2651 msvc_copts = xnnpack_msvc_std_copts(),
2652 deps = [
2653 ":tables",
2654 "@FP16",
2655 "@pthreadpool",
2656 ],
2657)
2658
2659xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002660 name = "neonfp16arith_ukernels",
2661 hdrs = INTERNAL_HDRS,
2662 aarch64_copts = ["-march=armv8.2-a+fp16"],
2663 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002664 gcc_copts = xnnpack_gcc_std_copts(),
2665 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002666 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002667 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002668 "@FP16",
2669 "@pthreadpool",
2670 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002671)
2672
2673xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002674 name = "neonfp16arith_ukernels_test_mode",
2675 hdrs = INTERNAL_HDRS,
2676 aarch64_copts = ["-march=armv8.2-a+fp16"],
2677 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2678 copts = [
2679 "-UNDEBUG",
2680 "-DXNN_TEST_MODE=1",
2681 ],
2682 gcc_copts = xnnpack_gcc_std_copts(),
2683 msvc_copts = xnnpack_msvc_std_copts(),
2684 deps = [
2685 ":tables",
2686 "@FP16",
2687 "@pthreadpool",
2688 ],
2689)
2690
2691xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002692 name = "sse2_ukernels",
2693 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002694 gcc_copts = xnnpack_gcc_std_copts(),
2695 gcc_x86_copts = ["-msse2"],
2696 msvc_copts = xnnpack_msvc_std_copts(),
2697 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002698 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002699 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002700 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002701 "@FP16",
2702 "@pthreadpool",
2703 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002704)
2705
2706xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002707 name = "sse2_ukernels_test_mode",
2708 hdrs = INTERNAL_HDRS,
2709 copts = [
2710 "-UNDEBUG",
2711 "-DXNN_TEST_MODE=1",
2712 ],
2713 gcc_copts = xnnpack_gcc_std_copts(),
2714 gcc_x86_copts = ["-msse2"],
2715 msvc_copts = xnnpack_msvc_std_copts(),
2716 msvc_x86_32_copts = ["/arch:SSE2"],
2717 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2718 deps = [
2719 ":tables",
2720 "@FP16",
2721 "@pthreadpool",
2722 ],
2723)
2724
2725xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002726 name = "ssse3_ukernels",
2727 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002728 gcc_copts = xnnpack_gcc_std_copts(),
2729 gcc_x86_copts = ["-mssse3"],
2730 msvc_copts = xnnpack_msvc_std_copts(),
2731 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002732 x86_srcs = SSSE3_UKERNELS,
2733 deps = [
2734 ":tables",
2735 "@FP16",
2736 "@pthreadpool",
2737 ],
2738)
2739
2740xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002741 name = "ssse3_ukernels_test_mode",
2742 hdrs = INTERNAL_HDRS,
2743 copts = [
2744 "-UNDEBUG",
2745 "-DXNN_TEST_MODE=1",
2746 ],
2747 gcc_copts = xnnpack_gcc_std_copts(),
2748 gcc_x86_copts = ["-mssse3"],
2749 msvc_copts = xnnpack_msvc_std_copts(),
2750 msvc_x86_32_copts = ["/arch:SSE2"],
2751 x86_srcs = SSSE3_UKERNELS,
2752 deps = [
2753 ":tables",
2754 "@FP16",
2755 "@pthreadpool",
2756 ],
2757)
2758
2759xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002760 name = "sse41_ukernels",
2761 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002762 gcc_copts = xnnpack_gcc_std_copts(),
2763 gcc_x86_copts = ["-msse4.1"],
2764 msvc_copts = xnnpack_msvc_std_copts(),
2765 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002766 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002767 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002768 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002769 "@FP16",
2770 "@pthreadpool",
2771 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002772)
2773
2774xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002775 name = "sse41_ukernels_test_mode",
2776 hdrs = INTERNAL_HDRS,
2777 copts = [
2778 "-UNDEBUG",
2779 "-DXNN_TEST_MODE=1",
2780 ],
2781 gcc_copts = xnnpack_gcc_std_copts(),
2782 gcc_x86_copts = ["-msse4.1"],
2783 msvc_copts = xnnpack_msvc_std_copts(),
2784 msvc_x86_32_copts = ["/arch:SSE2"],
2785 x86_srcs = SSE41_UKERNELS,
2786 deps = [
2787 ":tables",
2788 "@FP16",
2789 "@pthreadpool",
2790 ],
2791)
2792
2793xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002794 name = "avx_ukernels",
2795 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002796 gcc_copts = xnnpack_gcc_std_copts(),
2797 gcc_x86_copts = ["-mavx"],
2798 msvc_copts = xnnpack_msvc_std_copts(),
2799 msvc_x86_32_copts = ["/arch:AVX"],
2800 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002801 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002802 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002803 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002804 "@FP16",
2805 "@pthreadpool",
2806 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002807)
2808
2809xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002810 name = "avx_ukernels_test_mode",
2811 hdrs = INTERNAL_HDRS,
2812 copts = [
2813 "-UNDEBUG",
2814 "-DXNN_TEST_MODE=1",
2815 ],
2816 gcc_copts = xnnpack_gcc_std_copts(),
2817 gcc_x86_copts = ["-mavx"],
2818 msvc_copts = xnnpack_msvc_std_copts(),
2819 msvc_x86_32_copts = ["/arch:AVX"],
2820 msvc_x86_64_copts = ["/arch:AVX"],
2821 x86_srcs = AVX_UKERNELS,
2822 deps = [
2823 ":tables",
2824 "@FP16",
2825 "@pthreadpool",
2826 ],
2827)
2828
2829xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002830 name = "fma3_ukernels",
2831 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002832 gcc_copts = xnnpack_gcc_std_copts(),
2833 gcc_x86_copts = ["-mfma"],
2834 msvc_copts = xnnpack_msvc_std_copts(),
2835 msvc_x86_32_copts = ["/arch:AVX"],
2836 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002837 x86_srcs = FMA3_UKERNELS,
2838 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002839 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002840 "@FP16",
2841 "@pthreadpool",
2842 ],
2843)
2844
2845xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002846 name = "fma3_ukernels_test_mode",
2847 hdrs = INTERNAL_HDRS,
2848 copts = [
2849 "-UNDEBUG",
2850 "-DXNN_TEST_MODE=1",
2851 ],
2852 gcc_copts = xnnpack_gcc_std_copts(),
2853 gcc_x86_copts = ["-mfma"],
2854 msvc_copts = xnnpack_msvc_std_copts(),
2855 msvc_x86_32_copts = ["/arch:AVX"],
2856 msvc_x86_64_copts = ["/arch:AVX"],
2857 x86_srcs = FMA3_UKERNELS,
2858 deps = [
2859 ":tables",
2860 "@FP16",
2861 "@pthreadpool",
2862 ],
2863)
2864
2865xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002866 name = "avx2_ukernels",
2867 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002868 gcc_copts = xnnpack_gcc_std_copts(),
2869 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002870 "-mfma",
2871 "-mavx2",
2872 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002873 msvc_copts = xnnpack_msvc_std_copts(),
2874 msvc_x86_32_copts = ["/arch:AVX2"],
2875 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002876 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002877 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002878 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002879 "@FP16",
2880 "@pthreadpool",
2881 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002882)
2883
2884xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002885 name = "avx2_ukernels_test_mode",
2886 hdrs = INTERNAL_HDRS,
2887 copts = [
2888 "-UNDEBUG",
2889 "-DXNN_TEST_MODE=1",
2890 ],
2891 gcc_copts = xnnpack_gcc_std_copts(),
2892 gcc_x86_copts = [
2893 "-mfma",
2894 "-mavx2",
2895 ],
2896 msvc_copts = xnnpack_msvc_std_copts(),
2897 msvc_x86_32_copts = ["/arch:AVX2"],
2898 msvc_x86_64_copts = ["/arch:AVX2"],
2899 x86_srcs = AVX2_UKERNELS,
2900 deps = [
2901 ":tables",
2902 "@FP16",
2903 "@pthreadpool",
2904 ],
2905)
2906
2907xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002908 name = "avx512f_ukernels",
2909 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002910 gcc_copts = xnnpack_gcc_std_copts(),
2911 gcc_x86_copts = ["-mavx512f"],
2912 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2913 msvc_copts = xnnpack_msvc_std_copts(),
2914 msvc_x86_32_copts = ["/arch:AVX512"],
2915 msvc_x86_64_copts = ["/arch:AVX512"],
2916 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002917 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002918 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002919 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002920 "@FP16",
2921 "@pthreadpool",
2922 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002923)
2924
2925xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002926 name = "avx512f_ukernels_test_mode",
2927 hdrs = INTERNAL_HDRS,
2928 copts = [
2929 "-UNDEBUG",
2930 "-DXNN_TEST_MODE=1",
2931 ],
2932 gcc_copts = xnnpack_gcc_std_copts(),
2933 gcc_x86_copts = ["-mavx512f"],
2934 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2935 msvc_copts = xnnpack_msvc_std_copts(),
2936 msvc_x86_32_copts = ["/arch:AVX512"],
2937 msvc_x86_64_copts = ["/arch:AVX512"],
2938 msys_copts = ["-fno-asynchronous-unwind-tables"],
2939 x86_srcs = AVX512F_UKERNELS,
2940 deps = [
2941 ":tables",
2942 "@FP16",
2943 "@pthreadpool",
2944 ],
2945)
2946
2947xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002948 name = "asm_ukernels",
2949 hdrs = ["src/xnnpack/assembly.h"],
2950 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002951 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002952 aarch64_srcs = AARCH64_ASM_UKERNELS,
2953)
2954
Marat Dukhan3b59de22020-06-03 20:15:19 -07002955xnnpack_cc_library(
2956 name = "logging_utils",
2957 srcs = LOGGING_SRCS,
2958 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2959 copts = LOGGING_COPTS + [
2960 "-Isrc",
2961 "-Iinclude",
2962 ] + select({
2963 ":debug_build": [],
2964 "//conditions:default": xnnpack_min_size_copts(),
2965 }),
2966 gcc_copts = xnnpack_gcc_std_copts(),
2967 msvc_copts = xnnpack_msvc_std_copts(),
2968 visibility = xnnpack_visibility(),
2969 deps = [
2970 "@FP16",
2971 "@clog",
2972 "@pthreadpool",
2973 ],
2974)
2975
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976xnnpack_aggregate_library(
2977 name = "ukernels",
2978 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 ":neon_ukernels",
2980 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002981 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 ":asm_ukernels",
2983 ],
2984 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 ":neon_ukernels",
2986 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002987 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002988 ":neonfp16arith_ukernels",
2989 ":asm_ukernels",
2990 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002991 generic_deps = [
2992 ":scalar_ukernels",
2993 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002994 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002995 ":psimd_fastmath_ukernels",
2996 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002997 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002998 wasm_deps = [
2999 ":wasm_ukernels",
3000 ],
3001 wasmsimd_deps = [
3002 ":wasm_ukernels",
3003 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003004 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003006 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003007 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003009 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003010 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011 ":avx512f_ukernels",
3012 ],
3013)
3014
Marat Dukhan33fcf782020-05-24 14:27:15 -07003015xnnpack_aggregate_library(
3016 name = "ukernels_test_mode",
3017 aarch32_deps = [
3018 ":neon_ukernels_test_mode",
3019 ":neonfma_ukernels_test_mode",
3020 ":neonv8_ukernels_test_mode",
3021 ":asm_ukernels",
3022 ],
3023 aarch64_deps = [
3024 ":neon_ukernels_test_mode",
3025 ":neonfma_ukernels_test_mode",
3026 ":neonv8_ukernels_test_mode",
3027 ":neonfp16arith_ukernels_test_mode",
3028 ":asm_ukernels",
3029 ],
3030 generic_deps = [
3031 ":scalar_ukernels_test_mode",
3032 ],
3033 psimd_deps = [
3034 ":psimd_fastmath_ukernels_test_mode",
3035 ":psimd_accmath_ukernels_test_mode",
3036 ],
3037 wasm_deps = [
3038 ":wasm_ukernels_test_mode",
3039 ],
3040 wasmsimd_deps = [
3041 ":wasm_ukernels_test_mode",
3042 ],
3043 x86_deps = [
3044 ":sse2_ukernels_test_mode",
3045 ":ssse3_ukernels_test_mode",
3046 ":sse41_ukernels_test_mode",
3047 ":avx_ukernels_test_mode",
3048 ":fma3_ukernels_test_mode",
3049 ":avx2_ukernels_test_mode",
3050 ":avx512f_ukernels_test_mode",
3051 ],
3052)
3053
Marat Dukhan86df3842020-06-17 14:03:31 -07003054# This target must be compatible with Android NDK r17 and compile with Clang 6.
3055# It is similar to :ukernels target, but excludes microkernels which require
3056# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003057xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003058 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003059 aarch32_deps = [
3060 ":neon_ukernels",
3061 ":neonfma_ukernels",
3062 ":neonv8_ukernels",
3063 ":asm_ukernels",
3064 ],
3065 aarch64_deps = [
3066 ":neon_ukernels",
3067 ":neonfma_ukernels",
3068 ":neonv8_ukernels",
3069 ":asm_ukernels",
3070 ],
3071 generic_deps = [
3072 ":scalar_ukernels",
3073 ],
3074 psimd_deps = [
3075 ":psimd_fastmath_ukernels",
3076 ":psimd_accmath_ukernels",
3077 ],
3078 wasm_deps = [
3079 ":wasm_ukernels",
3080 ],
3081 wasmsimd_deps = [
3082 ":wasm_ukernels",
3083 ],
3084 x86_deps = [
3085 ":sse2_ukernels",
3086 ":ssse3_ukernels",
3087 ":sse41_ukernels",
3088 ":avx_ukernels",
3089 ":fma3_ukernels",
3090 ":avx2_ukernels",
3091 ":avx512f_ukernels",
3092 ],
3093)
3094
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095xnnpack_cc_library(
3096 name = "im2col",
3097 srcs = ["src/im2col.c"],
3098 hdrs = [
3099 "src/xnnpack/common.h",
3100 "src/xnnpack/im2col.h",
3101 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003102 gcc_copts = xnnpack_gcc_std_copts(),
3103 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003104)
3105
3106xnnpack_cc_library(
3107 name = "indirection",
3108 srcs = ["src/indirection.c"],
3109 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003110 gcc_copts = xnnpack_gcc_std_copts(),
3111 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003112 deps = [
3113 "@FP16",
3114 "@FXdiv",
3115 "@pthreadpool",
3116 ],
3117)
3118
3119xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003120 name = "indirection_test_mode",
3121 srcs = ["src/indirection.c"],
3122 hdrs = INTERNAL_HDRS,
3123 copts = [
3124 "-UNDEBUG",
3125 "-DXNN_TEST_MODE=1",
3126 ],
3127 gcc_copts = xnnpack_gcc_std_copts(),
3128 msvc_copts = xnnpack_msvc_std_copts(),
3129 deps = [
3130 "@FP16",
3131 "@FXdiv",
3132 "@pthreadpool",
3133 ],
3134)
3135
3136xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003137 name = "packing",
3138 srcs = ["src/packing.c"],
3139 hdrs = INTERNAL_HDRS,
3140 gcc_copts = xnnpack_gcc_std_copts(),
3141 msvc_copts = xnnpack_msvc_std_copts(),
3142 deps = [
3143 "@FP16",
3144 "@FXdiv",
3145 "@pthreadpool",
3146 ],
3147)
3148
3149xnnpack_cc_library(
3150 name = "packing_test_mode",
3151 srcs = ["src/packing.c"],
3152 hdrs = INTERNAL_HDRS,
3153 copts = [
3154 "-UNDEBUG",
3155 "-DXNN_TEST_MODE=1",
3156 ],
3157 gcc_copts = xnnpack_gcc_std_copts(),
3158 msvc_copts = xnnpack_msvc_std_copts(),
3159 deps = [
3160 "@FP16",
3161 "@FXdiv",
3162 "@pthreadpool",
3163 ],
3164)
3165
3166xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 name = "operator_run",
3168 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003169 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003170 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003171 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3172 "//conditions:default": [],
3173 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003174 gcc_copts = xnnpack_gcc_std_copts(),
3175 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003176 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003177 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003178 "@FP16",
3179 "@FXdiv",
3180 "@clog",
3181 "@pthreadpool",
3182 ],
3183)
3184
Chao Mei6ddfc602020-05-13 22:29:36 -07003185xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003186 name = "operator_run_test_mode",
3187 srcs = ["src/operator-run.c"],
3188 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3189 copts = LOGGING_COPTS + [
3190 "-UNDEBUG",
3191 "-DXNN_TEST_MODE=1",
3192 ] + select({
3193 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3194 "//conditions:default": [],
3195 }),
3196 gcc_copts = xnnpack_gcc_std_copts(),
3197 msvc_copts = xnnpack_msvc_std_copts(),
3198 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003199 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003200 "@FP16",
3201 "@FXdiv",
3202 "@clog",
3203 "@pthreadpool",
3204 ],
3205)
3206
3207xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003208 name = "memory_planner",
3209 srcs = ["src/memory-planner.c"],
3210 hdrs = INTERNAL_HDRS,
3211 defines = select({
3212 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3213 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3214 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3215 }),
3216 gcc_copts = xnnpack_gcc_std_copts(),
3217 msvc_copts = xnnpack_msvc_std_copts(),
3218 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003219 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003220 "@pthreadpool",
3221 ],
3222)
3223
Marat Dukhan33fcf782020-05-24 14:27:15 -07003224xnnpack_cc_library(
3225 name = "memory_planner_test_mode",
3226 srcs = ["src/memory-planner.c"],
3227 hdrs = INTERNAL_HDRS,
3228 copts = [
3229 "-UNDEBUG",
3230 "-DXNN_TEST_MODE=1",
3231 ],
3232 defines = select({
3233 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3234 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3235 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3236 }),
3237 gcc_copts = xnnpack_gcc_std_copts(),
3238 msvc_copts = xnnpack_msvc_std_copts(),
3239 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003240 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003241 "@pthreadpool",
3242 ],
3243)
3244
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245cc_library(
3246 name = "enable_assembly",
3247 defines = select({
3248 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3249 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003250 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003251 }),
3252)
3253
Marat Dukhan9de90e02020-06-18 16:04:12 -07003254cc_library(
3255 name = "enable_sparse",
3256 defines = select({
3257 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3258 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003259 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003260 }),
3261)
3262
Marat Dukhancf056b22019-10-07 10:26:29 -07003263xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003264 name = "operators",
3265 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003266 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003267 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003268 ],
3269 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003270 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271 "-Isrc",
3272 "-Iinclude",
3273 ] + select({
3274 ":debug_build": [],
3275 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003276 }) + select({
3277 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3278 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003279 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003280 gcc_copts = xnnpack_gcc_std_copts(),
3281 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003282 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003283 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003284 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003285 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003286 "@FP16",
3287 "@FXdiv",
3288 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003289 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003290 ],
3291)
3292
Marat Dukhan10a38082020-04-17 03:58:35 -07003293xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003294 name = "operators_test_mode",
3295 srcs = OPERATOR_SRCS + [
3296 "src/memory.c",
3297 "src/operator-delete.c",
3298 ],
3299 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3300 copts = LOGGING_COPTS + [
3301 "-Isrc",
3302 "-Iinclude",
3303 "-UNDEBUG",
3304 "-DXNN_TEST_MODE=1",
3305 ] + select({
3306 ":debug_build": [],
3307 "//conditions:default": xnnpack_min_size_copts(),
3308 }) + select({
3309 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3310 "//conditions:default": [],
3311 }),
3312 gcc_copts = xnnpack_gcc_std_copts(),
3313 msvc_copts = xnnpack_msvc_std_copts(),
3314 deps = [
3315 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003316 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003317 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003318 "@FP16",
3319 "@FXdiv",
3320 "@clog",
3321 "@pthreadpool",
3322 ],
3323)
3324
3325xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003326 name = "XNNPACK",
3327 srcs = [
3328 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003329 "src/runtime.c",
3330 "src/subgraph.c",
3331 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003332 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003333 hdrs = ["include/xnnpack.h"],
3334 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003335 "-Isrc",
3336 "-Iinclude",
3337 ] + select({
3338 ":debug_build": [],
3339 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003340 }) + select({
3341 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3342 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003343 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003344 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003345 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003346 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003347 visibility = xnnpack_visibility(),
3348 deps = [
3349 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003350 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003351 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003352 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003353 ":operator_run",
3354 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003355 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003356 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003357 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003358 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003359 ] + select({
3360 ":emscripten": [],
3361 "//conditions:default": ["@cpuinfo"],
3362 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003363)
3364
Marat Dukhan10a38082020-04-17 03:58:35 -07003365xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003366 name = "XNNPACK_test_mode",
3367 srcs = [
3368 "src/init.c",
3369 "src/runtime.c",
3370 "src/subgraph.c",
3371 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003372 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003373 hdrs = ["include/xnnpack.h"],
3374 copts = LOGGING_COPTS + [
3375 "-Isrc",
3376 "-Iinclude",
3377 "-UNDEBUG",
3378 "-DXNN_TEST_MODE=1",
3379 ] + select({
3380 ":debug_build": [],
3381 "//conditions:default": xnnpack_min_size_copts(),
3382 }) + select({
3383 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3384 "//conditions:default": [],
3385 }),
3386 gcc_copts = xnnpack_gcc_std_copts(),
3387 includes = ["include"],
3388 msvc_copts = xnnpack_msvc_std_copts(),
3389 visibility = xnnpack_visibility(),
3390 deps = [
3391 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003392 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003393 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003394 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003395 ":operator_run_test_mode",
3396 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003397 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003398 "@clog",
3399 "@FP16",
3400 "@pthreadpool",
3401 ] + select({
3402 ":emscripten": [],
3403 "//conditions:default": ["@cpuinfo"],
3404 }),
3405)
3406
3407xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003408 name = "xnnpack_f32",
3409 srcs = [
3410 "src/init.c",
3411 "src/runtime.c",
3412 "src/subgraph.c",
3413 "src/tensor.c",
3414 ] + SUBGRAPH_SRCS,
3415 hdrs = ["include/xnnpack.h"],
3416 copts = LOGGING_COPTS + [
3417 "-Isrc",
3418 "-Iinclude",
3419 ] + select({
3420 ":debug_build": [],
3421 "//conditions:default": xnnpack_min_size_copts(),
3422 }) + select({
3423 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3424 "//conditions:default": [],
3425 }),
3426 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003427 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003428 "XNN_NO_U8_OPERATORS",
3429 "XNN_NO_X8_OPERATORS",
3430 "XNN_NO_F16_OPERATORS",
3431 "XNN_NO_X16_OPERATORS",
3432 ],
3433 gcc_copts = xnnpack_gcc_std_copts(),
3434 includes = ["include"],
3435 msvc_copts = xnnpack_msvc_std_copts(),
3436 visibility = xnnpack_visibility(),
3437 deps = [
3438 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003439 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003440 ":logging_utils",
3441 ":memory_planner",
3442 ":operator_run",
3443 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003444 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003445 "@clog",
3446 "@FP16",
3447 "@pthreadpool",
3448 ] + select({
3449 ":emscripten": [],
3450 "//conditions:default": ["@cpuinfo"],
3451 }),
3452)
3453
3454xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003455 name = "xnnpack_operators_nhwc_f32",
3456 srcs = [
3457 "src/init.c",
3458 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003459 hdrs = ["include/xnnpack.h"],
3460 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003461 "-Isrc",
3462 "-Iinclude",
3463 ] + select({
3464 ":debug_build": [],
3465 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003466 }) + select({
3467 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3468 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003469 }),
3470 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003471 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003472 "XNN_NO_U8_OPERATORS",
3473 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003474 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003475 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003476 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003477 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003478 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479 visibility = xnnpack_visibility(),
3480 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003481 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003482 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003483 ":operator_run",
3484 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003485 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003486 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003487 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003488 ] + select({
3489 ":emscripten": [],
3490 "//conditions:default": ["@cpuinfo"],
3491 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003492)
3493
Marat Dukhancf056b22019-10-07 10:26:29 -07003494xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003495 name = "bench_utils",
3496 srcs = ["bench/utils.cc"],
3497 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003498 deps = [
3499 "@com_google_benchmark//:benchmark",
3500 "@cpuinfo",
3501 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003502)
3503
Frank Barchard7e955972019-10-11 10:34:25 -07003504######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505
3506xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003507 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 srcs = [
3509 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003510 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511 "src/xnnpack/AlignedAllocator.h",
3512 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003513 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003514 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515)
3516
3517xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003518 name = "f16_igemm_bench",
3519 srcs = [
3520 "bench/f16-igemm.cc",
3521 "bench/conv.h",
3522 "bench/google/conv.h",
3523 "src/xnnpack/AlignedAllocator.h",
3524 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003525 deps = MICROKERNEL_BENCHMARK_DEPS + [
3526 ":indirection",
3527 ":packing",
3528 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003529)
3530
3531xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003532 name = "f16_gemm_bench",
3533 srcs = [
3534 "bench/f16-gemm.cc",
3535 "bench/gemm.h",
3536 "src/xnnpack/AlignedAllocator.h",
3537 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003538 deps = MICROKERNEL_BENCHMARK_DEPS + [
3539 ":packing",
3540 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003541)
3542
3543xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003544 name = "f16_spmm_bench",
3545 srcs = [
3546 "bench/f16-spmm.cc",
3547 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003548 "src/xnnpack/AlignedAllocator.h",
3549 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003550 deps = MICROKERNEL_BENCHMARK_DEPS,
3551)
3552
3553xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554 name = "f32_igemm_bench",
3555 srcs = [
3556 "bench/f32-igemm.cc",
3557 "bench/conv.h",
3558 "src/xnnpack/AlignedAllocator.h",
3559 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003560 deps = MICROKERNEL_BENCHMARK_DEPS + [
3561 ":indirection",
3562 ":packing",
3563 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564)
3565
3566xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003567 name = "f16_relu_bench",
3568 srcs = [
3569 "bench/f16-relu.cc",
3570 "src/xnnpack/AlignedAllocator.h",
3571 ] + MICROKERNEL_BENCHMARK_HDRS,
3572 deps = MICROKERNEL_BENCHMARK_DEPS,
3573)
3574
3575xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576 name = "f32_conv_hwc_bench",
3577 srcs = [
3578 "bench/f32-conv-hwc.cc",
3579 "bench/dconv.h",
3580 "src/xnnpack/AlignedAllocator.h",
3581 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003582 deps = MICROKERNEL_BENCHMARK_DEPS + [
3583 ":packing",
3584 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585)
3586
3587xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003588 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003589 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003590 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003591 "bench/dconv.h",
3592 "src/xnnpack/AlignedAllocator.h",
3593 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003594 deps = MICROKERNEL_BENCHMARK_DEPS + [
3595 ":packing",
3596 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003597)
3598
3599xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003600 name = "f16_dwconv_bench",
3601 srcs = [
3602 "bench/f16-dwconv.cc",
3603 "bench/dwconv.h",
3604 "bench/google/dwconv.h",
3605 "src/xnnpack/AlignedAllocator.h",
3606 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003607 deps = MICROKERNEL_BENCHMARK_DEPS + [
3608 ":indirection",
3609 ":packing",
3610 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003611)
3612
3613xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003614 name = "f32_dwconv_bench",
3615 srcs = [
3616 "bench/f32-dwconv.cc",
3617 "bench/dwconv.h",
3618 "src/xnnpack/AlignedAllocator.h",
3619 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003620 deps = MICROKERNEL_BENCHMARK_DEPS + [
3621 ":indirection",
3622 ":packing",
3623 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624)
3625
3626xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003627 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003628 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003629 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003630 "bench/dwconv.h",
3631 "src/xnnpack/AlignedAllocator.h",
3632 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003633 deps = MICROKERNEL_BENCHMARK_DEPS + [
3634 ":indirection",
3635 ":packing",
3636 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637)
3638
3639xnnpack_benchmark(
3640 name = "f32_gemm_bench",
3641 srcs = [
3642 "bench/f32-gemm.cc",
3643 "bench/gemm.h",
3644 "src/xnnpack/AlignedAllocator.h",
3645 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003646 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003647 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648)
3649
3650xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003651 name = "f32_hswish_bench",
3652 srcs = [
3653 "bench/f32-hswish.cc",
3654 "src/xnnpack/AlignedAllocator.h",
3655 ] + MICROKERNEL_BENCHMARK_HDRS,
3656 deps = MICROKERNEL_BENCHMARK_DEPS,
3657)
3658
3659xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003660 name = "f32_raddexpminusmax_bench",
3661 srcs = [
3662 "bench/f32-raddexpminusmax.cc",
3663 "src/xnnpack/AlignedAllocator.h",
3664 ] + MICROKERNEL_BENCHMARK_HDRS,
3665 deps = MICROKERNEL_BENCHMARK_DEPS,
3666)
3667
3668xnnpack_benchmark(
3669 name = "f32_raddextexp_bench",
3670 srcs = [
3671 "bench/f32-raddextexp.cc",
3672 "src/xnnpack/AlignedAllocator.h",
3673 ] + MICROKERNEL_BENCHMARK_HDRS,
3674 deps = MICROKERNEL_BENCHMARK_DEPS,
3675)
3676
3677xnnpack_benchmark(
3678 name = "f32_raddstoreexpminusmax_bench",
3679 srcs = [
3680 "bench/f32-raddstoreexpminusmax.cc",
3681 "src/xnnpack/AlignedAllocator.h",
3682 ] + MICROKERNEL_BENCHMARK_HDRS,
3683 deps = MICROKERNEL_BENCHMARK_DEPS,
3684)
3685
3686xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003687 name = "f32_relu_bench",
3688 srcs = [
3689 "bench/f32-relu.cc",
3690 "src/xnnpack/AlignedAllocator.h",
3691 ] + MICROKERNEL_BENCHMARK_HDRS,
3692 deps = MICROKERNEL_BENCHMARK_DEPS,
3693)
3694
3695xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 name = "f32_rmax_bench",
3697 srcs = [
3698 "bench/f32-rmax.cc",
3699 "src/xnnpack/AlignedAllocator.h",
3700 ] + MICROKERNEL_BENCHMARK_HDRS,
3701 deps = MICROKERNEL_BENCHMARK_DEPS,
3702)
3703
3704xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003705 name = "f32_sigmoid_bench",
3706 srcs = [
3707 "bench/f32-sigmoid.cc",
3708 "src/xnnpack/AlignedAllocator.h",
3709 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003710 deps = MICROKERNEL_BENCHMARK_DEPS,
3711)
3712
3713xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003714 name = "f32_spmm_bench",
3715 srcs = [
3716 "bench/f32-spmm.cc",
3717 "bench/gemm.h",
3718 "src/xnnpack/AlignedAllocator.h",
3719 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003720 deps = MICROKERNEL_BENCHMARK_DEPS,
3721)
3722
3723xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003724 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003725 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003726 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003727 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003728 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003729 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003730)
3731
3732xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003733 name = "f32_vscaleexpminusmax_bench",
3734 srcs = [
3735 "bench/f32-vscaleexpminusmax.cc",
3736 "src/xnnpack/AlignedAllocator.h",
3737 ] + MICROKERNEL_BENCHMARK_HDRS,
3738 deps = MICROKERNEL_BENCHMARK_DEPS,
3739)
3740
3741xnnpack_benchmark(
3742 name = "f32_vscaleextexp_bench",
3743 srcs = [
3744 "bench/f32-vscaleextexp.cc",
3745 "src/xnnpack/AlignedAllocator.h",
3746 ] + MICROKERNEL_BENCHMARK_HDRS,
3747 deps = MICROKERNEL_BENCHMARK_DEPS,
3748)
3749
3750xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003751 name = "f32_vsqrt_bench",
3752 srcs = [
3753 "bench/f32-vsqrt.cc",
3754 "src/xnnpack/AlignedAllocator.h",
3755 ] + MICROKERNEL_BENCHMARK_HDRS,
3756 deps = MICROKERNEL_BENCHMARK_DEPS,
3757)
3758
3759xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003760 name = "f32_im2col_gemm_bench",
3761 srcs = [
3762 "bench/f32-im2col-gemm.cc",
3763 "bench/conv.h",
3764 "src/xnnpack/AlignedAllocator.h",
3765 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003766 deps = MICROKERNEL_BENCHMARK_DEPS + [
3767 ":im2col",
3768 ":packing",
3769 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003770)
3771
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003772xnnpack_benchmark(
3773 name = "requantization_bench",
3774 srcs = [
3775 "bench/requantization.cc",
3776 "src/xnnpack/requantization-stubs.h",
3777 "src/xnnpack/AlignedAllocator.h",
3778 ] + MICROKERNEL_BENCHMARK_HDRS,
3779 deps = MICROKERNEL_BENCHMARK_DEPS,
3780)
3781
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003782xnnpack_benchmark(
3783 name = "rounding_bench",
3784 srcs = [
3785 "bench/rounding.cc",
3786 "src/xnnpack/math-stubs.h",
3787 "src/xnnpack/AlignedAllocator.h",
3788 ] + MICROKERNEL_BENCHMARK_HDRS,
3789 deps = MICROKERNEL_BENCHMARK_DEPS,
3790)
3791
Marat Dukhan08c4a432019-10-03 09:29:21 -07003792########################### Benchmarks for operators ###########################
3793
3794xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003795 name = "average_pooling_bench",
3796 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003797 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003798 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003799 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003800)
3801
3802xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003803 name = "bankers_rounding_bench",
3804 srcs = ["bench/bankers-rounding.cc"],
3805 copts = xnnpack_optional_tflite_copts(),
3806 tags = ["nowin32"],
3807 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3808)
3809
3810xnnpack_benchmark(
3811 name = "ceiling_bench",
3812 srcs = ["bench/ceiling.cc"],
3813 copts = xnnpack_optional_tflite_copts(),
3814 tags = ["nowin32"],
3815 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3816)
3817
3818xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819 name = "channel_shuffle_bench",
3820 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003821 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003822)
3823
3824xnnpack_benchmark(
3825 name = "convolution_bench",
3826 srcs = ["bench/convolution.cc"],
3827 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003828 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003829 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003830)
3831
3832xnnpack_benchmark(
3833 name = "deconvolution_bench",
3834 srcs = ["bench/deconvolution.cc"],
3835 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003836 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003837 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003838)
3839
3840xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003841 name = "floor_bench",
3842 srcs = ["bench/floor.cc"],
3843 copts = xnnpack_optional_tflite_copts(),
3844 tags = ["nowin32"],
3845 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3846)
3847
3848xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003849 name = "global_average_pooling_bench",
3850 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003851 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003852)
3853
3854xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003855 name = "hardswish_bench",
3856 srcs = ["bench/hardswish.cc"],
3857 copts = xnnpack_optional_tflite_copts(),
3858 tags = ["nowin32"],
3859 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3860)
3861
3862xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003863 name = "max_pooling_bench",
3864 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003865 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866)
3867
3868xnnpack_benchmark(
3869 name = "sigmoid_bench",
3870 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003871 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003872 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003873 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003874)
3875
3876xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003877 name = "prelu_bench",
3878 srcs = ["bench/prelu.cc"],
3879 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003880 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003881 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003882)
3883
3884xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003885 name = "softmax_bench",
3886 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003887 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003888 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003889 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890)
3891
Marat Dukhan87727142020-06-24 15:24:10 -07003892xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003893 name = "square_root_bench",
3894 srcs = ["bench/square-root.cc"],
3895 copts = xnnpack_optional_tflite_copts(),
3896 tags = ["nowin32"],
3897 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3898)
3899
3900xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003901 name = "truncation_bench",
3902 srcs = ["bench/truncation.cc"],
3903 deps = OPERATOR_BENCHMARK_DEPS,
3904)
3905
Marat Dukhanc068bb62019-10-04 13:24:39 -07003906############################# End-to-end benchmarks ############################
3907
3908cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003909 name = "fp32_mobilenet_v1",
3910 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003911 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003912 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003913 linkstatic = True,
3914 deps = [
3915 ":XNNPACK",
3916 "@pthreadpool",
3917 ],
3918)
3919
3920cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003921 name = "fp16_mobilenet_v1",
3922 srcs = ["models/fp16-mobilenet-v1.cc"],
3923 hdrs = ["models/models.h"],
3924 copts = xnnpack_std_cxxopts(),
3925 linkstatic = True,
3926 deps = [
3927 ":XNNPACK",
3928 "@FP16",
3929 "@pthreadpool",
3930 ],
3931)
3932
3933cc_library(
3934 name = "fp32_mobilenet_v2",
3935 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003936 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003937 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003938 linkstatic = True,
3939 deps = [
3940 ":XNNPACK",
3941 "@pthreadpool",
3942 ],
3943)
3944
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003945cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003946 name = "fp16_mobilenet_v2",
3947 srcs = ["models/fp16-mobilenet-v2.cc"],
3948 hdrs = ["models/models.h"],
3949 copts = xnnpack_std_cxxopts(),
3950 linkstatic = True,
3951 deps = [
3952 ":XNNPACK",
3953 "@FP16",
3954 "@pthreadpool",
3955 ],
3956)
3957
3958cc_library(
3959 name = "fp32_mobilenet_v3_large",
3960 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003961 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003962 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003963 linkstatic = True,
3964 deps = [
3965 ":XNNPACK",
3966 "@pthreadpool",
3967 ],
3968)
3969
3970cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003971 name = "fp32_mobilenet_v3_small",
3972 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003973 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003974 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003975 linkstatic = True,
3976 deps = [
3977 ":XNNPACK",
3978 "@pthreadpool",
3979 ],
3980)
3981
Marat Dukhanc068bb62019-10-04 13:24:39 -07003982xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003983 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003984 srcs = [
3985 "bench/f32-dwconv-e2e.cc",
3986 "bench/end2end.h",
3987 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003988 deps = MICROKERNEL_BENCHMARK_DEPS + [
3989 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003990 ":fp32_mobilenet_v1",
3991 ":fp32_mobilenet_v2",
3992 ":fp32_mobilenet_v3_large",
3993 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003994 ],
3995)
3996
3997xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003998 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003999 srcs = [
4000 "bench/f32-gemm-e2e.cc",
4001 "bench/end2end.h",
4002 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004003 deps = MICROKERNEL_BENCHMARK_DEPS + [
4004 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004005 ":fp32_mobilenet_v1",
4006 ":fp32_mobilenet_v2",
4007 ":fp32_mobilenet_v3_large",
4008 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004009 ],
4010)
4011
4012xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004013 name = "end2end_bench",
4014 srcs = ["bench/end2end.cc"],
4015 deps = [
4016 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004017 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004018 ":fp16_mobilenet_v1",
4019 ":fp16_mobilenet_v2",
4020 ":fp32_mobilenet_v1",
4021 ":fp32_mobilenet_v2",
4022 ":fp32_mobilenet_v3_large",
4023 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004024 "@pthreadpool",
4025 ],
4026)
4027
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004028#################### Accuracy evaluation for math functions ####################
4029
4030xnnpack_benchmark(
4031 name = "f32_exp_eval",
4032 srcs = [
4033 "eval/f32-exp.cc",
4034 "src/xnnpack/AlignedAllocator.h",
4035 ] + ACCURACY_EVAL_HDRS,
4036 deps = ACCURACY_EVAL_DEPS,
4037)
4038
Marat Dukhan515c9772019-10-17 18:07:57 -07004039xnnpack_benchmark(
4040 name = "f32_expminus_eval",
4041 srcs = [
4042 "eval/f32-expminus.cc",
4043 "src/xnnpack/AlignedAllocator.h",
4044 ] + ACCURACY_EVAL_HDRS,
4045 deps = ACCURACY_EVAL_DEPS,
4046)
4047
Marat Dukhan98ba4412019-10-23 02:14:28 -07004048xnnpack_benchmark(
4049 name = "f32_extexp_eval",
4050 srcs = [
4051 "eval/f32-extexp.cc",
4052 "src/xnnpack/AlignedAllocator.h",
4053 ] + ACCURACY_EVAL_HDRS,
4054 deps = ACCURACY_EVAL_DEPS,
4055)
4056
Marat Dukhan8853b822020-05-07 12:19:01 -07004057xnnpack_unit_test(
4058 name = "f32_roundne_eval",
4059 srcs = [
4060 "eval/f32-roundne.cc",
4061 "src/xnnpack/AlignedAllocator.h",
4062 "src/xnnpack/math-stubs.h",
4063 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004064 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004065 deps = MICROKERNEL_TEST_DEPS,
4066)
4067
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004068xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004069 name = "f32_roundd_eval",
4070 srcs = [
4071 "eval/f32-roundd.cc",
4072 "src/xnnpack/AlignedAllocator.h",
4073 "src/xnnpack/math-stubs.h",
4074 ] + MICROKERNEL_TEST_HDRS,
4075 automatic = False,
4076 deps = MICROKERNEL_TEST_DEPS,
4077)
4078
4079xnnpack_unit_test(
4080 name = "f32_roundu_eval",
4081 srcs = [
4082 "eval/f32-roundu.cc",
4083 "src/xnnpack/AlignedAllocator.h",
4084 "src/xnnpack/math-stubs.h",
4085 ] + MICROKERNEL_TEST_HDRS,
4086 automatic = False,
4087 deps = MICROKERNEL_TEST_DEPS,
4088)
4089
4090xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004091 name = "f32_roundz_eval",
4092 srcs = [
4093 "eval/f32-roundz.cc",
4094 "src/xnnpack/AlignedAllocator.h",
4095 "src/xnnpack/math-stubs.h",
4096 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004097 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004098 deps = MICROKERNEL_TEST_DEPS,
4099)
4100
Marat Dukhan346a9e52019-11-15 09:06:30 -08004101xnnpack_benchmark(
4102 name = "f32_sigmoid_eval",
4103 srcs = [
4104 "eval/f32-sigmoid.cc",
4105 "src/xnnpack/AlignedAllocator.h",
4106 ] + ACCURACY_EVAL_HDRS,
4107 deps = ACCURACY_EVAL_DEPS,
4108)
4109
Marat Dukhan84000762020-06-29 18:38:43 -07004110xnnpack_benchmark(
4111 name = "f32_sqrt_eval",
4112 srcs = [
4113 "eval/f32-sqrt.cc",
4114 "src/xnnpack/AlignedAllocator.h",
4115 ] + ACCURACY_EVAL_HDRS,
4116 deps = ACCURACY_EVAL_DEPS + [
4117 ":bench_utils",
4118 ],
4119)
4120
Marat Dukhan08c4a432019-10-03 09:29:21 -07004121######################### Unit tests for micro-kernels #########################
4122
4123xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004124 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004125 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004126 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127 "test/gemm-microkernel-tester.h",
4128 "src/xnnpack/AlignedAllocator.h",
4129 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004130 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131)
4132
4133xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004134 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004135 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004136 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004137 "test/spmm-microkernel-tester.h",
4138 "src/xnnpack/AlignedAllocator.h",
4139 ] + MICROKERNEL_TEST_HDRS,
4140 deps = MICROKERNEL_TEST_DEPS,
4141)
4142
4143xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004144 name = "f32_argmaxpool_test",
4145 srcs = [
4146 "test/f32-argmaxpool.cc",
4147 "test/argmaxpool-microkernel-tester.h",
4148 "src/xnnpack/AlignedAllocator.h",
4149 ] + MICROKERNEL_TEST_HDRS,
4150 deps = MICROKERNEL_TEST_DEPS,
4151)
4152
4153xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004154 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004155 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004156 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004157 "test/avgpool-microkernel-tester.h",
4158 "src/xnnpack/AlignedAllocator.h",
4159 ] + MICROKERNEL_TEST_HDRS,
4160 deps = MICROKERNEL_TEST_DEPS,
4161)
4162
4163xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004164 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004165 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004166 "test/f32-ibilinear.cc",
4167 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004168 "src/xnnpack/AlignedAllocator.h",
4169 ] + MICROKERNEL_TEST_HDRS,
4170 deps = MICROKERNEL_TEST_DEPS,
4171)
4172
4173xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004174 name = "f16_clamp_test",
4175 srcs = [
4176 "test/f16-clamp.cc",
4177 "test/clamp-microkernel-tester.h",
4178 ] + MICROKERNEL_TEST_HDRS,
4179 deps = MICROKERNEL_TEST_DEPS,
4180)
4181
4182xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004183 name = "f32_clamp_test",
4184 srcs = [
4185 "test/f32-clamp.cc",
4186 "test/clamp-microkernel-tester.h",
4187 ] + MICROKERNEL_TEST_HDRS,
4188 deps = MICROKERNEL_TEST_DEPS,
4189)
4190
4191xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004192 name = "f32_igemm_test",
4193 srcs = [
4194 "test/f32-igemm.cc",
4195 "test/gemm-microkernel-tester.h",
4196 "src/xnnpack/AlignedAllocator.h",
4197 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004198 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004199)
4200
4201xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004202 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004204 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205 "test/gemm-microkernel-tester.h",
4206 "src/xnnpack/AlignedAllocator.h",
4207 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004208 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004209)
4210
4211xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004212 name = "f32_igemm_minmax_test",
4213 srcs = [
4214 "test/f32-igemm-minmax.cc",
4215 "test/gemm-microkernel-tester.h",
4216 "src/xnnpack/AlignedAllocator.h",
4217 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004218 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004219)
4220
4221xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004222 name = "f16_igemm_minmax_test",
4223 srcs = [
4224 "test/f16-igemm-minmax.cc",
4225 "test/gemm-microkernel-tester.h",
4226 "src/xnnpack/AlignedAllocator.h",
4227 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004228 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004229)
4230
4231xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004232 name = "f32_conv_hwc_test",
4233 srcs = [
4234 "test/f32-conv-hwc.cc",
4235 "test/conv-hwc-microkernel-tester.h",
4236 "src/xnnpack/AlignedAllocator.h",
4237 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004238 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004239)
4240
4241xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004242 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004243 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004244 "test/f32-conv-hwc2chw.cc",
4245 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004246 "src/xnnpack/AlignedAllocator.h",
4247 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004248 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004249)
4250
4251xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004252 name = "f32_dwconv_test",
4253 srcs = [
4254 "test/f32-dwconv.cc",
4255 "test/dwconv-microkernel-tester.h",
4256 "src/xnnpack/AlignedAllocator.h",
4257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004258 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004259)
4260
4261xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004262 name = "f16_dwconv_minmax_test",
4263 srcs = [
4264 "test/f16-dwconv-minmax.cc",
4265 "test/dwconv-microkernel-tester.h",
4266 "src/xnnpack/AlignedAllocator.h",
4267 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004268 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004269)
4270
4271xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004272 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004273 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004274 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275 "test/dwconv-microkernel-tester.h",
4276 "src/xnnpack/AlignedAllocator.h",
4277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004278 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004279)
4280
4281xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004282 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004284 "test/f32-dwconv-chw.cc",
4285 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004286 "src/xnnpack/AlignedAllocator.h",
4287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004288 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289)
4290
4291xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004292 name = "f16_gavgpool_minmax_test",
4293 srcs = [
4294 "test/f16-gavgpool-minmax.cc",
4295 "test/gavgpool-microkernel-tester.h",
4296 "src/xnnpack/AlignedAllocator.h",
4297 ] + MICROKERNEL_TEST_HDRS,
4298 deps = MICROKERNEL_TEST_DEPS,
4299)
4300
4301xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004302 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004303 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004304 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305 "test/gavgpool-microkernel-tester.h",
4306 "src/xnnpack/AlignedAllocator.h",
4307 ] + MICROKERNEL_TEST_HDRS,
4308 deps = MICROKERNEL_TEST_DEPS,
4309)
4310
4311xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004312 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004313 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004314 "test/f32-gavgpool-cw.cc",
4315 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004316 "src/xnnpack/AlignedAllocator.h",
4317 ] + MICROKERNEL_TEST_HDRS,
4318 deps = MICROKERNEL_TEST_DEPS,
4319)
4320
4321xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004322 name = "f32_gemm_test",
4323 srcs = [
4324 "test/f32-gemm.cc",
4325 "test/gemm-microkernel-tester.h",
4326 "src/xnnpack/AlignedAllocator.h",
4327 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004328 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004329)
4330
4331xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004332 name = "f32_gemm_relu_test",
4333 srcs = [
4334 "test/f32-gemm-relu.cc",
4335 "test/gemm-microkernel-tester.h",
4336 "src/xnnpack/AlignedAllocator.h",
4337 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004338 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004339)
4340
4341xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004342 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004343 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004344 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 "test/gemm-microkernel-tester.h",
4346 "src/xnnpack/AlignedAllocator.h",
4347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004348 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349)
4350
4351xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004352 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004353 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004354 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 "test/gemm-microkernel-tester.h",
4356 "src/xnnpack/AlignedAllocator.h",
4357 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004358 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004359)
4360
4361xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004362 name = "f16_hswish_test",
4363 srcs = [
4364 "test/f16-hswish.cc",
4365 "test/hswish-microkernel-tester.h",
4366 ] + MICROKERNEL_TEST_HDRS,
4367 deps = MICROKERNEL_TEST_DEPS,
4368)
4369
4370xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371 name = "f32_hswish_test",
4372 srcs = [
4373 "test/f32-hswish.cc",
4374 "test/hswish-microkernel-tester.h",
4375 ] + MICROKERNEL_TEST_HDRS,
4376 deps = MICROKERNEL_TEST_DEPS,
4377)
4378
4379xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004380 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004381 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004382 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 "test/maxpool-microkernel-tester.h",
4384 ] + MICROKERNEL_TEST_HDRS,
4385 deps = MICROKERNEL_TEST_DEPS,
4386)
4387
4388xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004389 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004391 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004392 "test/avgpool-microkernel-tester.h",
4393 "src/xnnpack/AlignedAllocator.h",
4394 ] + MICROKERNEL_TEST_HDRS,
4395 deps = MICROKERNEL_TEST_DEPS,
4396)
4397
4398xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004399 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004400 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004401 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004402 "test/gemm-microkernel-tester.h",
4403 "src/xnnpack/AlignedAllocator.h",
4404 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004405 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004406)
4407
4408xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004409 name = "f16_prelu_test",
4410 srcs = [
4411 "test/f16-prelu.cc",
4412 "test/prelu-microkernel-tester.h",
4413 "src/xnnpack/AlignedAllocator.h",
4414 ] + MICROKERNEL_TEST_HDRS,
4415 deps = MICROKERNEL_TEST_DEPS,
4416)
4417
4418xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004419 name = "f32_prelu_test",
4420 srcs = [
4421 "test/f32-prelu.cc",
4422 "test/prelu-microkernel-tester.h",
4423 "src/xnnpack/AlignedAllocator.h",
4424 ] + MICROKERNEL_TEST_HDRS,
4425 deps = MICROKERNEL_TEST_DEPS,
4426)
4427
4428xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004429 name = "f32_raddexpminusmax_test",
4430 srcs = [
4431 "test/f32-raddexpminusmax.cc",
4432 "test/raddexpminusmax-microkernel-tester.h",
4433 ] + MICROKERNEL_TEST_HDRS,
4434 deps = MICROKERNEL_TEST_DEPS,
4435)
4436
4437xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004438 name = "f32_raddextexp_test",
4439 srcs = [
4440 "test/f32-raddextexp.cc",
4441 "test/raddextexp-microkernel-tester.h",
4442 ] + MICROKERNEL_TEST_HDRS,
4443 deps = MICROKERNEL_TEST_DEPS,
4444)
4445
4446xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004447 name = "f32_raddstoreexpminusmax_test",
4448 srcs = [
4449 "test/f32-raddstoreexpminusmax.cc",
4450 "test/raddstoreexpminusmax-microkernel-tester.h",
4451 ] + MICROKERNEL_TEST_HDRS,
4452 deps = MICROKERNEL_TEST_DEPS,
4453)
4454
4455xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004456 name = "f32_relu_test",
4457 srcs = [
4458 "test/f32-relu.cc",
4459 "test/vunary-microkernel-tester.h",
4460 ] + MICROKERNEL_TEST_HDRS,
4461 deps = MICROKERNEL_TEST_DEPS,
4462)
4463
4464xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465 name = "f32_rmax_test",
4466 srcs = [
4467 "test/f32-rmax.cc",
4468 "test/rmax-microkernel-tester.h",
4469 ] + MICROKERNEL_TEST_HDRS,
4470 deps = MICROKERNEL_TEST_DEPS,
4471)
4472
4473xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004474 name = "f32_sigmoid_test",
4475 srcs = [
4476 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004477 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004478 ] + MICROKERNEL_TEST_HDRS,
4479 deps = MICROKERNEL_TEST_DEPS,
4480)
4481
4482xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004483 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004484 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004485 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486 "test/spmm-microkernel-tester.h",
4487 "src/xnnpack/AlignedAllocator.h",
4488 ] + MICROKERNEL_TEST_HDRS,
4489 deps = MICROKERNEL_TEST_DEPS,
4490)
4491
4492xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004493 name = "f16_vadd_minmax_test",
4494 srcs = [
4495 "test/f16-vadd-minmax.cc",
4496 "test/vbinary-microkernel-tester.h",
4497 ] + MICROKERNEL_TEST_HDRS,
4498 deps = MICROKERNEL_TEST_DEPS,
4499)
4500
4501xnnpack_unit_test(
4502 name = "f16_vaddc_minmax_test",
4503 srcs = [
4504 "test/f16-vaddc-minmax.cc",
4505 "test/vbinaryc-microkernel-tester.h",
4506 ] + MICROKERNEL_TEST_HDRS,
4507 deps = MICROKERNEL_TEST_DEPS,
4508)
4509
4510xnnpack_unit_test(
4511 name = "f16_vdiv_minmax_test",
4512 srcs = [
4513 "test/f16-vdiv-minmax.cc",
4514 "test/vbinary-microkernel-tester.h",
4515 ] + MICROKERNEL_TEST_HDRS,
4516 deps = MICROKERNEL_TEST_DEPS,
4517)
4518
4519xnnpack_unit_test(
4520 name = "f16_vdivc_minmax_test",
4521 srcs = [
4522 "test/f16-vdivc-minmax.cc",
4523 "test/vbinaryc-microkernel-tester.h",
4524 ] + MICROKERNEL_TEST_HDRS,
4525 deps = MICROKERNEL_TEST_DEPS,
4526)
4527
4528xnnpack_unit_test(
4529 name = "f16_vrdivc_minmax_test",
4530 srcs = [
4531 "test/f16-vrdivc-minmax.cc",
4532 "test/vbinaryc-microkernel-tester.h",
4533 ] + MICROKERNEL_TEST_HDRS,
4534 deps = MICROKERNEL_TEST_DEPS,
4535)
4536
4537xnnpack_unit_test(
4538 name = "f16_vmax_test",
4539 srcs = [
4540 "test/f16-vmax.cc",
4541 "test/vbinary-microkernel-tester.h",
4542 ] + MICROKERNEL_TEST_HDRS,
4543 deps = MICROKERNEL_TEST_DEPS,
4544)
4545
4546xnnpack_unit_test(
4547 name = "f16_vmaxc_test",
4548 srcs = [
4549 "test/f16-vmaxc.cc",
4550 "test/vbinaryc-microkernel-tester.h",
4551 ] + MICROKERNEL_TEST_HDRS,
4552 deps = MICROKERNEL_TEST_DEPS,
4553)
4554
4555xnnpack_unit_test(
4556 name = "f16_vmin_test",
4557 srcs = [
4558 "test/f16-vmin.cc",
4559 "test/vbinary-microkernel-tester.h",
4560 ] + MICROKERNEL_TEST_HDRS,
4561 deps = MICROKERNEL_TEST_DEPS,
4562)
4563
4564xnnpack_unit_test(
4565 name = "f16_vminc_test",
4566 srcs = [
4567 "test/f16-vminc.cc",
4568 "test/vbinaryc-microkernel-tester.h",
4569 ] + MICROKERNEL_TEST_HDRS,
4570 deps = MICROKERNEL_TEST_DEPS,
4571)
4572
4573xnnpack_unit_test(
4574 name = "f16_vmul_minmax_test",
4575 srcs = [
4576 "test/f16-vmul-minmax.cc",
4577 "test/vbinary-microkernel-tester.h",
4578 ] + MICROKERNEL_TEST_HDRS,
4579 deps = MICROKERNEL_TEST_DEPS,
4580)
4581
4582xnnpack_unit_test(
4583 name = "f16_vmulc_minmax_test",
4584 srcs = [
4585 "test/f16-vmulc-minmax.cc",
4586 "test/vbinaryc-microkernel-tester.h",
4587 ] + MICROKERNEL_TEST_HDRS,
4588 deps = MICROKERNEL_TEST_DEPS,
4589)
4590
4591xnnpack_unit_test(
4592 name = "f16_vsub_minmax_test",
4593 srcs = [
4594 "test/f16-vsub-minmax.cc",
4595 "test/vbinary-microkernel-tester.h",
4596 ] + MICROKERNEL_TEST_HDRS,
4597 deps = MICROKERNEL_TEST_DEPS,
4598)
4599
4600xnnpack_unit_test(
4601 name = "f16_vsubc_minmax_test",
4602 srcs = [
4603 "test/f16-vsubc-minmax.cc",
4604 "test/vbinaryc-microkernel-tester.h",
4605 ] + MICROKERNEL_TEST_HDRS,
4606 deps = MICROKERNEL_TEST_DEPS,
4607)
4608
4609xnnpack_unit_test(
4610 name = "f16_vrsubc_minmax_test",
4611 srcs = [
4612 "test/f16-vrsubc-minmax.cc",
4613 "test/vbinaryc-microkernel-tester.h",
4614 ] + MICROKERNEL_TEST_HDRS,
4615 deps = MICROKERNEL_TEST_DEPS,
4616)
4617
4618xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004619 name = "f32_vabs_test",
4620 srcs = [
4621 "test/f32-vabs.cc",
4622 "test/vunary-microkernel-tester.h",
4623 ] + MICROKERNEL_TEST_HDRS,
4624 deps = MICROKERNEL_TEST_DEPS,
4625)
4626
4627xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004628 name = "f32_vadd_test",
4629 srcs = [
4630 "test/f32-vadd.cc",
4631 "test/vbinary-microkernel-tester.h",
4632 ] + MICROKERNEL_TEST_HDRS,
4633 deps = MICROKERNEL_TEST_DEPS,
4634)
4635
4636xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004637 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004638 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004639 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004640 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004641 ] + MICROKERNEL_TEST_HDRS,
4642 deps = MICROKERNEL_TEST_DEPS,
4643)
4644
4645xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004646 name = "f32_vaddc_test",
4647 srcs = [
4648 "test/f32-vaddc.cc",
4649 "test/vbinaryc-microkernel-tester.h",
4650 ] + MICROKERNEL_TEST_HDRS,
4651 deps = MICROKERNEL_TEST_DEPS,
4652)
4653
4654xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004655 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004656 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004657 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004658 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004659 ] + MICROKERNEL_TEST_HDRS,
4660 deps = MICROKERNEL_TEST_DEPS,
4661)
4662
4663xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004664 name = "f32_vdiv_test",
4665 srcs = [
4666 "test/f32-vdiv.cc",
4667 "test/vbinary-microkernel-tester.h",
4668 ] + MICROKERNEL_TEST_HDRS,
4669 deps = MICROKERNEL_TEST_DEPS,
4670)
4671
4672xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004673 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004674 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004675 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004676 "test/vbinary-microkernel-tester.h",
4677 ] + MICROKERNEL_TEST_HDRS,
4678 deps = MICROKERNEL_TEST_DEPS,
4679)
4680
4681xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004682 name = "f32_vdivc_test",
4683 srcs = [
4684 "test/f32-vdivc.cc",
4685 "test/vbinaryc-microkernel-tester.h",
4686 ] + MICROKERNEL_TEST_HDRS,
4687 deps = MICROKERNEL_TEST_DEPS,
4688)
4689
4690xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004691 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004692 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004693 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004694 "test/vbinaryc-microkernel-tester.h",
4695 ] + MICROKERNEL_TEST_HDRS,
4696 deps = MICROKERNEL_TEST_DEPS,
4697)
4698
4699xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004700 name = "f32_vrdivc_test",
4701 srcs = [
4702 "test/f32-vrdivc.cc",
4703 "test/vbinaryc-microkernel-tester.h",
4704 ] + MICROKERNEL_TEST_HDRS,
4705 deps = MICROKERNEL_TEST_DEPS,
4706)
4707
4708xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004709 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004710 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004711 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004712 "test/vbinaryc-microkernel-tester.h",
4713 ] + MICROKERNEL_TEST_HDRS,
4714 deps = MICROKERNEL_TEST_DEPS,
4715)
4716
4717xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004718 name = "f32_vmax_test",
4719 srcs = [
4720 "test/f32-vmax.cc",
4721 "test/vbinary-microkernel-tester.h",
4722 ] + MICROKERNEL_TEST_HDRS,
4723 deps = MICROKERNEL_TEST_DEPS,
4724)
4725
4726xnnpack_unit_test(
4727 name = "f32_vmaxc_test",
4728 srcs = [
4729 "test/f32-vmaxc.cc",
4730 "test/vbinaryc-microkernel-tester.h",
4731 ] + MICROKERNEL_TEST_HDRS,
4732 deps = MICROKERNEL_TEST_DEPS,
4733)
4734
4735xnnpack_unit_test(
4736 name = "f32_vmin_test",
4737 srcs = [
4738 "test/f32-vmin.cc",
4739 "test/vbinary-microkernel-tester.h",
4740 ] + MICROKERNEL_TEST_HDRS,
4741 deps = MICROKERNEL_TEST_DEPS,
4742)
4743
4744xnnpack_unit_test(
4745 name = "f32_vminc_test",
4746 srcs = [
4747 "test/f32-vminc.cc",
4748 "test/vbinaryc-microkernel-tester.h",
4749 ] + MICROKERNEL_TEST_HDRS,
4750 deps = MICROKERNEL_TEST_DEPS,
4751)
4752
4753xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004754 name = "f32_vmul_test",
4755 srcs = [
4756 "test/f32-vmul.cc",
4757 "test/vbinary-microkernel-tester.h",
4758 ] + MICROKERNEL_TEST_HDRS,
4759 deps = MICROKERNEL_TEST_DEPS,
4760)
4761
4762xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004763 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004764 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004765 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004766 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004767 ] + MICROKERNEL_TEST_HDRS,
4768 deps = MICROKERNEL_TEST_DEPS,
4769)
4770
4771xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004772 name = "f32_vmulc_test",
4773 srcs = [
4774 "test/f32-vmulc.cc",
4775 "test/vbinaryc-microkernel-tester.h",
4776 ] + MICROKERNEL_TEST_HDRS,
4777 deps = MICROKERNEL_TEST_DEPS,
4778)
4779
4780xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004781 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004782 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004783 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004784 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004785 ] + MICROKERNEL_TEST_HDRS,
4786 deps = MICROKERNEL_TEST_DEPS,
4787)
4788
4789xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004790 name = "f16_vmulcaddc_minmax_test",
4791 srcs = [
4792 "test/f16-vmulcaddc-minmax.cc",
4793 "test/vmulcaddc-microkernel-tester.h",
4794 "src/xnnpack/AlignedAllocator.h",
4795 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004796 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004797)
4798
4799xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004800 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004801 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004802 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004803 "test/vmulcaddc-microkernel-tester.h",
4804 "src/xnnpack/AlignedAllocator.h",
4805 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004806 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004807)
4808
4809xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004810 name = "f32_vlrelu_test",
4811 srcs = [
4812 "test/f32-vlrelu.cc",
4813 "test/vunary-microkernel-tester.h",
4814 ] + MICROKERNEL_TEST_HDRS,
4815 deps = MICROKERNEL_TEST_DEPS,
4816)
4817
4818xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004819 name = "f32_vneg_test",
4820 srcs = [
4821 "test/f32-vneg.cc",
4822 "test/vunary-microkernel-tester.h",
4823 ] + MICROKERNEL_TEST_HDRS,
4824 deps = MICROKERNEL_TEST_DEPS,
4825)
4826
4827xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004828 name = "f32_vrndne_test",
4829 srcs = [
4830 "test/f32-vrndne.cc",
4831 "test/vunary-microkernel-tester.h",
4832 ] + MICROKERNEL_TEST_HDRS,
4833 deps = MICROKERNEL_TEST_DEPS,
4834)
4835
4836xnnpack_unit_test(
4837 name = "f32_vrndz_test",
4838 srcs = [
4839 "test/f32-vrndz.cc",
4840 "test/vunary-microkernel-tester.h",
4841 ] + MICROKERNEL_TEST_HDRS,
4842 deps = MICROKERNEL_TEST_DEPS,
4843)
4844
4845xnnpack_unit_test(
4846 name = "f32_vrndu_test",
4847 srcs = [
4848 "test/f32-vrndu.cc",
4849 "test/vunary-microkernel-tester.h",
4850 ] + MICROKERNEL_TEST_HDRS,
4851 deps = MICROKERNEL_TEST_DEPS,
4852)
4853
4854xnnpack_unit_test(
4855 name = "f32_vrndd_test",
4856 srcs = [
4857 "test/f32-vrndd.cc",
4858 "test/vunary-microkernel-tester.h",
4859 ] + MICROKERNEL_TEST_HDRS,
4860 deps = MICROKERNEL_TEST_DEPS,
4861)
4862
4863xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004864 name = "f32_vscale_test",
4865 srcs = [
4866 "test/f32-vscale.cc",
4867 "test/vscale-microkernel-tester.h",
4868 ] + MICROKERNEL_TEST_HDRS,
4869 deps = MICROKERNEL_TEST_DEPS,
4870)
4871
4872xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004873 name = "f32_vscaleexpminusmax_test",
4874 srcs = [
4875 "test/f32-vscaleexpminusmax.cc",
4876 "test/vscaleexpminusmax-microkernel-tester.h",
4877 ] + MICROKERNEL_TEST_HDRS,
4878 deps = MICROKERNEL_TEST_DEPS,
4879)
4880
4881xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004882 name = "f32_vscaleextexp_test",
4883 srcs = [
4884 "test/f32-vscaleextexp.cc",
4885 "test/vscaleextexp-microkernel-tester.h",
4886 ] + MICROKERNEL_TEST_HDRS,
4887 deps = MICROKERNEL_TEST_DEPS,
4888)
4889
4890xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004891 name = "f32_vsqr_test",
4892 srcs = [
4893 "test/f32-vsqr.cc",
4894 "test/vunary-microkernel-tester.h",
4895 ] + MICROKERNEL_TEST_HDRS,
4896 deps = MICROKERNEL_TEST_DEPS,
4897)
4898
4899xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004900 name = "f32_vsqrdiff_test",
4901 srcs = [
4902 "test/f32-vsqrdiff.cc",
4903 "test/vbinary-microkernel-tester.h",
4904 ] + MICROKERNEL_TEST_HDRS,
4905 deps = MICROKERNEL_TEST_DEPS,
4906)
4907
4908xnnpack_unit_test(
4909 name = "f32_vsqrdiffc_test",
4910 srcs = [
4911 "test/f32-vsqrdiffc.cc",
4912 "test/vbinaryc-microkernel-tester.h",
4913 ] + MICROKERNEL_TEST_HDRS,
4914 deps = MICROKERNEL_TEST_DEPS,
4915)
4916
4917xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004918 name = "f32_vsqrt_test",
4919 srcs = [
4920 "test/f32-vsqrt.cc",
4921 "test/vunary-microkernel-tester.h",
4922 ] + MICROKERNEL_TEST_HDRS,
4923 deps = MICROKERNEL_TEST_DEPS,
4924)
4925
4926xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004927 name = "f32_vsub_test",
4928 srcs = [
4929 "test/f32-vsub.cc",
4930 "test/vbinary-microkernel-tester.h",
4931 ] + MICROKERNEL_TEST_HDRS,
4932 deps = MICROKERNEL_TEST_DEPS,
4933)
4934
4935xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004936 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004937 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004938 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004939 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004940 ] + MICROKERNEL_TEST_HDRS,
4941 deps = MICROKERNEL_TEST_DEPS,
4942)
4943
4944xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004945 name = "f32_vsubc_test",
4946 srcs = [
4947 "test/f32-vsubc.cc",
4948 "test/vbinaryc-microkernel-tester.h",
4949 ] + MICROKERNEL_TEST_HDRS,
4950 deps = MICROKERNEL_TEST_DEPS,
4951)
4952
4953xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004954 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004955 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004956 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004957 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004958 ] + MICROKERNEL_TEST_HDRS,
4959 deps = MICROKERNEL_TEST_DEPS,
4960)
4961
4962xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004963 name = "f32_vrsubc_test",
4964 srcs = [
4965 "test/f32-vrsubc.cc",
4966 "test/vbinaryc-microkernel-tester.h",
4967 ] + MICROKERNEL_TEST_HDRS,
4968 deps = MICROKERNEL_TEST_DEPS,
4969)
4970
4971xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004972 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004973 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004974 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004975 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004976 ] + MICROKERNEL_TEST_HDRS,
4977 deps = MICROKERNEL_TEST_DEPS,
4978)
4979
4980xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004981 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004982 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004983 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004984 "test/avgpool-microkernel-tester.h",
4985 "src/xnnpack/AlignedAllocator.h",
4986 ] + MICROKERNEL_TEST_HDRS,
4987 deps = MICROKERNEL_TEST_DEPS,
4988)
4989
4990xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004991 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004992 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004993 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004994 "test/gemm-microkernel-tester.h",
4995 "src/xnnpack/AlignedAllocator.h",
4996 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004997 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004998)
4999
5000xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005001 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005002 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005003 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005004 "test/dwconv-microkernel-tester.h",
5005 "src/xnnpack/AlignedAllocator.h",
5006 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005007 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005008)
5009
5010xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005011 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005013 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005014 "test/gavgpool-microkernel-tester.h",
5015 "src/xnnpack/AlignedAllocator.h",
5016 ] + MICROKERNEL_TEST_HDRS,
5017 deps = MICROKERNEL_TEST_DEPS,
5018)
5019
5020xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005021 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005022 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005023 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005024 "test/gemm-microkernel-tester.h",
5025 "src/xnnpack/AlignedAllocator.h",
5026 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005027 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005028)
5029
5030xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005031 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005032 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005033 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005034 "test/vadd-microkernel-tester.h",
5035 ] + MICROKERNEL_TEST_HDRS,
5036 deps = MICROKERNEL_TEST_DEPS,
5037)
5038
5039xnnpack_unit_test(
5040 name = "u8_clamp_test",
5041 srcs = [
5042 "test/u8-clamp.cc",
5043 "test/clamp-microkernel-tester.h",
5044 ] + MICROKERNEL_TEST_HDRS,
5045 deps = MICROKERNEL_TEST_DEPS,
5046)
5047
5048xnnpack_unit_test(
5049 name = "u8_lut32norm_test",
5050 srcs = [
5051 "test/u8-lut32norm.cc",
5052 "test/lut-norm-microkernel-tester.h",
5053 ] + MICROKERNEL_TEST_HDRS,
5054 deps = MICROKERNEL_TEST_DEPS,
5055)
5056
5057xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005058 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005059 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005060 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005061 "test/maxpool-microkernel-tester.h",
5062 ] + MICROKERNEL_TEST_HDRS,
5063 deps = MICROKERNEL_TEST_DEPS,
5064)
5065
5066xnnpack_unit_test(
5067 name = "u8_rmax_test",
5068 srcs = [
5069 "test/u8-rmax.cc",
5070 "test/rmax-microkernel-tester.h",
5071 ] + MICROKERNEL_TEST_HDRS,
5072 deps = MICROKERNEL_TEST_DEPS,
5073)
5074
5075xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005076 name = "x32_fill_test",
5077 srcs = [
5078 "test/x32-fill.cc",
5079 "test/fill-microkernel-tester.h",
5080 ] + MICROKERNEL_TEST_HDRS,
5081 deps = MICROKERNEL_TEST_DEPS,
5082)
5083
5084xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085 name = "x32_packx_test",
5086 srcs = [
5087 "test/x32-packx.cc",
5088 "test/pack-microkernel-tester.h",
5089 "src/xnnpack/AlignedAllocator.h",
5090 ] + MICROKERNEL_TEST_HDRS,
5091 deps = MICROKERNEL_TEST_DEPS,
5092)
5093
5094xnnpack_unit_test(
5095 name = "x32_pad_test",
5096 srcs = [
5097 "test/x32-pad.cc",
5098 "test/pad-microkernel-tester.h",
5099 ] + MICROKERNEL_TEST_HDRS,
5100 deps = MICROKERNEL_TEST_DEPS,
5101)
5102
5103xnnpack_unit_test(
5104 name = "x32_unpool_test",
5105 srcs = [
5106 "test/x32-unpool.cc",
5107 "test/unpool-microkernel-tester.h",
5108 ] + MICROKERNEL_TEST_HDRS,
5109 deps = MICROKERNEL_TEST_DEPS,
5110)
5111
5112xnnpack_unit_test(
5113 name = "x32_zip_test",
5114 srcs = [
5115 "test/x32-zip.cc",
5116 "test/zip-microkernel-tester.h",
5117 ] + MICROKERNEL_TEST_HDRS,
5118 deps = MICROKERNEL_TEST_DEPS,
5119)
5120
5121xnnpack_unit_test(
5122 name = "x8_lut_test",
5123 srcs = [
5124 "test/x8-lut.cc",
5125 "test/lut-microkernel-tester.h",
5126 ] + MICROKERNEL_TEST_HDRS,
5127 deps = MICROKERNEL_TEST_DEPS,
5128)
5129
5130xnnpack_unit_test(
5131 name = "x8_zip_test",
5132 srcs = [
5133 "test/x8-zip.cc",
5134 "test/zip-microkernel-tester.h",
5135 ] + MICROKERNEL_TEST_HDRS,
5136 deps = MICROKERNEL_TEST_DEPS,
5137)
5138
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005139xnnpack_unit_test(
5140 name = "requantization_test",
5141 srcs = [
5142 "src/xnnpack/requantization-stubs.h",
5143 "test/requantization.cc",
5144 "test/requantization-tester.h",
5145 ] + MICROKERNEL_TEST_HDRS,
5146 deps = MICROKERNEL_TEST_DEPS,
5147)
5148
Marat Dukhan20c3b922020-03-10 03:45:06 -07005149########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005150
5151xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005152 name = "operator_size_test",
5153 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005154 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005155)
5156
Marat Dukhan20c3b922020-03-10 03:45:06 -07005157xnnpack_binary(
5158 name = "subgraph_size_test",
5159 srcs = ["test/subgraph-size.c"],
5160 deps = [":XNNPACK"],
5161)
5162
5163########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005164
5165xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005166 name = "abs_nc_test",
5167 srcs = [
5168 "test/abs-nc.cc",
5169 "test/abs-operator-tester.h",
5170 ],
5171 deps = OPERATOR_TEST_DEPS,
5172)
5173
5174xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005175 name = "add_nd_test",
5176 srcs = [
5177 "test/add-nd.cc",
5178 "test/binary-elementwise-operator-tester.h",
5179 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005180 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005181)
5182
5183xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005184 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005185 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005186 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 "test/argmax-pooling-operator-tester.h",
5188 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005189 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190)
5191
5192xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005193 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005195 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005196 "test/average-pooling-operator-tester.h",
5197 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005198 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005199)
5200
5201xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005202 name = "bankers_rounding_nc_test",
5203 srcs = [
5204 "test/bankers-rounding-nc.cc",
5205 "test/bankers-rounding-operator-tester.h",
5206 ],
5207 deps = OPERATOR_TEST_DEPS,
5208)
5209
5210xnnpack_unit_test(
5211 name = "ceiling_nc_test",
5212 srcs = [
5213 "test/ceiling-nc.cc",
5214 "test/ceiling-operator-tester.h",
5215 ],
5216 deps = OPERATOR_TEST_DEPS,
5217)
5218
5219xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005220 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005221 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005222 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223 "test/channel-shuffle-operator-tester.h",
5224 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005225 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005226)
5227
5228xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005229 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005231 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005232 "test/clamp-operator-tester.h",
5233 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005234 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235)
5236
5237xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005238 name = "constant_pad_nd_test",
5239 srcs = [
5240 "test/constant-pad-nd.cc",
5241 "test/constant-pad-operator-tester.h",
5242 ],
5243 deps = OPERATOR_TEST_DEPS,
5244)
5245
5246xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005247 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005248 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005249 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005250 "test/convolution-operator-tester.h",
5251 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005252 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253)
5254
5255xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005256 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005257 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005258 "test/convolution-nchw.cc",
5259 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005261 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005262)
5263
5264xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005265 name = "copy_nc_test",
5266 srcs = [
5267 "test/copy-nc.cc",
5268 "test/copy-operator-tester.h",
5269 ],
5270 deps = OPERATOR_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005274 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005275 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005276 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005277 "test/deconvolution-operator-tester.h",
5278 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005279 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005280)
5281
5282xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005283 name = "divide_nd_test",
5284 srcs = [
5285 "test/binary-elementwise-operator-tester.h",
5286 "test/divide-nd.cc",
5287 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005288 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005289)
5290
5291xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005292 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005294 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 "test/fully-connected-operator-tester.h",
5296 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005297 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005298)
5299
5300xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005301 name = "floor_nc_test",
5302 srcs = [
5303 "test/floor-nc.cc",
5304 "test/floor-operator-tester.h",
5305 ],
5306 deps = OPERATOR_TEST_DEPS,
5307)
5308
5309xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005310 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005311 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005312 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005313 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005314 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005315 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316)
5317
5318xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005319 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005320 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005321 "test/global-average-pooling-ncw.cc",
5322 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005324 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325)
5326
5327xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005328 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005329 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005330 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005331 "test/hardswish-operator-tester.h",
5332 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005333 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334)
5335
5336xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005337 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005338 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005339 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005340 "test/leaky-relu-operator-tester.h",
5341 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005342 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005343)
5344
5345xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005346 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005347 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005348 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005349 "test/max-pooling-operator-tester.h",
5350 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005351 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005352)
5353
5354xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005355 name = "maximum_nd_test",
5356 srcs = [
5357 "test/binary-elementwise-operator-tester.h",
5358 "test/maximum-nd.cc",
5359 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005360 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005361)
5362
5363xnnpack_unit_test(
5364 name = "minimum_nd_test",
5365 srcs = [
5366 "test/binary-elementwise-operator-tester.h",
5367 "test/minimum-nd.cc",
5368 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005369 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005370)
5371
5372xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005373 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005374 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005375 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005376 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005377 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005378 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005379)
5380
5381xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005382 name = "negate_nc_test",
5383 srcs = [
5384 "test/negate-nc.cc",
5385 "test/negate-operator-tester.h",
5386 ],
5387 deps = OPERATOR_TEST_DEPS,
5388)
5389
5390xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005391 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005392 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005393 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005394 "test/prelu-operator-tester.h",
5395 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005396 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005397)
5398
5399xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005400 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005401 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005402 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005403 "test/resize-bilinear-operator-tester.h",
5404 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005405 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005406)
5407
5408xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005409 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005410 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005411 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412 "test/sigmoid-operator-tester.h",
5413 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005414 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415)
5416
5417xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005418 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005419 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005420 "test/softmax-nc.cc",
5421 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005422 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005423 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005424)
5425
5426xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005427 name = "square_nc_test",
5428 srcs = [
5429 "test/square-nc.cc",
5430 "test/square-operator-tester.h",
5431 ],
5432 deps = OPERATOR_TEST_DEPS,
5433)
5434
5435xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005436 name = "square_root_nc_test",
5437 srcs = [
5438 "test/square-root-nc.cc",
5439 "test/square-root-operator-tester.h",
5440 ],
5441 deps = OPERATOR_TEST_DEPS,
5442)
5443
5444xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005445 name = "squared_difference_nd_test",
5446 srcs = [
5447 "test/binary-elementwise-operator-tester.h",
5448 "test/squared-difference-nd.cc",
5449 ],
5450 deps = OPERATOR_TEST_DEPS,
5451)
5452
5453xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005454 name = "subtract_nd_test",
5455 srcs = [
5456 "test/binary-elementwise-operator-tester.h",
5457 "test/subtract-nd.cc",
5458 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005459 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005460)
5461
5462xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005463 name = "truncation_nc_test",
5464 srcs = [
5465 "test/truncation-nc.cc",
5466 "test/truncation-operator-tester.h",
5467 ],
5468 deps = OPERATOR_TEST_DEPS,
5469)
5470
5471xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005472 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005474 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475 "test/unpooling-operator-tester.h",
5476 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005477 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478)
5479
Chao Mei6ddfc602020-05-13 22:29:36 -07005480############################### Misc unit tests ###############################
5481
5482xnnpack_unit_test(
5483 name = "memory_planner_test",
5484 srcs = [
5485 "test/memory-planner-test.cc",
5486 ],
5487 deps = [
5488 ":XNNPACK",
5489 ":memory_planner",
5490 ],
5491)
5492
Marat Dukhan08c4a432019-10-03 09:29:21 -07005493############################# Build configurations #############################
5494
Marat Dukhanb8642352019-10-30 15:43:02 -07005495# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005497 name = "xnn_enable_assembly_explicit_true",
5498 define_values = {"xnn_enable_assembly": "true"},
5499)
5500
5501# Disables usage of assembly kernels.
5502config_setting(
5503 name = "xnn_enable_assembly_explicit_false",
5504 define_values = {"xnn_enable_assembly": "false"},
5505)
5506
Marat Dukhan9de90e02020-06-18 16:04:12 -07005507# Enables usage of sparse inference.
5508config_setting(
5509 name = "xnn_enable_sparse_explicit_true",
5510 define_values = {"xnn_enable_sparse": "true"},
5511)
5512
5513# Disables usage of sparse inference.
5514config_setting(
5515 name = "xnn_enable_sparse_explicit_false",
5516 define_values = {"xnn_enable_sparse": "false"},
5517)
5518
Marat Dukhan05702cf2020-03-26 15:41:33 -07005519# Disables usage of HMP-aware optimizations.
5520config_setting(
5521 name = "xnn_enable_hmp_explicit_false",
5522 define_values = {"xnn_enable_hmp": "false"},
5523)
5524
Chao Mei6ddfc602020-05-13 22:29:36 -07005525# Enable usage of optimized memory allocation
5526config_setting(
5527 name = "xnn_enable_memopt_explicit_true",
5528 define_values = {"xnnpack_enable_memopt": "true"},
5529)
5530
5531# Disable usage of optimized memory allocation
5532config_setting(
5533 name = "xnn_enable_memopt_explicit_false",
5534 define_values = {"xnnpack_enable_memopt": "false"},
5535)
5536
Marat Dukhanb8642352019-10-30 15:43:02 -07005537# Builds with -c dbg
5538config_setting(
5539 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005540 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005541 "compilation_mode": "dbg",
5542 },
5543)
5544
5545# Builds with -c opt
5546config_setting(
5547 name = "optimized_build",
5548 values = {
5549 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005550 },
5551)
5552
5553config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005554 name = "linux_k8",
5555 values = {"cpu": "k8"},
5556)
5557
5558config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005559 name = "linux_arm",
5560 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005561)
5562
5563config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005564 name = "linux_armeabi",
5565 values = {"cpu": "armeabi"},
5566)
5567
5568config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005569 name = "linux_armhf",
5570 values = {"cpu": "armhf"},
5571)
5572
5573config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005574 name = "linux_armv7a",
5575 values = {"cpu": "armv7a"},
5576)
5577
5578config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005579 name = "linux_aarch64",
5580 values = {"cpu": "aarch64"},
5581)
5582
5583config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005584 name = "android",
5585 values = {"crosstool_top": "//external:android/crosstool"},
5586)
5587
5588config_setting(
5589 name = "android_armv7",
5590 values = {
5591 "crosstool_top": "//external:android/crosstool",
5592 "cpu": "armeabi-v7a",
5593 },
5594)
5595
5596config_setting(
5597 name = "android_arm64",
5598 values = {
5599 "crosstool_top": "//external:android/crosstool",
5600 "cpu": "arm64-v8a",
5601 },
5602)
5603
5604config_setting(
5605 name = "android_x86",
5606 values = {
5607 "crosstool_top": "//external:android/crosstool",
5608 "cpu": "x86",
5609 },
5610)
5611
5612config_setting(
5613 name = "android_x86_64",
5614 values = {
5615 "crosstool_top": "//external:android/crosstool",
5616 "cpu": "x86_64",
5617 },
5618)
5619
5620config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005621 name = "windows_x86_64",
5622 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005623)
5624
5625config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005626 name = "windows_x86_64_clang",
5627 values = {
5628 "compiler": "clang-cl",
5629 "cpu": "x64_windows",
5630 },
5631)
5632
5633config_setting(
5634 name = "windows_x86_64_mingw",
5635 values = {
5636 "compiler": "mingw-gcc",
5637 "cpu": "x64_windows",
5638 },
5639)
5640
5641config_setting(
5642 name = "windows_x86_64_msys",
5643 values = {
5644 "compiler": "msys-gcc",
5645 "cpu": "x64_windows",
5646 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005647)
5648
5649config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005650 name = "macos_x86_64",
5651 values = {
5652 "apple_platform_type": "macos",
5653 "cpu": "darwin",
5654 },
5655)
5656
5657config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005658 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005659 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005660)
5661
5662config_setting(
5663 name = "emscripten_wasm",
5664 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005665 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005666 "cpu": "wasm",
5667 },
5668)
5669
5670config_setting(
5671 name = "emscripten_wasmsimd",
5672 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005673 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005674 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005675 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005676 },
5677)
5678
5679config_setting(
5680 name = "emscripten_asmjs",
5681 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005682 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005683 "cpu": "asmjs",
5684 },
5685)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005686
5687config_setting(
5688 name = "ios_armv7",
5689 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005690 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005691 "cpu": "ios_armv7",
5692 },
5693)
5694
5695config_setting(
5696 name = "ios_arm64",
5697 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005698 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005699 "cpu": "ios_arm64",
5700 },
5701)
5702
5703config_setting(
5704 name = "ios_arm64e",
5705 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005706 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005707 "cpu": "ios_arm64e",
5708 },
5709)
5710
5711config_setting(
5712 name = "ios_x86",
5713 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005714 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005715 "cpu": "ios_i386",
5716 },
5717)
5718
5719config_setting(
5720 name = "ios_x86_64",
5721 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005722 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005723 "cpu": "ios_x86_64",
5724 },
5725)
5726
5727config_setting(
5728 name = "watchos_armv7k",
5729 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005730 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005731 "cpu": "watchos_armv7k",
5732 },
5733)
5734
5735config_setting(
5736 name = "watchos_arm64_32",
5737 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005738 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005739 "cpu": "watchos_arm64_32",
5740 },
5741)
5742
5743config_setting(
5744 name = "watchos_x86",
5745 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005746 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005747 "cpu": "watchos_i386",
5748 },
5749)
5750
5751config_setting(
5752 name = "watchos_x86_64",
5753 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005754 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005755 "cpu": "watchos_x86_64",
5756 },
5757)
5758
5759config_setting(
5760 name = "tvos_arm64",
5761 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005762 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005763 "cpu": "tvos_arm64",
5764 },
5765)
5766
5767config_setting(
5768 name = "tvos_x86_64",
5769 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005770 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005771 "cpu": "tvos_x86_64",
5772 },
5773)