blob: a5a7c466a8078571233566e654c2ff116e1e43bf [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800124 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800125 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
133 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-scalar.c",
135 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-scalar.c",
137 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-scalar.c",
139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
141 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
146 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700158 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
160 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700161 "src/f32-gemm/gen/1x4-scalar.c",
162 "src/f32-gemm/gen/2x4-scalar.c",
163 "src/f32-gemm/gen/4x2-scalar.c",
164 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700165 "src/f32-gemm/gen/1x4-relu-scalar.c",
166 "src/f32-gemm/gen/2x4-relu-scalar.c",
167 "src/f32-gemm/gen/4x2-relu-scalar.c",
168 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700169 "src/f32-gemm/gen/1x4-minmax-scalar.c",
170 "src/f32-gemm/gen/2x4-minmax-scalar.c",
171 "src/f32-gemm/gen/4x2-minmax-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700191 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
193 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700194 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
195 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
197 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800198 "src/f32-prelu/gen/scalar-2x1.c",
199 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700212 "src/f32-relu/gen/scalar-x1.c",
213 "src/f32-relu/gen/scalar-x2.c",
214 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700215 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700225 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/1x1-minmax-scalar.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
228 "src/f32-spmm/gen/2x1-minmax-scalar.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/4x1-minmax-scalar.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/8x1-minmax-scalar.c",
233 "src/f32-spmm/gen/8x2-minmax-scalar.c",
234 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vadd-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
246 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
252 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700256 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
258 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800259 "src/f32-vbinary/gen/vmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
264 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
265 "src/f32-vbinary/gen/vmin-scalar-x1.c",
266 "src/f32-vbinary/gen/vmin-scalar-x2.c",
267 "src/f32-vbinary/gen/vmin-scalar-x4.c",
268 "src/f32-vbinary/gen/vminc-scalar-x1.c",
269 "src/f32-vbinary/gen/vminc-scalar-x2.c",
270 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmul-scalar-x1.c",
275 "src/f32-vbinary/gen/vmul-scalar-x2.c",
276 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
282 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
294 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700295 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
297 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
300 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsub-scalar-x1.c",
305 "src/f32-vbinary/gen/vsub-scalar-x2.c",
306 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
312 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700313 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
315 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
318 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700319 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
327 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
330 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700331 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
333 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700334 "src/f32-vunary/gen/vabs-scalar-x1.c",
335 "src/f32-vunary/gen/vabs-scalar-x2.c",
336 "src/f32-vunary/gen/vabs-scalar-x4.c",
337 "src/f32-vunary/gen/vneg-scalar-x1.c",
338 "src/f32-vunary/gen/vneg-scalar-x2.c",
339 "src/f32-vunary/gen/vneg-scalar-x4.c",
340 "src/f32-vunary/gen/vsqr-scalar-x1.c",
341 "src/f32-vunary/gen/vsqr-scalar-x2.c",
342 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800343 "src/math/expminus-scalar-lut2048-p1.c",
344 "src/math/expminus-scalar-lut64-p2.c",
345 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700346 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundne-scalar-nearbyint.c",
348 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700349 "src/math/roundd-scalar-addsub.c",
350 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700352 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700354 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700355 "src/math/roundz-scalar-addsub.c",
356 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700357 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800358 "src/math/sigmoid-scalar-lut2048-p1-div.c",
359 "src/math/sigmoid-scalar-lut64-p2-div.c",
360 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700361 "src/qs8-requantization/fp32-scalar-lrintf.c",
362 "src/qs8-requantization/fp32-scalar-magic.c",
363 "src/qs8-requantization/precise-scalar-signed64.c",
364 "src/qs8-requantization/precise-scalar-unsigned32.c",
365 "src/qs8-requantization/precise-scalar-unsigned64.c",
366 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700367 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
368 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
369 "src/qu8-dwconv/up1x9-minmax-scalar.c",
370 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
371 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
372 "src/qu8-gemm/2x2-minmax-scalar.c",
373 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700374 "src/qu8-requantization/fp32-scalar-lrintf.c",
375 "src/qu8-requantization/fp32-scalar-magic.c",
376 "src/qu8-requantization/precise-scalar-signed64.c",
377 "src/qu8-requantization/precise-scalar-unsigned32.c",
378 "src/qu8-requantization/precise-scalar-unsigned64.c",
379 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700380 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700381 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700382 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700384 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700385 "src/x32-fill/scalar-float.c",
386 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700387 "src/x32-packx/x2-scalar.c",
388 "src/x32-packx/x3-scalar.c",
389 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700390 "src/x32-pad/scalar-int.c",
391 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700392 "src/x32-unpool/scalar.c",
393 "src/x32-zip/x2-scalar.c",
394 "src/x32-zip/x3-scalar.c",
395 "src/x32-zip/x4-scalar.c",
396 "src/x32-zip/xm-scalar.c",
397 "src/x8-lut/scalar.c",
398 "src/x8-zip/x2-scalar.c",
399 "src/x8-zip/x3-scalar.c",
400 "src/x8-zip/x4-scalar.c",
401 "src/x8-zip/xm-scalar.c",
402]
403
Marat Dukhan436ebe62019-12-04 15:10:12 -0800404WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700405 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
406 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700407 "src/f32-clamp/gen/wasm-x1.c",
408 "src/f32-clamp/gen/wasm-x2.c",
409 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700410 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x4-wasm.c",
412 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x9-wasm.c",
414 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x25-wasm.c",
416 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x4-wasm.c",
418 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x9-wasm.c",
420 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
424 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700426 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
427 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700428 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
429 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
430 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
431 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700432 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
433 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700434 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
435 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700436 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
437 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
438 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700439 "src/f32-gemm/gen/1x4-wasm.c",
440 "src/f32-gemm/gen/2x4-wasm.c",
441 "src/f32-gemm/gen/4x2-wasm.c",
442 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700443 "src/f32-gemm/gen/1x4-relu-wasm.c",
444 "src/f32-gemm/gen/2x4-relu-wasm.c",
445 "src/f32-gemm/gen/4x2-relu-wasm.c",
446 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700447 "src/f32-gemm/gen/1x4-minmax-wasm.c",
448 "src/f32-gemm/gen/2x4-minmax-wasm.c",
449 "src/f32-gemm/gen/4x2-minmax-wasm.c",
450 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700451 "src/f32-hswish/gen/hswish-wasm-x1.c",
452 "src/f32-hswish/gen/hswish-wasm-x2.c",
453 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700454 "src/f32-igemm/gen/1x4-wasm.c",
455 "src/f32-igemm/gen/2x4-wasm.c",
456 "src/f32-igemm/gen/4x2-wasm.c",
457 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700458 "src/f32-igemm/gen/1x4-relu-wasm.c",
459 "src/f32-igemm/gen/2x4-relu-wasm.c",
460 "src/f32-igemm/gen/4x2-relu-wasm.c",
461 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700462 "src/f32-igemm/gen/1x4-minmax-wasm.c",
463 "src/f32-igemm/gen/2x4-minmax-wasm.c",
464 "src/f32-igemm/gen/4x2-minmax-wasm.c",
465 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700466 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
467 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
468 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700469 "src/f32-prelu/gen/wasm-2x1.c",
470 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700471 "src/f32-relu/gen/wasm-x1.c",
472 "src/f32-relu/gen/wasm-x2.c",
473 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700474 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
477 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
481 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
482 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
483 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
484 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
485 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800486 "src/f32-vbinary/gen/vmax-wasm-x1.c",
487 "src/f32-vbinary/gen/vmax-wasm-x2.c",
488 "src/f32-vbinary/gen/vmax-wasm-x4.c",
489 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
490 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
491 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
492 "src/f32-vbinary/gen/vmin-wasm-x1.c",
493 "src/f32-vbinary/gen/vmin-wasm-x2.c",
494 "src/f32-vbinary/gen/vmin-wasm-x4.c",
495 "src/f32-vbinary/gen/vminc-wasm-x1.c",
496 "src/f32-vbinary/gen/vminc-wasm-x2.c",
497 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700498 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
507 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
508 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
509 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
510 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
511 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
512 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
513 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
514 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
515 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700516 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
517 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
518 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
520 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
521 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800522]
523
Marat Dukhan290055c2020-06-09 12:24:29 -0700524WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700525 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
526 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
527 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700528 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
529 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
530 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
531 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700532 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
533 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
534 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
535 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700536 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
537 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
538 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
539 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
540 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
541 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700542 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
547 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
548 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
549 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
550 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
551 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
552 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
553 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
554 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
558 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
559 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
560 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
561 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
562 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
563 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
564 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
565 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700566 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
567 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
568 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700569 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700570 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
571 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
572 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
573 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700574 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
581 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
582 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
583 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
584 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
585 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
586 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
587 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
588 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
589 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
595 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
596 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
597 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
598 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
599 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
600 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
601 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
602 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
603 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700604 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
605 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700606 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
613 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
614 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
615 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
616 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
617 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
618 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
619 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
620 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
621 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
627 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
628 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
629 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
630 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
631 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
632 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
633 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
634 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
635 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700636 "src/f32-ibilinear/gen/wasmsimd-c4.c",
637 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700638 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
639 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
640 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700641 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700642 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
643 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
644 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
645 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700646 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
653 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
654 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
655 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
656 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
657 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
658 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
659 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
660 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
661 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
667 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
668 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
669 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
670 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
671 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
672 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
673 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
674 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
675 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700676 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
677 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700678 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700682 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
683 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700684 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
685 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
686 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700687 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
688 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700689 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
690 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
691 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
692 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700693 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
694 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700695 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
696 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
697 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700698 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
699 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700700 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
701 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
702 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
703 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
704 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
705 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
706 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700707 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
708 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700709 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
710 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
711 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
712 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700713 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
714 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
715 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
716 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
717 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
718 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
719 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
720 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
722 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
723 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700725 "src/f32-relu/gen/wasmsimd-x4.c",
726 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700727 "src/f32-rmax/wasmsimd-arm.c",
728 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700729 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
730 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
731 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
732 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
733 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
734 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700735 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
736 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
737 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
738 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
739 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
740 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700741 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700747 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700751 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
752 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700753 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
754 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
755 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
756 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700757 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700763 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
764 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700765 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
766 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
767 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
768 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700769 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
770 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
771 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
772 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700773 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700777 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
778 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
779 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
780 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700781 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
782 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
783 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
784 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700785 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
786 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700787 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
788 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
789 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
790 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700791 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
792 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700793 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
794 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
795 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
796 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700797 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
798 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700799 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
800 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
801 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
802 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700803 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
804 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700805 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
806 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700807 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
808 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700809 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
810 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
811 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
812 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700813 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
814 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700815 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
816 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
817 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
818 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700819 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
820 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700821 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
822 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
823 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
824 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700825 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
826 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
827 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
828 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700829 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
830 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
831 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
832 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
833 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
834 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
835 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
836 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
837 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
838 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
839 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
840 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
841 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
842 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700843 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
844 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700845 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
846 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
847 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
848 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
849 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
850 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700851 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700852 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700853 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700854 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700855 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700856 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700857 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700858 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700859 "src/x32-zip/x2-wasmsimd.c",
860 "src/x32-zip/x3-wasmsimd.c",
861 "src/x32-zip/x4-wasmsimd.c",
862 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700863 "src/math/roundne-wasmsimd-addsub.c",
864 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700865 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700866 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700867 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700868 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700869 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700870 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
871 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700872]
873
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800874PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700875 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700876 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700877 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700878 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700879 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700880 "src/f32-spmm/gen/4x1-minmax-psimd.c",
881 "src/f32-spmm/gen/8x1-minmax-psimd.c",
882 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700883]
884
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800885PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700886 "src/qs8-requantization/precise-psimd.c",
887 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700888 "src/qu8-requantization/precise-psimd.c",
889 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800890]
891
Marat Dukhan08c4a432019-10-03 09:29:21 -0700892# ISA-specific micro-kernels
893NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700894 "src/f32-argmaxpool/4x-neon-c4.c",
895 "src/f32-argmaxpool/9p8x-neon-c4.c",
896 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700897 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
898 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700899 "src/f32-clamp/gen/neon-x4.c",
900 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700903 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
904 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700905 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
906 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700907 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
908 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700909 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
910 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
911 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
912 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700913 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
914 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
915 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
916 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700917 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
918 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
919 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
920 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700921 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700922 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
923 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700924 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
925 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
926 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
927 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
928 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
929 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
930 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
931 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
932 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
933 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
934 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
935 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
936 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
937 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
938 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
939 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
940 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
941 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
942 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
944 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
945 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
946 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
947 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
948 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
949 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
950 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
951 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
952 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
953 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
954 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700955 "src/f32-hswish/gen/hswish-neon-x4.c",
956 "src/f32-hswish/gen/hswish-neon-x8.c",
957 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700958 "src/f32-ibilinear/gen/neon-c4.c",
959 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700960 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
961 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
962 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
963 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
964 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
965 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
966 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
967 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
968 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
969 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
970 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
971 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
972 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
973 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
974 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
975 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700976 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
977 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
978 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700979 "src/f32-ppmm/gen/4x8-minmax-neon.c",
980 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700981 "src/f32-prelu/gen/neon-1x4.c",
982 "src/f32-prelu/gen/neon-1x8.c",
983 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800984 "src/f32-prelu/gen/neon-2x4.c",
985 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700986 "src/f32-prelu/gen/neon-2x16.c",
987 "src/f32-prelu/gen/neon-4x4.c",
988 "src/f32-prelu/gen/neon-4x8.c",
989 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1005 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1013 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001014 "src/f32-relu/gen/neon-x4.c",
1015 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001016 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001017 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001018 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1019 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1020 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1021 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1022 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1023 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1026 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1027 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1028 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1029 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1030 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1031 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1032 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1033 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1034 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1035 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001036 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1037 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1038 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1039 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001040 "src/f32-vbinary/gen/vmax-neon-x4.c",
1041 "src/f32-vbinary/gen/vmax-neon-x8.c",
1042 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1043 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1044 "src/f32-vbinary/gen/vmin-neon-x4.c",
1045 "src/f32-vbinary/gen/vmin-neon-x8.c",
1046 "src/f32-vbinary/gen/vminc-neon-x4.c",
1047 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001048 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1049 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1050 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1051 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1052 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1053 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001054 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1055 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1056 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1057 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001058 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1059 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1060 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1061 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001062 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1063 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001064 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1065 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001066 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1067 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1068 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1069 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1070 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1071 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1072 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1073 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001074 "src/f32-vunary/gen/vabs-neon-x4.c",
1075 "src/f32-vunary/gen/vabs-neon-x8.c",
1076 "src/f32-vunary/gen/vneg-neon-x4.c",
1077 "src/f32-vunary/gen/vneg-neon-x8.c",
1078 "src/f32-vunary/gen/vsqr-neon-x4.c",
1079 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001080 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1081 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1082 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1083 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1084 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1085 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1086 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1087 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001088 "src/qs8-requantization/precise-neon.c",
1089 "src/qs8-requantization/fp32-neon.c",
1090 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001091 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1092 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1093 "src/qu8-dwconv/up8x9-minmax-neon.c",
1094 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1095 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1096 "src/qu8-gemm/4x8-minmax-neon.c",
1097 "src/qu8-gemm/8x8-minmax-neon.c",
1098 "src/qu8-igemm/4x8-minmax-neon.c",
1099 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001100 "src/qu8-requantization/precise-neon.c",
1101 "src/qu8-requantization/fp32-neon.c",
1102 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001103 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001104 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001105 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001106 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001107 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001108 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001109 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001110 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001111 "src/x32-zip/x2-neon.c",
1112 "src/x32-zip/x3-neon.c",
1113 "src/x32-zip/x4-neon.c",
1114 "src/x32-zip/xm-neon.c",
1115 "src/x8-zip/x2-neon.c",
1116 "src/x8-zip/x3-neon.c",
1117 "src/x8-zip/x4-neon.c",
1118 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001119 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001120 "src/math/roundd-neon-addsub.c",
1121 "src/math/roundd-neon-cvt.c",
1122 "src/math/roundu-neon-addsub.c",
1123 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001124 "src/math/roundz-neon-addsub.c",
1125 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001126 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001127 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1128 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1129 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1130 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1131 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1132 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001133 "src/math/sqrt-neon-nr1rsqrts.c",
1134 "src/math/sqrt-neon-nr2rsqrts.c",
1135 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001136]
1137
1138NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001139 "src/f32-ibilinear/gen/neonfma-c4.c",
1140 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001141 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1142 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1143 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1144 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1145 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1146 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1147 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1148 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1149 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001150 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1151 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1152 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1153 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001154 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1155 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1156 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1157 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001158 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1159 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1160 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1161 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001162 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1163 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1164 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1165 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1166 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1167 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1168 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1169 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1170 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1171 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1172 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1173 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1174 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1175 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1176 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1177 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1178 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1179 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001180 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1181 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001182 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1185 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1187 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1188 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1189 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1190 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1191 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1196 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1197 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1198 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1199 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1200 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1201 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1202 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1203 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1204 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1205 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001206 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001260 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1261 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001262 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1263 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1264 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1265 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1266 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1267 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1268 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1269 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1270 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1271 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1272 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1273 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1274 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1275 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1276 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1277 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1278 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1279 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1280 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1281 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001282 "src/math/exp-neonfma-lut64-p2.c",
1283 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001284 "src/math/expminus-neonfma-lut2048-p1.c",
1285 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001286 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001287 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1288 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1289 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1290 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1291 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1292 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1293 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1294 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1295 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1296 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1297 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1298 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1299 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1300 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1301 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1302 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1303 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1304 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001305 "src/math/sqrt-neonfma-nr1fma.c",
1306 "src/math/sqrt-neonfma-nr2fma.c",
1307 "src/math/sqrt-neonfma-nr3fma.c",
1308 "src/math/sqrt-neonfma-nr2fma1adj.c",
1309 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001310]
1311
1312AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001313 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1314 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1315 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1316 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1317 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1318 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001319 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1320 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1321 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1322 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1323 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1324 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1325 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1326 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1327 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1328 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1329 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1330 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1331 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1332 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1333 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1334 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1335 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1336 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1337 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1338 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001339 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1340 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001341 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1342 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001343 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1344 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001345 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1346 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001347 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1348 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1349 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1350 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1351 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001352 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1361 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1362 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1363 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1364 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1365 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1366 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1367 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1368 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1369 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001370 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1371 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1372 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1373 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1374 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1375 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1376 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1377 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1378 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1379 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1380 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1381 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1382 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1383 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1384 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1385 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1386 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1387 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001388 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1389 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001390 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1391 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1392 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1393 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1394 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1395 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001396]
1397
Marat Dukhan8853b822020-05-07 12:19:01 -07001398NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001399 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1400 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1401 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1402 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1403 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1404 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1405 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1406 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001407 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001408 "src/math/roundd-neonv8.c",
1409 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001410 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001411]
1412
Marat Dukhan08c4a432019-10-03 09:29:21 -07001413AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001414 "src/f16-clamp/gen/neonfp16arith-x16.c",
1415 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001416 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1417 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1418 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1419 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1420 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1421 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1422 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1423 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1424 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1425 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1426 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1427 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001428 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1429 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001430 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1431 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001432 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001433 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001434 "src/f16-relu/gen/neonfp16arith-x16.c",
1435 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001436 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1437 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1438 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1439 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1440 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1441 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1442 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1443 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1444 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1445 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1446 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1447 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1448 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1449 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1450 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1451 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1452 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1453 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1454 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1455 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1456 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1457 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1458 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1459 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1460 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1461 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1462 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1463 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001464 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1465 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001466 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1467 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1468 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1469 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1470 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1471 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1472 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1473 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001474 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1475 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1476 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1477 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001478 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1479 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1480 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1481 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1482 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1483 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1484 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1485 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1486 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1487 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1488 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1489 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001490 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1491 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1492 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1493 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1494 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1495 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1496 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1497 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001498]
1499
1500SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001501 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1502 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001503 "src/f32-clamp/gen/sse-x4.c",
1504 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001505 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1506 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001507 "src/f32-dwconv-chw/3x3p1-sse.c",
1508 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001509 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1510 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1511 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1512 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1513 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1514 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1515 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1516 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1517 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1518 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1519 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1520 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001521 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001522 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1523 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001524 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1525 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1526 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1527 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1528 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1529 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1530 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1531 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1532 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1533 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1534 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1535 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1536 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001537 "src/f32-hswish/gen/hswish-sse-x4.c",
1538 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001539 "src/f32-ibilinear/gen/sse-c4.c",
1540 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001541 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1542 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1543 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1544 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1545 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1546 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1547 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001548 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1549 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1550 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001551 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001552 "src/f32-prelu/gen/sse-2x4.c",
1553 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001554 "src/f32-relu/gen/sse-x4.c",
1555 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001556 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001557 "src/f32-spmm/gen/4x1-minmax-sse.c",
1558 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001559 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001560 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1561 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1562 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1563 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1564 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1565 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1566 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1567 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001568 "src/f32-vbinary/gen/vmax-sse-x4.c",
1569 "src/f32-vbinary/gen/vmax-sse-x8.c",
1570 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1571 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1572 "src/f32-vbinary/gen/vmin-sse-x4.c",
1573 "src/f32-vbinary/gen/vmin-sse-x8.c",
1574 "src/f32-vbinary/gen/vminc-sse-x4.c",
1575 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001576 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1577 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1578 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1579 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1580 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1581 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1582 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1583 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001584 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1585 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1586 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1587 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001588 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1589 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1590 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1591 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001592 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1593 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001594 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1595 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001596 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1597 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001598 "src/f32-vunary/gen/vabs-sse-x4.c",
1599 "src/f32-vunary/gen/vabs-sse-x8.c",
1600 "src/f32-vunary/gen/vneg-sse-x4.c",
1601 "src/f32-vunary/gen/vneg-sse-x8.c",
1602 "src/f32-vunary/gen/vsqr-sse-x4.c",
1603 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001604 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001605 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001606 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001607 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001608 "src/math/roundd-sse-addsub.c",
1609 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001610 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001611 "src/math/sqrt-sse-hh1mac.c",
1612 "src/math/sqrt-sse-nr1mac.c",
1613 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001614]
1615
1616SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001617 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1618 "src/f32-argmaxpool/4x-sse2-c4.c",
1619 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001620 "src/f32-prelu/gen/sse2-2x4.c",
1621 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001622 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1623 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1624 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1625 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1626 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1627 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1628 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1629 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1630 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1631 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1632 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1633 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001634 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001635 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001636 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001637 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001638 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1639 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001640 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1641 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001642 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1643 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1644 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1645 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1646 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1647 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1648 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1649 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001650 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1651 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1652 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001653 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1654 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001655 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1656 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001657 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1658 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001659 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1660 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001661 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001662 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1663 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001664 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001665 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1666 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1667 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001668 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1669 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001670 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1671 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001672 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1673 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001674 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001675 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1676 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001677 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001678 "src/qs8-requantization/precise-sse2.c",
1679 "src/qs8-requantization/fp32-sse2.c",
1680 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001681 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1682 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1683 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1684 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1685 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1686 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1687 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1688 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001689 "src/qu8-requantization/precise-sse2.c",
1690 "src/qu8-requantization/fp32-sse2.c",
1691 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001692 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001693 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001694 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001695 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001696 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001697 "src/x32-zip/x2-sse2.c",
1698 "src/x32-zip/x3-sse2.c",
1699 "src/x32-zip/x4-sse2.c",
1700 "src/x32-zip/xm-sse2.c",
1701 "src/x8-zip/x2-sse2.c",
1702 "src/x8-zip/x3-sse2.c",
1703 "src/x8-zip/x4-sse2.c",
1704 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001705 "src/math/exp-sse2-p5.c",
1706 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001707 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001708 "src/math/roundd-sse2-cvt.c",
1709 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001710 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001711 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001712]
1713
1714SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001715 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1716 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1717 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001718 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1719 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001720 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1721 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001722 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1723 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001724 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1725 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001726 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001727 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1728 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001729 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001730 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1731 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1732 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001733 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1734 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001735 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1736 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001737 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1738 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001739 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001740 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1741 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001742 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001743 "src/qs8-requantization/precise-ssse3.c",
1744 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001745 "src/qu8-requantization/precise-ssse3.c",
1746 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747]
1748
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001749SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001750 "src/f32-prelu/gen/sse41-2x4.c",
1751 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001752 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001753 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001754 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001755 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001756 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1757 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001758 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1759 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001760 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1761 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1762 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1763 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1764 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1765 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1766 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1767 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001768 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1769 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1770 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001771 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1772 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001773 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1774 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001775 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1776 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001777 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1778 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001779 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001780 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1781 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001782 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001783 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1784 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1785 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001786 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1787 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001788 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1789 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001790 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1791 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001792 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001793 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1794 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001795 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001796 "src/qs8-requantization/fp32-sse4.c",
1797 "src/qs8-requantization/precise-sse4.c",
1798 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001799 "src/qu8-requantization/precise-sse4.c",
1800 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001801 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001802 "src/math/roundd-sse41.c",
1803 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001804 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001805]
1806
Marat Dukhan08c4a432019-10-03 09:29:21 -07001807AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001808 "src/f32-clamp/gen/avx-x8.c",
1809 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001810 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1811 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1812 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1813 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1814 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1815 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1816 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1817 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1818 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1819 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1820 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1821 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1822 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1823 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1824 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1825 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1826 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1827 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1828 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1829 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1830 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1831 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1832 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1833 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1834 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1835 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1836 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1837 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1838 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1839 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001840 "src/f32-hswish/gen/hswish-avx-x8.c",
1841 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001842 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1843 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1844 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1845 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1846 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1847 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1848 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1849 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1850 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001851 "src/f32-prelu/gen/avx-2x8.c",
1852 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001853 "src/f32-relu/gen/avx-x16.c",
1854 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001855 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001856 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1857 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1858 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1859 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1860 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1861 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1862 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1863 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001864 "src/f32-vbinary/gen/vmax-avx-x8.c",
1865 "src/f32-vbinary/gen/vmax-avx-x16.c",
1866 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1867 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1868 "src/f32-vbinary/gen/vmin-avx-x8.c",
1869 "src/f32-vbinary/gen/vmin-avx-x16.c",
1870 "src/f32-vbinary/gen/vminc-avx-x8.c",
1871 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001872 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1873 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1874 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1875 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1876 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1877 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1878 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1879 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001880 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1881 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1882 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1883 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001884 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1885 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1886 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1887 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001888 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1889 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001890 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001891 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1892 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1893 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1894 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1895 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1896 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1897 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1898 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001899 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1900 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001901 "src/f32-vunary/gen/vabs-avx-x8.c",
1902 "src/f32-vunary/gen/vabs-avx-x16.c",
1903 "src/f32-vunary/gen/vneg-avx-x8.c",
1904 "src/f32-vunary/gen/vneg-avx-x16.c",
1905 "src/f32-vunary/gen/vsqr-avx-x8.c",
1906 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001907]
1908
Marat Dukhan1566fee2020-08-02 21:55:41 -07001909XOP_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001910 "src/qs8-dwconv/gen/up8x9-minmax-xop-mul16.c",
1911 "src/qs8-dwconv/gen/up16x9-minmax-xop-mul16.c",
1912 "src/qs8-dwconv/gen/up24x9-minmax-xop-mul16.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001913 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1914 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1915 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1916 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001917 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
1918 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001919 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1920 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001921 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001922 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1923 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001924 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001925 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
1926 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
1927 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001928 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1929 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1930 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1931 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1932 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1933 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001934 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001935 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1936 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001937 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001938]
1939
Marat Dukhanfda12b82019-11-21 12:27:59 -08001940FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001941 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1942 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1943 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1944 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1945 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1946 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1947 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1948 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1949 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1950 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1951 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1952 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1953 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1954 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1955 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1956 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1957 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1958 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1959 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1960 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1961 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1962 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1963 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1964 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1965 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1966 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1967 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1968 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1969 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1970 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1971 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1972 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1973 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1974 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1975 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1976 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1977 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1978 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1979 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1980 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001981 "src/f32-hswish/gen/hswish-fma3-x8.c",
1982 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001983 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1984 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1985 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1986 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1987 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1988 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1989 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1990 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1991 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1992 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1993 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1994 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1995 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1996 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001997 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1998 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1999 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2000 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2001 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2002 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2003 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2004 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002005 "src/math/sqrt-fma3-nr1fma.c",
2006 "src/math/sqrt-fma3-nr2fma.c",
2007 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002008]
2009
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002010AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002011 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2012 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2013 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2014 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2015 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2016 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2017 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2018 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2019 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2020 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2021 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2022 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2023 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2024 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2025 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2026 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2027 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2028 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2029 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2030 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2031 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2032 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2033 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2034 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2035 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2036 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2037 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2038 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2039 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2040 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2041 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2042 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2043 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2044 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2045 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2046 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002047 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2048 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2049 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2050 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2051 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2052 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2053 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2054 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2055 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2056 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2057 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2058 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2059 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2060 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2061 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2062 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2063 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2064 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2065 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2066 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2067 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2068 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2069 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2070 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2071 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2072 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2073 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2074 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2075 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2076 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002077 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2078 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2079 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2080 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2081 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2082 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2083 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2084 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2085 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2086 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2087 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2088 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2089 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2090 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2091 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2092 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2093 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2094 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2095 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2096 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2097 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2098 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2099 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2100 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002101 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2102 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2103 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002104 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2105 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2106 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002107 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2108 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2109 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002110 "src/math/exp-avx2-p5.c",
2111 "src/math/exp-avx2-perm-p3.c",
2112 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002113 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002114 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002115 "src/math/sigmoid-avx2-rr2-p5-div.c",
2116 "src/math/sigmoid-avx2-rr1-p5-div.c",
2117 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2118 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2119 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2120 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002121]
2122
Marat Dukhan08c4a432019-10-03 09:29:21 -07002123AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002124 "src/f32-clamp/gen/avx512f-x16.c",
2125 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002126 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2127 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2128 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2129 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2130 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2131 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2132 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2133 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2134 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2135 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2136 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2137 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2138 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2139 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2140 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2141 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2142 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2143 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2144 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2145 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2146 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2147 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2148 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2149 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002150 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2151 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002152 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2153 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2154 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2155 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2156 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2157 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002158 "src/f32-prelu/gen/avx512f-2x16.c",
2159 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002160 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2161 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2162 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2163 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2164 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2165 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2166 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2167 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2168 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2169 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2170 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2171 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2172 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2173 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2174 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2175 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2176 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2177 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2178 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2179 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2180 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2181 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2182 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2183 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2184 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2185 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2186 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2187 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2188 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2189 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2190 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2191 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2192 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2193 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2194 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2195 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002196 "src/f32-relu/gen/avx512f-x16.c",
2197 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002198 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002199 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2200 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2201 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2202 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2203 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2204 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2205 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2206 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002207 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2208 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2209 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2210 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2211 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2212 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2213 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2214 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002215 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2216 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2217 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2218 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2219 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2220 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2221 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2222 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002223 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2224 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2225 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2226 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002227 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2228 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2229 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2230 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002231 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2232 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002233 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002234 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2235 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2236 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2237 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2238 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2239 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2240 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2241 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2242 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2243 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2244 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2245 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2246 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2247 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2248 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2249 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2250 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2251 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2252 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2253 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2254 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2255 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2256 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2257 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002258 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2259 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2260 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2261 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2262 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2263 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2264 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2265 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002266 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2267 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2268 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2269 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2270 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2271 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2272 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2273 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002274 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2275 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2276 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2277 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2278 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2279 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002280 "src/math/exp-avx512f-p5-scalef.c",
2281 "src/math/exp-avx512f-p5.c",
2282 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002283 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002284 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002285 "src/math/sqrt-avx512f-nr1fma.c",
2286 "src/math/sqrt-avx512f-nr2fma.c",
2287 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002288]
2289
2290AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002291 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002292 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002293 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2294 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002295 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002296 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002297 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002298 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002299 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2300 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002301 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2302 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2303 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2304 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002305]
2306
2307AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002308 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002309 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2310 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002311 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002312 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2313 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002314 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2315 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2316 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002317 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002318 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2319 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2320 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002321 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002322 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2323 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002324 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002325 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2326 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2327 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2328 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2329 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2330 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2331 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2332 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2333 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2334 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2335 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2336 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2337 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2338 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2339 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2340 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2341 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2342 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002343 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2344 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002345 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002346 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2347 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2348 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2349 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2350 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2351 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2352 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2353 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2354 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2355 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2356 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2357 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2358 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2359 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2360 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2361 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2362 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2363 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002364 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2365 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2366 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2367 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2368 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2369 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2370 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2371 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2372 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2373 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2374 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2375 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2376 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2377 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2378 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2379 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2380 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2381 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002382]
2383
Marat Dukhan1b354632020-03-23 12:50:22 -07002384INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002385 "src/xnnpack/argmaxpool.h",
2386 "src/xnnpack/avgpool.h",
2387 "src/xnnpack/clamp.h",
2388 "src/xnnpack/common.h",
2389 "src/xnnpack/conv.h",
2390 "src/xnnpack/dwconv.h",
2391 "src/xnnpack/gavgpool.h",
2392 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002393 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002394 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002395 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002396 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002397 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002398 "src/xnnpack/lut.h",
2399 "src/xnnpack/math.h",
2400 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002401 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002402 "src/xnnpack/packx.h",
2403 "src/xnnpack/pad.h",
2404 "src/xnnpack/params.h",
2405 "src/xnnpack/pavgpool.h",
2406 "src/xnnpack/ppmm.h",
2407 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002408 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002409 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002410 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002411 "src/xnnpack/rmax.h",
2412 "src/xnnpack/scalar-utils.h",
2413 "src/xnnpack/spmm.h",
2414 "src/xnnpack/unpool.h",
2415 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002416 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002417 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002418 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002419 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002420 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002421 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002422 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002423]
2424
2425INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002426 "include/xnnpack.h",
2427 "src/xnnpack/allocator.h",
2428 "src/xnnpack/compute.h",
2429 "src/xnnpack/im2col.h",
2430 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002431 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002432 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002433 "src/xnnpack/operator.h",
2434 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002435 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002436 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002437 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002438 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002439]
2440
Marat Dukhan1b354632020-03-23 12:50:22 -07002441ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002442 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002443]
2444
Marat Dukhan1b354632020-03-23 12:50:22 -07002445MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002446 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002447 "include/xnnpack.h",
2448]
2449
Marat Dukhan1b354632020-03-23 12:50:22 -07002450MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002451 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002452 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002453 "src/xnnpack/requantization.h",
2454 "include/xnnpack.h",
2455]
2456
2457OPERATOR_TEST_PARAMS_HDRS = [
2458 "src/xnnpack/params.h",
2459 "src/xnnpack/common.h",
2460]
2461
2462WEIGHTS_PACK_HDRS = [
2463 "src/xnnpack/pack.h",
2464 "src/xnnpack/operator.h",
2465 "src/xnnpack/compute.h",
2466]
2467
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002468LOGGING_COPTS = select({
2469 # No logging in optimized mode
2470 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2471 # Full logging in debug mode
2472 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2473 # Error-only logging in default (fastbuild) mode
2474 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2475})
2476
Marat Dukhan3b59de22020-06-03 20:15:19 -07002477LOGGING_SRCS = select({
2478 # No logging in optimized mode
2479 ":optimized_build": [],
2480 "//conditions:default": [
2481 "src/operator-strings.c",
2482 "src/subgraph-strings.c",
2483 ],
2484})
2485
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002486LOGGING_HDRS = [
2487 "src/xnnpack/log.h",
2488]
2489
Marat Dukhan08c4a432019-10-03 09:29:21 -07002490xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002491 name = "tables",
2492 srcs = TABLE_SRCS,
2493 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002494 gcc_copts = xnnpack_gcc_std_copts(),
2495 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002496)
2497
2498xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002499 name = "scalar_ukernels",
2500 srcs = SCALAR_UKERNELS,
2501 hdrs = INTERNAL_HDRS,
2502 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002503 gcc_copts = xnnpack_gcc_std_copts(),
2504 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002505 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002506 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002507 "@FP16",
2508 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002509 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002510 ],
2511)
2512
2513xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002514 name = "scalar_ukernels_test_mode",
2515 srcs = SCALAR_UKERNELS,
2516 hdrs = INTERNAL_HDRS,
2517 aarch32_copts = ["-marm"],
2518 copts = [
2519 "-UNDEBUG",
2520 "-DXNN_TEST_MODE=1",
2521 ],
2522 gcc_copts = xnnpack_gcc_std_copts(),
2523 msvc_copts = xnnpack_msvc_std_copts(),
2524 deps = [
2525 ":tables",
2526 "@FP16",
2527 "@FXdiv",
2528 "@pthreadpool",
2529 ],
2530)
2531
2532xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002533 name = "wasm_ukernels",
2534 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002535 gcc_copts = xnnpack_gcc_std_copts(),
2536 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002537 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002538 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002539 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002540 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002541 "@FP16",
2542 "@FXdiv",
2543 "@pthreadpool",
2544 ],
2545)
2546
2547xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002548 name = "wasm_ukernels_test_mode",
2549 hdrs = INTERNAL_HDRS,
2550 copts = [
2551 "-UNDEBUG",
2552 "-DXNN_TEST_MODE=1",
2553 ],
2554 gcc_copts = xnnpack_gcc_std_copts(),
2555 msvc_copts = xnnpack_msvc_std_copts(),
2556 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002557 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002558 deps = [
2559 ":tables",
2560 "@FP16",
2561 "@FXdiv",
2562 "@pthreadpool",
2563 ],
2564)
2565
2566xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002567 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002568 hdrs = INTERNAL_HDRS,
2569 aarch32_copts = [
2570 "-marm",
2571 "-mfpu=neon",
2572 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002573 gcc_copts = xnnpack_gcc_std_copts(),
2574 gcc_x86_copts = ["-msse2"],
2575 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002576 optimized_copts = [
2577 "-O3",
2578 "-ffast-math",
2579 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002580 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002581 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002582 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002583 "@FP16",
2584 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002585 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002586 ],
2587)
2588
2589xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002590 name = "psimd_fastmath_ukernels_test_mode",
2591 hdrs = INTERNAL_HDRS,
2592 aarch32_copts = [
2593 "-marm",
2594 "-mfpu=neon",
2595 ],
2596 copts = [
2597 "-UNDEBUG",
2598 "-DXNN_TEST_MODE=1",
2599 ],
2600 gcc_copts = xnnpack_gcc_std_copts(),
2601 gcc_x86_copts = ["-msse2"],
2602 msvc_copts = xnnpack_msvc_std_copts(),
2603 optimized_copts = [
2604 "-O3",
2605 "-ffast-math",
2606 ],
2607 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2608 deps = [
2609 ":tables",
2610 "@FP16",
2611 "@psimd",
2612 "@pthreadpool",
2613 ],
2614)
2615
2616xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002617 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002618 hdrs = INTERNAL_HDRS,
2619 aarch32_copts = [
2620 "-marm",
2621 "-mfpu=neon",
2622 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002623 gcc_copts = xnnpack_gcc_std_copts(),
2624 gcc_x86_copts = ["-msse2"],
2625 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002626 optimized_copts = [
2627 "-O3",
2628 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002629 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002630 deps = [
2631 ":tables",
2632 "@FP16",
2633 "@psimd",
2634 "@pthreadpool",
2635 ],
2636)
2637
2638xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002639 name = "psimd_accmath_ukernels_test_mode",
2640 hdrs = INTERNAL_HDRS,
2641 aarch32_copts = [
2642 "-marm",
2643 "-mfpu=neon",
2644 ],
2645 copts = [
2646 "-UNDEBUG",
2647 "-DXNN_TEST_MODE=1",
2648 ],
2649 gcc_copts = xnnpack_gcc_std_copts(),
2650 gcc_x86_copts = ["-msse2"],
2651 msvc_copts = xnnpack_msvc_std_copts(),
2652 optimized_copts = [
2653 "-O3",
2654 ],
2655 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2656 deps = [
2657 ":tables",
2658 "@FP16",
2659 "@psimd",
2660 "@pthreadpool",
2661 ],
2662)
2663
2664xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002665 name = "neon_ukernels",
2666 hdrs = INTERNAL_HDRS,
2667 aarch32_copts = [
2668 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002669 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002670 "-mfpu=neon",
2671 ],
2672 aarch32_srcs = NEON_UKERNELS,
2673 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002674 gcc_copts = xnnpack_gcc_std_copts(),
2675 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002676 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002677 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002678 "@FP16",
2679 "@pthreadpool",
2680 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002681)
2682
2683xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002684 name = "neon_ukernels_test_mode",
2685 hdrs = INTERNAL_HDRS,
2686 aarch32_copts = [
2687 "-marm",
2688 "-march=armv7-a",
2689 "-mfpu=neon",
2690 ],
2691 aarch32_srcs = NEON_UKERNELS,
2692 aarch64_srcs = NEON_UKERNELS,
2693 copts = [
2694 "-UNDEBUG",
2695 "-DXNN_TEST_MODE=1",
2696 ],
2697 gcc_copts = xnnpack_gcc_std_copts(),
2698 msvc_copts = xnnpack_msvc_std_copts(),
2699 deps = [
2700 ":tables",
2701 "@FP16",
2702 "@pthreadpool",
2703 ],
2704)
2705
2706xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002707 name = "neonfma_ukernels",
2708 hdrs = INTERNAL_HDRS,
2709 aarch32_copts = [
2710 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002711 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002712 "-mfpu=neon-vfpv4",
2713 ],
2714 aarch32_srcs = NEONFMA_UKERNELS,
2715 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002716 apple_aarch32_copts = [
2717 "-mcpu=swift",
2718 "-mtune=generic",
2719 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002720 gcc_copts = xnnpack_gcc_std_copts(),
2721 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002722 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002723 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002724 "@FP16",
2725 "@pthreadpool",
2726 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002727)
2728
2729xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002730 name = "neonfma_ukernels_test_mode",
2731 hdrs = INTERNAL_HDRS,
2732 aarch32_copts = [
2733 "-marm",
2734 "-march=armv7-a",
2735 "-mfpu=neon-vfpv4",
2736 ],
2737 aarch32_srcs = NEONFMA_UKERNELS,
2738 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002739 apple_aarch32_copts = [
2740 "-mcpu=swift",
2741 "-mtune=generic",
2742 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002743 copts = [
2744 "-UNDEBUG",
2745 "-DXNN_TEST_MODE=1",
2746 ],
2747 gcc_copts = xnnpack_gcc_std_copts(),
2748 msvc_copts = xnnpack_msvc_std_copts(),
2749 deps = [
2750 ":tables",
2751 "@FP16",
2752 "@pthreadpool",
2753 ],
2754)
2755
2756xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002757 name = "neonv8_ukernels",
2758 hdrs = INTERNAL_HDRS,
2759 aarch32_copts = [
2760 "-marm",
2761 "-march=armv8-a",
2762 "-mfpu=neon-fp-armv8",
2763 ],
2764 aarch32_srcs = NEONV8_UKERNELS,
2765 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002766 apple_aarch32_copts = [
2767 "-mcpu=cyclone",
2768 "-mtune=generic",
2769 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002770 gcc_copts = xnnpack_gcc_std_copts(),
2771 msvc_copts = xnnpack_msvc_std_copts(),
2772 deps = [
2773 ":tables",
2774 "@FP16",
2775 "@pthreadpool",
2776 ],
2777)
2778
2779xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002780 name = "neonv8_ukernels_test_mode",
2781 hdrs = INTERNAL_HDRS,
2782 aarch32_copts = [
2783 "-marm",
2784 "-march=armv8-a",
2785 "-mfpu=neon-fp-armv8",
2786 ],
2787 aarch32_srcs = NEONV8_UKERNELS,
2788 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002789 apple_aarch32_copts = [
2790 "-mcpu=cyclone",
2791 "-mtune=generic",
2792 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002793 copts = [
2794 "-UNDEBUG",
2795 "-DXNN_TEST_MODE=1",
2796 ],
2797 gcc_copts = xnnpack_gcc_std_copts(),
2798 msvc_copts = xnnpack_msvc_std_copts(),
2799 deps = [
2800 ":tables",
2801 "@FP16",
2802 "@pthreadpool",
2803 ],
2804)
2805
2806xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002807 name = "neonfp16arith_ukernels",
2808 hdrs = INTERNAL_HDRS,
2809 aarch64_copts = ["-march=armv8.2-a+fp16"],
2810 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002811 gcc_copts = xnnpack_gcc_std_copts(),
2812 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002813 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002814 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002815 "@FP16",
2816 "@pthreadpool",
2817 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002818)
2819
2820xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002821 name = "neonfp16arith_ukernels_test_mode",
2822 hdrs = INTERNAL_HDRS,
2823 aarch64_copts = ["-march=armv8.2-a+fp16"],
2824 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2825 copts = [
2826 "-UNDEBUG",
2827 "-DXNN_TEST_MODE=1",
2828 ],
2829 gcc_copts = xnnpack_gcc_std_copts(),
2830 msvc_copts = xnnpack_msvc_std_copts(),
2831 deps = [
2832 ":tables",
2833 "@FP16",
2834 "@pthreadpool",
2835 ],
2836)
2837
2838xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002839 name = "sse2_ukernels",
2840 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002841 gcc_copts = xnnpack_gcc_std_copts(),
2842 gcc_x86_copts = ["-msse2"],
2843 msvc_copts = xnnpack_msvc_std_copts(),
2844 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002845 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002846 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002847 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002848 "@FP16",
2849 "@pthreadpool",
2850 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002851)
2852
2853xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002854 name = "sse2_ukernels_test_mode",
2855 hdrs = INTERNAL_HDRS,
2856 copts = [
2857 "-UNDEBUG",
2858 "-DXNN_TEST_MODE=1",
2859 ],
2860 gcc_copts = xnnpack_gcc_std_copts(),
2861 gcc_x86_copts = ["-msse2"],
2862 msvc_copts = xnnpack_msvc_std_copts(),
2863 msvc_x86_32_copts = ["/arch:SSE2"],
2864 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2865 deps = [
2866 ":tables",
2867 "@FP16",
2868 "@pthreadpool",
2869 ],
2870)
2871
2872xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002873 name = "ssse3_ukernels",
2874 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002875 gcc_copts = xnnpack_gcc_std_copts(),
2876 gcc_x86_copts = ["-mssse3"],
2877 msvc_copts = xnnpack_msvc_std_copts(),
2878 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002879 x86_srcs = SSSE3_UKERNELS,
2880 deps = [
2881 ":tables",
2882 "@FP16",
2883 "@pthreadpool",
2884 ],
2885)
2886
2887xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002888 name = "ssse3_ukernels_test_mode",
2889 hdrs = INTERNAL_HDRS,
2890 copts = [
2891 "-UNDEBUG",
2892 "-DXNN_TEST_MODE=1",
2893 ],
2894 gcc_copts = xnnpack_gcc_std_copts(),
2895 gcc_x86_copts = ["-mssse3"],
2896 msvc_copts = xnnpack_msvc_std_copts(),
2897 msvc_x86_32_copts = ["/arch:SSE2"],
2898 x86_srcs = SSSE3_UKERNELS,
2899 deps = [
2900 ":tables",
2901 "@FP16",
2902 "@pthreadpool",
2903 ],
2904)
2905
2906xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002907 name = "sse41_ukernels",
2908 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002909 gcc_copts = xnnpack_gcc_std_copts(),
2910 gcc_x86_copts = ["-msse4.1"],
2911 msvc_copts = xnnpack_msvc_std_copts(),
2912 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002913 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002914 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002915 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002916 "@FP16",
2917 "@pthreadpool",
2918 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002919)
2920
2921xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002922 name = "sse41_ukernels_test_mode",
2923 hdrs = INTERNAL_HDRS,
2924 copts = [
2925 "-UNDEBUG",
2926 "-DXNN_TEST_MODE=1",
2927 ],
2928 gcc_copts = xnnpack_gcc_std_copts(),
2929 gcc_x86_copts = ["-msse4.1"],
2930 msvc_copts = xnnpack_msvc_std_copts(),
2931 msvc_x86_32_copts = ["/arch:SSE2"],
2932 x86_srcs = SSE41_UKERNELS,
2933 deps = [
2934 ":tables",
2935 "@FP16",
2936 "@pthreadpool",
2937 ],
2938)
2939
2940xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002941 name = "avx_ukernels",
2942 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002943 gcc_copts = xnnpack_gcc_std_copts(),
2944 gcc_x86_copts = ["-mavx"],
2945 msvc_copts = xnnpack_msvc_std_copts(),
2946 msvc_x86_32_copts = ["/arch:AVX"],
2947 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002948 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002949 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002950 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002951 "@FP16",
2952 "@pthreadpool",
2953 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002954)
2955
2956xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002957 name = "avx_ukernels_test_mode",
2958 hdrs = INTERNAL_HDRS,
2959 copts = [
2960 "-UNDEBUG",
2961 "-DXNN_TEST_MODE=1",
2962 ],
2963 gcc_copts = xnnpack_gcc_std_copts(),
2964 gcc_x86_copts = ["-mavx"],
2965 msvc_copts = xnnpack_msvc_std_copts(),
2966 msvc_x86_32_copts = ["/arch:AVX"],
2967 msvc_x86_64_copts = ["/arch:AVX"],
2968 x86_srcs = AVX_UKERNELS,
2969 deps = [
2970 ":tables",
2971 "@FP16",
2972 "@pthreadpool",
2973 ],
2974)
2975
2976xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07002977 name = "xop_ukernels",
2978 hdrs = INTERNAL_HDRS,
2979 gcc_copts = xnnpack_gcc_std_copts(),
2980 gcc_x86_copts = ["-mxop"],
2981 msvc_copts = xnnpack_msvc_std_copts(),
2982 msvc_x86_32_copts = ["/arch:AVX"],
2983 msvc_x86_64_copts = ["/arch:AVX"],
2984 x86_srcs = XOP_UKERNELS,
2985 deps = [
2986 ":tables",
2987 "@FP16",
2988 "@pthreadpool",
2989 ],
2990)
2991
2992xnnpack_cc_library(
2993 name = "xop_ukernels_test_mode",
2994 hdrs = INTERNAL_HDRS,
2995 copts = [
2996 "-UNDEBUG",
2997 "-DXNN_TEST_MODE=1",
2998 ],
2999 gcc_copts = xnnpack_gcc_std_copts(),
3000 gcc_x86_copts = ["-mxop"],
3001 msvc_copts = xnnpack_msvc_std_copts(),
3002 msvc_x86_32_copts = ["/arch:AVX"],
3003 msvc_x86_64_copts = ["/arch:AVX"],
3004 x86_srcs = XOP_UKERNELS,
3005 deps = [
3006 ":tables",
3007 "@FP16",
3008 "@pthreadpool",
3009 ],
3010)
3011
3012xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003013 name = "fma3_ukernels",
3014 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003015 gcc_copts = xnnpack_gcc_std_copts(),
3016 gcc_x86_copts = ["-mfma"],
3017 msvc_copts = xnnpack_msvc_std_copts(),
3018 msvc_x86_32_copts = ["/arch:AVX"],
3019 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003020 x86_srcs = FMA3_UKERNELS,
3021 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003022 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003023 "@FP16",
3024 "@pthreadpool",
3025 ],
3026)
3027
3028xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003029 name = "fma3_ukernels_test_mode",
3030 hdrs = INTERNAL_HDRS,
3031 copts = [
3032 "-UNDEBUG",
3033 "-DXNN_TEST_MODE=1",
3034 ],
3035 gcc_copts = xnnpack_gcc_std_copts(),
3036 gcc_x86_copts = ["-mfma"],
3037 msvc_copts = xnnpack_msvc_std_copts(),
3038 msvc_x86_32_copts = ["/arch:AVX"],
3039 msvc_x86_64_copts = ["/arch:AVX"],
3040 x86_srcs = FMA3_UKERNELS,
3041 deps = [
3042 ":tables",
3043 "@FP16",
3044 "@pthreadpool",
3045 ],
3046)
3047
3048xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003049 name = "avx2_ukernels",
3050 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003051 gcc_copts = xnnpack_gcc_std_copts(),
3052 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003053 "-mfma",
3054 "-mavx2",
3055 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003056 msvc_copts = xnnpack_msvc_std_copts(),
3057 msvc_x86_32_copts = ["/arch:AVX2"],
3058 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003059 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003060 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003061 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003062 "@FP16",
3063 "@pthreadpool",
3064 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003065)
3066
3067xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003068 name = "avx2_ukernels_test_mode",
3069 hdrs = INTERNAL_HDRS,
3070 copts = [
3071 "-UNDEBUG",
3072 "-DXNN_TEST_MODE=1",
3073 ],
3074 gcc_copts = xnnpack_gcc_std_copts(),
3075 gcc_x86_copts = [
3076 "-mfma",
3077 "-mavx2",
3078 ],
3079 msvc_copts = xnnpack_msvc_std_copts(),
3080 msvc_x86_32_copts = ["/arch:AVX2"],
3081 msvc_x86_64_copts = ["/arch:AVX2"],
3082 x86_srcs = AVX2_UKERNELS,
3083 deps = [
3084 ":tables",
3085 "@FP16",
3086 "@pthreadpool",
3087 ],
3088)
3089
3090xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003091 name = "avx512f_ukernels",
3092 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003093 gcc_copts = xnnpack_gcc_std_copts(),
3094 gcc_x86_copts = ["-mavx512f"],
3095 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3096 msvc_copts = xnnpack_msvc_std_copts(),
3097 msvc_x86_32_copts = ["/arch:AVX512"],
3098 msvc_x86_64_copts = ["/arch:AVX512"],
3099 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003100 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003101 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003102 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003103 "@FP16",
3104 "@pthreadpool",
3105 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106)
3107
3108xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003109 name = "avx512f_ukernels_test_mode",
3110 hdrs = INTERNAL_HDRS,
3111 copts = [
3112 "-UNDEBUG",
3113 "-DXNN_TEST_MODE=1",
3114 ],
3115 gcc_copts = xnnpack_gcc_std_copts(),
3116 gcc_x86_copts = ["-mavx512f"],
3117 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3118 msvc_copts = xnnpack_msvc_std_copts(),
3119 msvc_x86_32_copts = ["/arch:AVX512"],
3120 msvc_x86_64_copts = ["/arch:AVX512"],
3121 msys_copts = ["-fno-asynchronous-unwind-tables"],
3122 x86_srcs = AVX512F_UKERNELS,
3123 deps = [
3124 ":tables",
3125 "@FP16",
3126 "@pthreadpool",
3127 ],
3128)
3129
3130xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003131 name = "asm_ukernels",
3132 hdrs = ["src/xnnpack/assembly.h"],
3133 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003134 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003135 aarch64_srcs = AARCH64_ASM_UKERNELS,
3136)
3137
Marat Dukhan3b59de22020-06-03 20:15:19 -07003138xnnpack_cc_library(
3139 name = "logging_utils",
3140 srcs = LOGGING_SRCS,
3141 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3142 copts = LOGGING_COPTS + [
3143 "-Isrc",
3144 "-Iinclude",
3145 ] + select({
3146 ":debug_build": [],
3147 "//conditions:default": xnnpack_min_size_copts(),
3148 }),
3149 gcc_copts = xnnpack_gcc_std_copts(),
3150 msvc_copts = xnnpack_msvc_std_copts(),
3151 visibility = xnnpack_visibility(),
3152 deps = [
3153 "@FP16",
3154 "@clog",
3155 "@pthreadpool",
3156 ],
3157)
3158
Marat Dukhan08c4a432019-10-03 09:29:21 -07003159xnnpack_aggregate_library(
3160 name = "ukernels",
3161 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003162 ":neon_ukernels",
3163 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003164 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165 ":asm_ukernels",
3166 ],
3167 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003168 ":neon_ukernels",
3169 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003170 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003171 ":neonfp16arith_ukernels",
3172 ":asm_ukernels",
3173 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003174 generic_deps = [
3175 ":scalar_ukernels",
3176 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003177 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003178 ":psimd_fastmath_ukernels",
3179 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003181 wasm_deps = [
3182 ":wasm_ukernels",
3183 ],
3184 wasmsimd_deps = [
3185 ":wasm_ukernels",
3186 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003188 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003189 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003190 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003191 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003192 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003193 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003194 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003195 ":avx512f_ukernels",
3196 ],
3197)
3198
Marat Dukhan33fcf782020-05-24 14:27:15 -07003199xnnpack_aggregate_library(
3200 name = "ukernels_test_mode",
3201 aarch32_deps = [
3202 ":neon_ukernels_test_mode",
3203 ":neonfma_ukernels_test_mode",
3204 ":neonv8_ukernels_test_mode",
3205 ":asm_ukernels",
3206 ],
3207 aarch64_deps = [
3208 ":neon_ukernels_test_mode",
3209 ":neonfma_ukernels_test_mode",
3210 ":neonv8_ukernels_test_mode",
3211 ":neonfp16arith_ukernels_test_mode",
3212 ":asm_ukernels",
3213 ],
3214 generic_deps = [
3215 ":scalar_ukernels_test_mode",
3216 ],
3217 psimd_deps = [
3218 ":psimd_fastmath_ukernels_test_mode",
3219 ":psimd_accmath_ukernels_test_mode",
3220 ],
3221 wasm_deps = [
3222 ":wasm_ukernels_test_mode",
3223 ],
3224 wasmsimd_deps = [
3225 ":wasm_ukernels_test_mode",
3226 ],
3227 x86_deps = [
3228 ":sse2_ukernels_test_mode",
3229 ":ssse3_ukernels_test_mode",
3230 ":sse41_ukernels_test_mode",
3231 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003232 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003233 ":fma3_ukernels_test_mode",
3234 ":avx2_ukernels_test_mode",
3235 ":avx512f_ukernels_test_mode",
3236 ],
3237)
3238
Marat Dukhan86df3842020-06-17 14:03:31 -07003239# This target must be compatible with Android NDK r17 and compile with Clang 6.
3240# It is similar to :ukernels target, but excludes microkernels which require
3241# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003242xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003243 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003244 aarch32_deps = [
3245 ":neon_ukernels",
3246 ":neonfma_ukernels",
3247 ":neonv8_ukernels",
3248 ":asm_ukernels",
3249 ],
3250 aarch64_deps = [
3251 ":neon_ukernels",
3252 ":neonfma_ukernels",
3253 ":neonv8_ukernels",
3254 ":asm_ukernels",
3255 ],
3256 generic_deps = [
3257 ":scalar_ukernels",
3258 ],
3259 psimd_deps = [
3260 ":psimd_fastmath_ukernels",
3261 ":psimd_accmath_ukernels",
3262 ],
3263 wasm_deps = [
3264 ":wasm_ukernels",
3265 ],
3266 wasmsimd_deps = [
3267 ":wasm_ukernels",
3268 ],
3269 x86_deps = [
3270 ":sse2_ukernels",
3271 ":ssse3_ukernels",
3272 ":sse41_ukernels",
3273 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003274 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003275 ":fma3_ukernels",
3276 ":avx2_ukernels",
3277 ":avx512f_ukernels",
3278 ],
3279)
3280
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281xnnpack_cc_library(
3282 name = "im2col",
3283 srcs = ["src/im2col.c"],
3284 hdrs = [
3285 "src/xnnpack/common.h",
3286 "src/xnnpack/im2col.h",
3287 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003288 gcc_copts = xnnpack_gcc_std_copts(),
3289 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003290)
3291
3292xnnpack_cc_library(
3293 name = "indirection",
3294 srcs = ["src/indirection.c"],
3295 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003296 gcc_copts = xnnpack_gcc_std_copts(),
3297 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003298 deps = [
3299 "@FP16",
3300 "@FXdiv",
3301 "@pthreadpool",
3302 ],
3303)
3304
3305xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003306 name = "indirection_test_mode",
3307 srcs = ["src/indirection.c"],
3308 hdrs = INTERNAL_HDRS,
3309 copts = [
3310 "-UNDEBUG",
3311 "-DXNN_TEST_MODE=1",
3312 ],
3313 gcc_copts = xnnpack_gcc_std_copts(),
3314 msvc_copts = xnnpack_msvc_std_copts(),
3315 deps = [
3316 "@FP16",
3317 "@FXdiv",
3318 "@pthreadpool",
3319 ],
3320)
3321
3322xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003323 name = "packing",
3324 srcs = ["src/packing.c"],
3325 hdrs = INTERNAL_HDRS,
3326 gcc_copts = xnnpack_gcc_std_copts(),
3327 msvc_copts = xnnpack_msvc_std_copts(),
3328 deps = [
3329 "@FP16",
3330 "@FXdiv",
3331 "@pthreadpool",
3332 ],
3333)
3334
3335xnnpack_cc_library(
3336 name = "packing_test_mode",
3337 srcs = ["src/packing.c"],
3338 hdrs = INTERNAL_HDRS,
3339 copts = [
3340 "-UNDEBUG",
3341 "-DXNN_TEST_MODE=1",
3342 ],
3343 gcc_copts = xnnpack_gcc_std_copts(),
3344 msvc_copts = xnnpack_msvc_std_copts(),
3345 deps = [
3346 "@FP16",
3347 "@FXdiv",
3348 "@pthreadpool",
3349 ],
3350)
3351
3352xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003353 name = "operator_run",
3354 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003355 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003356 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003357 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3358 "//conditions:default": [],
3359 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003360 gcc_copts = xnnpack_gcc_std_copts(),
3361 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003362 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003363 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364 "@FP16",
3365 "@FXdiv",
3366 "@clog",
3367 "@pthreadpool",
3368 ],
3369)
3370
Chao Mei6ddfc602020-05-13 22:29:36 -07003371xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003372 name = "operator_run_test_mode",
3373 srcs = ["src/operator-run.c"],
3374 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3375 copts = LOGGING_COPTS + [
3376 "-UNDEBUG",
3377 "-DXNN_TEST_MODE=1",
3378 ] + select({
3379 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3380 "//conditions:default": [],
3381 }),
3382 gcc_copts = xnnpack_gcc_std_copts(),
3383 msvc_copts = xnnpack_msvc_std_copts(),
3384 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003385 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003386 "@FP16",
3387 "@FXdiv",
3388 "@clog",
3389 "@pthreadpool",
3390 ],
3391)
3392
3393xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003394 name = "memory_planner",
3395 srcs = ["src/memory-planner.c"],
3396 hdrs = INTERNAL_HDRS,
3397 defines = select({
3398 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3399 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3400 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3401 }),
3402 gcc_copts = xnnpack_gcc_std_copts(),
3403 msvc_copts = xnnpack_msvc_std_copts(),
3404 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003405 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003406 "@pthreadpool",
3407 ],
3408)
3409
Marat Dukhan33fcf782020-05-24 14:27:15 -07003410xnnpack_cc_library(
3411 name = "memory_planner_test_mode",
3412 srcs = ["src/memory-planner.c"],
3413 hdrs = INTERNAL_HDRS,
3414 copts = [
3415 "-UNDEBUG",
3416 "-DXNN_TEST_MODE=1",
3417 ],
3418 defines = select({
3419 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3420 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3421 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3422 }),
3423 gcc_copts = xnnpack_gcc_std_copts(),
3424 msvc_copts = xnnpack_msvc_std_copts(),
3425 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003426 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003427 "@pthreadpool",
3428 ],
3429)
3430
Marat Dukhan08c4a432019-10-03 09:29:21 -07003431cc_library(
3432 name = "enable_assembly",
3433 defines = select({
3434 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3435 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003436 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003437 }),
3438)
3439
Marat Dukhan9de90e02020-06-18 16:04:12 -07003440cc_library(
3441 name = "enable_sparse",
3442 defines = select({
3443 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3444 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003445 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003446 }),
3447)
3448
Marat Dukhancf056b22019-10-07 10:26:29 -07003449xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003450 name = "operators",
3451 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003452 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003453 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003454 ],
3455 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003456 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003457 "-Isrc",
3458 "-Iinclude",
3459 ] + select({
3460 ":debug_build": [],
3461 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003462 }) + select({
3463 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3464 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003465 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003466 gcc_copts = xnnpack_gcc_std_copts(),
3467 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003468 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003469 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003470 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003471 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003472 "@FP16",
3473 "@FXdiv",
3474 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003475 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003476 ],
3477)
3478
Marat Dukhan10a38082020-04-17 03:58:35 -07003479xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003480 name = "operators_test_mode",
3481 srcs = OPERATOR_SRCS + [
3482 "src/memory.c",
3483 "src/operator-delete.c",
3484 ],
3485 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3486 copts = LOGGING_COPTS + [
3487 "-Isrc",
3488 "-Iinclude",
3489 "-UNDEBUG",
3490 "-DXNN_TEST_MODE=1",
3491 ] + select({
3492 ":debug_build": [],
3493 "//conditions:default": xnnpack_min_size_copts(),
3494 }) + select({
3495 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3496 "//conditions:default": [],
3497 }),
3498 gcc_copts = xnnpack_gcc_std_copts(),
3499 msvc_copts = xnnpack_msvc_std_copts(),
3500 deps = [
3501 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003502 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003503 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003504 "@FP16",
3505 "@FXdiv",
3506 "@clog",
3507 "@pthreadpool",
3508 ],
3509)
3510
3511xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003512 name = "XNNPACK",
3513 srcs = [
3514 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003515 "src/runtime.c",
3516 "src/subgraph.c",
3517 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003518 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003519 hdrs = ["include/xnnpack.h"],
3520 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003521 "-Isrc",
3522 "-Iinclude",
3523 ] + select({
3524 ":debug_build": [],
3525 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003526 }) + select({
3527 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3528 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003529 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003530 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003531 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003532 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003533 visibility = xnnpack_visibility(),
3534 deps = [
3535 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003536 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003537 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003538 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003539 ":operator_run",
3540 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003541 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003542 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003543 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003544 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003545 ] + select({
3546 ":emscripten": [],
3547 "//conditions:default": ["@cpuinfo"],
3548 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003549)
3550
Marat Dukhan10a38082020-04-17 03:58:35 -07003551xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003552 name = "XNNPACK_test_mode",
3553 srcs = [
3554 "src/init.c",
3555 "src/runtime.c",
3556 "src/subgraph.c",
3557 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003558 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003559 hdrs = ["include/xnnpack.h"],
3560 copts = LOGGING_COPTS + [
3561 "-Isrc",
3562 "-Iinclude",
3563 "-UNDEBUG",
3564 "-DXNN_TEST_MODE=1",
3565 ] + select({
3566 ":debug_build": [],
3567 "//conditions:default": xnnpack_min_size_copts(),
3568 }) + select({
3569 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3570 "//conditions:default": [],
3571 }),
3572 gcc_copts = xnnpack_gcc_std_copts(),
3573 includes = ["include"],
3574 msvc_copts = xnnpack_msvc_std_copts(),
3575 visibility = xnnpack_visibility(),
3576 deps = [
3577 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003578 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003579 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003580 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003581 ":operator_run_test_mode",
3582 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003583 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003584 "@clog",
3585 "@FP16",
3586 "@pthreadpool",
3587 ] + select({
3588 ":emscripten": [],
3589 "//conditions:default": ["@cpuinfo"],
3590 }),
3591)
3592
3593xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003594 name = "xnnpack_f32",
3595 srcs = [
3596 "src/init.c",
3597 "src/runtime.c",
3598 "src/subgraph.c",
3599 "src/tensor.c",
3600 ] + SUBGRAPH_SRCS,
3601 hdrs = ["include/xnnpack.h"],
3602 copts = LOGGING_COPTS + [
3603 "-Isrc",
3604 "-Iinclude",
3605 ] + select({
3606 ":debug_build": [],
3607 "//conditions:default": xnnpack_min_size_copts(),
3608 }) + select({
3609 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3610 "//conditions:default": [],
3611 }),
3612 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003613 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003614 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003615 "XNN_NO_U8_OPERATORS",
3616 "XNN_NO_X8_OPERATORS",
3617 "XNN_NO_F16_OPERATORS",
3618 "XNN_NO_X16_OPERATORS",
3619 ],
3620 gcc_copts = xnnpack_gcc_std_copts(),
3621 includes = ["include"],
3622 msvc_copts = xnnpack_msvc_std_copts(),
3623 visibility = xnnpack_visibility(),
3624 deps = [
3625 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003626 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003627 ":logging_utils",
3628 ":memory_planner",
3629 ":operator_run",
3630 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003631 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003632 "@clog",
3633 "@FP16",
3634 "@pthreadpool",
3635 ] + select({
3636 ":emscripten": [],
3637 "//conditions:default": ["@cpuinfo"],
3638 }),
3639)
3640
3641xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003642 name = "xnnpack_operators_nhwc_f32",
3643 srcs = [
3644 "src/init.c",
3645 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003646 hdrs = ["include/xnnpack.h"],
3647 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003648 "-Isrc",
3649 "-Iinclude",
3650 ] + select({
3651 ":debug_build": [],
3652 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003653 }) + select({
3654 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3655 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003656 }),
3657 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003658 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003659 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003660 "XNN_NO_U8_OPERATORS",
3661 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003662 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003663 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003664 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003665 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003666 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003667 visibility = xnnpack_visibility(),
3668 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003669 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003670 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003671 ":operator_run",
3672 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003673 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003674 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003675 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003676 ] + select({
3677 ":emscripten": [],
3678 "//conditions:default": ["@cpuinfo"],
3679 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003680)
3681
Marat Dukhancf056b22019-10-07 10:26:29 -07003682xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003683 name = "bench_utils",
3684 srcs = ["bench/utils.cc"],
3685 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003686 deps = [
3687 "@com_google_benchmark//:benchmark",
3688 "@cpuinfo",
3689 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003690)
3691
Frank Barchard7e955972019-10-11 10:34:25 -07003692######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003693
3694xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003695 name = "qs8_gemm_bench",
3696 srcs = [
3697 "bench/gemm.h",
3698 "bench/qs8-gemm.cc",
3699 "src/xnnpack/AlignedAllocator.h",
3700 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3701 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3702)
3703
3704xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003705 name = "qs8_requantization_bench",
3706 srcs = [
3707 "bench/qs8-requantization.cc",
3708 "src/xnnpack/requantization-stubs.h",
3709 "src/xnnpack/AlignedAllocator.h",
3710 ] + MICROKERNEL_BENCHMARK_HDRS,
3711 deps = MICROKERNEL_BENCHMARK_DEPS,
3712)
3713
3714xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003715 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003716 srcs = [
3717 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003718 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719 "src/xnnpack/AlignedAllocator.h",
3720 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003721 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003722 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003723)
3724
3725xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003726 name = "qu8_requantization_bench",
3727 srcs = [
3728 "bench/qu8-requantization.cc",
3729 "src/xnnpack/requantization-stubs.h",
3730 "src/xnnpack/AlignedAllocator.h",
3731 ] + MICROKERNEL_BENCHMARK_HDRS,
3732 deps = MICROKERNEL_BENCHMARK_DEPS,
3733)
3734
3735xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003736 name = "f16_igemm_bench",
3737 srcs = [
3738 "bench/f16-igemm.cc",
3739 "bench/conv.h",
3740 "bench/google/conv.h",
3741 "src/xnnpack/AlignedAllocator.h",
3742 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003743 deps = MICROKERNEL_BENCHMARK_DEPS + [
3744 ":indirection",
3745 ":packing",
3746 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003747)
3748
3749xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 name = "f16_gemm_bench",
3751 srcs = [
3752 "bench/f16-gemm.cc",
3753 "bench/gemm.h",
3754 "src/xnnpack/AlignedAllocator.h",
3755 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003756 deps = MICROKERNEL_BENCHMARK_DEPS + [
3757 ":packing",
3758 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759)
3760
3761xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003762 name = "f16_spmm_bench",
3763 srcs = [
3764 "bench/f16-spmm.cc",
3765 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003766 "src/xnnpack/AlignedAllocator.h",
3767 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003768 deps = MICROKERNEL_BENCHMARK_DEPS,
3769)
3770
3771xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772 name = "f32_igemm_bench",
3773 srcs = [
3774 "bench/f32-igemm.cc",
3775 "bench/conv.h",
3776 "src/xnnpack/AlignedAllocator.h",
3777 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003778 deps = MICROKERNEL_BENCHMARK_DEPS + [
3779 ":indirection",
3780 ":packing",
3781 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003782)
3783
3784xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003785 name = "f16_relu_bench",
3786 srcs = [
3787 "bench/f16-relu.cc",
3788 "src/xnnpack/AlignedAllocator.h",
3789 ] + MICROKERNEL_BENCHMARK_HDRS,
3790 deps = MICROKERNEL_BENCHMARK_DEPS,
3791)
3792
3793xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794 name = "f32_conv_hwc_bench",
3795 srcs = [
3796 "bench/f32-conv-hwc.cc",
3797 "bench/dconv.h",
3798 "src/xnnpack/AlignedAllocator.h",
3799 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003800 deps = MICROKERNEL_BENCHMARK_DEPS + [
3801 ":packing",
3802 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003803)
3804
3805xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003806 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003807 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003808 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003809 "bench/dconv.h",
3810 "src/xnnpack/AlignedAllocator.h",
3811 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003812 deps = MICROKERNEL_BENCHMARK_DEPS + [
3813 ":packing",
3814 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003815)
3816
3817xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003818 name = "f16_dwconv_bench",
3819 srcs = [
3820 "bench/f16-dwconv.cc",
3821 "bench/dwconv.h",
3822 "bench/google/dwconv.h",
3823 "src/xnnpack/AlignedAllocator.h",
3824 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003825 deps = MICROKERNEL_BENCHMARK_DEPS + [
3826 ":indirection",
3827 ":packing",
3828 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003829)
3830
3831xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003832 name = "f32_dwconv_bench",
3833 srcs = [
3834 "bench/f32-dwconv.cc",
3835 "bench/dwconv.h",
3836 "src/xnnpack/AlignedAllocator.h",
3837 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003838 deps = MICROKERNEL_BENCHMARK_DEPS + [
3839 ":indirection",
3840 ":packing",
3841 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003842)
3843
3844xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003845 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003846 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003847 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003848 "bench/dwconv.h",
3849 "src/xnnpack/AlignedAllocator.h",
3850 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003851 deps = MICROKERNEL_BENCHMARK_DEPS + [
3852 ":indirection",
3853 ":packing",
3854 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003855)
3856
3857xnnpack_benchmark(
3858 name = "f32_gemm_bench",
3859 srcs = [
3860 "bench/f32-gemm.cc",
3861 "bench/gemm.h",
3862 "src/xnnpack/AlignedAllocator.h",
3863 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003864 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003865 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866)
3867
3868xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003869 name = "f32_hswish_bench",
3870 srcs = [
3871 "bench/f32-hswish.cc",
3872 "src/xnnpack/AlignedAllocator.h",
3873 ] + MICROKERNEL_BENCHMARK_HDRS,
3874 deps = MICROKERNEL_BENCHMARK_DEPS,
3875)
3876
3877xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003878 name = "f32_raddexpminusmax_bench",
3879 srcs = [
3880 "bench/f32-raddexpminusmax.cc",
3881 "src/xnnpack/AlignedAllocator.h",
3882 ] + MICROKERNEL_BENCHMARK_HDRS,
3883 deps = MICROKERNEL_BENCHMARK_DEPS,
3884)
3885
3886xnnpack_benchmark(
3887 name = "f32_raddextexp_bench",
3888 srcs = [
3889 "bench/f32-raddextexp.cc",
3890 "src/xnnpack/AlignedAllocator.h",
3891 ] + MICROKERNEL_BENCHMARK_HDRS,
3892 deps = MICROKERNEL_BENCHMARK_DEPS,
3893)
3894
3895xnnpack_benchmark(
3896 name = "f32_raddstoreexpminusmax_bench",
3897 srcs = [
3898 "bench/f32-raddstoreexpminusmax.cc",
3899 "src/xnnpack/AlignedAllocator.h",
3900 ] + MICROKERNEL_BENCHMARK_HDRS,
3901 deps = MICROKERNEL_BENCHMARK_DEPS,
3902)
3903
3904xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003905 name = "f32_relu_bench",
3906 srcs = [
3907 "bench/f32-relu.cc",
3908 "src/xnnpack/AlignedAllocator.h",
3909 ] + MICROKERNEL_BENCHMARK_HDRS,
3910 deps = MICROKERNEL_BENCHMARK_DEPS,
3911)
3912
3913xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 name = "f32_rmax_bench",
3915 srcs = [
3916 "bench/f32-rmax.cc",
3917 "src/xnnpack/AlignedAllocator.h",
3918 ] + MICROKERNEL_BENCHMARK_HDRS,
3919 deps = MICROKERNEL_BENCHMARK_DEPS,
3920)
3921
3922xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003923 name = "f32_sigmoid_bench",
3924 srcs = [
3925 "bench/f32-sigmoid.cc",
3926 "src/xnnpack/AlignedAllocator.h",
3927 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003928 deps = MICROKERNEL_BENCHMARK_DEPS,
3929)
3930
3931xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003932 name = "f32_spmm_bench",
3933 srcs = [
3934 "bench/f32-spmm.cc",
3935 "bench/gemm.h",
3936 "src/xnnpack/AlignedAllocator.h",
3937 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003938 deps = MICROKERNEL_BENCHMARK_DEPS,
3939)
3940
3941xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003942 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003943 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003944 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003945 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003946 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003947 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003948)
3949
3950xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003951 name = "f32_vscaleexpminusmax_bench",
3952 srcs = [
3953 "bench/f32-vscaleexpminusmax.cc",
3954 "src/xnnpack/AlignedAllocator.h",
3955 ] + MICROKERNEL_BENCHMARK_HDRS,
3956 deps = MICROKERNEL_BENCHMARK_DEPS,
3957)
3958
3959xnnpack_benchmark(
3960 name = "f32_vscaleextexp_bench",
3961 srcs = [
3962 "bench/f32-vscaleextexp.cc",
3963 "src/xnnpack/AlignedAllocator.h",
3964 ] + MICROKERNEL_BENCHMARK_HDRS,
3965 deps = MICROKERNEL_BENCHMARK_DEPS,
3966)
3967
3968xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003969 name = "f32_vsqrt_bench",
3970 srcs = [
3971 "bench/f32-vsqrt.cc",
3972 "src/xnnpack/AlignedAllocator.h",
3973 ] + MICROKERNEL_BENCHMARK_HDRS,
3974 deps = MICROKERNEL_BENCHMARK_DEPS,
3975)
3976
3977xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003978 name = "f32_im2col_gemm_bench",
3979 srcs = [
3980 "bench/f32-im2col-gemm.cc",
3981 "bench/conv.h",
3982 "src/xnnpack/AlignedAllocator.h",
3983 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003984 deps = MICROKERNEL_BENCHMARK_DEPS + [
3985 ":im2col",
3986 ":packing",
3987 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003988)
3989
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003990xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003991 name = "rounding_bench",
3992 srcs = [
3993 "bench/rounding.cc",
3994 "src/xnnpack/math-stubs.h",
3995 "src/xnnpack/AlignedAllocator.h",
3996 ] + MICROKERNEL_BENCHMARK_HDRS,
3997 deps = MICROKERNEL_BENCHMARK_DEPS,
3998)
3999
Marat Dukhan08c4a432019-10-03 09:29:21 -07004000########################### Benchmarks for operators ###########################
4001
4002xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004003 name = "average_pooling_bench",
4004 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004005 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004006 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004007 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004008)
4009
4010xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004011 name = "bankers_rounding_bench",
4012 srcs = ["bench/bankers-rounding.cc"],
4013 copts = xnnpack_optional_tflite_copts(),
4014 tags = ["nowin32"],
4015 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4016)
4017
4018xnnpack_benchmark(
4019 name = "ceiling_bench",
4020 srcs = ["bench/ceiling.cc"],
4021 copts = xnnpack_optional_tflite_copts(),
4022 tags = ["nowin32"],
4023 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4024)
4025
4026xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004027 name = "channel_shuffle_bench",
4028 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004029 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004030)
4031
4032xnnpack_benchmark(
4033 name = "convolution_bench",
4034 srcs = ["bench/convolution.cc"],
4035 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004036 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004037 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004038)
4039
4040xnnpack_benchmark(
4041 name = "deconvolution_bench",
4042 srcs = ["bench/deconvolution.cc"],
4043 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004044 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004045 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004046)
4047
4048xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004049 name = "floor_bench",
4050 srcs = ["bench/floor.cc"],
4051 copts = xnnpack_optional_tflite_copts(),
4052 tags = ["nowin32"],
4053 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4054)
4055
4056xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004057 name = "global_average_pooling_bench",
4058 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004059 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004060)
4061
4062xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004063 name = "hardswish_bench",
4064 srcs = ["bench/hardswish.cc"],
4065 copts = xnnpack_optional_tflite_copts(),
4066 tags = ["nowin32"],
4067 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4068)
4069
4070xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004071 name = "max_pooling_bench",
4072 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004073 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004074)
4075
4076xnnpack_benchmark(
4077 name = "sigmoid_bench",
4078 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004079 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004080 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004081 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004082)
4083
4084xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004085 name = "prelu_bench",
4086 srcs = ["bench/prelu.cc"],
4087 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004088 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004089 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004090)
4091
4092xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004093 name = "softmax_bench",
4094 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004095 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004096 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004097 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004098)
4099
Marat Dukhan87727142020-06-24 15:24:10 -07004100xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004101 name = "square_root_bench",
4102 srcs = ["bench/square-root.cc"],
4103 copts = xnnpack_optional_tflite_copts(),
4104 tags = ["nowin32"],
4105 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4106)
4107
4108xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004109 name = "truncation_bench",
4110 srcs = ["bench/truncation.cc"],
4111 deps = OPERATOR_BENCHMARK_DEPS,
4112)
4113
Marat Dukhanc068bb62019-10-04 13:24:39 -07004114############################# End-to-end benchmarks ############################
4115
4116cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004117 name = "fp32_mobilenet_v1",
4118 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004119 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004120 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004121 linkstatic = True,
4122 deps = [
4123 ":XNNPACK",
4124 "@pthreadpool",
4125 ],
4126)
4127
4128cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004129 name = "fp16_mobilenet_v1",
4130 srcs = ["models/fp16-mobilenet-v1.cc"],
4131 hdrs = ["models/models.h"],
4132 copts = xnnpack_std_cxxopts(),
4133 linkstatic = True,
4134 deps = [
4135 ":XNNPACK",
4136 "@FP16",
4137 "@pthreadpool",
4138 ],
4139)
4140
4141cc_library(
4142 name = "fp32_mobilenet_v2",
4143 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004144 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004145 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004146 linkstatic = True,
4147 deps = [
4148 ":XNNPACK",
4149 "@pthreadpool",
4150 ],
4151)
4152
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004153cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004154 name = "fp16_mobilenet_v2",
4155 srcs = ["models/fp16-mobilenet-v2.cc"],
4156 hdrs = ["models/models.h"],
4157 copts = xnnpack_std_cxxopts(),
4158 linkstatic = True,
4159 deps = [
4160 ":XNNPACK",
4161 "@FP16",
4162 "@pthreadpool",
4163 ],
4164)
4165
4166cc_library(
4167 name = "fp32_mobilenet_v3_large",
4168 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004169 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004170 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004171 linkstatic = True,
4172 deps = [
4173 ":XNNPACK",
4174 "@pthreadpool",
4175 ],
4176)
4177
4178cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004179 name = "fp32_mobilenet_v3_small",
4180 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004181 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004182 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004183 linkstatic = True,
4184 deps = [
4185 ":XNNPACK",
4186 "@pthreadpool",
4187 ],
4188)
4189
Marat Dukhanc068bb62019-10-04 13:24:39 -07004190xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004191 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004192 srcs = [
4193 "bench/f32-dwconv-e2e.cc",
4194 "bench/end2end.h",
4195 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004196 deps = MICROKERNEL_BENCHMARK_DEPS + [
4197 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004198 ":fp32_mobilenet_v1",
4199 ":fp32_mobilenet_v2",
4200 ":fp32_mobilenet_v3_large",
4201 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004202 ],
4203)
4204
4205xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004206 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004207 srcs = [
4208 "bench/f32-gemm-e2e.cc",
4209 "bench/end2end.h",
4210 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004211 deps = MICROKERNEL_BENCHMARK_DEPS + [
4212 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004213 ":fp32_mobilenet_v1",
4214 ":fp32_mobilenet_v2",
4215 ":fp32_mobilenet_v3_large",
4216 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004217 ],
4218)
4219
4220xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004221 name = "end2end_bench",
4222 srcs = ["bench/end2end.cc"],
4223 deps = [
4224 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004225 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004226 ":fp16_mobilenet_v1",
4227 ":fp16_mobilenet_v2",
4228 ":fp32_mobilenet_v1",
4229 ":fp32_mobilenet_v2",
4230 ":fp32_mobilenet_v3_large",
4231 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004232 "@pthreadpool",
4233 ],
4234)
4235
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004236#################### Accuracy evaluation for math functions ####################
4237
4238xnnpack_benchmark(
4239 name = "f32_exp_eval",
4240 srcs = [
4241 "eval/f32-exp.cc",
4242 "src/xnnpack/AlignedAllocator.h",
4243 ] + ACCURACY_EVAL_HDRS,
4244 deps = ACCURACY_EVAL_DEPS,
4245)
4246
Marat Dukhan515c9772019-10-17 18:07:57 -07004247xnnpack_benchmark(
4248 name = "f32_expminus_eval",
4249 srcs = [
4250 "eval/f32-expminus.cc",
4251 "src/xnnpack/AlignedAllocator.h",
4252 ] + ACCURACY_EVAL_HDRS,
4253 deps = ACCURACY_EVAL_DEPS,
4254)
4255
Marat Dukhan98ba4412019-10-23 02:14:28 -07004256xnnpack_benchmark(
4257 name = "f32_extexp_eval",
4258 srcs = [
4259 "eval/f32-extexp.cc",
4260 "src/xnnpack/AlignedAllocator.h",
4261 ] + ACCURACY_EVAL_HDRS,
4262 deps = ACCURACY_EVAL_DEPS,
4263)
4264
Marat Dukhan8853b822020-05-07 12:19:01 -07004265xnnpack_unit_test(
4266 name = "f32_roundne_eval",
4267 srcs = [
4268 "eval/f32-roundne.cc",
4269 "src/xnnpack/AlignedAllocator.h",
4270 "src/xnnpack/math-stubs.h",
4271 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004272 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004273 deps = MICROKERNEL_TEST_DEPS,
4274)
4275
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004276xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004277 name = "f32_roundd_eval",
4278 srcs = [
4279 "eval/f32-roundd.cc",
4280 "src/xnnpack/AlignedAllocator.h",
4281 "src/xnnpack/math-stubs.h",
4282 ] + MICROKERNEL_TEST_HDRS,
4283 automatic = False,
4284 deps = MICROKERNEL_TEST_DEPS,
4285)
4286
4287xnnpack_unit_test(
4288 name = "f32_roundu_eval",
4289 srcs = [
4290 "eval/f32-roundu.cc",
4291 "src/xnnpack/AlignedAllocator.h",
4292 "src/xnnpack/math-stubs.h",
4293 ] + MICROKERNEL_TEST_HDRS,
4294 automatic = False,
4295 deps = MICROKERNEL_TEST_DEPS,
4296)
4297
4298xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004299 name = "f32_roundz_eval",
4300 srcs = [
4301 "eval/f32-roundz.cc",
4302 "src/xnnpack/AlignedAllocator.h",
4303 "src/xnnpack/math-stubs.h",
4304 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004305 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004306 deps = MICROKERNEL_TEST_DEPS,
4307)
4308
Marat Dukhan346a9e52019-11-15 09:06:30 -08004309xnnpack_benchmark(
4310 name = "f32_sigmoid_eval",
4311 srcs = [
4312 "eval/f32-sigmoid.cc",
4313 "src/xnnpack/AlignedAllocator.h",
4314 ] + ACCURACY_EVAL_HDRS,
4315 deps = ACCURACY_EVAL_DEPS,
4316)
4317
Marat Dukhan84000762020-06-29 18:38:43 -07004318xnnpack_benchmark(
4319 name = "f32_sqrt_eval",
4320 srcs = [
4321 "eval/f32-sqrt.cc",
4322 "src/xnnpack/AlignedAllocator.h",
4323 ] + ACCURACY_EVAL_HDRS,
4324 deps = ACCURACY_EVAL_DEPS + [
4325 ":bench_utils",
4326 ],
4327)
4328
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329######################### Unit tests for micro-kernels #########################
4330
4331xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004332 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004333 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004334 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335 "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 Dukhan08c4a432019-10-03 09:29:21 -07004339)
4340
4341xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004342 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004343 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004344 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004345 "test/spmm-microkernel-tester.h",
4346 "src/xnnpack/AlignedAllocator.h",
4347 ] + MICROKERNEL_TEST_HDRS,
4348 deps = MICROKERNEL_TEST_DEPS,
4349)
4350
4351xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352 name = "f32_argmaxpool_test",
4353 srcs = [
4354 "test/f32-argmaxpool.cc",
4355 "test/argmaxpool-microkernel-tester.h",
4356 "src/xnnpack/AlignedAllocator.h",
4357 ] + MICROKERNEL_TEST_HDRS,
4358 deps = MICROKERNEL_TEST_DEPS,
4359)
4360
4361xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004362 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004363 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004364 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 "test/avgpool-microkernel-tester.h",
4366 "src/xnnpack/AlignedAllocator.h",
4367 ] + MICROKERNEL_TEST_HDRS,
4368 deps = MICROKERNEL_TEST_DEPS,
4369)
4370
4371xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004372 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004373 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004374 "test/f32-ibilinear.cc",
4375 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004376 "src/xnnpack/AlignedAllocator.h",
4377 ] + MICROKERNEL_TEST_HDRS,
4378 deps = MICROKERNEL_TEST_DEPS,
4379)
4380
4381xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004382 name = "f16_clamp_test",
4383 srcs = [
4384 "test/f16-clamp.cc",
4385 "test/clamp-microkernel-tester.h",
4386 ] + MICROKERNEL_TEST_HDRS,
4387 deps = MICROKERNEL_TEST_DEPS,
4388)
4389
4390xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391 name = "f32_clamp_test",
4392 srcs = [
4393 "test/f32-clamp.cc",
4394 "test/clamp-microkernel-tester.h",
4395 ] + MICROKERNEL_TEST_HDRS,
4396 deps = MICROKERNEL_TEST_DEPS,
4397)
4398
4399xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004400 name = "f32_igemm_test",
4401 srcs = [
4402 "test/f32-igemm.cc",
4403 "test/gemm-microkernel-tester.h",
4404 "src/xnnpack/AlignedAllocator.h",
4405 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004406 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004407)
4408
4409xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004410 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004411 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004412 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413 "test/gemm-microkernel-tester.h",
4414 "src/xnnpack/AlignedAllocator.h",
4415 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004416 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004417)
4418
4419xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004420 name = "f32_igemm_minmax_test",
4421 srcs = [
4422 "test/f32-igemm-minmax.cc",
4423 "test/gemm-microkernel-tester.h",
4424 "src/xnnpack/AlignedAllocator.h",
4425 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004426 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004427)
4428
4429xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004430 name = "f16_igemm_minmax_test",
4431 srcs = [
4432 "test/f16-igemm-minmax.cc",
4433 "test/gemm-microkernel-tester.h",
4434 "src/xnnpack/AlignedAllocator.h",
4435 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004436 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004437)
4438
4439xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004440 name = "f32_conv_hwc_test",
4441 srcs = [
4442 "test/f32-conv-hwc.cc",
4443 "test/conv-hwc-microkernel-tester.h",
4444 "src/xnnpack/AlignedAllocator.h",
4445 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004446 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004447)
4448
4449xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004450 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004452 "test/f32-conv-hwc2chw.cc",
4453 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004454 "src/xnnpack/AlignedAllocator.h",
4455 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004456 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004457)
4458
4459xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004460 name = "f32_dwconv_test",
4461 srcs = [
4462 "test/f32-dwconv.cc",
4463 "test/dwconv-microkernel-tester.h",
4464 "src/xnnpack/AlignedAllocator.h",
4465 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004466 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004467)
4468
4469xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004470 name = "f16_dwconv_minmax_test",
4471 srcs = [
4472 "test/f16-dwconv-minmax.cc",
4473 "test/dwconv-microkernel-tester.h",
4474 "src/xnnpack/AlignedAllocator.h",
4475 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004476 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004477)
4478
4479xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004480 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004481 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004482 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004483 "test/dwconv-microkernel-tester.h",
4484 "src/xnnpack/AlignedAllocator.h",
4485 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004486 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004487)
4488
4489xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004490 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004491 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004492 "test/f32-dwconv-chw.cc",
4493 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004494 "src/xnnpack/AlignedAllocator.h",
4495 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004496 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004497)
4498
4499xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004500 name = "f16_gavgpool_minmax_test",
4501 srcs = [
4502 "test/f16-gavgpool-minmax.cc",
4503 "test/gavgpool-microkernel-tester.h",
4504 "src/xnnpack/AlignedAllocator.h",
4505 ] + MICROKERNEL_TEST_HDRS,
4506 deps = MICROKERNEL_TEST_DEPS,
4507)
4508
4509xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004510 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004511 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004512 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004513 "test/gavgpool-microkernel-tester.h",
4514 "src/xnnpack/AlignedAllocator.h",
4515 ] + MICROKERNEL_TEST_HDRS,
4516 deps = MICROKERNEL_TEST_DEPS,
4517)
4518
4519xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004520 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004521 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004522 "test/f32-gavgpool-cw.cc",
4523 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004524 "src/xnnpack/AlignedAllocator.h",
4525 ] + MICROKERNEL_TEST_HDRS,
4526 deps = MICROKERNEL_TEST_DEPS,
4527)
4528
4529xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004530 name = "f32_gemm_test",
4531 srcs = [
4532 "test/f32-gemm.cc",
4533 "test/gemm-microkernel-tester.h",
4534 "src/xnnpack/AlignedAllocator.h",
4535 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004536 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004537)
4538
4539xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004540 name = "f32_gemm_relu_test",
4541 srcs = [
4542 "test/f32-gemm-relu.cc",
4543 "test/gemm-microkernel-tester.h",
4544 "src/xnnpack/AlignedAllocator.h",
4545 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004546 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004547)
4548
4549xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004550 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004552 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004553 "test/gemm-microkernel-tester.h",
4554 "src/xnnpack/AlignedAllocator.h",
4555 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004556 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004557)
4558
4559xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004560 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004561 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004562 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004563 "test/gemm-microkernel-tester.h",
4564 "src/xnnpack/AlignedAllocator.h",
4565 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004566 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004567)
4568
4569xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004570 name = "f16_hswish_test",
4571 srcs = [
4572 "test/f16-hswish.cc",
4573 "test/hswish-microkernel-tester.h",
4574 ] + MICROKERNEL_TEST_HDRS,
4575 deps = MICROKERNEL_TEST_DEPS,
4576)
4577
4578xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004579 name = "f32_hswish_test",
4580 srcs = [
4581 "test/f32-hswish.cc",
4582 "test/hswish-microkernel-tester.h",
4583 ] + MICROKERNEL_TEST_HDRS,
4584 deps = MICROKERNEL_TEST_DEPS,
4585)
4586
4587xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004588 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004589 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004590 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004591 "test/maxpool-microkernel-tester.h",
4592 ] + MICROKERNEL_TEST_HDRS,
4593 deps = MICROKERNEL_TEST_DEPS,
4594)
4595
4596xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004597 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004598 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004599 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004600 "test/avgpool-microkernel-tester.h",
4601 "src/xnnpack/AlignedAllocator.h",
4602 ] + MICROKERNEL_TEST_HDRS,
4603 deps = MICROKERNEL_TEST_DEPS,
4604)
4605
4606xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004607 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004608 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004609 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004610 "test/gemm-microkernel-tester.h",
4611 "src/xnnpack/AlignedAllocator.h",
4612 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004613 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004614)
4615
4616xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004617 name = "f16_prelu_test",
4618 srcs = [
4619 "test/f16-prelu.cc",
4620 "test/prelu-microkernel-tester.h",
4621 "src/xnnpack/AlignedAllocator.h",
4622 ] + MICROKERNEL_TEST_HDRS,
4623 deps = MICROKERNEL_TEST_DEPS,
4624)
4625
4626xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004627 name = "f32_prelu_test",
4628 srcs = [
4629 "test/f32-prelu.cc",
4630 "test/prelu-microkernel-tester.h",
4631 "src/xnnpack/AlignedAllocator.h",
4632 ] + MICROKERNEL_TEST_HDRS,
4633 deps = MICROKERNEL_TEST_DEPS,
4634)
4635
4636xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004637 name = "f32_raddexpminusmax_test",
4638 srcs = [
4639 "test/f32-raddexpminusmax.cc",
4640 "test/raddexpminusmax-microkernel-tester.h",
4641 ] + MICROKERNEL_TEST_HDRS,
4642 deps = MICROKERNEL_TEST_DEPS,
4643)
4644
4645xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004646 name = "f32_raddextexp_test",
4647 srcs = [
4648 "test/f32-raddextexp.cc",
4649 "test/raddextexp-microkernel-tester.h",
4650 ] + MICROKERNEL_TEST_HDRS,
4651 deps = MICROKERNEL_TEST_DEPS,
4652)
4653
4654xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004655 name = "f32_raddstoreexpminusmax_test",
4656 srcs = [
4657 "test/f32-raddstoreexpminusmax.cc",
4658 "test/raddstoreexpminusmax-microkernel-tester.h",
4659 ] + MICROKERNEL_TEST_HDRS,
4660 deps = MICROKERNEL_TEST_DEPS,
4661)
4662
4663xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004664 name = "f32_relu_test",
4665 srcs = [
4666 "test/f32-relu.cc",
4667 "test/vunary-microkernel-tester.h",
4668 ] + MICROKERNEL_TEST_HDRS,
4669 deps = MICROKERNEL_TEST_DEPS,
4670)
4671
4672xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004673 name = "f32_rmax_test",
4674 srcs = [
4675 "test/f32-rmax.cc",
4676 "test/rmax-microkernel-tester.h",
4677 ] + MICROKERNEL_TEST_HDRS,
4678 deps = MICROKERNEL_TEST_DEPS,
4679)
4680
4681xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004682 name = "f32_sigmoid_test",
4683 srcs = [
4684 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004685 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004686 ] + MICROKERNEL_TEST_HDRS,
4687 deps = MICROKERNEL_TEST_DEPS,
4688)
4689
4690xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004691 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004692 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004693 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004694 "test/spmm-microkernel-tester.h",
4695 "src/xnnpack/AlignedAllocator.h",
4696 ] + MICROKERNEL_TEST_HDRS,
4697 deps = MICROKERNEL_TEST_DEPS,
4698)
4699
4700xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004701 name = "f16_vadd_minmax_test",
4702 srcs = [
4703 "test/f16-vadd-minmax.cc",
4704 "test/vbinary-microkernel-tester.h",
4705 ] + MICROKERNEL_TEST_HDRS,
4706 deps = MICROKERNEL_TEST_DEPS,
4707)
4708
4709xnnpack_unit_test(
4710 name = "f16_vaddc_minmax_test",
4711 srcs = [
4712 "test/f16-vaddc-minmax.cc",
4713 "test/vbinaryc-microkernel-tester.h",
4714 ] + MICROKERNEL_TEST_HDRS,
4715 deps = MICROKERNEL_TEST_DEPS,
4716)
4717
4718xnnpack_unit_test(
4719 name = "f16_vdiv_minmax_test",
4720 srcs = [
4721 "test/f16-vdiv-minmax.cc",
4722 "test/vbinary-microkernel-tester.h",
4723 ] + MICROKERNEL_TEST_HDRS,
4724 deps = MICROKERNEL_TEST_DEPS,
4725)
4726
4727xnnpack_unit_test(
4728 name = "f16_vdivc_minmax_test",
4729 srcs = [
4730 "test/f16-vdivc-minmax.cc",
4731 "test/vbinaryc-microkernel-tester.h",
4732 ] + MICROKERNEL_TEST_HDRS,
4733 deps = MICROKERNEL_TEST_DEPS,
4734)
4735
4736xnnpack_unit_test(
4737 name = "f16_vrdivc_minmax_test",
4738 srcs = [
4739 "test/f16-vrdivc-minmax.cc",
4740 "test/vbinaryc-microkernel-tester.h",
4741 ] + MICROKERNEL_TEST_HDRS,
4742 deps = MICROKERNEL_TEST_DEPS,
4743)
4744
4745xnnpack_unit_test(
4746 name = "f16_vmax_test",
4747 srcs = [
4748 "test/f16-vmax.cc",
4749 "test/vbinary-microkernel-tester.h",
4750 ] + MICROKERNEL_TEST_HDRS,
4751 deps = MICROKERNEL_TEST_DEPS,
4752)
4753
4754xnnpack_unit_test(
4755 name = "f16_vmaxc_test",
4756 srcs = [
4757 "test/f16-vmaxc.cc",
4758 "test/vbinaryc-microkernel-tester.h",
4759 ] + MICROKERNEL_TEST_HDRS,
4760 deps = MICROKERNEL_TEST_DEPS,
4761)
4762
4763xnnpack_unit_test(
4764 name = "f16_vmin_test",
4765 srcs = [
4766 "test/f16-vmin.cc",
4767 "test/vbinary-microkernel-tester.h",
4768 ] + MICROKERNEL_TEST_HDRS,
4769 deps = MICROKERNEL_TEST_DEPS,
4770)
4771
4772xnnpack_unit_test(
4773 name = "f16_vminc_test",
4774 srcs = [
4775 "test/f16-vminc.cc",
4776 "test/vbinaryc-microkernel-tester.h",
4777 ] + MICROKERNEL_TEST_HDRS,
4778 deps = MICROKERNEL_TEST_DEPS,
4779)
4780
4781xnnpack_unit_test(
4782 name = "f16_vmul_minmax_test",
4783 srcs = [
4784 "test/f16-vmul-minmax.cc",
4785 "test/vbinary-microkernel-tester.h",
4786 ] + MICROKERNEL_TEST_HDRS,
4787 deps = MICROKERNEL_TEST_DEPS,
4788)
4789
4790xnnpack_unit_test(
4791 name = "f16_vmulc_minmax_test",
4792 srcs = [
4793 "test/f16-vmulc-minmax.cc",
4794 "test/vbinaryc-microkernel-tester.h",
4795 ] + MICROKERNEL_TEST_HDRS,
4796 deps = MICROKERNEL_TEST_DEPS,
4797)
4798
4799xnnpack_unit_test(
4800 name = "f16_vsub_minmax_test",
4801 srcs = [
4802 "test/f16-vsub-minmax.cc",
4803 "test/vbinary-microkernel-tester.h",
4804 ] + MICROKERNEL_TEST_HDRS,
4805 deps = MICROKERNEL_TEST_DEPS,
4806)
4807
4808xnnpack_unit_test(
4809 name = "f16_vsubc_minmax_test",
4810 srcs = [
4811 "test/f16-vsubc-minmax.cc",
4812 "test/vbinaryc-microkernel-tester.h",
4813 ] + MICROKERNEL_TEST_HDRS,
4814 deps = MICROKERNEL_TEST_DEPS,
4815)
4816
4817xnnpack_unit_test(
4818 name = "f16_vrsubc_minmax_test",
4819 srcs = [
4820 "test/f16-vrsubc-minmax.cc",
4821 "test/vbinaryc-microkernel-tester.h",
4822 ] + MICROKERNEL_TEST_HDRS,
4823 deps = MICROKERNEL_TEST_DEPS,
4824)
4825
4826xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004827 name = "f32_vabs_test",
4828 srcs = [
4829 "test/f32-vabs.cc",
4830 "test/vunary-microkernel-tester.h",
4831 ] + MICROKERNEL_TEST_HDRS,
4832 deps = MICROKERNEL_TEST_DEPS,
4833)
4834
4835xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004836 name = "f32_vadd_test",
4837 srcs = [
4838 "test/f32-vadd.cc",
4839 "test/vbinary-microkernel-tester.h",
4840 ] + MICROKERNEL_TEST_HDRS,
4841 deps = MICROKERNEL_TEST_DEPS,
4842)
4843
4844xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004845 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004846 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004847 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004848 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004849 ] + MICROKERNEL_TEST_HDRS,
4850 deps = MICROKERNEL_TEST_DEPS,
4851)
4852
4853xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004854 name = "f32_vaddc_test",
4855 srcs = [
4856 "test/f32-vaddc.cc",
4857 "test/vbinaryc-microkernel-tester.h",
4858 ] + MICROKERNEL_TEST_HDRS,
4859 deps = MICROKERNEL_TEST_DEPS,
4860)
4861
4862xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004863 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004864 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004865 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004866 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004867 ] + MICROKERNEL_TEST_HDRS,
4868 deps = MICROKERNEL_TEST_DEPS,
4869)
4870
4871xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004872 name = "f32_vdiv_test",
4873 srcs = [
4874 "test/f32-vdiv.cc",
4875 "test/vbinary-microkernel-tester.h",
4876 ] + MICROKERNEL_TEST_HDRS,
4877 deps = MICROKERNEL_TEST_DEPS,
4878)
4879
4880xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004881 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004882 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004883 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004884 "test/vbinary-microkernel-tester.h",
4885 ] + MICROKERNEL_TEST_HDRS,
4886 deps = MICROKERNEL_TEST_DEPS,
4887)
4888
4889xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004890 name = "f32_vdivc_test",
4891 srcs = [
4892 "test/f32-vdivc.cc",
4893 "test/vbinaryc-microkernel-tester.h",
4894 ] + MICROKERNEL_TEST_HDRS,
4895 deps = MICROKERNEL_TEST_DEPS,
4896)
4897
4898xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004899 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004900 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004901 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004902 "test/vbinaryc-microkernel-tester.h",
4903 ] + MICROKERNEL_TEST_HDRS,
4904 deps = MICROKERNEL_TEST_DEPS,
4905)
4906
4907xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004908 name = "f32_vrdivc_test",
4909 srcs = [
4910 "test/f32-vrdivc.cc",
4911 "test/vbinaryc-microkernel-tester.h",
4912 ] + MICROKERNEL_TEST_HDRS,
4913 deps = MICROKERNEL_TEST_DEPS,
4914)
4915
4916xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004917 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004918 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004919 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004920 "test/vbinaryc-microkernel-tester.h",
4921 ] + MICROKERNEL_TEST_HDRS,
4922 deps = MICROKERNEL_TEST_DEPS,
4923)
4924
4925xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004926 name = "f32_vmax_test",
4927 srcs = [
4928 "test/f32-vmax.cc",
4929 "test/vbinary-microkernel-tester.h",
4930 ] + MICROKERNEL_TEST_HDRS,
4931 deps = MICROKERNEL_TEST_DEPS,
4932)
4933
4934xnnpack_unit_test(
4935 name = "f32_vmaxc_test",
4936 srcs = [
4937 "test/f32-vmaxc.cc",
4938 "test/vbinaryc-microkernel-tester.h",
4939 ] + MICROKERNEL_TEST_HDRS,
4940 deps = MICROKERNEL_TEST_DEPS,
4941)
4942
4943xnnpack_unit_test(
4944 name = "f32_vmin_test",
4945 srcs = [
4946 "test/f32-vmin.cc",
4947 "test/vbinary-microkernel-tester.h",
4948 ] + MICROKERNEL_TEST_HDRS,
4949 deps = MICROKERNEL_TEST_DEPS,
4950)
4951
4952xnnpack_unit_test(
4953 name = "f32_vminc_test",
4954 srcs = [
4955 "test/f32-vminc.cc",
4956 "test/vbinaryc-microkernel-tester.h",
4957 ] + MICROKERNEL_TEST_HDRS,
4958 deps = MICROKERNEL_TEST_DEPS,
4959)
4960
4961xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004962 name = "f32_vmul_test",
4963 srcs = [
4964 "test/f32-vmul.cc",
4965 "test/vbinary-microkernel-tester.h",
4966 ] + MICROKERNEL_TEST_HDRS,
4967 deps = MICROKERNEL_TEST_DEPS,
4968)
4969
4970xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004971 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004972 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004973 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004974 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004975 ] + MICROKERNEL_TEST_HDRS,
4976 deps = MICROKERNEL_TEST_DEPS,
4977)
4978
4979xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004980 name = "f32_vmulc_test",
4981 srcs = [
4982 "test/f32-vmulc.cc",
4983 "test/vbinaryc-microkernel-tester.h",
4984 ] + MICROKERNEL_TEST_HDRS,
4985 deps = MICROKERNEL_TEST_DEPS,
4986)
4987
4988xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004989 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004990 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004991 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004992 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993 ] + MICROKERNEL_TEST_HDRS,
4994 deps = MICROKERNEL_TEST_DEPS,
4995)
4996
4997xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004998 name = "f16_vmulcaddc_minmax_test",
4999 srcs = [
5000 "test/f16-vmulcaddc-minmax.cc",
5001 "test/vmulcaddc-microkernel-tester.h",
5002 "src/xnnpack/AlignedAllocator.h",
5003 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005004 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005005)
5006
5007xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005008 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005010 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005011 "test/vmulcaddc-microkernel-tester.h",
5012 "src/xnnpack/AlignedAllocator.h",
5013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005014 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015)
5016
5017xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005018 name = "f32_vlrelu_test",
5019 srcs = [
5020 "test/f32-vlrelu.cc",
5021 "test/vunary-microkernel-tester.h",
5022 ] + MICROKERNEL_TEST_HDRS,
5023 deps = MICROKERNEL_TEST_DEPS,
5024)
5025
5026xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005027 name = "f32_vneg_test",
5028 srcs = [
5029 "test/f32-vneg.cc",
5030 "test/vunary-microkernel-tester.h",
5031 ] + MICROKERNEL_TEST_HDRS,
5032 deps = MICROKERNEL_TEST_DEPS,
5033)
5034
5035xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005036 name = "f32_vrndne_test",
5037 srcs = [
5038 "test/f32-vrndne.cc",
5039 "test/vunary-microkernel-tester.h",
5040 ] + MICROKERNEL_TEST_HDRS,
5041 deps = MICROKERNEL_TEST_DEPS,
5042)
5043
5044xnnpack_unit_test(
5045 name = "f32_vrndz_test",
5046 srcs = [
5047 "test/f32-vrndz.cc",
5048 "test/vunary-microkernel-tester.h",
5049 ] + MICROKERNEL_TEST_HDRS,
5050 deps = MICROKERNEL_TEST_DEPS,
5051)
5052
5053xnnpack_unit_test(
5054 name = "f32_vrndu_test",
5055 srcs = [
5056 "test/f32-vrndu.cc",
5057 "test/vunary-microkernel-tester.h",
5058 ] + MICROKERNEL_TEST_HDRS,
5059 deps = MICROKERNEL_TEST_DEPS,
5060)
5061
5062xnnpack_unit_test(
5063 name = "f32_vrndd_test",
5064 srcs = [
5065 "test/f32-vrndd.cc",
5066 "test/vunary-microkernel-tester.h",
5067 ] + MICROKERNEL_TEST_HDRS,
5068 deps = MICROKERNEL_TEST_DEPS,
5069)
5070
5071xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005072 name = "f32_vscale_test",
5073 srcs = [
5074 "test/f32-vscale.cc",
5075 "test/vscale-microkernel-tester.h",
5076 ] + MICROKERNEL_TEST_HDRS,
5077 deps = MICROKERNEL_TEST_DEPS,
5078)
5079
5080xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005081 name = "f32_vscaleexpminusmax_test",
5082 srcs = [
5083 "test/f32-vscaleexpminusmax.cc",
5084 "test/vscaleexpminusmax-microkernel-tester.h",
5085 ] + MICROKERNEL_TEST_HDRS,
5086 deps = MICROKERNEL_TEST_DEPS,
5087)
5088
5089xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005090 name = "f32_vscaleextexp_test",
5091 srcs = [
5092 "test/f32-vscaleextexp.cc",
5093 "test/vscaleextexp-microkernel-tester.h",
5094 ] + MICROKERNEL_TEST_HDRS,
5095 deps = MICROKERNEL_TEST_DEPS,
5096)
5097
5098xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005099 name = "f32_vsqr_test",
5100 srcs = [
5101 "test/f32-vsqr.cc",
5102 "test/vunary-microkernel-tester.h",
5103 ] + MICROKERNEL_TEST_HDRS,
5104 deps = MICROKERNEL_TEST_DEPS,
5105)
5106
5107xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005108 name = "f32_vsqrdiff_test",
5109 srcs = [
5110 "test/f32-vsqrdiff.cc",
5111 "test/vbinary-microkernel-tester.h",
5112 ] + MICROKERNEL_TEST_HDRS,
5113 deps = MICROKERNEL_TEST_DEPS,
5114)
5115
5116xnnpack_unit_test(
5117 name = "f32_vsqrdiffc_test",
5118 srcs = [
5119 "test/f32-vsqrdiffc.cc",
5120 "test/vbinaryc-microkernel-tester.h",
5121 ] + MICROKERNEL_TEST_HDRS,
5122 deps = MICROKERNEL_TEST_DEPS,
5123)
5124
5125xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005126 name = "f32_vsqrt_test",
5127 srcs = [
5128 "test/f32-vsqrt.cc",
5129 "test/vunary-microkernel-tester.h",
5130 ] + MICROKERNEL_TEST_HDRS,
5131 deps = MICROKERNEL_TEST_DEPS,
5132)
5133
5134xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005135 name = "f32_vsub_test",
5136 srcs = [
5137 "test/f32-vsub.cc",
5138 "test/vbinary-microkernel-tester.h",
5139 ] + MICROKERNEL_TEST_HDRS,
5140 deps = MICROKERNEL_TEST_DEPS,
5141)
5142
5143xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005144 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005145 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005146 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005147 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005148 ] + MICROKERNEL_TEST_HDRS,
5149 deps = MICROKERNEL_TEST_DEPS,
5150)
5151
5152xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005153 name = "f32_vsubc_test",
5154 srcs = [
5155 "test/f32-vsubc.cc",
5156 "test/vbinaryc-microkernel-tester.h",
5157 ] + MICROKERNEL_TEST_HDRS,
5158 deps = MICROKERNEL_TEST_DEPS,
5159)
5160
5161xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005162 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005163 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005164 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005165 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005166 ] + MICROKERNEL_TEST_HDRS,
5167 deps = MICROKERNEL_TEST_DEPS,
5168)
5169
5170xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005171 name = "f32_vrsubc_test",
5172 srcs = [
5173 "test/f32-vrsubc.cc",
5174 "test/vbinaryc-microkernel-tester.h",
5175 ] + MICROKERNEL_TEST_HDRS,
5176 deps = MICROKERNEL_TEST_DEPS,
5177)
5178
5179xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005180 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005181 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005182 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005183 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005184 ] + MICROKERNEL_TEST_HDRS,
5185 deps = MICROKERNEL_TEST_DEPS,
5186)
5187
5188xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005189 name = "qs8_dwconv_minmax_test",
5190 srcs = [
5191 "test/qs8-dwconv-minmax.cc",
5192 "test/dwconv-microkernel-tester.h",
5193 "src/xnnpack/AlignedAllocator.h",
5194 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5195 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5196)
5197
5198xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005199 name = "qs8_gemm_minmax_test",
5200 srcs = [
5201 "test/qs8-gemm-minmax.cc",
5202 "test/gemm-microkernel-tester.h",
5203 "src/xnnpack/AlignedAllocator.h",
5204 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5205 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5206)
5207
5208xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005209 name = "qs8_igemm_minmax_test",
5210 srcs = [
5211 "test/qs8-igemm-minmax.cc",
5212 "test/gemm-microkernel-tester.h",
5213 "src/xnnpack/AlignedAllocator.h",
5214 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5215 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5216)
5217
5218xnnpack_unit_test(
5219 name = "qs8_requantization_test",
5220 srcs = [
5221 "src/xnnpack/requantization-stubs.h",
5222 "test/qs8-requantization.cc",
5223 "test/requantization-tester.h",
5224 ] + MICROKERNEL_TEST_HDRS,
5225 deps = MICROKERNEL_TEST_DEPS,
5226)
5227
5228xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005229 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005231 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005232 "test/avgpool-microkernel-tester.h",
5233 "src/xnnpack/AlignedAllocator.h",
5234 ] + MICROKERNEL_TEST_HDRS,
5235 deps = MICROKERNEL_TEST_DEPS,
5236)
5237
5238xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005239 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005240 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005241 "test/qu8-dwconv-minmax.cc",
5242 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005243 "src/xnnpack/AlignedAllocator.h",
5244 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005245 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246)
5247
5248xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005249 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005250 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005251 "test/qu8-igemm-minmax.cc",
5252 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253 "src/xnnpack/AlignedAllocator.h",
5254 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005255 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005256)
5257
5258xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005259 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005261 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005262 "test/gavgpool-microkernel-tester.h",
5263 "src/xnnpack/AlignedAllocator.h",
5264 ] + MICROKERNEL_TEST_HDRS,
5265 deps = MICROKERNEL_TEST_DEPS,
5266)
5267
5268xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005269 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005270 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005271 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272 "test/gemm-microkernel-tester.h",
5273 "src/xnnpack/AlignedAllocator.h",
5274 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005275 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005276)
5277
5278xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005279 name = "qu8_requantization_test",
5280 srcs = [
5281 "src/xnnpack/requantization-stubs.h",
5282 "test/qu8-requantization.cc",
5283 "test/requantization-tester.h",
5284 ] + MICROKERNEL_TEST_HDRS,
5285 deps = MICROKERNEL_TEST_DEPS,
5286)
5287
5288xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005289 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005290 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005291 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005292 "test/vadd-microkernel-tester.h",
5293 ] + MICROKERNEL_TEST_HDRS,
5294 deps = MICROKERNEL_TEST_DEPS,
5295)
5296
5297xnnpack_unit_test(
5298 name = "u8_clamp_test",
5299 srcs = [
5300 "test/u8-clamp.cc",
5301 "test/clamp-microkernel-tester.h",
5302 ] + MICROKERNEL_TEST_HDRS,
5303 deps = MICROKERNEL_TEST_DEPS,
5304)
5305
5306xnnpack_unit_test(
5307 name = "u8_lut32norm_test",
5308 srcs = [
5309 "test/u8-lut32norm.cc",
5310 "test/lut-norm-microkernel-tester.h",
5311 ] + MICROKERNEL_TEST_HDRS,
5312 deps = MICROKERNEL_TEST_DEPS,
5313)
5314
5315xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005316 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005318 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005319 "test/maxpool-microkernel-tester.h",
5320 ] + MICROKERNEL_TEST_HDRS,
5321 deps = MICROKERNEL_TEST_DEPS,
5322)
5323
5324xnnpack_unit_test(
5325 name = "u8_rmax_test",
5326 srcs = [
5327 "test/u8-rmax.cc",
5328 "test/rmax-microkernel-tester.h",
5329 ] + MICROKERNEL_TEST_HDRS,
5330 deps = MICROKERNEL_TEST_DEPS,
5331)
5332
5333xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005334 name = "x32_fill_test",
5335 srcs = [
5336 "test/x32-fill.cc",
5337 "test/fill-microkernel-tester.h",
5338 ] + MICROKERNEL_TEST_HDRS,
5339 deps = MICROKERNEL_TEST_DEPS,
5340)
5341
5342xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005343 name = "x32_packx_test",
5344 srcs = [
5345 "test/x32-packx.cc",
5346 "test/pack-microkernel-tester.h",
5347 "src/xnnpack/AlignedAllocator.h",
5348 ] + MICROKERNEL_TEST_HDRS,
5349 deps = MICROKERNEL_TEST_DEPS,
5350)
5351
5352xnnpack_unit_test(
5353 name = "x32_pad_test",
5354 srcs = [
5355 "test/x32-pad.cc",
5356 "test/pad-microkernel-tester.h",
5357 ] + MICROKERNEL_TEST_HDRS,
5358 deps = MICROKERNEL_TEST_DEPS,
5359)
5360
5361xnnpack_unit_test(
5362 name = "x32_unpool_test",
5363 srcs = [
5364 "test/x32-unpool.cc",
5365 "test/unpool-microkernel-tester.h",
5366 ] + MICROKERNEL_TEST_HDRS,
5367 deps = MICROKERNEL_TEST_DEPS,
5368)
5369
5370xnnpack_unit_test(
5371 name = "x32_zip_test",
5372 srcs = [
5373 "test/x32-zip.cc",
5374 "test/zip-microkernel-tester.h",
5375 ] + MICROKERNEL_TEST_HDRS,
5376 deps = MICROKERNEL_TEST_DEPS,
5377)
5378
5379xnnpack_unit_test(
5380 name = "x8_lut_test",
5381 srcs = [
5382 "test/x8-lut.cc",
5383 "test/lut-microkernel-tester.h",
5384 ] + MICROKERNEL_TEST_HDRS,
5385 deps = MICROKERNEL_TEST_DEPS,
5386)
5387
5388xnnpack_unit_test(
5389 name = "x8_zip_test",
5390 srcs = [
5391 "test/x8-zip.cc",
5392 "test/zip-microkernel-tester.h",
5393 ] + MICROKERNEL_TEST_HDRS,
5394 deps = MICROKERNEL_TEST_DEPS,
5395)
5396
Marat Dukhan20c3b922020-03-10 03:45:06 -07005397########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005398
5399xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005400 name = "operator_size_test",
5401 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005402 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403)
5404
Marat Dukhan20c3b922020-03-10 03:45:06 -07005405xnnpack_binary(
5406 name = "subgraph_size_test",
5407 srcs = ["test/subgraph-size.c"],
5408 deps = [":XNNPACK"],
5409)
5410
5411########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412
5413xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005414 name = "abs_nc_test",
5415 srcs = [
5416 "test/abs-nc.cc",
5417 "test/abs-operator-tester.h",
5418 ],
5419 deps = OPERATOR_TEST_DEPS,
5420)
5421
5422xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005423 name = "add_nd_test",
5424 srcs = [
5425 "test/add-nd.cc",
5426 "test/binary-elementwise-operator-tester.h",
5427 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005428 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005429)
5430
5431xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005432 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005433 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005434 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435 "test/argmax-pooling-operator-tester.h",
5436 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005437 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005438)
5439
5440xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005441 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005442 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005443 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005444 "test/average-pooling-operator-tester.h",
5445 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005446 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005447)
5448
5449xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005450 name = "bankers_rounding_nc_test",
5451 srcs = [
5452 "test/bankers-rounding-nc.cc",
5453 "test/bankers-rounding-operator-tester.h",
5454 ],
5455 deps = OPERATOR_TEST_DEPS,
5456)
5457
5458xnnpack_unit_test(
5459 name = "ceiling_nc_test",
5460 srcs = [
5461 "test/ceiling-nc.cc",
5462 "test/ceiling-operator-tester.h",
5463 ],
5464 deps = OPERATOR_TEST_DEPS,
5465)
5466
5467xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005468 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005469 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005470 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005471 "test/channel-shuffle-operator-tester.h",
5472 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005473 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005474)
5475
5476xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005477 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005479 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005480 "test/clamp-operator-tester.h",
5481 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005482 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005483)
5484
5485xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005486 name = "constant_pad_nd_test",
5487 srcs = [
5488 "test/constant-pad-nd.cc",
5489 "test/constant-pad-operator-tester.h",
5490 ],
5491 deps = OPERATOR_TEST_DEPS,
5492)
5493
5494xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005495 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005497 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005498 "test/convolution-operator-tester.h",
5499 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005500 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005501)
5502
5503xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005504 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005505 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005506 "test/convolution-nchw.cc",
5507 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005508 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005509 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005510)
5511
5512xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005513 name = "copy_nc_test",
5514 srcs = [
5515 "test/copy-nc.cc",
5516 "test/copy-operator-tester.h",
5517 ],
5518 deps = OPERATOR_TEST_DEPS,
5519)
5520
5521xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005522 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005523 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005524 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005525 "test/deconvolution-operator-tester.h",
5526 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005527 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005528)
5529
5530xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005531 name = "divide_nd_test",
5532 srcs = [
5533 "test/binary-elementwise-operator-tester.h",
5534 "test/divide-nd.cc",
5535 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005536 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005537)
5538
5539xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005540 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005541 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005542 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005543 "test/fully-connected-operator-tester.h",
5544 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005545 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005546)
5547
5548xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005549 name = "floor_nc_test",
5550 srcs = [
5551 "test/floor-nc.cc",
5552 "test/floor-operator-tester.h",
5553 ],
5554 deps = OPERATOR_TEST_DEPS,
5555)
5556
5557xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005558 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005559 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005560 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005561 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005562 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005563 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005564)
5565
5566xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005567 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005568 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005569 "test/global-average-pooling-ncw.cc",
5570 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005571 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005572 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573)
5574
5575xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005576 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005577 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005578 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005579 "test/hardswish-operator-tester.h",
5580 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005581 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005582)
5583
5584xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005585 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005586 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005587 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005588 "test/leaky-relu-operator-tester.h",
5589 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005590 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005591)
5592
5593xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005594 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005595 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005596 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005597 "test/max-pooling-operator-tester.h",
5598 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005599 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005600)
5601
5602xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005603 name = "maximum_nd_test",
5604 srcs = [
5605 "test/binary-elementwise-operator-tester.h",
5606 "test/maximum-nd.cc",
5607 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005608 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005609)
5610
5611xnnpack_unit_test(
5612 name = "minimum_nd_test",
5613 srcs = [
5614 "test/binary-elementwise-operator-tester.h",
5615 "test/minimum-nd.cc",
5616 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005617 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005618)
5619
5620xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005621 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005622 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005623 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005624 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005625 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005626 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005627)
5628
5629xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005630 name = "negate_nc_test",
5631 srcs = [
5632 "test/negate-nc.cc",
5633 "test/negate-operator-tester.h",
5634 ],
5635 deps = OPERATOR_TEST_DEPS,
5636)
5637
5638xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005639 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005640 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005641 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005642 "test/prelu-operator-tester.h",
5643 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005644 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005645)
5646
5647xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005648 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005649 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005650 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005651 "test/resize-bilinear-operator-tester.h",
5652 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005653 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005654)
5655
5656xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005657 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005658 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005659 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005660 "test/sigmoid-operator-tester.h",
5661 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005662 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005663)
5664
5665xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005666 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005667 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005668 "test/softmax-nc.cc",
5669 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005670 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005671 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005672)
5673
5674xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005675 name = "square_nc_test",
5676 srcs = [
5677 "test/square-nc.cc",
5678 "test/square-operator-tester.h",
5679 ],
5680 deps = OPERATOR_TEST_DEPS,
5681)
5682
5683xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005684 name = "square_root_nc_test",
5685 srcs = [
5686 "test/square-root-nc.cc",
5687 "test/square-root-operator-tester.h",
5688 ],
5689 deps = OPERATOR_TEST_DEPS,
5690)
5691
5692xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005693 name = "squared_difference_nd_test",
5694 srcs = [
5695 "test/binary-elementwise-operator-tester.h",
5696 "test/squared-difference-nd.cc",
5697 ],
5698 deps = OPERATOR_TEST_DEPS,
5699)
5700
5701xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005702 name = "subtract_nd_test",
5703 srcs = [
5704 "test/binary-elementwise-operator-tester.h",
5705 "test/subtract-nd.cc",
5706 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005707 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005708)
5709
5710xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005711 name = "truncation_nc_test",
5712 srcs = [
5713 "test/truncation-nc.cc",
5714 "test/truncation-operator-tester.h",
5715 ],
5716 deps = OPERATOR_TEST_DEPS,
5717)
5718
5719xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005720 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005721 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005722 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005723 "test/unpooling-operator-tester.h",
5724 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005725 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005726)
5727
Chao Mei6ddfc602020-05-13 22:29:36 -07005728############################### Misc unit tests ###############################
5729
5730xnnpack_unit_test(
5731 name = "memory_planner_test",
5732 srcs = [
5733 "test/memory-planner-test.cc",
5734 ],
5735 deps = [
5736 ":XNNPACK",
5737 ":memory_planner",
5738 ],
5739)
5740
Marat Dukhan08c4a432019-10-03 09:29:21 -07005741############################# Build configurations #############################
5742
Marat Dukhanb8642352019-10-30 15:43:02 -07005743# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005744config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005745 name = "xnn_enable_assembly_explicit_true",
5746 define_values = {"xnn_enable_assembly": "true"},
5747)
5748
5749# Disables usage of assembly kernels.
5750config_setting(
5751 name = "xnn_enable_assembly_explicit_false",
5752 define_values = {"xnn_enable_assembly": "false"},
5753)
5754
Marat Dukhan9de90e02020-06-18 16:04:12 -07005755# Enables usage of sparse inference.
5756config_setting(
5757 name = "xnn_enable_sparse_explicit_true",
5758 define_values = {"xnn_enable_sparse": "true"},
5759)
5760
5761# Disables usage of sparse inference.
5762config_setting(
5763 name = "xnn_enable_sparse_explicit_false",
5764 define_values = {"xnn_enable_sparse": "false"},
5765)
5766
Marat Dukhan05702cf2020-03-26 15:41:33 -07005767# Disables usage of HMP-aware optimizations.
5768config_setting(
5769 name = "xnn_enable_hmp_explicit_false",
5770 define_values = {"xnn_enable_hmp": "false"},
5771)
5772
Chao Mei6ddfc602020-05-13 22:29:36 -07005773# Enable usage of optimized memory allocation
5774config_setting(
5775 name = "xnn_enable_memopt_explicit_true",
5776 define_values = {"xnnpack_enable_memopt": "true"},
5777)
5778
5779# Disable usage of optimized memory allocation
5780config_setting(
5781 name = "xnn_enable_memopt_explicit_false",
5782 define_values = {"xnnpack_enable_memopt": "false"},
5783)
5784
Marat Dukhanb8642352019-10-30 15:43:02 -07005785# Builds with -c dbg
5786config_setting(
5787 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005788 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005789 "compilation_mode": "dbg",
5790 },
5791)
5792
5793# Builds with -c opt
5794config_setting(
5795 name = "optimized_build",
5796 values = {
5797 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005798 },
5799)
5800
5801config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005802 name = "linux_k8",
5803 values = {"cpu": "k8"},
5804)
5805
5806config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005807 name = "linux_arm",
5808 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005809)
5810
5811config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005812 name = "linux_armeabi",
5813 values = {"cpu": "armeabi"},
5814)
5815
5816config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005817 name = "linux_armhf",
5818 values = {"cpu": "armhf"},
5819)
5820
5821config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005822 name = "linux_armv7a",
5823 values = {"cpu": "armv7a"},
5824)
5825
5826config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005827 name = "linux_aarch64",
5828 values = {"cpu": "aarch64"},
5829)
5830
5831config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005832 name = "android",
5833 values = {"crosstool_top": "//external:android/crosstool"},
5834)
5835
5836config_setting(
5837 name = "android_armv7",
5838 values = {
5839 "crosstool_top": "//external:android/crosstool",
5840 "cpu": "armeabi-v7a",
5841 },
5842)
5843
5844config_setting(
5845 name = "android_arm64",
5846 values = {
5847 "crosstool_top": "//external:android/crosstool",
5848 "cpu": "arm64-v8a",
5849 },
5850)
5851
5852config_setting(
5853 name = "android_x86",
5854 values = {
5855 "crosstool_top": "//external:android/crosstool",
5856 "cpu": "x86",
5857 },
5858)
5859
5860config_setting(
5861 name = "android_x86_64",
5862 values = {
5863 "crosstool_top": "//external:android/crosstool",
5864 "cpu": "x86_64",
5865 },
5866)
5867
5868config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005869 name = "windows_x86_64",
5870 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005871)
5872
5873config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005874 name = "windows_x86_64_clang",
5875 values = {
5876 "compiler": "clang-cl",
5877 "cpu": "x64_windows",
5878 },
5879)
5880
5881config_setting(
5882 name = "windows_x86_64_mingw",
5883 values = {
5884 "compiler": "mingw-gcc",
5885 "cpu": "x64_windows",
5886 },
5887)
5888
5889config_setting(
5890 name = "windows_x86_64_msys",
5891 values = {
5892 "compiler": "msys-gcc",
5893 "cpu": "x64_windows",
5894 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005895)
5896
5897config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005898 name = "macos_x86_64",
5899 values = {
5900 "apple_platform_type": "macos",
5901 "cpu": "darwin",
5902 },
5903)
5904
5905config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005906 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005907 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005908)
5909
5910config_setting(
5911 name = "emscripten_wasm",
5912 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005913 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005914 "cpu": "wasm",
5915 },
5916)
5917
5918config_setting(
5919 name = "emscripten_wasmsimd",
5920 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005921 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005922 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005923 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005924 },
5925)
5926
5927config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005928 name = "ios_armv7",
5929 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005930 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005931 "cpu": "ios_armv7",
5932 },
5933)
5934
5935config_setting(
5936 name = "ios_arm64",
5937 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005938 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005939 "cpu": "ios_arm64",
5940 },
5941)
5942
5943config_setting(
5944 name = "ios_arm64e",
5945 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005946 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005947 "cpu": "ios_arm64e",
5948 },
5949)
5950
5951config_setting(
5952 name = "ios_x86",
5953 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005954 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005955 "cpu": "ios_i386",
5956 },
5957)
5958
5959config_setting(
5960 name = "ios_x86_64",
5961 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005962 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005963 "cpu": "ios_x86_64",
5964 },
5965)
5966
5967config_setting(
5968 name = "watchos_armv7k",
5969 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005970 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005971 "cpu": "watchos_armv7k",
5972 },
5973)
5974
5975config_setting(
5976 name = "watchos_arm64_32",
5977 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005978 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005979 "cpu": "watchos_arm64_32",
5980 },
5981)
5982
5983config_setting(
5984 name = "watchos_x86",
5985 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005986 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005987 "cpu": "watchos_i386",
5988 },
5989)
5990
5991config_setting(
5992 name = "watchos_x86_64",
5993 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005994 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005995 "cpu": "watchos_x86_64",
5996 },
5997)
5998
5999config_setting(
6000 name = "tvos_arm64",
6001 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006002 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006003 "cpu": "tvos_arm64",
6004 },
6005)
6006
6007config_setting(
6008 name = "tvos_x86_64",
6009 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006010 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006011 "cpu": "tvos_x86_64",
6012 },
6013)